APPI News — Technology

OpenAI cuts Luna and Terra prices in widening AI cost race

APPI News Editorial — Sat, 15 Aug 2026 09:05:49 GMT

OpenAI cut standard application programming interface (API) prices for GPT-5.6 Luna by 80 percent and GPT-5.6 Terra by 20 percent on July 30, 2026. The company left GPT-5.6 Sol's standard token rates unchanged.

OpenAI attributed the reductions to lower serving costs and more efficient token generation. Its announcement did not name Chinese competitors, although lower-cost and open-weight models from Chinese developers have widened the choices available to buyers.

Luna now starts at US$0.20 per million input tokens

OpenAI's July 30 announcement set Luna at US$0.20 per million input tokens and US$1.20 per million output tokens, while Terra moved to US$2 and US$12 respectively. Sol remained at US$5 for input and US$30 for output under the comparable standard rate.

OpenAI's current pricing table identifies those figures as standard short-context rates and lists separate prices for cached input, cache writes and long-context use. The page also shows lower Batch and Flex rates, higher Fast mode rates and a 10 percent surcharge for eligible regional-processing endpoints.

For a monthly workload totaling 100 million uncached input tokens and 20 million output tokens, the listed short-context rates produce an estimated token charge of US$44 on Luna, US$440 on Terra or US$1,100 on Sol. That calculation excludes tools, cache writes, regional processing and any requests billed under another service tier.

OpenAI points to serving efficiency, not Chinese competition

OpenAI said GPT-5.6 Sol helped optimize production kernels that reduced end-to-end serving costs by 20 percent, while related experiments improved token-generation efficiency by more than 15 percent. The company presented those gains as the reason it could lower Luna and Terra prices.

The official announcement did not say the cuts were a response to Moonshot AI, MiniMax, Z.ai or another Chinese company. The South China Morning Post framed the move as a defense against lower-cost Chinese rivals and reported that Luna moved ahead of Z.ai's GLM-5.2 and MiniMax's M3 in Artificial Analysis's intelligence-per-dollar ranking.

Chinese vendors are competing through price and access

Token rates are only one part of the competitive pressure. Some Chinese developers also publish model weights, allowing customers or hosting companies to operate a model without sending every request to the developer's own API.

Moonshot AI's Kimi K3 paper describes a mixture-of-experts model with 2.8 trillion total parameters, 104 billion active parameters and a one-million-token context window, and says the full weights were released. Access to weights does not supply the hardware or engineering needed to run a model of that size, so “free weights” and “free inference” are different propositions.

MiniMax introduced M3 on June 1, 2026, with a one-million-token context window and API access, and said it planned to release the corresponding weights. The company also published its own coding and agent benchmarks, but vendor-run results need comparison with independent tests and the buyer's actual workload.

Price per token is not price per completed task

A cheaper token lowers the bill only if the model completes the work at an acceptable quality level. Longer outputs, retries, failed tool calls and extra review can offset a lower posted rate, while prompt caching can reduce the cost of repeated context.

Comparisons therefore need a fixed test set and a common success threshold. Useful measurements include completion rate, input and output volume, response time, retry frequency, human review time and total cost for each accepted result.

Model routing can reduce spending when routine work goes to a lower-cost model and difficult cases move to a more capable one. The saving depends on the routing error rate: an inexpensive first attempt adds cost if most requests must be repeated elsewhere.

API and open-weight deployments carry different costs

A hosted API turns model operation into a usage charge and leaves capacity planning to the provider. A self-hosted model replaces some of that charge with servers or cloud accelerators, deployment engineers, monitoring, security controls, upgrades and idle capacity.

Data handling also differs by provider, hosting location and customer configuration. Open weights can offer more control over where requests are processed, but the operator then assumes responsibility for access controls, logs, model updates and the rules that apply in each country where the service runs.

A four-part check before changing models

Token profile: Forecast uncached input, cached input, cache writes and output separately, then apply the correct context length and service tier.
Task quality: Run the same representative cases against each candidate and count accepted results rather than relying on a composite benchmark alone.
Operating conditions: Measure latency, rate limits, outages, tool charges, regional processing and the staff time needed to review failures.
Switching path: Keep prompts, evaluations and application interfaces portable enough to test another provider without rebuilding the product around one model.

The July price cut changes the baseline for high-volume GPT-5.6 workloads, especially where Luna meets the required quality threshold. It does not establish that one model is cheapest for every task or that an open-weight alternative will cost less after infrastructure and operations are included.

Frequently asked questions

Did OpenAI say Chinese rivals caused the price cuts?
No. OpenAI's July 30 announcement credited serving and token-generation efficiencies; the link to Chinese competition is an interpretation reported by the South China Morning Post.

Did ChatGPT or Codex subscription prices fall?
No. OpenAI said subscription prices and quota budgets stayed unchanged, while Terra and Luna began consuming fewer credits in ChatGPT Work and Codex.

Does open-weight mean a model is free to run?
No. Weight access can remove dependence on the developer's hosted API, but computing, electricity, deployment, monitoring and security still carry costs.

Which figure matters most when comparing models?
No single figure settles the choice. Total cost per accepted result combines token rates with output length, retries, latency, review work and the share of tasks that meet the required quality level.

US law sets cyber-device duties as Taiwan publishes guidance

APPI News Editorial — Fri, 14 Aug 2026 17:09:18 GMT

The US Food and Drug Administration (FDA) issued its current medical-device cybersecurity guidance in February 2026, replacing a version published in June 2025. The revision explains how manufacturers can document security design and meet statutory duties that have applied to defined cyber devices since March 29, 2023.

Taiwan's Food and Drug Administration, the health ministry agency responsible for medical-product oversight, publishes a 2021 manufacturer guide and five assessment templates. Those materials cover design, registration and post-market maintenance, but their legal footing differs from the minimum submission duties written into US law.

Device security is narrower than hospital IT security

Medical-device cybersecurity concerns software, firmware and communications that are part of a regulated device or its connected system. The risks include unauthorized commands, altered clinical data, unavailable functions and vulnerable third-party components. A hospital scheduling platform or general records database presents serious security issues too, but it is not automatically a medical device.

Taiwan's English-language guidance applies to manufacturers of devices containing software or programmable logic and to medical-device software, while excluding hospital administration software, general health-management software and medication-record software from its scope. The guide treats confidentiality, integrity and availability as device-safety concerns when a failure could impair performance or expose protected information.

The boundary does not mean devices operate independently of hospital networks. Cloud services, wireless links, maintenance tools and integration software can all sit inside the device system. A security assessment therefore has to trace interfaces and dependencies instead of checking only the hardware at the bedside.

US law sets minimum duties for defined cyber devices

Section 524B of the US Federal Food, Drug, and Cosmetic Act applies when a product meets three conditions: it contains sponsor-authorized software, can connect to the internet and has technological characteristics that could be vulnerable to cyber threats. The FDA says the rule covers specified premarket submissions for those cyber devices and has applied to submissions filed since March 29, 2023.

The same law requires three minimum elements. A sponsor must submit a plan for monitoring and addressing post-market vulnerabilities, maintain processes that provide reasonable assurance of cybersecurity and make updates and patches available, and provide a software bill of materials (SBOM) covering commercial, open-source and off-the-shelf components.

The FDA's filing process is more precise than a claim that every incomplete package is immediately rejected. The agency says a 510(k) filed through its eSTAR system will be placed on technical-screening hold when the cybersecurity section lacks accurate responses or relevant attachments. The separate transition policy that delayed refuse-to-accept decisions expired on October 1, 2023.

The FDA's February 2026 guidance supersedes its June 2025 document and recommends evidence covering device design, labeling and premarket review. The underlying statutory obligations remain distinct from the agency's broader, nonbinding recommendations.

The recommended evidence connects threat models, architecture diagrams, risk assessments, testing, component support information and unresolved software anomalies. It also extends beyond market entry: management plans address vulnerability monitoring, coordinated disclosure, patch-development timelines, update delivery and communication with customers.

An SBOM is an inventory, not a security verdict

An SBOM records the software components and dependencies inside a device. When a library vulnerability is disclosed, manufacturers and health-care providers can use that inventory to identify products containing the affected component rather than searching every device from scratch.

The inventory still needs context. A listed component may not expose the vulnerable function, a compensating control may reduce the risk, or the component may sit on a path that an attacker cannot reach. Conversely, a complete component list does not reveal a design flaw in code written specifically for the device.

The International Medical Device Regulators Forum's April 2023 final document says an SBOM can help health-care providers identify obsolete components before purchase and manage vulnerabilities after deployment, but it is not a substitute for a comprehensive device-level security risk assessment. The forum is a group of medical-device regulators; its document supplies an international practice framework rather than one country's market authorization.

Taiwan publishes lifecycle guidance and five templates

Taiwan's regulator issued its manufacturer guidance in April 2021 and announced the English version on May 5. It asks manufacturers to address cybersecurity from product design and development through registration, post-market monitoring and the end of support. The document covers threat modeling, risk controls, security testing, an SBOM, labeling, disclosure and updates.

The Chinese original describes the document as administrative guidance and says reviewers may request additional verification material based on a product's software architecture and design. The English translation says it is for reference only and that the Chinese text prevails, which limits how confidently an English-only comparison can characterize Taiwan's legal requirements.

On December 6, 2021, the regulator published one general assessment template and four product-specific templates covering an implantable pacemaker pulse generator, a glucose test system, oxygen-saturation software and a cloud electrocardiogram management system. The page records a maintenance date of June 10, 2022 and describes the documents as references for preparing cybersecurity assessments.

The list should not be read as a five-device whitelist. A product outside those examples can still present cybersecurity risk, and a completed form does not prove that a device will pass registration. The guidance expressly leaves room for reviewers to ask for evidence suited to a product's architecture.

APPI News could not find public Taiwan statistics showing how often cybersecurity material leads to a request for more evidence, a delay or a refusal. The cited documents establish what guidance and templates are available; they do not support a measured comparison of enforcement outcomes with the United States.

Procurement has to test the maintenance promise

Premarket documentation captures a product at one point in time, while software dependencies continue to change. A useful procurement record therefore identifies the SBOM format, component support dates, the manufacturer's vulnerability contact, the patch process and the conditions under which a device reaches end of support.

The deployment plan matters as much as the document list. Hospitals need to know which network services a device requires, how it authenticates updates, whether logs can be monitored, what happens when cloud access fails and how a faulty update can be rolled back. Those questions connect device security to the surrounding network without treating every hospital IT incident as proof of a device exploit.

The published US and Taiwan materials converge on lifecycle risk management, threat modeling, software transparency and post-market maintenance. They diverge in legal form and filing mechanics. That distinction can be documented from primary sources, while comparative rejection rates and real-world security outcomes remain unverified.

Frequently asked questions

Does an SBOM prove that a medical device is secure?
No. It identifies software components and supports vulnerability triage, but exploitability, system architecture, controls and possible patient harm still require a separate assessment.

Do the US requirements apply to every medical device?
No. Section 524B applies to submissions for products that meet the US statutory definition of a cyber device. The FDA's broader guidance also discusses devices with cybersecurity risk, so the scope of a recommendation and the scope of the statute are not identical.

Do Taiwan's five templates cover every connected device?
No. The regulator labels them reference templates: one is general and four address named product types. Its guidance also allows reviewers to request evidence based on the architecture and design of a particular product.

AOC raises China monitor prices as KTC warns of higher costs

APPI News Editorial — Fri, 14 Aug 2026 17:05:32 GMT

KTC warned in a notice reported by Fast Technology on August 14, 2026, that higher component costs could force later price increases across its monitor range. AOC had already raised prices across its China monitor range on August 1, but neither report gave a percentage or confirmed a matching global change.

The evidence supports one implemented adjustment and one warning, both reported in China. It does not support a single worldwide increase or a fixed timetable for every model.

AOC made a price change; KTC signaled one

Edmond Technology (Wuhan) Co., Ltd., which operates AOC's monitor business in mainland China, issued the AOC notice reproduced by VideoCardz. The notice covered the full product range and said the amount would vary by product, without listing percentages or new suggested prices.

Fast Technology said KTC issued a notice titled “Explanation and commitment regarding product price adjustments.” The report said higher core-component costs had become difficult for KTC to absorb and that prices across its monitor range might rise later, but it supplied no increase, effective date or model list.

The distinction matters: AOC communicated an effective price change, while KTC communicated the possibility of one. A warning from a vendor is not evidence that every retailer has changed its price.

APPI News could not locate a matching release on either brand's public website at the time of writing. The available record is therefore VideoCardz's reproduction and translation of the AOC notice, plus Fast Technology's account of the KTC notice.

The reports do not establish a global increase

The AOC report explicitly places the adjustment in the company's China distribution channel. Fast Technology's KTC report did not provide an overseas timetable or a market-by-market price list.

Neither source supports treating a China notice as a price change in Europe, Africa, the Americas or the rest of Asia. Local taxes, exchange rates, distributor inventory, promotions and warranty terms can produce different prices for the same model in different countries.

The notices attributed pressure to components and raw materials. Their silence on currencies or tariffs does not prove those factors have no effect on a particular seller; it only limits what can be attributed to the brands from these documents.

Panel concentration is real, but the closure scenario is conditional

TrendForce estimated that China's three largest panel makers would supply 74 percent of global LCD monitor panels in 2026, while LG Display continued reducing LCD production in favor of OLED. The research firm said the LCD monitor panel supply-demand balance was expected to become increasingly stable.

TrendForce also said global LCD monitor panel supply would fall by nearly 5 million units if Taiwan-based panel maker AUO closed its L6A fab. The word “if” is central: the report modeled a possible closure rather than reporting a final decision.

Those figures describe supplier concentration and a potential removal of capacity. They do not show that monitor panels were already in a global shortage in August, and they do not forecast a retail price for any AOC or KTC model.

Component pressure does not set a retail percentage

AOC's notice identified chips and motherboards, while the KTC report referred more broadly to core components. Neither disclosed how much those inputs had risen or what share of a monitor's total cost they represented.

A manufacturer can absorb part of an increase, change discounts, alter a configuration or pass the cost through after older inventory is sold. That is why an upstream increase cannot be applied directly to a store price without a model-specific notice or dated retail evidence.

Neither notice tied the change to DRAM, NAND Flash or AI data centers. Separate memory-price forecasts for computers and graphics cards cannot establish the cause or size of a monitor adjustment without data linking those markets.

A purchase decision needs local, model-specific evidence

A usable comparison starts with the same model number, screen technology, resolution, refresh rate and warranty in the same country. Prices should include taxes and delivery, because a lower headline figure may reflect different terms rather than a cheaper monitor.

Dated quotes or price histories from several authorized sellers can show whether a local price has moved. A promotion on older stock can coexist with higher replacement costs, while a brand warning can appear before any retailer changes a listing.

An urgent replacement and an optional upgrade create different timing constraints. The reports justify monitoring local offers, but they do not support a universal instruction to buy immediately or a prediction that prices will rise by a particular amount.

Frequently asked questions

Did both brands confirm a monitor price increase?
No. AOC's reproduced notice described an effective increase for its China range from August 1, while the KTC report described a possible later increase.

How large is the increase?
Neither report supplied a percentage, a new price list or model-level amounts. AOC said the adjustment varied by product, and KTC did not give an effective date.

Are AOC and KTC monitors becoming more expensive worldwide?
The cited evidence does not establish that. AOC's notice was specific to China, and the KTC report documented no matching rollout in other markets.

Does the TrendForce report prove an immediate panel shortage?
No. It projected high supplier concentration and modeled tighter conditions if AUO closed L6A, while also saying the LCD monitor panel balance was expected to become increasingly stable.

Older phone buyers need long support and usable controls, not novelty

APPI News Editorial — Thu, 13 Aug 2026 18:06:50 GMT

Samsung announced the Galaxy S26 series in February 2026, and the company says it will receive seven generations of operating-system upgrades and seven years of security updates. For an older user, that support clock matters only alongside grip, screen readability and the ability to complete familiar tasks without help.

The newest model is not automatically the strongest choice. A useful comparison starts with the support time left on the exact model, then tests whether the intended user can hold, unlock and operate it with the chosen case and display settings. Local repair terms and the price difference belong in the same decision.

Count remaining update years, not the headline promise

Samsung's Galaxy S25 policy provides seven generations of operating-system upgrades and seven years of security updates from the global launch date. The Galaxy S26 series carries the same seven-generation and seven-year terms, also counted from its initial global launch. Buying either model later does not reset its support clock, so a discounted S25 must be assessed against the support time already used.

Google promised the Pixel 8 and Pixel 8 Pro seven years of Android operating-system upgrades, security updates and feature releases, with security support counted from first availability on the US Google Store. Google said those two phones would remain supported into 2030. That fixed endpoint makes the remaining term visible when either model is sold used or refurbished.

An Apple filing under the United Kingdom's product security rules defines a minimum five-year support period from September 22, 2023, for model A3106, and Apple identifies A3106 as an iPhone 15 Pro Max variant. The filing does not promise five major iOS versions or state a worldwide support policy for every iPhone 15 model. APPI News could not verify a matching Apple document that would support a simple two-year comparison with Samsung and Google.

Support length also differs from rollout timing. Samsung says the timing of Galaxy S25 operating-system and security updates may vary by model, network provider and market. A long policy reduces the risk of an early support cutoff, but it does not mean every carrier releases each update simultaneously.

IP ratings do not predict a dropped phone

Samsung's US page lists the Galaxy S25 and S25+ with Armor Aluminum frames, Gorilla Glass Victus 2 displays and IP68 water and dust resistance. Its US Galaxy S26 page lists Armor Aluminum, Gorilla Glass Victus 2 on the front and back, and IP68 resistance for the S26 and S26+. The baseline specifications are close, but they are manufacturer declarations rather than independent evidence of how either phone survives years of falls.

The S26 page says its IP68 test used up to 1.5 meters of freshwater for up to 30 minutes and that resistance can diminish with ordinary wear. Samsung also says chips and cracks can reduce water and dust resistance and advises against salt-water exposure. Because the company defines the IP rating around dust and liquid ingress, it should not be read as a drop certification.

A store test should use the case that will stay on the phone. The intended user can then check whether the combined weight is comfortable, the side buttons remain easy to press and the screen can be reached without shifting to an unstable grip. Local prices for screen replacement and accidental-damage coverage provide information that glass branding alone cannot.

Test the interface after accessibility settings are enabled

Samsung says Easy mode enlarges on-screen items, provides a high-contrast keyboard and helps prevent accidental touches. The setting is under Display on the support page, which also warns that available screens and options vary by carrier, software version and device model. Its presence should be checked on the exact demonstration unit rather than assumed from the Galaxy name.

Apple describes Assistive Access as an iPhone interface for people with cognitive disabilities, with bigger items, more focused features and a choice of large grid or list layouts. Age alone does not establish that a person needs this mode. For users who need only a more readable display, iPhone also provides Larger Text, Bold Text and Display Zoom.

Font size is only one part of usability. The intended user should try unlocking the phone, answering a call, finding a frequent contact, sending a message, opening a familiar news or video app, taking a photo and returning to the home screen. The test should use the preferred font size and icon layout because a default store display does not show whether enlarged content clips or hides controls inside the apps that matter.

Choose the person first, then the model

The same short task list should be used on every candidate phone. A larger display may improve reading but can also add width and weight, while a smaller device may be easier to hold but leave less room after text enlargement. The result depends on the person and the case, not age alone.

Account recovery deserves attention before the old phone is erased. The intended user should retain control of the device passcode and primary account, while any agreed recovery contact and backup method are documented. This preserves access to future updates without making a relative's memory or phone number the only route back into the account.

A newer model earns a higher price when its later support endpoint or a tested usability difference matters to the buyer. If the S25 and S26 both pass the same tasks, the older model can still be the better value when its lower local price offsets the earlier support endpoint. A midrange phone may also qualify, but its update policy, water-resistance rating and simplified interface must be verified for that model rather than inferred from a flagship series.

Frequently asked questions

Does an older phone user need the latest Galaxy S26?
No. Samsung gives the S25 and S26 series the same headline update term, and its US pages list similar basic materials and IP68 resistance for the base and plus models. The S26 has a later support endpoint, but price, grip, screen size and local repair terms may outweigh that difference.

Do Samsung, Google and Apple promise the same type of support?
No. Samsung states numbers of operating-system generations and security years for the S25 and S26, while Google's Pixel 8 commitment covers Android upgrades, security updates and feature releases. Apple's five-year figure cited here comes from a UK product security filing for one iPhone 15 Pro Max model and is not a verified promise of five major iOS upgrades worldwide.

Can an IP68 phone be treated as waterproof?
No. Samsung's cited S26 test used controlled freshwater immersion at up to 1.5 meters for up to 30 minutes, and the company says resistance can weaken with wear or physical damage. IP68 does not establish drop survival or protection in every liquid.

Which simplified interface works best for an older adult?
There is no single answer based on age. Samsung Easy mode changes item size, keyboard contrast and touch behavior, while Apple designed Assistive Access for people with cognitive disabilities and provides separate display controls for larger text. The appropriate choice is the one that lets the intended user complete familiar tasks on the exact model without unwanted restrictions.

Japan launch leaves Taiwan status of ADHD therapy app unverified

APPI News Editorial — Wed, 12 Aug 2026 17:08:50 GMT

Shionogi launched ENDEAVORRIDE in Japan on June 5, 2026 and said it holds the product’s development and commercialization rights in Japan and Taiwan. The announcement identifies Japan as the launch market and does not report a Taiwan authorization or launch. APPI News could not verify an issued Taiwan license or a public filing under ENDEAVORRIDE, EndeavorRx or the named companies as of August 12, 2026. A commercial agreement for a territory does not itself authorize a medical device for sale there.

The gap illustrates why “digital therapeutic” does not have one regulatory status worldwide. Each jurisdiction separately defines the product category, reviews the evidence, authorizes the device and decides whether a public or private payer will cover it.

Digital therapeutic is not a single global category

A 2025 cross-sectional study used a narrow US definition: software-only medical devices with explicit prescription use. It identified 13 prescription digital therapeutics cleared by the US Food and Drug Administration (FDA) through May 2025, with eight using the 510(k) route and five using De Novo classification. The study also found that the authorized wording ranged from direct treatment claims to narrower claims about symptom management or functional improvement.

The total is not a count of every therapeutic app available worldwide. The study excluded nonprescription products, general-wellness software and devices that require hardware beyond a smartphone. Germany’s digital health application directory and Japan’s treatment-support software category therefore cannot be added to the US figure as if all three systems used the same boundary.

Japan’s file combined local and US evidence

Japan’s Pharmaceuticals and Medical Devices Agency (PMDA), which reviews medicines and medical devices for the country, received Shionogi’s ENDEAVORRIDE application on February 26, 2024. The PMDA review classified it as software supporting treatment of attention-deficit/hyperactivity disorder and examined two Japanese studies and three US studies. Japan granted marketing authorization in February 2025, and Shionogi launched the product 16 months later.

The Japanese phase 3 study enrolled 164 participants aged six to 17, according to the review record and the company’s launch announcement. PMDA also limited the authorized patient group and required Shionogi to distribute proper-use guidance prepared by relevant Japanese academic societies. This product-specific file shows a local regulator reassessing foreign and domestic evidence, but it does not establish a universal requirement to repeat the same trial in every country.

Germany links review to a national payment route

Germany created a different gate for eligible digital health applications, known by the German abbreviation DiGA. Germany’s health ministry says physicians can prescribe qualifying apps and statutory health insurance pays for them after the Federal Institute for Drugs and Medical Devices assesses safety, function, quality, data security and data protection. An app may receive provisional reimbursement for one year while its manufacturer supplies evidence that it improves care.

This arrangement connects assessment with a defined payer, but only inside Germany’s statutory health insurance system. It does not transfer German coverage to Japan, Taiwan, the United States or another market.

Taiwan rights leave the public status unresolved

Shionogi’s June announcement confirms Taiwan development and commercialization rights but reports a launch only in Japan. The Taiwan Food and Drug Administration, the health ministry unit that registers medical devices for sale in the territory, publishes an issued-license dataset that it says is synchronized weekly with its registration system. APPI News could not find ENDEAVORRIDE or EndeavorRx in the public sources it could access. It also could not verify whether Shionogi or another party had submitted an application that had not produced a public license record.

The distinction matters because rights, an application and an issued license are three different stages. Public license data can establish that a named product has a license, but a missing entry cannot establish what may be happening in a nonpublic review.

Regulatory clearance does not secure commercial continuity

Pear Therapeutics shows the difference between regulatory entry and a durable business. Its US annual report recorded $4.2 million in 2021 revenue, more than 14,000 prescriptions, a fulfillment rate of about 51 percent and 31.7 million covered lives. Those measures did not mean that every prescription became a paid, activated course of treatment.

Pear reported filing for Chapter 11 protection on April 7, 2023 and said it would seek a sale of the company or its assets. The sequence shows that authorization, payer reach, clinician prescribing, patient activation and company financing remain separate constraints. It is one company’s record and does not establish that other digital-therapeutics businesses will follow the same course.

Frequently asked questions

Is every health app a digital therapeutic?
No. The US study used a narrow definition covering prescription, software-only medical devices, while general-wellness apps and products requiring extra hardware were outside its count. Other countries use their own legal categories and evidence rules.

Is ENDEAVORRIDE authorized for sale in Taiwan?
APPI News could not verify an issued Taiwan license or a public filing as of August 12, 2026. The absence of a discoverable license does not prove that no application exists because the public dataset does not describe every possible step before a license is issued.

Does authorization in one country carry into another?
No. Shionogi’s Japan authorization, its Taiwan commercial rights and Germany’s reimbursement system are separate legal and payment arrangements. Availability and regulatory status vary by country.

AI scribes reduce paperwork as liability rules stay fragmented

APPI News Editorial — Tue, 11 Aug 2026 17:09:55 GMT

VA Providence, a health system within the US Department of Veterans Affairs, began rolling out an ambient artificial intelligence scribe in primary care on March 24, 2026. The health system said veterans could decline its use without affecting their care or benefits.

Ambient AI scribes record clinical conversations, convert speech into text and generate draft notes for clinicians to edit. They can reduce documentation work, but errors in a signed note raise separate questions about consent, organizational controls, product regulation and liability.

The system writes a draft, not the final record

An ambient scribe does more than basic dictation. Speech recognition produces a transcript, and a language model selects and organizes information into a clinical note, letter or other structured output. A clinician then reviews, edits and approves the draft before it enters the formal record.

That generation step can introduce errors even when the transcript is broadly accurate. A 2025 commentary in npj Digital Medicine identified fabricated details, omissions, contextual misinterpretation and incorrect speaker attribution as distinct failure modes. The authors also said reported error rates are difficult to compare because evaluations use different definitions and methods.

The burnout result is encouraging but narrow

A JAMA Network Open quality-improvement study published on October 2, 2025, analyzed 263 ambulatory clinicians at six US health systems after 30 days with one Abridge scribe platform. An adjusted model using 184 respondents estimated that self-reported burnout fell from 51.9 percent to 38.8 percent, while all 263 participants reported an average reduction of 0.90 hours per week in after-hours documentation.

The study shows an association, not proof that the software caused the change. Of 451 people enrolled, 272 completed both surveys and 263 met the final eligibility criteria. The authors reported no control group, a convenience sample, nonanonymous self-reports and no paired electronic-record activity data; they also disclosed vendor ties, while Abridge facilitated data collection.

Human approval does not settle liability

A 2025 JAMA Network Open invited commentary by US legal and risk-management authors said responsibility for the accuracy of patient records has traditionally rested with clinicians under US practice. The authors also urged health systems to perform due diligence, disclose ambient recording to patients, offer a way to decline and set protocols for review and data handling.

The same commentary warned that automation bias may make users less likely to spot an error in a polished draft. Its analysis is not a court judgment or a rule for every US state, and it does not remove possible duties held by a hospital or vendor. Human sign-off is therefore one control in the workflow, not a universal answer to who bears a loss.

Medical-device boundaries move with product functions

The US Food and Drug Administration (FDA) says certain software functions that serve as electronic patient records can fall outside the US device definition when they meet statutory criteria and do not interpret or analyze records for diagnosis or treatment. Its policy navigator does not name ambient scribes, so it cannot establish the status of every product carrying that label.

Australia provides a more explicit example. Australia's Therapeutic Goods Administration (TGA) says a digital scribe used only to transcribe and translate a clinical conversation is not a medical device, while a product that generates a diagnosis or treatment recommendation not stated by the clinician is a medical device. The TGA also says Australian health professionals are responsible for obtaining informed consent and verifying information entered into a patient's record.

The term “AI scribe” is therefore not a regulatory status. Intended purpose, actual functions and country-specific law determine whether device rules apply, and a software update that adds clinical recommendations can move a product across that boundary.

Hospitals and vendors retain separate duties

NHS England's guidance, updated on July 29, 2026, tells health organizations in England to ensure that users review and approve outputs before further action, audit clinical documentation and collect monitoring data independently of the manufacturer. It also calls for checks on data security, medical-device status and changes in product functionality.

The same guidance says NHS organizations may still face claims arising from AI products and should use supplier contracts to define roles, responsibilities and liability. A contract can allocate risk between organizations, but it does not establish one liability rule for patients and clinicians outside England.

A defensible deployment needs controls around the whole record path. Those include a clear purpose for the tool, a meaningful choice for the patient, access to the source transcript or audio where retained, a documented review step, error monitoring and terms covering storage, security, updates and incident handling.

Taiwan's basic law leaves the specific question open

Taiwan's Artificial Intelligence Basic Act was promulgated on January 14, 2026, and directs the government to clarify the attribution and conditions of liability for high-risk AI and to establish relief, compensation or insurance mechanisms. The law gives authorities two years to review rules and address gaps, while leaving sector authorities to identify high-risk applications in consultation with Taiwan's digital ministry.

The Act does not mention ambient scribes or classify them as high-risk. APPI News could not find an official English-language Taiwan rule that specifically classifies these tools or assigns liability for an erroneous generated note at the time of writing. That verification gap does not show that no existing privacy, medical or civil law could apply to a dispute.

The source report also cited performance figures from two Taiwan deployments that came from hospital and vendor publicity rather than independent evaluations. Those figures are not used here as evidence that ambient scribes deliver the same results across institutions, languages or products.

What the available evidence supports

Ambient scribes were associated with lower self-reported documentation burden and burnout in the cited short US study. The evidence does not establish long-term effects, equal performance across products or a universal allocation of liability.

Accountability remains a chain rather than a single signature. Vendors design and update the software, health organizations select and monitor it, clinicians review drafts, and regulators or courts apply the rules of a particular country to the facts of a particular case.

Windows OEM fees reportedly rise as PC component costs climb

APPI News Editorial — Tue, 11 Aug 2026 12:05:56 GMT

Windows Central reported on August 10, 2026, that an unnamed PC executive put Microsoft's average July Windows original equipment manufacturer (OEM) fee increase at 7 to 10 percent. Microsoft has not published the rate schedule, while Framework separately said higher Windows license costs were among several pressures on its new-order prices.

The reported percentage is a supplier-side cost change, not a 7 to 10 percent increase in every computer's retail price. Manufacturers set final prices around licensing, processors, memory, storage, distribution, taxes, exchange rates and margins, so the effect can differ by model and country.

The 7 to 10 percent figure remains unconfirmed

Windows Central reported on August 10 that an unnamed PC brand executive put the average Windows OEM fee increase at 7 to 10 percent from July and said Microsoft declined to comment. The report also said Microsoft charges different manufacturers different fees, which means the percentage does not reveal the amount paid for any one computer.

APPI News could not find a public Microsoft rate card that confirms the percentage. It also could not verify the source report's claim that fees are assigned through Intel Core i3, i5 and i7 tiers, so that detail is not used as a basis for comparing models.

OEM licensing is separate from a retail Windows purchase

Microsoft's licensing guide distinguishes a full Windows license preinstalled by a computer manufacturer, the OEM channel, from a Full Packaged Product bought at retail. The reported July increase concerns the manufacturer channel and does not establish that Microsoft changed the price of a retail Windows license.

A manufacturer may fold the OEM charge into a configured computer's price alongside the hardware and other software. Without a vendor's cost disclosure, a retail listing cannot show how much of a change came from Windows, and a manufacturer may absorb part of a cost increase or pass it through at a later date.

Framework confirms the pressure, not the reported rate

Framework said on July 22 that prices for new system and Mainboard orders reflected an upcoming Core Ultra Series 3 processor increase, higher Windows license costs, and higher memory, storage and other silicon costs in some configurations. The company did not disclose the Windows component in dollars or percent, and its statement covers its own product configurations rather than the PC market as a whole.

This cost mix makes a one-to-one link between an OEM fee and a store price especially difficult. A separate APPI News review found that conventional DRAM and NAND Flash contract prices were also forecast to rise in the third quarter of 2026, although contract prices do not dictate the price of a specific computer.

The 5 percent forecast is limited to Taiwan

TechRadar reported that ASUS and Acer computer prices in Taiwan were expected to rise by about 5 percent in the third quarter of 2026, while warning that other regions and vendors could differ. That forecast came from the same United Daily News report behind the 7 to 10 percent licensing figure.

The estimate is therefore a regional outlook, not an observed global price series. It also combines pressure from Windows licensing and hardware components, so it cannot show what share of a final price change came from either category.

Comparable configurations matter more than the headline rate

A defensible price comparison starts with the same model, processor, memory, storage, Windows edition and warranty in the same country. A lower price paired with less memory or storage does not demonstrate a lower licensing charge, and prices from different countries can reflect taxes, exchange rates and support terms.

Price histories and dated quotes from several retailers can show whether a specific configuration has moved. They cannot identify the cause unless the manufacturer itemizes it, but they are more useful for a purchase decision than applying the reported 7 to 10 percent fee increase directly to a computer's retail price.

An optional replacement can also be tested against the existing computer's actual limitation. APPI News's Windows 11 memory guide explains how to separate application pressure from cached, kernel and driver memory before treating a high usage figure as evidence that new hardware is required.

None of the cited sources gives a reliable date for Windows OEM fees or memory costs to fall. The evidence therefore supports comparing current, like-for-like offers, but it does not support a universal instruction to buy immediately or wait for a particular quarter.

Frequently asked questions

Did Microsoft confirm a 7 to 10 percent OEM fee increase?
No. The figure comes from an unnamed PC executive cited by United Daily News and repeated by English-language technology publications; Windows Central said Microsoft declined to comment.

Will every Windows computer become 7 to 10 percent more expensive?
No such conclusion follows from the report. Licensing is one cost among many, and each manufacturer decides how costs affect configurations, discounts and retail prices in each market.

Does the report show that retail Windows licenses also became more expensive?
No. Microsoft's own guidance treats manufacturer-preinstalled OEM licenses and Full Packaged Product retail licenses as separate channels, while the reported change concerns OEM fees.

Does the Taiwan forecast apply worldwide?
No. The roughly 5 percent estimate covered ASUS and Acer products in Taiwan during the third quarter of 2026, and the cited report did not provide a comparable forecast for every country, brand or model.

Deepfake scams make voice and video poor proof of identity

APPI News Editorial — Mon, 10 Aug 2026 17:12:04 GMT

INTERPOL and the United Nations Office on Drugs and Crime said on March 17, 2026, that generative AI tools, including deepfake video, images and audio, were making it easier for criminals to impersonate trusted people. A familiar face or voice can support a story, but it cannot authenticate a request for money, credentials or sensitive data.

The US National Institute of Standards and Technology (NIST) defines synthetic content as images, video, audio or text that algorithms, including AI, have substantially altered or generated. Deepfakes use that capability to make an identity appear to say or do something that did not happen. The fraud risk comes from what the media is used to establish: a recognizable identity, a plausible story and a demand designed to prevent independent checks.

Complaints filed with the US Federal Bureau of Investigation's (FBI) Internet Crime Complaint Center (IC3) in 2025 referenced AI in 22,364 cases and reported US$893,346,472 in adjusted losses. The report also recorded more than US$5 million in distress-scam losses, a category in which it said voice cloning can imitate loved ones, and more than US$632 million in investment complaints with a reported AI nexus. IC3 defines AI-related only as a complaint containing a reference to AI, so the figures do not measure deepfake losses alone or provide a global estimate.

The media supplies credibility; social engineering moves the money

A December 3, 2024, US FBI alert described voice clones impersonating relatives in a crisis, synthetic audio used to seek access to bank accounts, and video calls presenting supposed executives or public officials. The underlying requests remain familiar: send money now, disclose account information or accept an investment pitch before checking it elsewhere.

Voice and face generation can be only one part of the approach. Details collected from social profiles, breached accounts or previous conversations can make the story fit the target, while caller-ID spoofing and copied websites add separate signs of legitimacy.

On July 20, 2026, IC3 described scammers who used AI-generated videos of a senior US FBI official to direct previous fraud victims toward a spoofed IC3 website. The video did not carry out the fraud by itself: the recovery claim targeted people who had already lost money, and the copied site collected contact and financial information.

Visible glitches are clues, not an identity check

Lagging speech, mismatched lip movement, unnatural facial details and unusual word choice can justify closer scrutiny. The US FBI lists those irregularities as warning signs, but its July 2026 notice also says cloned voices can sound nearly identical to a known contact. A clip with no obvious defect is therefore not verified; it is only a clip with no defect the recipient noticed.

NIST's 2024 review says provenance records, watermarks, labels and content-based detection can provide evidence, but none offers a comprehensive solution by itself. The report says transparency may contribute to trustworthiness without guaranteeing it, and notes that watermarks can be removed, altered or placed on untrustworthy content. Authentic media can also mislead when removed from its original context.

European Union labels add information, not proof

The European Commission said the European Union's AI Act transparency requirements, which began to apply on August 2, 2026, require clear labels for deepfakes and for AI-generated or AI-manipulated text published on matters of public interest in covered cases. The Commission also published a voluntary code to help AI providers and deployers implement those legal duties.

The rule applies within the European Union framework rather than worldwide, and national obligations elsewhere vary. A visible disclosure can help a viewer identify synthetic media, but an absent label cannot make an unsolicited payment request authentic. A criminal impersonator has no reason to follow a disclosure rule.

Verification should leave the caller's channel

The US Federal Trade Commission says a short audio clip, including material posted online, can supply a voice-cloning program and that emergency scammers often combine the imitation with urgency and demands for secrecy. Its guidance tells recipients to end the exchange, call a number already known to belong to the person and consult another trusted contact if the person cannot be reached.

The US FBI recommends a private family word or phrase, while the US FTC suggests asking a personal question whose answer is not publicly available. Both measures add a check inside the conversation; a callback through a previously confirmed route moves the check outside the channel controlled by the caller.

Organizations can apply the same separation to payment and account requests. An employee can locate the requester in an existing corporate directory and follow the established approval process; a number, link or video room supplied in the suspicious message is not an independent route.

The available evidence does not yield a global loss figure

APPI News could not find published official data that separates worldwide losses from voice- and face-cloning scams from broader impersonation fraud. The IC3 figures are complaint data collected by a US agency, and its AI-related label records a reference in the complaint rather than a verified technical finding. INTERPOL's March 2026 statement establishes that international law-enforcement bodies are seeing the method, but it does not give a deepfake case count.

The evidence supports a narrower operational rule. Voice and video can trigger a verification process, but a payment, credential disclosure or account change should wait until the identity and the request have been confirmed through a separate channel.

Steam policy allows review after its two-hour refund window

APPI News Editorial — Mon, 10 Aug 2026 12:05:23 GMT

A Battlefield 6 player said on August 7, 2026, that Steam refunded the Phantom Edition after 470 hours when several modes left standard matchmaking. Valve's policy allows requests outside its 14-day and two-hour limits, but promises only that they will be considered.

The reported case does not change Steam's published rules or create a general right to a late refund. It shows the distinction between Valve's standard offer, an exceptional request that the company may examine, and any statutory remedy available under the law of a buyer's country.

The 470-hour refund remains a reported claim

Reddit user ExplorEverythingOnce said Steam granted a full Battlefield 6 Phantom Edition refund after 470 hours of play. The user attributed the request to Rush, King of the Hill and Squad Deathmatch leaving standard matchmaking, and said the submission included the game's public mode listings and a community notice.

PC Gamer reported on August 10 that a redacted image appeared to show a Phantom Edition refund, but did not prove the claimed playtime or the exchange with Steam support. APPI News could not independently verify either point and found no public confirmation from Valve, so the episode cannot establish that another account or complaint will receive the same result.

The published rule is 14 days and less than two hours

Valve says its standard offer covers Steam games and software requested within 14 days of purchase and used for less than two hours. Once approved, a full refund is issued within one week to the original payment method or Steam Wallet, although some payment methods cannot receive a reversal.

The same policy says a customer outside those limits may still submit a request and Valve will consider it. The page gives no separate approval test for those requests, so the sentence permits a review rather than establishing a broader refund entitlement.

Valve also says it may stop offering refunds to an account if it believes the system is being abused. Buying a title shortly before a sale, refunding it and immediately buying it again at the lower price is expressly excluded from the company's example of abuse.

Different purchases have different clocks

Steam's two-hour and 14-day test is the main rule for games and software, not every product sold through the service. Valve publishes separate conditions for downloadable content, pre-release titles, in-game purchases, bundles, gifts and subscriptions.

A request therefore needs the rule for the specific transaction rather than the game rule applied to every item. The store page may also mark some third-party downloadable content as nonrefundable before purchase.

Platform policy and statutory rights are separate

Valve's subscriber agreement says its refund policy applies without prejudice to statutory rights and notes withdrawal rights for European Union and United Kingdom consumers buying some digital content. This wording does not create one worldwide consumer rule, and it does not mean every downloaded or played game carries an unconditional withdrawal period.

Valve's refund page separately says consumers in some jurisdictions may have added rights when a game is faulty. Whether a removed feature, technical fault or store-page statement supports a legal remedy depends on the applicable country, the contract and the evidence, not solely on Steam's two-hour counter.

An over-limit request needs a dated record

Valve directs purchase-refund requests through Steam Support. A record can include the order receipt, recorded playtime, the feature or fault at issue, when the problem began, dated store descriptions or developer announcements, and earlier support messages.

The explanation should distinguish a technical fault from a later change to the product and identify the requested remedy. The 470-hour post may illustrate the type of material one claimant said was submitted, but it is not a precedent or an approval formula.

The current policy should be checked before filing because product rules and menu paths can change. If an initial response omits material evidence, a later contact can identify the missing record without representing that repeated submissions must receive a different outcome.

Frequently asked questions

Must a Steam game request be filed within 14 days and two hours?
Those are Valve's standard eligibility limits for games and software. The company accepts requests outside the limits for consideration, but its policy gives no assurance that they will be approved.

Does the 470-hour report mean other long-playtime refunds will succeed?
No general result follows from one self-reported case. The available image did not verify the claimed playtime or Steam support conversation, and Valve has not publicly confirmed the transaction.

Is Valve's policy the full measure of a buyer's rights?
Not in every country. Valve acknowledges that statutory rights may apply separately, while the scope and procedure depend on the law governing the individual purchase.

AI accelerator buyers need workload tests before choosing chips

APPI News Editorial — Mon, 10 Aug 2026 00:00:00 GMT

NVIDIA's Certified Systems Configuration Guide, updated August 6, 2026, separates inference and deep-learning training systems and says the workload, dataset, model and use case affect hardware selection. Those variables need to be measured before a team compares graphics processing units (GPUs), application-specific integrated circuits (ASICs) or accelerator services offered by cloud providers.

The selection starts with six questions: how often the model changes, whether the job is training or inference, the acceptable latency, peak throughput, memory demand and software portability. Chip specifications can narrow a shortlist, but they do not show how preprocessing, storage, networking and model loading will affect the complete service. A controlled trial with the buyer's own workload provides the evidence for a purchase or cloud commitment.

Measure each workload before comparing hardware

Create one record for every service rather than averaging several jobs into a single infrastructure profile. Record the model and version, parameter count, numerical precision, input and output sizes, batch size, concurrency, daily request volume and peak requests per second. Add the required model-quality threshold, availability target, data sensitivity and expected growth over the next three years.

Training, online inference and offline batch processing need separate records. Training runs may be constrained by accelerator memory, interconnect bandwidth and scaling efficiency, while an online service can fail its objective because of tail latency even when average throughput looks strong. An offline job can often trade completion time against price by running when capacity is cheaper or otherwise idle.

If measured traffic and latency data do not exist, the first procurement step is a small trial rather than a full rack purchase. The trial should capture normal and peak demand over a representative period. It should also separate warm-up time from steady-state performance so a fast sustained run does not conceal slow startup or model loading.

Workload stability determines the value of specialization

GPUs usually keep more options open when model architectures, frameworks or numerical formats are still changing. Their programmable hardware and established software tools can also make it easier to move capacity among unrelated jobs. Cloud rental can extend that flexibility when future demand is uncertain, although service availability, data-transfer charges and regional prices still need review.

A specialized accelerator becomes a candidate when the workload is stable, runs at enough volume to keep the hardware occupied and faces a clear limit on latency, energy or unit cost. imec IC-Link identifies throughput, latency and energy efficiency as three comparison metrics and says an ASIC reaches its strongest results only when the workload matches its design and data keeps the hardware utilized. The source is an ASIC design service, so its guidance is useful for defining questions but does not replace an independent run.

Peak operations per second should not decide the purchase. The host processor, accelerator memory, network path, storage system, compiler and runtime can each become the bottleneck. NVIDIA's guide likewise treats processor capacity, system memory, PCI Express topology, networking and storage as parts of an inference or training configuration rather than accessories to the accelerator.

Build a three-year cost range

Total cost of ownership (TCO) starts with purchased hardware or cloud usage, reserved capacity and the cost of meeting traffic above the reservation. For owned systems, the model also needs electricity, cooling, rack space, networking, storage and any facility work needed to support the equipment. For cloud services, it needs idle commitments, temporary peak capacity, data transfer and price differences among the regions the service can lawfully use.

Software and labor belong in the same calculation. Include drivers, compilers, monitoring, platform licenses, model conversion, performance tuning and on-call operations. A lower compute price can be offset by engineering work if unsupported operators, unfamiliar tools or a proprietary runtime slow deployment.

Reliability and exit costs complete the range. Buyers need to price spare capacity, replacement time, supply lead times, backups and recovery tests, then estimate the work required to move models and data to another platform. The result should show at least a base case and low- and high-utilization cases because a single utilization assumption can reverse the comparison between owned and rented capacity.

Run the same acceptance test on every candidate

Each candidate should run a normal-load case, a peak case and a failure case. Hold the model revision, dataset, numerical precision, software versions and quality threshold constant. Record any vendor-specific optimization because a comparison is not reproducible if only one system receives an undisclosed change to the model or runtime.

MLCommons says MLPerf Inference uses Offline, Server and Interactive scenarios for datacenter systems and SingleStream, MultiStream and Offline scenarios for edge systems; its load generator tracks latency and validates accuracy. The guide also describes its Closed Division as an apples-to-apples comparison using the same model and reference setup. A private procurement test need not reproduce MLPerf, but it can follow the same discipline of fixed inputs, declared scenarios and a quality check alongside speed.

For online inference, record 50th-, 95th- and 99th-percentile latency (P50, P95 and P99), throughput, error rate, output quality, power and cost per successful request. Generative-model tests may also need time to first token and the rate at which later tokens arrive. Training tests should record the time and cost required to reach a fixed quality target rather than comparing only the speed of individual steps.

The failure run should reflect the proposed deployment. It can remove a device or node, interrupt a dependency, restrict network capacity or force the service to recover from a saved checkpoint. Record dropped work, recovery time and whether the remaining capacity still meets the service-level objective (SLO).

Turn the measurements into a decision record

Set pass-or-fail gates first. A candidate that misses the quality threshold, peak-load SLO, data-control requirements or facility limits should not win through a weighted average. These conditions need written thresholds before results arrive.

Score the candidates that pass. Compare three-year cost, deployment time, software fit, operational skills, supply risk and portability. Weight each factor according to the service being purchased, then retain the raw measurements and assumptions with the score.

Record why the result could change. The decision should name the utilization, electricity price, demand forecast, software version and expected model life used in the calculation. It should also set a review trigger, such as a major model revision, a sustained traffic change or a new contract price.

Frequently asked questions

Is inference always better on an ASIC?
No. Stable, heavily used inference is a candidate for specialized hardware, but low or variable demand can leave that hardware idle. Frequent model changes can also make GPU flexibility worth more than a lower theoretical unit cost.

Does a higher FLOPS figure guarantee a faster service?
No. Floating-point operations per second (FLOPS) describes compute capacity under defined conditions, not end-to-end application time. Memory transfers, preprocessing, networking, software optimization and batching can determine the result.

Are owned accelerators always cheaper than cloud capacity?
No general answer applies. The comparison changes with utilization, electricity, cooling, staffing, financing, redundancy, cloud discounts and exit costs. A three-year range using the same demand forecast is more informative than comparing a server price with an hourly cloud rate.

Which metric should control the acceptance test?
The controlling metric is the service requirement that must not be missed, such as tail latency, completion time, output quality or cost per successful request. Throughput remains relevant, but it cannot compensate for a failed quality or latency threshold.

AI rack plans must account for peak power before servers arrive

APPI News Editorial — Mon, 10 Aug 2026 00:00:00 GMT

NVIDIA’s DGX H100 data center guide puts the maximum power demand of one system at 10.2 kW. Four systems bring the server portion of a rack to 40.8 kW before switches, storage and management hardware are added.

Those figures do not set a universal density for AI facilities. They show why planners should start with the final equipment list and peak demand instead of multiplying an older server-room average by the number of new racks.

Build a capacity record for every rack

List every server, switch, storage appliance, rack power distribution unit and cooling component planned for the cabinet. For each item, record quantity, rated and expected power, input-voltage range, plug type, power-supply count, weight, rack units, airflow direction, cooling connection and network ports.

Keep two totals: the credible peak IT load used to size the facility path, and the measured or estimated operating load used for energy and cost forecasts. Add the power drawn by rack-level cooling and network equipment, then reserve explicitly approved space and capacity for expansion rather than leaving an undefined margin.

NVIDIA’s planning guide lists a DGX H100 at a maximum of 10.2 kW and 130.45 kg, and shows rack configurations ranging from 10.2 kW for one system to 40.8 kW for four. The same guide says a power or cooling constraint can force equipment into more racks, which can change cable lengths and network performance.

Trace redundant power to the upstream source

Map each electrical path from the utility or on-site generation through transformers, uninterruptible power supplies (UPSs), switchgear, busways, floor distribution and rack power distribution units to each server power supply. Record the rated capacity, usable capacity, protection device and maintenance state at every stage.

Two plugs on a server do not prove end-to-end redundancy. Paths labeled A and B may converge at one UPS, switchboard or transformer, leaving a common failure point. The design review should show which loads transfer during maintenance or failure and whether the surviving equipment can carry them.

NVIDIA’s H100 electrical guide specifies three rack power paths for its N+1 design, requires each path to support half of expected peak rack demand and tells installers to verify redundancy by switching off the breakers feeding each rack power strip. That topology is specific to the cited system; other hardware must be checked against its own power-supply and support requirements.

Match cooling to the heat that remains in the room

For air-cooled equipment, calculate airflow and heat rejection at the proposed inlet temperature, altitude and rack density. Confirm aisle layout, blanking panels and containment, then test for recirculation and hot spots at the intended load.

A liquid-cooling plan needs supply and return temperatures, required flow, pressure limits, fluid chemistry, material compatibility, filtration and service clearances. It should also identify the coolant distribution unit, manifolds, couplings, isolation valves, leak sensors, containment and the response to pump or power loss.

The Open Compute Project’s cold-plate requirements say hybrid systems still need room air conditioning, while coolant distribution units manage pressure, flow, temperature, cleanliness and leak detection. The document also calls for a spill and leak plan covering detection, intervention, containment and pump failures.

Do not assume that adding cold plates removes every air-side load. Memory, power electronics, network devices and other components may still reject heat into the room unless the rack design captures it elsewhere.

Check weight, delivery routes and network geometry

Add the cabinet, compute hardware, power equipment, coolant, manifolds and cabling when calculating rack weight. A structural professional should assess both the installed load and the rolling load along the route from the loading area to the final position. Door widths, ramps, lifts, turning space and floor transitions belong in the delivery review.

Place compute, storage and switches as one system rather than treating the network as a later cable order. Document port counts, link speeds, oversubscription, cable types, maximum path lengths, management networks and the effect of losing a link or switch. Spreading machines across extra rows may solve a cooling limit while adding cable distance or switching layers.

Commission the facility in a fixed sequence

Freeze the bill of materials, firmware versions and rack positions before integrated testing. Record the expected result, acceptance threshold, responsible person and recovery procedure for every test.

Verify equipment labels, power connections, cooling connections and network ports against the final drawings.
Measure voltage, current and phase balance, then remove each claimed redundant power path in turn.
Apply the planned load and record inlet temperatures, component temperatures, coolant flow, pressure and room hot spots.
Test compute, storage and network throughput with the intended topology and software versions.
Simulate the loss of one rack, one switch and each cooling or power component that the design claims to tolerate.
Trigger power, temperature, flow, leak and network alarms and confirm that they reach the assigned operators.
Save electrical single-line diagrams, piping diagrams, cable maps, test results, change records and recovery steps.

Problems found before delivery can change a purchase order or rack layout. Problems found after installation may require new electrical work, piping, network paths or structural review while expensive equipment remains idle.

Frequently asked questions

Should power capacity use the rated or average load?
Use a defensible peak figure from the equipment specification and intended configuration when checking capacity and failure states. Keep measured or estimated operating load separate for energy and cost forecasts.

Do dual power supplies provide full redundancy?
Not by themselves. Trace both feeds to their upstream sources and identify every point where they share a UPS, switchboard, transformer or generator.

Does liquid cooling eliminate room air conditioning?
Not for a typical hybrid system. The facility must still remove heat from components and equipment that do not transfer all of their heat to the liquid loop.

Can commissioning wait until all servers arrive?
Final integrated tests require the installed system or representative loads, but power, cooling, weight and delivery constraints should be resolved during procurement and facility design.

Apple tests CXMT memory as US senators challenge China supply plan

APPI News Editorial — Sun, 09 Aug 2026 18:05:40 GMT

The Wall Street Journal reported on August 9, 2026, that Apple had tested dynamic random-access memory (DRAM) from China's ChangXin Memory Technologies (CXMT) across product lines including iPhones and MacBooks, citing people familiar with the matter. The report said early supply talks were aimed at using the components in some devices sold in China.

The report described testing and early talks, not a purchase order. Reuters' July 7 summary of an earlier Financial Times report likewise said Apple had not committed to commercial use. Neither Apple nor CXMT had publicly confirmed the reported tests or announced a supply agreement at the time of writing.

AI and server demand have tightened DRAM supply for PC and smartphone manufacturers, raising the value of another qualified supplier. (Illustrative image)

Testing is not a purchase decision

Component qualification checks whether a part meets a manufacturer's electrical, performance, reliability, thermal and production requirements. Passing those tests can make a component eligible for future orders. It does not establish that the buyer will place an order, how large that order would be or which finished products would use the part.

The reports do not identify a CXMT part number, a test result or a proposed volume. They also do not name particular iPhone or MacBook models. The reported China-only supply goal narrows the potential deployment, but it does not turn a test across product lines into a confirmed product plan.

Qualifying another supplier could give Apple more flexibility in a shortage and another option in price negotiations. Those are possible effects, not a motive Apple has confirmed. The public record does not show that Samsung Electronics, SK hynix or Micron Technology has lost an Apple order or changed a price because of the CXMT tests.

The memory squeeze is the commercial backdrop

TrendForce said conventional DRAM contract prices rose about 93 to 98 percent quarter on quarter in the first quarter of 2026. The research firm said supplier inventories remained extremely low in the second quarter and that manufacturers were directing incremental supply toward high-capacity server memory, limiting availability for PC and smartphone makers.

AI servers do not use exactly the same memory products as every phone or laptop. The pressure spreads because the same large suppliers allocate investment, wafer starts and production upgrades among high-bandwidth memory, server DRAM and consumer products. A buyer that qualifies another mobile or PC DRAM source gains a supply option even if that supplier cannot replace every product made by the three market leaders.

Price and availability are separate questions. A fourth supplier may have available capacity without offering a lower price, and a lower quote would not help if the component failed qualification. The cited reports disclose neither CXMT's proposed price to Apple nor the amount of capacity it could reserve.

CXMT entered 2026 as the fourth-largest DRAM producer and completed a Shanghai listing in July. (Illustrative image)

CXMT's scale depends on the metric

CXMT says it was founded in 2016 and manufactures DRAM for phones, PCs, tablets and servers. DRAM provides working memory while a device is powered. It is distinct from the nonvolatile storage supplied by flash-memory producers.

Reuters reported from CXMT's prospectus that the company was the world's fourth-largest DRAM producer with a 7.7 percent market share in 2025. CXMT raised 57.92 billion yuan (US$8.6 billion) in its July 2026 initial public offering in Shanghai.

An earlier Financial Times account, summarized by Reuters, put CXMT at about 11 percent of global DRAM wafer capacity in 2025. Wafer-capacity share and market share are different measures, and the available summaries do not provide enough methodology to convert one into the other. This report therefore does not present 11 percent as CXMT's share of DRAM sales.

CXMT and YMTC are not the same supplier

CXMT and Yangtze Memory Technologies Co. (YMTC) appear together in the US policy debate, but their core products differ. CXMT makes DRAM, while YMTC identifies itself as a designer and manufacturer of 3D NAND flash memory. The August 9 report concerns CXMT DRAM, not YMTC NAND.

The distinction also matters when reading policy documents. A restriction or licensing requirement attached to one company does not automatically apply to the other. Likewise, a test of working-memory chips does not establish that Apple is reviving its reported 2022 evaluation of YMTC storage components.

Seven US senators asked Apple to abandon the reported CXMT tests and to respond by August 21, 2026. (Illustrative image)

US designations create political and compliance risk

A US Department of Defense notice published in January 2025 listed both CXMT and YMTC as “Chinese military companies” under Section 1260H. That wording is a US government designation. It should not be restated as an independently proven ownership or operational relationship.

Seven US senators wrote to Apple on July 29, 2026, saying both companies remained on an updated June list and urging Apple to reject their memory products. The letter asked Apple to describe its qualification work, any technical information shared with the suppliers and its supply-chain risk assessment by August 21. Those questions document congressional scrutiny; they do not establish that Apple transferred controlled technology. The deadline had not passed at publication, so this report does not assess an Apple reply.

The Defense Department designation is not the same as an entry on the US Commerce Department's Entity List. The US Bureau of Industry and Security added YMTC to the Entity List in December 2022 and said listed parties face specific licensing requirements under the Export Administration Regulations. The senators' letter cites the Entity List designation for YMTC but not for CXMT, which shows why describing both companies as subject to the same trade restriction would be inaccurate.

A future Apple-CXMT transaction would have to be checked against the rules, products and technology transfers in effect at that time. The public documents reviewed for this report do not supply a final licensing determination for the reported test or for a hypothetical purchase. They establish political opposition and policy exposure, not a completed enforcement action against Apple.

Samsung Electronics, SK hynix and Micron remained the three largest DRAM vendors in the first quarter of 2026. Apple's tests do not show that any of them lost an order. (Illustrative image)

What the test could change for existing suppliers

A successful qualification could give Apple the option to source some China-market memory from CXMT. The commercial effect would depend on the approved part, available volume, contract price and duration. None of those terms is public.

Without those terms, claims that the test will force Samsung Electronics, SK hynix or Micron to cut prices remain market interpretation. The same applies to predictions that Apple will move a large share of orders to CXMT. Testing creates an option; it does not show that Apple has exercised it.

The evidence supports a narrower conclusion. Apple is examining a Chinese DRAM supplier during a supply squeeze, while US lawmakers are trying to prevent that supplier from entering Apple's production chain. Technical qualification, commercial negotiations and US policy could each stop or reshape the plan before any device ships with CXMT memory.

Frequently asked questions

Has Apple decided to buy CXMT memory?
No decision has been announced. The reports describe component testing and early talks for some devices sold in China, and the earlier Financial Times account said Apple had not committed to commercial use.

Is CXMT banned from supplying Apple?
The cited documents do not answer that question with a blanket yes or no. They establish a US Defense Department designation, a congressional request and separate Commerce Department restrictions on YMTC; the licensing outcome for a specific Apple-CXMT transaction has not been published.

Will the tests reduce Apple's orders from Samsung Electronics, SK hynix or Micron?
There is no public order volume or allocation change to measure. CXMT could become an additional option if it passes qualification and clears policy hurdles, but testing alone does not show that an incumbent supplier has lost business.

Smartwatch atrial fibrillation alerts still need ECG confirmation

APPI News Editorial — Sun, 09 Aug 2026 17:06:52 GMT

The 2019 Apple Heart Study sent irregular-pulse notifications to 2,161 of 419,297 self-enrolled US adults, or 0.52 percent. Atrial fibrillation (AF) appeared on later electrocardiogram (ECG) patches in 34 percent of the 450 participants with analyzable recordings. The study's 84 percent positive predictive value measured a narrower event: another notification that occurred while the patch was recording and showed AF at the same time.

Severe symptoms override a watch reading

The United Kingdom's National Health Service advises immediate emergency care when a fast or irregular heartbeat occurs with chest pain, shortness of breath, sweating, nausea, fainting, a severe headache, one-sided weakness or numbness, sight loss, confusion or difficulty speaking. A normal watch reading or the absence of a notification should not delay that assessment. Emergency numbers and services differ by country.

Background pulse checks and watch ECGs do different jobs

The original Apple feature reviewed by the US Food and Drug Administration (FDA) used photoplethysmography (PPG), which pairs green light with photodiodes to measure changes in blood flow and calculate intervals between pulses. The 2018 decision summary says that version attempted a one-minute background reading about every four hours when activity allowed. It required five of six sequential readings to be classified as irregular within 48 hours before sending a notification.

That background process was opportunistic rather than continuous, and motion or poor skin contact could prevent a usable reading. The US FDA document says the absence of an alert does not rule out a disease process and that the feature was not intended to diagnose AF. Its sampling and notification details describe the software version reviewed in 2018, not every later release.

The US FDA cleared Samsung's ECG Monitor Application with Irregular Heart Rhythm Notification on May 2, 2023. The software uses background PPG analysis and can prompt the wearer to record a single-channel ECG similar to Lead I. The labeling describes the output as informational, says the function will not identify every episode and directs users not to take clinical action without consulting a qualified health professional.

The Apple study produced two different percentages

The Apple Heart Study recruited participants over eight months and monitored them for a median of 117 days. Participants already owned compatible Apple devices, enrolled themselves and reported no previous AF diagnosis, while the cohort skewed younger than the population at highest risk of the condition.

The 450 analyzable ECG patches were applied an average of 13 days after the first notification, and 153 recorded AF, giving a diagnostic yield of 34 percent. AF can occur intermittently, so a later patch that did not capture it could not establish that the earlier notification was false.

A separate analysis covered 86 people who received another irregular-pulse notification while wearing the ECG patch. Seventy-two had AF on the patch at that same time, producing the 84 percent positive predictive value. The 34 percent figure asks whether AF appeared at any point during delayed follow-up; the 84 percent figure asks whether a new alert matched a simultaneous ECG record.

Samsung's clearance data use other measures

Samsung's US 510(k) summary says 888 people at risk of AF entered its validation study and 810 were included in the analyzable dataset after wearing a watch and reference ECG patch at the same time for about seven days. The notification function had subject-level sensitivity of 68.0 percent and specificity of 98.8 percent. These measures describe how the tested software separated participants with and without AF under that protocol; they are not the percentage of consumer alerts later confirmed.

The same filing reports a 95.7 percent positive predictive value for individual positive PPG sequences selected from a subset with reference ECG data. That is a tachogram-level calculation, not the share of all watch owners who would receive a correct notification. Product, signal, study population and unit of analysis all have to match before percentages can be compared.

A small research system found costs from false alerts

A 2023 analysis followed 85 people aged 50 or older who had survived a stroke or transient ischemic attack and were assigned to wear an ECG patch with a smartwatch running the custom Pulsewatch system for 14 days. Fifteen received AF alerts, and 10 of them, or 67 percent, had no AF on the simultaneous patch. The researchers built the software used on the Samsung hardware, so the result is not a false-alert rate for Samsung's commercial notification feature.

The study counted 35 false alerts: 19 coincided with noisy PPG signals, 11 with other rhythms and five with ECG data too corrupted to determine a cause. False alerts were associated with declines in participants' perceived physical health and confidence in managing chronic symptoms. The false-alert subgroup contained only 10 people, and the association does not establish that the alerts caused those changes.

The researchers reported that an offline deep-learning approach reduced the number of false-positive alerts by 83 percent after training on 60 times as many signal segments as the embedded rule-based detector. The analysis was not a prospective test of a deployed commercial feature and did not show whether the change improved clinical outcomes.

Age, medical history and country change the boundary

The Apple feature in the 2018 US decision was not tested or intended for people under 22 or those already diagnosed with AF. Samsung's 2023 US labeling likewise specifies adults aged 22 or older and excludes people with other known arrhythmias. These limits belong to the named software versions and US decisions, not to every smartwatch or market.

The cited evidence does not establish separate interpretation rules for children, pregnant people, people with chronic conditions or those taking medication. People in these groups should discuss an alert with a clinician before using it to change care. Prescribed medication should not be started, stopped or adjusted because of a watch result.

An alert supports review, not self-diagnosis

The US FDA's Apple decision says notifications are potential findings that should be reviewed by a medical professional and that AF diagnosis still requires ECG confirmation. The United Kingdom's National Health Service also lists an ECG among the tests used to check heart rhythm when AF is suspected. A notification time and any saved watch tracing can inform that review, but neither should be treated as a diagnosis or a basis for changing treatment.

No notification is not an all-clear because background sampling can miss intermittent episodes and detection sensitivity is not 100 percent. The cited studies assessed detection and short-term follow-up rather than whether consumer screening prevents strokes or deaths. APPI News verified the cited US records but could not verify current feature availability or regulatory status in every country at the time of writing.

China clears Apple Intelligence with Alibaba and Baidu models

APPI News Editorial — Sun, 09 Aug 2026 06:05:40 GMT

China's internet regulator registered Apple Intelligence for use on iPhones on July 15, 2026, clearing a major barrier to the service's release in mainland China. The Cyberspace Administration of China (CAC), which oversees internet content and platform regulation, listed “Apple Intelligence” among seven newly registered on-device generative AI services.

The registration does not amount to a product launch. The CAC notice gave no release date, while Alibaba and Baidu confirmed that their models would contribute to Apple's service in China.

Apple is preparing a China-specific version of Apple Intelligence that uses models from Alibaba and Baidu. (Illustrative image)

What China approved

The CAC announcement records a regulatory filing for a mobile-side generative AI service. It does not describe the filing as a blanket approval for every Apple Intelligence feature, and it does not say that the service was immediately available to users.

That distinction matters because Apple still described the product as unavailable in mainland China shortly before the filing. Apple's support page, published July 7, said Apple Intelligence would not work on supported devices purchased in mainland China and would become available there at a later date. APPI News could not find a subsequent Apple announcement giving a public launch date at the time of writing.

What Alibaba and Baidu have confirmed

Alibaba told Reuters that Qwen would be integrated into Apple Intelligence on iOS, iPadOS, macOS and visionOS in China. The company said the integration would include text and image capabilities, and Baidu separately confirmed work on Apple Intelligence features for Chinese iPhone users.

Public statements do not provide a feature-by-feature allocation between the two companies. The source article cited a reported 65 percent share for Alibaba and 35 percent for Baidu, but APPI News could not verify those figures in company statements or the CAC notice. They are therefore omitted from this report.

Alibaba and Baidu have confirmed roles in the service, but neither has published a full technical division of work. (Illustrative image)

Why the regulatory filing matters

China's Interim Measures for the Administration of Generative Artificial Intelligence Services took effect on August 15, 2023. The rules apply to services that generate text, images, audio or video for the public in mainland China, including services provided from outside the country.

Article 17 requires providers of services with public-opinion or social-mobilization attributes to complete a security assessment and an algorithm filing. Article 20 allows Chinese authorities to order technical or other measures against noncompliant services supplied from outside mainland China.

Those provisions explain why the overseas version of Apple Intelligence could not simply be assumed to qualify for the Chinese market. They do not explicitly require every foreign company to select a Chinese model provider, however. Apple's use of Alibaba and Baidu is the confirmed structure of this product, not a general rule written into the regulation.

China requires qualifying public-facing generative AI services to pass a security assessment and file their algorithms. (Illustrative image)

What remains unresolved

The filing clears one regulatory step, but the public record does not establish the final software configuration, supported devices or release schedule in mainland China. Apple may also vary features by platform, language and region, as it does elsewhere.

The available evidence supports a narrower conclusion: Apple Intelligence has entered China's registration system, and Alibaba and Baidu are participating in the localized service. Claims about exact workload percentages or a specific release window remain unverified.

Frequently asked questions

When will Apple Intelligence launch in mainland China?
No date has been announced. The CAC registration opens a path to release, but the regulator's notice did not set a timetable.

Which models will the China service use?
Alibaba has confirmed that Qwen will contribute text and image capabilities, while Baidu has confirmed that it is developing features with Apple. The companies have not published a complete technical breakdown.

Does Chinese law require Apple to use local partners?
The cited rules require compliance measures for qualifying generative AI services, including security assessment and algorithm filing. They do not state that every foreign provider must use a Chinese partner.

Hospitals train shared AI models without pooling patient records

APPI News Editorial — Sat, 08 Aug 2026 17:08:14 GMT

National Taiwan University Hospital said in 2026 that it had built a federated-learning platform with partner hospitals to validate medical AI while keeping each institution's data on site. The hospital's clinical AI validation center lists Fu Jen Catholic University Hospital and Min-Sheng General Hospital among its partners and says the platform is designed for different validation needs without moving data out of the hospitals.

The project is one example of a broader approach to training artificial intelligence across organizations that cannot readily pool sensitive records. Federated learning changes where the computation occurs, but it does not by itself resolve privacy attacks, uneven data quality or responsibility for the resulting model.

How federated learning works

Researchers introduced the modern federated-learning framework at the 2017 International Conference on Artificial Intelligence and Statistics. The paper described a shared model trained by aggregating updates computed on distributed devices, leaving the underlying training data where they originated. Its experiments focused on mobile devices rather than hospitals, but the same architecture can connect institutions that hold separate datasets.

A coordinating server sends a copy of the current model to participating sites. Each site trains that copy on its local records and returns an update; the coordinator combines the updates into the next version of the shared model. The US National Institute of Standards and Technology (NIST) says this design allows participants to send model updates instead of data, which remain inside each organization.

Why hospitals are testing it

Hospital records are fragmented across institutions and subject to legal, ethical and operational controls. Centralized training requires the institutions to copy records into a common repository, creating another system that must govern access, security and retention. Federated learning can reduce that movement, although every participating site still needs compatible data definitions, computing capacity and an agreed training protocol.

Taiwan's Personal Data Protection Act supplies the legal context for the National Taiwan University Hospital project. Article 6 of Taiwan's law restricts the collection, processing and use of medical records, healthcare data, genetic data and physical-examination records, while specifying six legal bases under which those activities may occur. Federated learning is therefore a technical control within a wider compliance process, not an exemption from the law.

A four-hospital study in France

A peer-reviewed study has shown how the approach can work across hospital boundaries. Researchers reported in Nature Medicine in January 2023 that they trained models across four French hospitals to predict how triple-negative breast cancer would respond to preoperative chemotherapy, with patient data remaining behind each hospital's firewall.

The study used pathology images and clinical information from 650 patients. Collaborative training improved performance over local models in the proof-of-concept study, but the authors did not present the system as a deployed clinical service. Access to the underlying patient datasets also remained restricted and required approval from the ethics committee at each center.

Privacy does not end at the hospital firewall

Model updates reflect the records used to produce them. NIST warns that some attacks can recover information about training data from those updates, while other attacks can infer information from the completed model. Defenses include secure aggregation, which limits what the coordinator can see, and differential privacy, which adds controlled noise to reduce the information attributable to individual records.

Those protections bring trade-offs. A 2024 review of privacy preservation in healthcare federated learning found that exchanged information can leak institutional data and surveyed defenses including differential privacy, homomorphic encryption and secure multiparty computation. Hospitals must define the threat model before selecting controls, because no single technique addresses every attack or trust arrangement.

Uneven hospital data can weaken the model

Participating hospitals rarely collect interchangeable data. Patient populations, disease prevalence, scanners, laboratory practices and coding systems may differ, producing what researchers call non-identically distributed data. A global model can perform well on the largest contributors while missing patterns found at a smaller or more specialized hospital.

A 2024 systematic review covering 89 healthcare federated-learning papers identified class imbalance, missing data, distribution shifts and nonstandard variables among the field's recurring problems. The review also found methodological weaknesses across the literature, limiting how confidently results can be compared or generalized.

Governance remains a separate task

A hospital consortium still has to decide who may join, which records are eligible, how updates are audited and who responds when performance deteriorates. Contracts and technical controls must also cover a participant's withdrawal, security incidents and changes to the model after validation. Keeping raw records local narrows one category of exposure; it does not assign accountability.

The National Taiwan University Hospital page describes a platform for cross-hospital validation and a coordinated ethics-review process. It does not provide a nationwide adoption count, a mandatory rollout schedule or a published evaluation of clinical outcomes. APPI News could not verify how many hospitals outside research and validation projects use federated learning in routine clinical care at the time of writing.

Frequently asked questions

Does federated learning anonymize patient records?
No. It changes the training architecture so that records can remain with the institution that holds them. Hospitals still need access controls, legal authority, security measures and protections against information leaking through updates or the final model.

Is it the same as sharing de-identified records?
No. A centralized project transfers copies of records to a common location after applying its chosen de-identification process. A federated project trains locally and exchanges model updates, though those updates can still carry privacy risk.

Does a federated model work equally well at every hospital?
Not necessarily. Differences in populations, equipment and data collection can produce uneven performance. Each institution needs local validation and ongoing monitoring before relying on a shared model.

US clearance brings autonomous retinal AI to a handheld camera

APPI News Editorial — Fri, 07 Aug 2026 16:59:30 GMT

The US Food and Drug Administration (FDA) cleared AEYE-DS for use with the handheld Optomed Aurora retinal camera on April 23, 2024. The US regulator's database identifies the device as diabetic-retinopathy detection software and records the decision as a 510(k) substantial-equivalence clearance.

The decision did not introduce a wholly new algorithm. It expanded an earlier AEYE-DS clearance, which covered the desktop Topcon NW400 camera, by adding a portable camera. That distinction matters because “cleared” and “approved” are not interchangeable terms in the US medical-device system.

What the system does

A healthcare worker captures one macula-centered color image of each eye and submits the images through the AEYE-DS client. The software sends them over a secure internet connection to server-based analysis modules, then returns one of three outputs: more-than-mild diabetic retinopathy detected, not detected or insufficient image quality.

The FDA summary says the prescription-only system is authorized for automatic detection in adults with diabetes who have not previously been diagnosed with diabetic retinopathy. The output is a referral-screening result for that defined condition, not a complete assessment of the eye or a finding about every stage of retinal disease.

The handheld studies produced different error profiles

The FDA summary describes two prospective, multicenter, single-arm and blinded studies conducted in the United States. Both enrolled adults aged 22 or older who had diabetes, no prior diabetic-retinopathy diagnosis and no visual symptoms. Novice operators captured the images, while an independent reading center established the reference result from additional imaging.

One study reported sensitivity of 92 percent and specificity of 94 percent; the other reported sensitivity of 93 percent and specificity of 89 percent. Imageability was 99 percent in both. The lower specificity in the second study means more people without the target level of disease could receive a positive screening result, while sensitivity below 100 percent means the system can miss some cases.

Those figures do not establish identical performance in every clinic or country. Patient mix, disease prevalence, operator technique and image quality can alter real-world results. The studies also evaluated a narrow screening indication rather than every task performed during an eye examination.

Autonomous output changes the clinical workflow

Autonomous screening refers to who interprets the image at the screening step. AEYE-DS generates the defined result without requiring a clinician to grade the retinal photographs. A positive result still leads to evaluation elsewhere in the care pathway; autonomy at the screening step does not turn the software into a treatment system.

The World Health Organization's European office treats screening as a program rather than a stand-alone test. Its 2021 guide says diabetic-retinopathy screening is intended to identify people at higher risk of sight-threatening disease so that those who need further assessment and intervention can enter the next stage of care. A camera, trained operators, referral capacity, quality assurance and follow-up therefore remain part of the service around an autonomous result.

Taiwan's VeriSee DR follows an assisted model

VeriSee DR, developed by Acer with clinicians from National Taiwan University Hospital, illustrates a different allocation of responsibility. A peer-reviewed retrospective study used 1,875 retinal images for validation and reported sensitivity of 89.2 percent and specificity of 90.1 percent for referable diabetic retinopathy. The researchers concluded that the software could assist non-ophthalmologists, while noting that ophthalmologists achieved higher specificity in the comparison.

The product's registered study documents make that boundary explicit. The protocol says VeriSee DR supplies screening results for physicians' reference and is not intended to diagnose or treat diabetic retinopathy. Its assisted role is therefore not equivalent to the automatic output described in the US indication for AEYE-DS.

Taiwan's Ministry of Health and Welfare, which oversees the country's health system and medical-device regulator, is now studying how VeriSee works in practice. A ministry project page published on May 18, 2026 says VeriSee DR and a related macular-degeneration product hold Taiwan medical-device permits and are undergoing clinical-impact and cost-effectiveness evaluation. The page describes physician reference and final clinical decision-making, not an autonomous classification matching the US-cleared AEYE-DS configuration.

Authorization does not transfer across borders

The US and Taiwan records cannot be combined into a global approval claim. Each authorization applies to a named device, indication, workflow and jurisdiction. A performance result from one product also cannot be assigned to another merely because both analyze retinal photographs.

APPI News could not verify a publicly documented Taiwan authorization for an autonomous diabetic-retinopathy system equivalent to the US-cleared AEYE-DS configuration at the time of writing. That finding is limited to the public records reviewed for this report; it does not establish that no such application or evaluation exists.

Frequently asked questions

Did the FDA approve Aurora AEYE in 2024?
The FDA cleared AEYE-DS under the 510(k) pathway for use with the Optomed Aurora camera. The database records a substantial-equivalence decision, not a premarket approval.

Does AEYE-DS diagnose every diabetic eye problem?
No. Its US indication is automatic detection of more-than-mild diabetic retinopathy in a defined adult screening population. The authorized output does not cover every retinal condition or replace the rest of an eye-care pathway.

Are the US and Taiwan systems regulated for the same role?
The records reviewed here describe different roles. The US indication allows AEYE-DS to return the specified screening result automatically, while VeriSee DR's protocol describes results supplied to physicians as assistance for their assessment.

Tesla and SpaceX outline $16.8 billion Texas chip plan

APPI News Editorial — Fri, 07 Aug 2026 05:06:08 GMT

Tesla and SpaceX unveiled a US$16.8 billion first phase for their Terafab semiconductor campus in Grimes County, Texas, on August 6, 2026. The announced plan calls for more than 9.3 million square meters of manufacturing space and at least 3,000 jobs, although the companies have not published a production date.

The proposal is meant to combine advanced logic, memory, packaging and testing at one site. Its long-term target is enough hardware for one terawatt of computing capacity each year, but that figure does not disclose wafer starts, chip volumes or expected yields.

Why Musk's companies want their own fab

Elon Musk has framed Terafab as a response to the amount of computing hardware sought by Tesla, SpaceX and xAI. The proposed output would serve processors for Tesla's Optimus robots and vehicles, along with chips designed for SpaceX's planned orbital computing systems.

Internal production is not presented as a full replacement for foundries. SpaceX's June 2026 prospectus says the company expects to continue sourcing a substantial portion of its computing hardware from third-party suppliers and describes Terafab as complementary to those relationships. The same filing calls the Tesla collaboration a general framework and says individual projects require separate negotiations over capital spending, milestones and development schedules.

The public numbers cover different commitments

The US$16.8 billion figure applies to the newly announced first phase. It should not be combined with earlier estimates as though all the numbers were firm commitments under one contract. A May proposal described US$55 billion in initial projects and a possible US$119 billion after all planned expansion.

The local incentive agreements set a lower enforceable floor. Documents released by Grimes County require SpaceX to invest at least US$5 billion by 2030 and create 1,800 full-time-equivalent jobs by 2035. KBTX reported that falling short does not trigger penalties if the company reaches at least 90 percent of either target.

The county documents also describe infrastructure that would sit outside the regional power grid. They call for on-site natural-gas generation and wastewater treatment, with water intended to come from Gibbons Creek Reservoir rather than local groundwater. The agreements do not publish a firm water-use ceiling.

Intel supplies manufacturing experience

Intel joined the project in April. Intel said its contribution would draw on its ability to design, fabricate and package high-performance chips, while the scope of its participation remained undisclosed. No public statement reviewed by APPI News specifies Intel's investment, production allocation or contractual responsibilities.

That gap matters because Tesla and SpaceX do not operate a leading-edge commercial foundry today. Capital can pay for buildings and equipment, but process integration, stable yields and experienced staff determine whether a fab can deliver chips at volume.

TSMC and Samsung remain in the supply chain

Tesla is still developing AI5 with established foundries. Electrek reported that Tesla had taped out slightly different AI5 versions with Samsung and TSMC, while Samsung was preparing its version for the 2-nanometer process at Taylor, Texas. Tape-out marks the transfer of a finished design into manufacturing preparation; it is not evidence of high-volume output.

TrendForce said Terafab was unlikely to threaten TSMC's advanced-node lead in the short term and identified advanced packaging as a more plausible entry point. The research firm cited yield control, limited access to advanced lithography equipment and shortages of experienced US semiconductor workers as barriers.

What the plan does and does not establish

Terafab gives Musk's companies a route to add internal capacity and shorten feedback between chip design, fabrication and packaging. It could also reduce exposure to tight supplier capacity if the campus reaches commercial yields.

The available documents do not establish when that will happen. They provide no construction-to-production schedule, target wafer volume, yield forecast or complete process roadmap. The announced scale therefore describes an ambition and an initial spending plan, not a functioning alternative to TSMC, Samsung or other suppliers.

Frequently asked questions

When will Terafab begin production?
No public production date has been announced. SpaceX's prospectus says specific projects under the Terafab framework remain subject to separate agreements covering schedules, milestones and spending.

Is US$16.8 billion the total cost?
No total has been fixed in the documents reviewed for this report. US$16.8 billion covers the announced first phase, while earlier applications contemplated a much larger multi-stage buildout.

Will Terafab replace TSMC and Samsung?
The public record does not support that conclusion. SpaceX expects to keep purchasing a substantial share of its computing hardware from outside suppliers, and Tesla's AI5 work with Samsung and TSMC is continuing.

US pilot links premarket digital devices to Medicare payments

APPI News Editorial — Fri, 07 Aug 2026 00:00:00 GMT

The US Food and Drug Administration (FDA) announced its first TEMPO participant on July 22, 2026 and now lists Cadence Solutions and Dexcom in the pilot. Selected digital health devices can enter Medicare-supported chronic care before the agency grants marketing authorization for the uses under review. The FDA says the devices' effectiveness for those intended uses has not yet been evaluated by the agency, and each manufacturer must collect, monitor and report real-world data.

The Technology-Enabled Meaningful Patient Outcomes (TEMPO) pilot is paired with the US Centers for Medicare & Medicaid Services' Advancing Chronic Care with Effective, Scalable Solutions (ACCESS) model. TEMPO does not confer clearance or approval. It lets the FDA decide not to enforce specified requirements in defined circumstances while ACCESS supplies a payment route for the care in which a selected device is used.

Two agencies control different parts of the route

ACCESS began on July 5, 2026 as a 10-year voluntary model in Original Medicare, the US public insurance program in which the government pays participating providers directly. CMS says participating care organizations receive recurring payments for managing qualifying chronic conditions, with full payment tied to measurable patient outcomes rather than a list of individual services.

TEMPO addresses the regulatory side. A manufacturer may ask the FDA to exercise enforcement discretion when its device is offered to or by an ACCESS participant for an intended use connected to covered care. The FDA then determines which requirements it will not enforce and under what conditions for that device.

Enforcement discretion is not authorization

The distinction changes what the public can infer about a participating device. Marketing authorization follows an FDA review under an applicable pathway, such as a 510(k) premarket notification or premarket approval. Enforcement discretion is an agency decision about whether to enforce particular legal requirements in a specified setting; it is not an affirmative determination that a device is safe and effective for the use being evaluated.

The FDA says a TEMPO request may cover premarket authorization, investigational device exemption requirements and the informed-consent and institutional-review-board rules in Parts 50 and 56 of Title 21 of the US Code of Federal Regulations. The exact scope is participant-specific. The agency may also set conditions involving labeling, records, risk controls and reporting.

The boundary around the pilot is narrow. A manufacturer cannot use TEMPO to market the same unapproved intended use outside ACCESS. The FDA also expects participants eventually to seek the appropriate marketing authorization, using pilot data and any additional evidence needed for a formal submission.

Eligibility limits the size and scope of the test

The FDA plans to select up to about 10 US-based manufacturers in each of four ACCESS clinical areas. Those areas cover early cardio-kidney-metabolic conditions; diabetes, chronic kidney disease and atherosclerotic cardiovascular disease; chronic musculoskeletal pain; and depression or anxiety. Devices must be finished products intended for clinician-supervised outpatient care and must not present the potential for serious risk to patient health, safety or welfare.

Selection is not based on a brief product description alone. The FDA says it may examine evidence that a device can function as designed, the manufacturer's quality system, plans to mitigate risk, proposed outcome measures and a statistical plan. It may also seek a schedule for real-world data collection, interim reports and a later marketing submission.

The two listed participants illustrate the range. Cadence Solutions' HypertensionOS is under evaluation for a clinician-supervised role in starting and adjusting antihypertensive medication under predefined rules. Dexcom's Glucose Health Program is under evaluation for metabolic monitoring, tailored guidance, AI-supported insights and screening assistance related to prediabetes and type 2 diabetes.

Payment and oversight remain conditional

ACCESS pays for a package of technology-supported care, not simply for possession of a device. Participating organizations must enroll in Medicare Part B, comply with applicable US federal and state rules, and appoint a physician clinical director to oversee quality and compliance. CMS says it will monitor clinical outcomes, claims and program-integrity data and may remove organizations that fail quality, safety or compliance standards.

The model is limited to Original Medicare. Medicare Advantage plans, which deliver Medicare benefits through private insurers under contracts with the US government, may establish similar arrangements independently, but they are not participating in the ACCESS test itself. Patients enroll voluntarily and retain their existing Medicare rights and access to other Medicare providers.

This structure connects evidence collection, regulatory oversight and reimbursement more directly than a conventional sequence in which coverage is considered only after authorization. It also moves some evidence generation into routine care. The trade-off is that patients and clinicians may encounter a device before the FDA has evaluated its effectiveness for that use, making the pilot's reporting, supervision and exit controls central to its credibility.

Frequently asked questions

Are TEMPO devices FDA-approved?
No. The FDA may decline to enforce specified premarket and investigational requirements for a selected device in defined ACCESS use. The agency explicitly says it has not yet evaluated the listed devices' effectiveness for the intended uses under review.

Can a manufacturer sell the same use outside ACCESS?
Not under TEMPO. The FDA says offering the device for that intended use outside the ACCESS context falls beyond the pilot and generally requires the appropriate US marketing authorization.

Does ACCESS cover every person enrolled in Medicare?
No. ACCESS is being tested in Original Medicare. Medicare Advantage organizations may choose to build similar payment arrangements with their contracted providers, but those arrangements are separate from the federal model test.

Foldable phones pay off only when the larger screen gets used

APPI News Editorial — Thu, 06 Aug 2026 05:55:28 GMT

Samsung unveiled its eighth-generation Galaxy Z foldables on July 22, 2026, with the Galaxy Z Fold8 built around a 7.6-inch inner display. That larger workspace can reduce app switching, but buyers still carry a thicker phone and face market-specific repair costs.

The practical question is how often the phone will be opened for a task that benefits from more screen area. Reading long documents, keeping two supported apps visible and reviewing content beside a message can use the format. Photography, calls and one-app social feeds do not automatically gain the same advantage.

The larger display must replace a real daily step

Samsung says the Galaxy Z Fold8 has a 7.6-inch main display with a 4:3 aspect ratio, which becomes 3:4 when rotated, and a 5.5-inch cover display. The company presents the two orientations for video, games, e-books and web browsing. Those dimensions create more room than a conventional phone, but screen size alone does not establish that a buyer needs it.

A useful test is to record one week of phone use before buying. Tasks that repeatedly require switching between a document and a message, or zooming and scrolling through a PDF, are candidates for the inner screen. If most sessions stay inside one app, the foldable mechanism adds hardware without removing a recurring step.

App support also matters. Samsung warns that supported apps for split view may vary and that some features differ by country, language, carrier and software version. A buyer should test the exact work, reading and messaging apps involved rather than assume every app will use the larger canvas well.

Closed dimensions remain part of every trip

The main screen is available only after the phone is opened, while its closed dimensions follow the buyer throughout the day. Samsung lists the Fold8 at 201 grams, 4.5 millimeters thick when open and 9.7 millimeters thick when folded.

Apple lists the iPhone 17 at 177 grams and 7.95 millimeters thick. On those manufacturer specifications, the Fold8 is 24 grams heavier and 1.75 millimeters thicker when closed. Cases can widen that difference, so an in-store pocket and one-handed typing test provides more useful evidence than a specification table alone.

IP48 is not a dust-tight rating

Samsung gives the Fold8 an IP48 rating and says its water-resistance testing involved immersion in up to 1.5 meters of fresh water for up to 30 minutes. The company also says water resistance is not permanent and may decline with ordinary wear.

Samsung's guide to the International Electrotechnical Commission's IEC 60529 system says the first digit 4 protects against solid objects 1 millimeter or larger, while ratings 5 and 6 denote dust-protected and dust-tight enclosures. The Fold8's rating therefore should not be read as a promise of dust-tight construction. Sand or dusty work environments deserve more weight in the buying decision than water protection alone.

The crease claim still needs time

Samsung says its Flex Titanium display structure helps minimize the fold crease and improve durability. This is a manufacturer claim about a newly released device, not an independent finding that the crease disappears or stays unchanged over years of use.

APPI News could not find an independent long-term durability dataset for the Fold8 at the time of writing. Buyers who notice reflections or surface changes should inspect the screen under bright light and at an angle before committing. A short demonstration cannot predict years of hinge and display wear, but it can show whether the visible crease is already distracting.

Repair and resale costs cannot be globalized

Samsung's repair prices, warranty coverage and optional protection plans differ across markets. The source report cited a Taiwan repair schedule, but that figure does not describe costs in other countries and is not used here. Buyers should obtain the local out-of-warranty inner-screen price and read the accidental-damage terms before comparing the Fold8 with a conventional flagship.

The same limitation applies to resale value. APPI News could not locate the original dataset behind the source report's average one-year depreciation figures, so those numbers were omitted. Local trade-in quotes for the exact storage capacity provide a more defensible comparison than a cross-market average with an unknown sample.

A four-part test before buying

A foldable has a clear use case when the inner screen replaces frequent app switching, a separate small tablet or repeated zooming through documents. The value is weaker when the phone stays closed for most tasks. Count those openings for a week, test the required apps in split view, carry a device of similar weight, and obtain local repair and trade-in quotes.

The Fold8 shows that a book-style foldable can now weigh close to a conventional flagship while still offering a much larger inner display. It also remains thicker when closed, lacks a dust-tight rating and carries a new display structure without independent long-term evidence. The buying decision turns on repeated use of the open screen, not the novelty of opening it.

Frequently asked questions

Who is most likely to benefit from a foldable phone?
People who regularly read long documents, compare information between supported apps or replace a small tablet with their phone have the clearest use case. Buyers whose daily use centers on calls, photos and one app at a time may see less benefit from the inner screen.

Does the Galaxy Z Fold8 eliminate the screen crease?
No verified evidence supports that conclusion. Samsung says Flex Titanium minimizes the crease, while APPI News could not find independent long-term testing showing that it disappears or remains unchanged over time.

Is IP48 the same as the IP68 rating found on some conventional phones?
No. Under the IEC 60529 scale described by Samsung, the first digit 4 covers solid objects at least 1 millimeter across; the first digit 6 means dust-tight. Both ratings use 8 for water protection, but their certified protection against solids differs.

How should repair cost be compared?
Use the manufacturer's current out-of-warranty inner-display quote for the buyer's country, then add any insurance fee and deductible. Repair prices from Taiwan or another market should not be treated as a global price.

AI memory squeeze pushes up PC and graphics card prices

APPI News Editorial — Thu, 06 Aug 2026 05:55:28 GMT

TrendForce said on July 3 that conventional DRAM contract prices were likely to rise 13 to 18 percent quarter on quarter in the third quarter of 2026. The research firm also forecast a 10 to 15 percent increase for NAND Flash and said higher component costs would continue moving through notebook inventories.

The increases connect the expansion of AI data centers with the cost of consumer electronics, but they do not set one global retail price. Currency movements, taxes, local inventories, promotions and the amount of memory in each configuration can produce different results for buyers in different countries.

Memory prices sit upstream of PCs and graphics cards

Dynamic random-access memory (DRAM) provides working memory in computers and servers. Graphics cards use specialized graphics DRAM, including GDDR6 and GDDR7, while solid-state drives rely on NAND Flash. A higher contract price for those components can raise a manufacturer's bill of materials before the change reaches retail shelves.

The timing is uneven. Computer makers often buy components under contracts and hold finished inventory, so a higher chip price may appear in a new shipment before it affects stock already at a retailer. A sale on one model is therefore weak evidence that the broader supply cycle has turned.

AI servers use more wafer capacity per unit of memory

High-bandwidth memory (HBM) is built to move large volumes of data between memory and AI accelerators. Its manufacturing and packaging demands make a simple comparison by gigabyte misleading. TrendForce News, relaying industry estimates, said 1 GB of HBM used four times the wafer capacity of standard DRAM and 1 GB of GDDR7 used 1.7 times as much.

The same report estimated that AI-related memory could consume nearly 20 percent of global DRAM output in 2026 on a wafer-equivalent basis. That figure is an estimate rather than a consolidated disclosure from memory producers, but it explains why AI demand can tighten capacity faster than shipment volumes alone suggest.

Memory suppliers also have a commercial reason to prioritize these products. SK hynix's January outlook said HBM3E was expected to make up about two-thirds of HBM shipments in 2026 as demand expanded across graphics processors and custom AI chips. The company said investment focused on HBM was also affecting the supply-demand balance for general-purpose memory.

Higher component costs are reaching finished devices

Apple provides a documented example of the pass-through to retail hardware. On June 25, the company raised the US starting price of the 13-inch MacBook Air from US$1,099 to US$1,299 and the base iPad from US$349 to US$449. The report cited Apple Chief Executive Tim Cook as linking the move to rising memory and storage costs and the amount of HBM going into AI servers.

This does not mean every computer or graphics card will rise by the same percentage. Memory represents a different share of the total cost in an entry-level laptop, a workstation and a high-end graphics card. Vendors can also change specifications, reduce discounts or absorb part of the increase rather than lift the headline price immediately.

Why older memory can become expensive

Older memory is not automatically insulated from a shortage. When suppliers reduce output of a mature generation and buyers still need replacement parts, a smaller pool of available chips can produce abrupt price changes. The effect can be sharper for low-volume modules that cannot be substituted without changing a motherboard or device design.

That makes compatibility as important as the market cycle. A buyer considering an upgrade should identify the exact memory generation, capacity and supported module type before comparing prices. Moving from an older platform to a new one can require a new motherboard or processor, so a cheaper new-generation module may not reduce the total upgrade cost.

Signals that matter more than a single retail listing

Contract prices show the direction of large transactions, while spot prices reflect smaller and more immediate trades. TrendForce publishes a DRAM spot-price page with current quotes for selected chip specifications. Neither series is a direct forecast for a particular laptop or graphics card, but a sustained decline across several reporting periods would provide stronger evidence than one retailer's promotion.

Supplier production plans are another signal because new capacity takes time to install and qualify. Buyers can watch whether Samsung Electronics, SK hynix and Micron Technology describe added capacity as under construction, in qualification or in volume production. Reaching volume production is a stronger supply signal than announcing construction.

Forecasts also differ on when conditions may ease. TechCrunch reported in February that IDC expected RAM prices to stabilize by mid-2027, while Counterpoint expected the effect on smartphones to continue through the second half of 2027. Those estimates may move as suppliers change output and device demand responds to higher prices.

How to assess a purchase in a rising market

A required replacement and an optional upgrade call for different decisions. For a required purchase, compare the full configured price across several sellers and check whether memory or storage can be upgraded later. For an optional purchase, a working device leaves more room to wait for contract-price growth to slow or for new inventory to improve competition.

Price history is more useful than the manufacturer's suggested retail price when stock is constrained. The relevant comparison is the same model and configuration in the same country, including taxes and warranty coverage. Cross-border prices can look lower while reflecting different tax treatment, exchange rates or support terms.

Frequently asked questions

Are graphics card and memory price increases part of the same supply problem?
They can be connected because graphics cards contain graphics DRAM and compete within the wider memory manufacturing base. Retail graphics card prices also depend on the graphics processor, board design, inventory and local distribution, so memory alone does not explain every change.

Does a slower quarterly increase mean prices are falling?
No. TrendForce's third-quarter forecast called for smaller increases than in the preceding quarter, but both DRAM and NAND contract prices were still expected to rise.

Which indicator should buyers follow?
Contract-price direction, spot prices and supplier production updates answer different questions. Several periods of easing contract prices, followed by more supply reaching volume production, would be stronger evidence of a turn than a temporary discount on one product.

Passkeys replace SMS codes, but recovery still needs planning

APPI News Editorial — Thu, 06 Aug 2026 05:55:28 GMT

The FIDO Alliance estimated in May 2026 that five billion passkeys were in use worldwide as more services offered passwordless sign-in. Passkeys remove the reusable password or one-time code from the authentication exchange, but their convenience after a lost or replaced device depends on where the credential was saved.

The alliance said its research covered 11,000 consumers and 1,400 corporate decision-makers in 10 countries, with 75 percent of consumer respondents reporting that they had enabled a passkey on at least one account. The figures are industry-group estimates rather than a global census, and they do not measure adoption in Taiwan separately.

Why SMS codes remain a weaker option

An SMS one-time password can be entered into a fraudulent site and relayed to the real service before it expires. It also depends on control of a telephone number, which can change after a SIM replacement or number-porting attack.

The US National Institute of Standards and Technology's 2025 authentication guidelines classify verification over public telephone networks, including SMS, as restricted and tell verifiers to consider SIM changes, device swaps and number porting before sending a code. The guidance requires US federal services using this method to offer an unrestricted alternative, but it does not ban SMS or set rules for services in other countries.

Passkeys address the relay problem by binding each credential to the website or app for which it was created. A fraudulent domain cannot ask the authenticator to sign a challenge for the legitimate domain, while control of a phone number alone does not provide the private key.

How a passkey works

A service creates a public-private key pair during registration. The service keeps the public key, while a phone, computer, security key or credential manager protects the private key and uses it to sign a fresh challenge at login.

The FIDO Alliance says passkeys use FIDO2 standards and can be unlocked with the same biometric check, PIN or pattern used to unlock a device. A fingerprint or face scan normally authorizes the credential locally; the biometric information is not sent to the service.

Passkeys can be synced through a credential manager or kept on one device. That distinction matters more for recovery than the fingerprint or face prompt visible during sign-in, because the prompt does not reveal whether the credential has a cloud-backed copy.

Set up a passkey and record where it is saved

Start in the security or sign-in settings of a service that explicitly offers passkeys. After choosing to create one, check the account name and service shown by the operating system, select the intended credential manager or device when a choice appears, and approve creation with the device unlock method.

Do not create a passkey on a shared or public device unless the flow clearly saves it back to a credential manager under personal control. Anyone who can unlock a device may be able to use a passkey stored on it, subject to the protections and account policies of that platform.

The safest setup for device loss is service-specific. Where the account permits it, users can register passkeys on more than one personally controlled device, use a synced credential manager, or keep another recovery method that the service supports. A device-bound passkey on a computer or hardware security key has no automatic cloud recovery unless its provider explicitly offers one.

Moving to a new phone depends on the credential manager

Apple says passkeys stored in iCloud Keychain sync across a user's Apple devices and can be recovered through its protected keychain-recovery process. Apple requires two-factor authentication for accounts that use iCloud Keychain, so access to the Apple Account and its recovery process remains part of the security model.

Google says passkeys saved in Google Password Manager are backed up to the Google Account and become available on supported devices signed in to that account. A Google Password Manager PIN or an eligible Android device's screen-lock credential protects access to the saved keys.

Microsoft distinguishes passkeys saved to a synced credential manager from passkeys stored locally with Windows Hello or on a physical security key. Its setup flow may offer several storage locations, so the label selected during creation determines whether the passkey follows the user to another device.

A QR code signs in from nearby; it does not always transfer a key

When a passkey is on a phone but the sign-in screen is on another device, the service may display a QR code. Scanning it lets the nearby phone approve that session, usually with a proximity check, while the passkey remains in its original credential manager.

This process can bridge operating systems without automatically moving the credential between them. To keep signing in after the old phone is unavailable, the user may need to create another passkey on the new device through the service's account settings or rely on the service's recovery process.

Support in Taiwan remains difficult to compare

Biometric login inside a banking app is not by itself proof that the service offers a standards-based passkey for account authentication. The reviewed primary sources did not establish a current, comparable list of Taiwan banks and payment providers offering passkeys rather than device-bound app login, so this report does not reproduce the provider claims in the source article.

Availability must be checked in each service's current security settings and official support material. Menu labels, supported browsers, credential-manager choices and recovery routes can change independently, even when two services display a similar face or fingerprint prompt.

Phone updates need a backup and a plan for early problems

APPI News Editorial — Thu, 06 Aug 2026 05:55:28 GMT

Apple released iOS 26.6 on July 27, 2026, with fixes for flaws affecting components including WebKit, Wi-Fi and the operating-system kernel. Phone owners weighing an update should separate security exposure from early performance problems, then verify backups and recovery options before installing it.

A blanket rule to install every update immediately or delay every major release misses that distinction. Security patches close documented weaknesses, while a large feature release can also change app compatibility and performance. The safer decision starts with the release notes, the phone's support status and the apps that must keep working.

Security fixes set a limit on how long to wait

Apple's security notice says iOS 26.6 was released on July 27, 2026, for iPhone 11 and later and lists fixes across WebKit, Wi-Fi, the kernel and other components. The effects described by Apple include access to sensitive data, arbitrary code execution and unexpected app termination. Staying on an earlier release can therefore mean retaining flaws that the newer version addresses.

Android has a similar security consideration, but delivery is fragmented. Google's August 2026 Android Security Bulletin says patch levels dated August 5, 2026, or later address all issues covered by that bulletin. A phone's displayed Android version does not by itself show whether it has that patch level.

That does not make every available download equally urgent. Check whether the release is primarily a security patch, a large operating-system upgrade or both. If the vendor says a flaw is being exploited, the case for prompt installation is stronger than it is for an optional feature release with no stated security fix.

Higher battery use after an update can be temporary

Google says Pixel phones normally use more battery after an update while they download, optimize and set up the new software. The same page cautions that older devices cannot always run newer Android versions and says update schedules differ by device, manufacturer and mobile operator.

This official statement is specific to Pixel devices, not evidence for every Android model or iPhone. Apple separately says network conditions and individual use can affect battery and system performance. Short-term battery drain alone is not enough to diagnose a permanent fault, but severe heat, repeated crashes or lost connectivity warrants checking the manufacturer's support channel rather than waiting indefinitely.

Make four checks before installing

First, identify the type of update. Read the vendor's release and security notes. A monthly security patch presents a different tradeoff from the first public release of a redesigned operating system, and the two can arrive together.

Second, confirm that a current backup completed. Apple tells iPhone and iPad owners to back up with iCloud or a computer before updating, then connect the device to power and Wi-Fi. A backup protects data; it does not guarantee that the old operating system can be restored.

Google says Android backups can include apps and app data, call history, contacts, settings, and SMS and MMS messages. Google also warns that not every app backs up or restores all settings and data, and that a backup from a higher Android version cannot be restored to a phone running a lower version. Check the backup details instead of relying only on the word “On.”

Third, verify the apps that cannot be interrupted. Banking, payment, authentication and workplace apps may have their own supported operating-system ranges. Check each developer's current support page or release notes, especially when the phone is managed by an employer.

Fourth, prepare the installation. Charge the phone, use a stable network and leave enough time for the download, restart and post-update setup. Google recommends Wi-Fi and at least 75 percent battery for Pixel updates; Apple advises power and Wi-Fi for iPhone and iPad updates.

Rollback should not be treated as a safety net

There is no single rollback rule that covers every current iPhone and Android phone. Apple's public consumer guide describes updating to the latest iOS release, while Android installation tools and policies vary among manufacturers, carriers and models. APPI News could not verify an official source for the original report's claim that an iPhone rollback window normally lasts about one week, so that timetable is not used here.

A factory reset is also different from a downgrade. It erases user data and settings and is not by itself a supported procedure for choosing an older operating-system version. Before assuming recovery is possible, check the exact model and build with the manufacturer or an authorized support provider, and confirm whether the backup can be restored afterward.

Respond to problems in stages

After installation, note the time, battery level and affected apps so the pattern can be compared across several charging cycles. Install app updates, restart the phone and check free storage before considering a reset. For a work-managed phone, contact the administrator before removing accounts, profiles or security software.

Escalate sooner if the phone repeatedly shuts down, cannot connect to a required network, loses access to an account or becomes too hot to handle normally. A factory reset should come only after the backup has been checked and the vendor's recovery instructions have been reviewed because it removes local data.

Frequently asked questions

Should a major phone update be installed on release day?
Check the release notes first. A patch for a documented or actively exploited flaw favors prompt installation, while an optional feature release allows more room to confirm app support and early problem reports.

Does battery drain after an update mean the battery is damaged?
Not by itself. Google documents temporary extra battery use during post-update setup on Pixel phones, but continuing drain can also have other causes. Record the change, check battery usage by app and use the device maker's troubleshooting guidance if it persists.

Can a phone always return to its previous operating system?
No. Availability depends on the platform, device, build and vendor policy. Confirm a supported procedure for the exact phone before updating; a data backup and a factory reset do not establish that a downgrade is available.

What is the minimum preparation before an update?
Verify a completed backup, confirm support for required apps, charge the phone and use a stable network. Save access to account-recovery information separately in case an app asks for a new sign-in after the update.

Windows 11 memory fixes start with apps, not RAM cleaners

APPI News Editorial — Thu, 06 Aug 2026 05:55:28 GMT

Microsoft said on July 31, 2026, that it is reducing the Windows memory footprint on PCs with at least 8GB of RAM. Users can act sooner by identifying heavy processes and limiting unneeded startup, background and sync activity, without disabling security features or installing RAM cleaners.

The percentage shown in Task Manager is a starting point, not a diagnosis. Windows divides physical memory among applications, cached files, the kernel and device drivers, so the useful question is what is consuming memory and whether the computer slows under the workload that matters.

Separate memory use from memory pressure

Open Task Manager with Ctrl+Shift+Esc, select Processes and sort the Memory column to find the largest current users. Then open Performance > Memory and compare the figures while reproducing the slowdown. Microsoft's RAMMap documentation explains that Windows memory includes process working sets, cached file data, kernel and driver use, and prioritized standby lists.

A high percentage without stuttering, long waits or an application failure does not establish that the PC has run out of usable memory. Look for repeatable symptoms and a process whose usage grows with them. If Task Manager does not explain the allocation, RAMMap can separate the broad categories and save snapshots for comparison.

Microsoft targets the Windows footprint

Microsoft's July 31 quality update names memory optimization for PCs with 8GB or more as one of four new focus areas through the end of 2026. The company says it has been adopting a more efficient memory allocator, tuning WinUI 3 so applications use less memory by design, and improving Chromium and WebView2 components that appear in Windows.

The post says a broader rollout of quality improvements begins in fall 2026, but it does not assign a release date to the memory work or quantify the expected reduction. It should therefore be treated as a product plan, not evidence that a particular Windows update has already lowered memory use on every PC.

Operating-system changes and local cleanup address different parts of the problem. Microsoft's work may reduce baseline overhead, while a user can stop software that has no reason to launch, sync or run in the background on a particular machine.

Start with processes and automatic launches

Close an unused application from its own menu before ending its process in Task Manager. If the application returns after every sign-in, review Settings > Apps > Startup or the Startup apps page in Task Manager. Microsoft says both interfaces can disable registered startup applications, while Task Manager also shows each app's startup impact.

Disable only software that does not need to be ready at sign-in. Security software, device utilities and workplace management tools may serve functions that are not obvious from their names. On a managed computer, leave organization-controlled entries to the administrator.

Next, check background permissions for applications that keep running after their windows close. Microsoft's performance guide directs Windows 11 users to Settings > Apps > Installed apps > Advanced options to set an app's background permission to Never. The same guide recommends reviewing resource use in Task Manager, closing unused apps and browser tabs, and turning off unneeded notifications.

Pause OneDrive only when synchronization is the load

OneDrive can be paused when active synchronization coincides with the slowdown. Microsoft's OneDrive instructions offer pause periods of two, eight or 24 hours from the taskbar cloud icon, after which synchronization resumes automatically unless it is resumed sooner.

Pausing sync is a test and a temporary tradeoff, not a universal memory fix. Files changed during the pause will not synchronize until OneDrive resumes, and the step will have little effect when OneDrive was idle. Work or school accounts may also be governed by an organization's policies.

Reduce visual effects after removing unnecessary activity

Windows 11 animations and other visual effects consume system resources. Microsoft's performance guide says to search for “Adjust the appearance and performance of Windows,” open the Visual Effects tab and select Adjust for best performance. This changes the interface's appearance, so it is most useful when responsiveness matters more than animation and transparency.

The guide also supports disabling unneeded background applications and startup entries. None of these settings guarantees a fixed reduction in gigabytes because the result depends on the applications, hardware and workload. Change one group of settings at a time, restart if required and repeat the same workload before judging the result.

Do not trade kernel protection for a cleaner graph

Memory integrity is a security control, not an ordinary background application. Microsoft says the feature uses virtualization-based security to isolate kernel-mode code-integrity checks and restrict kernel memory allocations that malware could exploit. Windows 11 displays a security warning when Memory integrity is turned off.

Older processors can experience a larger performance effect from the feature, according to Microsoft, but disabling it removes a documented layer of kernel protection. This guide does not recommend turning it off to reduce a memory percentage. Compatibility problems should be handled through Microsoft's troubleshooting guidance, driver updates or the device administrator.

RAM cleaners can obscure the cause

A utility that makes the memory percentage fall immediately has not necessarily fixed the workload that caused a slowdown. It may alter cached or standby data while leaving the application, startup task or driver responsible for repeated pressure untouched. APPI News could not verify a reliable benchmark that supports a general performance claim for RAM-cleaner utilities across current Windows 11 hardware.

Use Task Manager or RAMMap to identify the allocation first, then remove the responsible workload through its own settings. Installing another always-running utility adds a process and expands the software that must be kept updated. The practical sequence is diagnosis, a targeted change and a repeatable before-and-after test.

Frequently asked questions

Does 70 percent memory use mean Windows 11 needs more RAM?
No conclusion follows from that percentage alone. Check whether the machine slows during a repeatable workload, identify the largest processes and separate application use from cached, kernel and driver allocations.

Which setting should be changed first?
Start with an application or browser tab that Task Manager shows consuming memory and that is not needed. Then review unnecessary startup apps, background permissions and active synchronization in that order.

When will Microsoft's memory optimization arrive?
Microsoft says broader Windows quality improvements start rolling out in fall 2026 and lists memory optimization among work continuing through the end of the year. It has not published a memory-specific completion date or a measured reduction.

Should Memory integrity be disabled on an 8GB PC?
This guide does not recommend it. Microsoft describes Memory integrity as protection against malware targeting the Windows kernel, and turning it off causes Windows 11 to display a security warning.

Taiwan care-device subsidy leaves performance details hard to compare

APPI News Editorial — Wed, 05 Aug 2026 17:09:07 GMT

Taiwan added 17 smart assistive products to a long-term-care rental benefit on July 1, 2026. Taiwan's Ministry of Health and Welfare, which administers the country's health and social-care policies, said the first group covers mobility, toileting and bathing, home care beds, and safety monitoring.

The benefit provides up to NT$60,000 every three years. That was about US$1,900 using Bank of Taiwan's quoted spot rate on August 9, 2026. The larger issue for families and care assessors is what the public documents do not show: product-level error rates, test conditions and an explanation of how each monitoring system reaches a decision.

The published list does not define the technology

The ministry uses “smart assistive products” as an administrative category. It says eligible equipment can be assessed for function, information and communications features, and safety certifications, but that wording does not mean every listed product uses artificial intelligence.

The ministry's one-page list identifies each product by name, model, supplier and payment code. Several names include “AI” or “AIoT,” while others use terms such as “smart” or “sensing.” The document does not state whether a device uses a learned model, fixed rules or both, and it does not report false alarms or missed events.

AI and fixed thresholds can operate in the same system

A simple AI-versus-sensor split can also misdescribe how fall-detection systems work. Sensors collect signals; software then interprets those signals through rules, statistical models or a combination of methods. The label attached to the product does not reveal that processing chain.

One research system, mmFall, shows the overlap. The system used millimeter-wave radar point clouds and a recurrent autoencoder to learn patterns in normal movement, then combined the model's anomaly score with a fixed 0.6-meter threshold for a drop in body-centroid height. Calling it either purely adaptive AI or purely threshold-based would leave out part of its design.

The researchers detected 49 of 50 staged falls with two false alarms in an apartment dataset. That result describes a prototype and its chosen test conditions, not the 17 products in Taiwan's program. It also illustrates why a percentage without the number and type of trials gives buyers little basis for comparison.

Test setting matters as much as the headline number

A separate peer-reviewed study published in February 2026 tested a multi-radar system in a 12-by-12-meter indoor area. The study reported 97.9 percent overall accuracy across simulated multi-person scenarios, while the false-negative rate rose as more people entered the monitored area.

The experiment involved 10 healthy adults, not frail older adults experiencing unplanned falls. Its authors said simulated soft falls could not fully represent real events and that obstacles may affect performance. Those limits are material in homes, where furniture, room layout, visitors and the position of the sensor vary.

Earlier evidence points to the same validation gap. A systematic review of 92 fall-detection projects found that only 7.1 percent of wearable-device projects had monitored older adults in real-world settings, while none of the nonwearable projects had tested with older adults in laboratory or real-world settings. The review covered literature through June 2013, so it does not measure the current market, but its call for standardized, real-world evaluation remains relevant to how performance claims are read.

Procurement needs specifications, not marketing categories

The distinction matters beyond whether a product carries an AI label. Monitoring equipment can differ in the events it detects, the rooms and positions in which it was tested, the time before an alert is sent, and the rates of false alarms and missed events. Camera use, local or cloud processing, retention periods and access permissions create separate privacy questions.

World Health Organization guidance for assistive-product procurement says specifications should set minimum requirements for technical performance and function. The guidance covers 26 product types and is not a standard for home fall detectors, but its procurement principle provides a useful benchmark: a category name alone is not a performance specification.

Taiwan's announcement says professional groups review product function, communications features and safety certifications before assigning payment codes. The public list does not show the resulting evidence at product level. APPI News could not find public false-alarm and missed-event results for all 17 products at the time of writing.

Frequently asked questions

Does Taiwan's “smart assistive product” label mean a device uses AI?
No such conclusion can be drawn from the published list. The list includes AI-branded and smart-branded products but does not describe each product's decision method.

Is a machine-learning fall detector always better than a fixed-rule device?
The cited evidence does not support that general claim. Performance depends on the sensor, algorithm, installation, monitored population and test conditions. Some systems combine learned models with fixed thresholds.

What information is missing from the subsidy list?
The one-page document does not publish product-level false-alarm rates, missed-event rates, evaluation populations, room configurations or decision methods. Those details would be needed for evidence-based comparisons.

ChatGPT Health and Taiwan take different paths to health data

APPI News Editorial — Wed, 05 Aug 2026 00:00:00 GMT

OpenAI began rolling out Health in ChatGPT to US users aged 18 and older on July 23, 2026, allowing them to connect Apple Health and supported medical records. Taiwan uses a different route for records held through its national insurance system: people authorize selected data to move from My Health Bank to an app reviewed by the government.

The two services are not direct substitutes. Health in ChatGPT is a consumer AI service whose operator manages the connection and its partners, while Taiwan's software development kit (SDK) is a government-controlled gateway used by multiple app operators. Their differences are clearest in who approves access, where data go and which organization can act when a connection is misused.

ChatGPT connects records inside one consumer service

OpenAI's July launch notice says US adults can connect Apple Health and supported records from US hospital systems, One Medical or Function Health. ChatGPT asks for permission by default before using connected information in a response, although an account holder can grant continuing access. The information can inform conversations outside the Health area after permission is given, reversing the early product design in which the dedicated space was the only place that could use it.

OpenAI says connected medical records, Apple Health information and conversations that use those data are not used to train its foundation models or target advertisements. Disconnecting a source stops future access and starts deletion of synchronized data from OpenAI's systems within 30 days. Information already placed in conversation history remains until those conversations are deleted.

Taiwan places a gateway between records and apps

Taiwan's National Health Insurance Administration (NHIA), which runs the territory's single-payer health insurance program, built My Health Bank to give insured people access to their claims-linked health records. The agency says its SDK lets a person select records covering visits, medications and test results for a specified period and authorize their transfer to a trusted third-party app.

The app, rather than the NHIA app, starts the connection. The person then verifies identity, chooses the categories and period to share, and grants a one-time authorization. This model does not put every participating app inside one government product; it controls the point at which an outside operator receives data from My Health Bank.

The NHIA's current service page says new first- and second-stage applications to connect through the SDK are temporarily suspended because of operational adjustments. The notice does not give a resumption date. Existing connected apps remain listed separately.

Review and consent do different jobs

OpenAI controls which record sources and apps its service supports. Its published privacy notice describes how Health Features collect and use linked records, but it also makes clear that processing is not confined to a single corporate system. OpenAI says authorized personnel and trusted service providers may access Health Feature data for model-safety work unless the user opts out, and vendors may process data for functions such as hosting, support and safety monitoring.

Taiwan's SDK adds a public gate before an app can connect. The NHIA says connected apps must pass its review and meet Taiwan's published mobile-app security requirements; transfers use encryption, each release requires authorization, and the agency retains access logs for investigations. The agency may terminate an app's SDK service if medical information is leaked or misused.

Those controls address access and traceability, not the clinical quality of an app's interpretation. An app can satisfy identity, encryption and recordkeeping requirements without proving that every AI-generated explanation is accurate. OpenAI likewise reports product testing with physicians, but its July notice says ChatGPT can still make mistakes and does not replace professional judgment.

Accountability follows the data path

Health in ChatGPT gives one account holder a set of controls managed through OpenAI: permission prompts, disconnection, memory settings and deletion. Its responsibility chain also includes medical-record sources and service providers described in OpenAI's privacy notice. Which consumer-health privacy rules apply can depend on where a user lives, so the US launch does not establish a single complaint or enforcement route for every future market.

The Taiwan model divides responsibility between the NHIA and each connected app. The agency can review access, retain transfer logs and end the SDK connection. Once a third-party app has received authorized records, however, a person seeking to stop processing or request deletion must contact that app's operator, according to the NHIA.

The current app count is less certain than the governance design. The source report cited 31 organizations and 64 apps in January 2026, but the NHIA's current public directory names connected apps without stating an overall total. APPI News could not reconcile the earlier figure with the live directory at the time of writing.

Two models, two limits

OpenAI's model can make linked information available within a broad AI service under one set of account controls. Taiwan's model gives a public agency control over entry to a national record gateway while leaving services and downstream storage to separate apps. Neither structure alone establishes the accuracy of an AI-generated health explanation.

The comparison also has a geographic limit. Health in ChatGPT currently supports connected medical records in the United States, while My Health Bank is tied to Taiwan's insurance system. Claims about which model offers stronger accountability therefore depend on the specific operator, data recipient and law involved, not only on whether the gateway is run by a company or a government agency.

Coldcard firmware bug routed seed generation through weak randomness

APPI News Editorial — Wed, 05 Aug 2026 00:00:00 GMT

Coinkite warned on July 30, 2026, that seeds created by several Coldcard hardware-wallet models on affected firmware may be exposed because seed generation used a general-purpose pseudorandom generator. The company said fixed firmware is available for every affected model and release track, but an update cannot repair a seed that was generated earlier.

The incident was not described as malicious code running on the device or a compromise of its firmware-signing system. The failure occurred when the firmware created the secret seed from which wallet keys and addresses are derived. A seed built from a small, predictable search space can still look valid and operate normally while allowing an attacker to calculate candidate keys away from the device.

A disabled feature still supplied the active function

Coldcard's design was meant to use a hardware true random-number generator for seed creation. The firmware set the MicroPython option MICROPY_HW_ENABLE_RNG to zero because Coldcard supplied its own hardware-backed implementation. Zero disabled MicroPython's hardware route, but it did not prevent a software implementation with the same interface from entering the build.

Coinkite's technical account says the 2021 migration to the libNgU library changed seed generation from ckcc.rng_bytes() to ngu.random.bytes(), and the resulting rng_get() symbol resolved to MicroPython's Yasmarang generator. Yasmarang was a general-purpose pseudorandom generator, not the hardware source Coldcard intended to use for cryptographic seed generation.

A related preprocessor check used #ifndef MICROPY_HW_ENABLE_RNG. That condition asks whether the macro exists, not whether its value is nonzero. Because the macro existed with a value of zero, the guard did not stop the build. The software fallback and the intended board-specific implementation also exposed the same function signature, allowing compilation and linking to finish without an obvious error.

The output looked random while the search space shrank

A pseudorandom generator can produce bytes that appear evenly distributed even when its internal state came from predictable inputs. Hashing those bytes can change their appearance and remove statistical bias, but it cannot create information that was absent from the input. An observer who can reconstruct the small set of possible starting states can generate the same candidate seeds and test which ones control funded addresses.

Coinkite estimated that affected Mk2 and Mk3 seeds had about 40 bits of effective search space under its attack assumptions. It estimated about 72 bits for Mk4, Mk5 and Q because later models mixed in entropy from two secure elements. The company labeled those figures preliminary, and APPI News has not independently reproduced them. They should be read as the vendor's current risk model rather than final measurements.

The US National Institute of Standards and Technology's SP 800-90B sets out design principles, validation requirements and health tests for entropy sources used with deterministic random-bit generators. This separation matters: a test of output format or apparent randomness does not by itself prove that the input supplied enough unpredictable information.

Review checked the right code but not the executed path

The intended hardware implementation was present in the firmware binary, according to Coinkite. Earlier review confirmed that implementation but did not verify which definition of rng_get() the seed-generation call actually reached across the linked components. Reading either codebase alone would not necessarily reveal the path selected in the finished artifact.

This is a build-integrity failure as much as a source-code defect. Security review that stops at individual functions can miss symbol resolution, compile-time flags and dependency defaults. Reproducible builds show that a published binary matches source and configuration; they do not prove that the configuration selects the security mechanism engineers intended.

Coinkite said its hotfix excludes the MicroPython fallback object and adds a build-time symbol check. The build now fails unless the board-specific object defines the global rng_get() symbol and the upstream fallback contributes no symbols. That converts an architectural assumption into a property the toolchain must verify.

Security-sensitive fallback should fail visibly

The incident points to three checks for systems that generate keys, tokens or identifiers. First, a missing cryptographic entropy source should stop the operation instead of selecting a weaker generator. Second, boot and integration tests should identify the implementation that actually serves the request. Third, release review should cover compiled configuration, linked symbols and call reachability, not only source-level algorithms.

These checks address different stages of the same failure. A build-time assertion prevents the wrong object from shipping. An integration test exercises the production call path. A runtime health test detects failures after deployment. None can infer lost entropy by looking only at well-formatted output after the fact.

Affected versions require a new seed, not only an update

The affected Mk2 and Mk3 range is firmware 4.0.1 through 4.1.9, with version 4.2.0 or later listed as fixed. For Mk4 and Mk5, the fixed releases are standard version 5.6.0 or later and Edge version 6.6.0X or later. For Q, they are standard version 1.5.0Q or later and Edge version 6.6.0QX or later. Coinkite cautions that standard and Edge are separate release tracks.

The vendor's migration guidance is to install the correct fixed firmware, generate a new seed, verify the backup, wallet fingerprint and receive address, and send a small test transaction before moving the remaining funds. It also says to retain the old backup until migration is confirmed. Updating the firmware without replacing an affected seed leaves the original key material unchanged.

Coinkite says seeds created with at least 50 fair, independent and private dice rolls are not at risk from this random-number issue alone. A strong, unique BIP-39 passphrase supplies a separate barrier, but the company still advises migration because the passphrase does not repair the seed. TAPSIGNER, OPENDIME and SATSCARD use different codebases and are not covered by the advisory.

The broader failure mode extends beyond wallets

Any system that depends on unpredictable values can fail in the same quiet way. Authentication tokens, password-reset links, cryptographic nonces and pseudonymous identifiers may remain syntactically correct when a weak generator supplies them. The practical question is not whether the output resembles random data, but whether the deployed call path reaches a validated entropy source with the expected security properties.

Coldcard's five-year exposure followed a March 2021 integration change, not a runtime hardware failure. The build kept succeeding, the intended function remained present and generated wallets continued to operate. That combination explains why ordinary functional testing offered little warning: every visible feature worked while the security margin depended on a different implementation than reviewers believed.

AI benchmarks reward confident guesses over admitted uncertainty

APPI News Editorial — Tue, 04 Aug 2026 17:07:03 GMT

A September 2025 paper argued that language-model training and widely used benchmarks reward guessing over acknowledging uncertainty. The problem becomes harder to detect when a model presents a false statement in fluent, confident language without signaling what it does not know.

The risk is not confined to factual trivia. A 2026 audit of five public chatbots found that 49.6 percent of their answers to health questions were somewhat or highly problematic, while Taiwan's health ministry has classified incorrect generative-AI output as a patient-safety risk for medical institutions.

Hallucination describes a plausible falsehood

An AI hallucination is a generated statement that appears plausible but is false. The term covers invented facts and citations, unsupported claims and answers that conflict with available evidence. Fluency is part of the risk because grammar and specificity can make an unsupported answer look verified.

The September 4, 2025, paper by Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala and Edwin Zhang argues that hallucinations emerge from statistical errors during pretraining and persist because common evaluations reward guessing. The analysis does not depend on a single chatbot or transformer design, and it does not claim that every model produces the same errors at the same rate.

The authors demonstrated the behavior by asking an open-source model for Kalai's birthday and telling it to answer only if it knew. Across three attempts, the model supplied three different dates, all wrong. The example shows that an instruction to withhold an uncertain answer does not necessarily overcome the model's learned incentive to provide one.

Pretraining creates errors before a chatbot answers a question

During pretraining, a language model learns a probability distribution from large collections of text. It does not store a verified ledger of every statement and then consult that ledger before generating a response. For facts with little or no learnable pattern, many plausible strings may fit the surrounding language even though only one is correct.

The 2025 paper models this as a classification problem: deciding whether a candidate output is valid can itself be difficult, and generation requires making that distinction across possible responses. Its statistical lower bounds apply even under an assumption of error-free training data. Real training collections can add ambiguity, contradictions and outdated material, but those are not required for the paper's core mechanism.

This does not mean that every error is unavoidable after training. The authors distinguish the behavior of a pretrained base model from a finished system that can abstain, retrieve information or use other controls. Their narrower conclusion is that the usual pretraining objective creates predictable error pressure, while later design choices determine how much of that pressure reaches users.

Binary benchmarks make guessing pay

Many evaluations give one point for a correct answer and zero for either a wrong answer or an admission of uncertainty. Under that rule, abstaining cannot improve the score, while guessing has some chance of earning a point. A model optimized for leaderboard performance can therefore appear stronger by attempting questions it cannot answer reliably.

The authors proposed adding explicit confidence targets to existing evaluations: a correct answer would earn one point, an abstention would earn zero, and an incorrect answer would incur a penalty tied to the stated threshold. At a 0.75 confidence target, for example, an error would lose two points; at 0.9, it would lose nine. The proposal is intended to make withholding an answer rational when estimated confidence falls below the threshold.

The reform remains a proposal, not evidence that major benchmark operators have changed their scoring. It also addresses one incentive rather than every source of false output. Retrieval can supply better evidence, and post-training can teach abstention, but either control can still fail when the source is poor or the system applies it incorrectly.

Health-chatbot audit found problems in half of responses

A BMJ Open study published in April 2026 tested free versions of ChatGPT, Gemini, Meta AI, Grok and DeepSeek with 50 health prompts per model and rated 49.6 percent of the 250 responses as problematic. Thirty percent were somewhat problematic and 19.6 percent were highly problematic. The questions covered cancer, vaccines, stem cells, nutrition and athletic performance and were designed to pressure the systems toward misinformation or contraindicated advice.

Two subject specialists rated answers in each category using predefined criteria. The study found no statistically significant difference in overall response quality among the five chatbots, although Grok produced more highly problematic responses than expected under the study's comparison. The tested versions were available between 2022 and 2024, and the prompts were run in February 2025, so the results do not establish the performance of later releases.

Citations did not supply a dependable shortcut for judging the answers. Across 25 prompts that requested 10 references from each model, the chatbots returned 1,013 of the requested 1,250 references. Median completeness was 40 percent, and no chatbot produced a fully complete and accurate reference list for any prompt under the study's method.

Taiwan guidance treats hallucination as an institutional risk

Taiwan's Ministry of Health and Welfare, the national authority responsible for health policy, issued guidance on May 29, 2026, for hospitals and clinics preparing to introduce or already using generative AI. The document is administrative guidance rather than a binding rule, and it does not govern consumer chatbots outside regulated medical settings.

The guidance lists plausible but incorrect output as one of six risk categories and says medical institutions should tell patients and families how a generative-AI system is involved, what it does and where its limits lie. When a system interacts directly with the public, the institution should disclose that AI is operating it and warn that its output may contain hallucinations or other errors. The document also leaves final confirmation and responsibility with qualified medical personnel when AI is used in clinical judgment, patient communication or medical records.

This is a national example, not a standard for all English-speaking markets. Its value for international readers lies in the control structure: identify the system's role, disclose its limits, monitor output and preserve a named human decision-maker. Legal duties and enforcement differ by country.

Controls need to address both evidence and incentives

A source link is useful only when it resolves to a real document and supports the statement attached to it. Systems that retrieve from a controlled collection can narrow the material available for an answer, but retrieval does not prove that the model interpreted the document correctly. Human review remains a separate control in settings where an error can affect health, legal rights or finances.

Evaluation design determines what developers optimize. If benchmarks reward every attempted answer and do not distinguish an honest abstention from an error, reported accuracy can favor overconfident behavior. Confidence-aware scoring can change that incentive, but its effect must be measured after adoption rather than inferred from the proposal alone.

The evidence supports a limited conclusion: hallucination is not adequately described as a rare software malfunction. It arises from how generative models learn and how their outputs are rewarded, while its practical harm depends on the safeguards around a particular use. A fluent answer, even one with citations, still requires verification against the underlying source.

AI-designed drugs advance in trials with no US drug approval

APPI News Editorial — Tue, 04 Aug 2026 00:00:00 GMT

AI-designed drug candidates have moved into mid- and late-stage trials, but APPI News found no publicly identified US approval for a novel drug originating from an AI-led discovery program as of August 4, 2026. A June 2026 industry commentary estimated that roughly US$60 billion had entered the field since 2019 and about 175 AI-originated programs had reached human trials, with none approved by the US Food and Drug Administration. The estimate is not an official FDA classification, and the agency's public approval lists do not identify how a drug was discovered.

The distinction matters because AI can affect several different steps: selecting a biological target, designing a molecule, predicting its properties or organizing a trial. Faster work before human testing does not remove the need for clinical evidence. Current results suggest that AI may improve the odds of producing a viable early candidate, but they do not yet show a higher approval rate.

Investment claims run ahead of approval evidence

A BCC Research release published July 14, 2026 said recent AI drug-discovery investment exceeded US$2 billion and claimed that discovery timelines could fall from four or five years to 12 to 18 months. It also projected research and development cost reductions of 30 to 40 percent. Those figures come from a commercial market report and describe the discovery phase, not the full route through human trials and regulatory review.

The reported US$60 billion cumulative total and 175-program count cover a wider period and use another publisher's definition of an AI-originated program. The two sets of figures should therefore not be combined into a single performance measure. Funding can expand the number of targets and molecules tested, but an approval requires evidence that a particular product meets the regulator's standards for its intended use.

Early trials show an advantage on limited evidence

A 2024 analysis in Drug Discovery Today reported Phase 1 success rates of 80 to 90 percent for molecules discovered by AI-native biotechnology companies, compared with historical industry ranges of 40 to 65 percent. Phase 1 trials mainly examine safety, tolerability and how a drug moves through the body. A higher rate at that stage is consistent with better selection of molecules that have usable drug properties, but it does not establish that they treat disease effectively.

The same analysis found a Phase 2 success rate of about 40 percent, comparable with historical industry results. That estimate came from only 10 Phase 2 trials, making broad conclusions premature. The paper's definition also included several ways AI could contribute to discovery, a boundary that is not standardized across companies or databases.

Rentosertib supplies a mid-stage test

Rentosertib, an experimental treatment for idiopathic pulmonary fibrosis, is among the most advanced examples. Its developers used generative AI to identify the TNIK target and design the molecule. A peer-reviewed Phase 2a trial published in Nature Medicine enrolled 71 people for 12 weeks and reported a mean forced-vital-capacity increase of 98.4 milliliters in the highest-dose group, compared with a 20.3-milliliter decline in the placebo group.

Lung function was a secondary endpoint, and each treatment group included only 17 or 18 participants. Treatment-emergent adverse events occurred at similar frequencies across the groups, while liver toxicity or diarrhea accounted for the most common events leading participants to stop treatment. The authors said the findings supported further investigation in larger, longer trials rather than a conclusion that the medicine works.

A failed ALS program shows where prediction stops

Other candidates have not carried an early scientific rationale into patient benefit. Verge Genomics used AI to identify VRG50635, which targeted the enzyme PIKfyve for amyotrophic lateral sclerosis. BioPharma Dive reported in May 2026 that the company ended the program after an early-stage trial failed to produce the intended benefit and a nerve-damage biomarker moved in the wrong direction.

The failure does not show that AI drug discovery as a whole is ineffective. It shows that a model's target selection and molecule design remain hypotheses until experiments and human trials test them. Biological response, toxicity, dosing, endpoints and patient selection can defeat a candidate after computational screening has finished.

The approval gap is real but hard to count

The FDA's public lists identify approved products and uses, not whether AI contributed to target selection or molecular design. APPI News could not independently recreate the 175-program pipeline or the zero-approval figure from agency records alone. The MedCity News estimate remains the clearest current public count located for this report, but its result depends on what qualifies as AI-originated.

That uncertainty does not erase the broader evidence gap. Published results support faster nomination of candidates and a possible Phase 1 advantage, while Phase 2 evidence remains sparse and mixed. Mid- and late-stage readouts will show whether more AI-originated candidates survive the parts of development that test clinical benefit rather than molecular promise.

Frequently asked questions

Has the FDA approved an AI-discovered drug?
APPI News found no publicly identified FDA approval for a novel drug originating from an AI-led discovery program as of August 4, 2026. The FDA does not categorize approvals by discovery method, so the answer relies on published pipeline tracking rather than an agency field that can be searched directly.

What has AI changed in drug development?
Its clearest demonstrated role is before and around early clinical testing: prioritizing targets, designing or selecting molecules and supporting trial planning. Evidence that it shortens the entire clinical and regulatory process has not been established.

Do higher Phase 1 success rates predict approval?
Not yet. The published comparison covered a small early cohort, and its Phase 2 results were similar to historical industry rates. Approval also requires later evidence of efficacy, safety and manufacturing quality.

Abbott's AI glucose forecast has no documented Taiwan release

APPI News Editorial — Mon, 03 Aug 2026 17:12:33 GMT

Taiwan-listed Bionime secured a local medical-device license for its iFree2 continuous glucose monitor in April 2026, but the public notice did not identify an artificial intelligence forecasting feature. Abbott introduced that kind of prediction in the United States three months earlier through Libre Assist, creating a gap between the AI function it promotes and the products documented in Taiwan.

The gap is not simply a difference in app design. Continuous glucose monitoring (CGM) combines a wearable sensor, software and country-specific regulatory authorization. A feature shown in one market does not establish that the same app and compatible sensor are available through official channels elsewhere.

Libre Assist predicts a meal's possible effect

Abbott announced Libre Assist on January 5, 2026 as a US feature within its Libre app that uses generative AI to estimate how a food choice could affect glucose levels. A user takes a photograph or enters a text description of a meal. The software identifies ingredients, assigns a color-coded estimate and later compares that estimate with readings from a Libre sensor.

The announcement describes a consumer-facing prediction layer rather than a new sensing method. The sensor still measures glucose in interstitial fluid, while the AI processes information about the meal and the user's later readings. Abbott says the feature is intended to inform food choices before eating, but its safety note says generative AI may be inaccurate and should not be used for treatment decisions.

The company release does not provide a peer-reviewed study, accuracy rate or clinical-validation dataset for the prediction feature. It also presents Libre Assist as a US launch and does not set out a country-by-country release schedule. The evidence available at launch therefore supports a description of how Abbott says the feature works, not a conclusion about its clinical accuracy or international availability.

Taiwan's documented products follow a different path

The Taiwan Stock Exchange's company record for Bionime says the manufacturer announced a Class II medical-device license from the Taiwan Food and Drug Administration for the RIGHTEST iFree2 CGM on April 29, 2026. The Taiwan Food and Drug Administration is the health ministry agency that registers medical devices for the Taiwanese market. The exchange record establishes the license announcement, but it does not describe a generative AI component.

Abbott's own Taiwan storefront points to a different generation of its product. The current Taiwan product page describes the FreeStyle Libre 2 sensor, a 14-day device that sends glucose readings to a phone app and can issue high- and low-glucose alerts. The page does not list Libre Assist or present meal-impact forecasting as a local feature.

Those records do not prove that no other CGM can be obtained in Taiwan through every possible route. They show what the manufacturers and the exchange publicly document through official local channels. APPI News did not treat overseas retail listings or cross-border sellers as evidence of Taiwan authorization.

A sensor recall sets a separate safety boundary

The distinction between software and sensor hardware also matters because specified newer Abbott sensors were already under a US device correction when Libre Assist was announced. The US Food and Drug Administration classified the removal of specified FreeStyle Libre 3 and Libre 3 Plus sensor lots as a Class I recall and reported 860 serious injuries and seven deaths associated with the issue as of January 7, 2026.

The affected sensors could report glucose readings lower than the actual level. The FDA says Abbott traced the problem to specified products and lots. The agency also states that Libre 3 readers and mobile apps were not affected, and that other Libre products were outside the recall.

The recall does not demonstrate a failure in Libre Assist's AI model. It does show why a forecast feature cannot be assessed apart from the sensor and distribution system around it. A meal estimate may be generated by software, but the later comparison depends on readings from regulated hardware and on the manufacturer reaching customers when particular lots require correction.

The missing evidence is as important as the feature list

Three questions remain separate in this market: whether a sensor has authorization in a country, whether the compatible app feature is offered there, and whether the AI prediction has published validation. The public records reviewed for this article answer only parts of those questions. Taiwan has documented CGM options, while Abbott's January announcement locates Libre Assist in the United States and supplies no clinical-performance dataset.

That makes “AI-enabled CGM” an incomplete description on its own. The label can refer to a sensor paired with an AI feature rather than intelligence inside the sensing hardware. It also says nothing about where the combination is authorized, whether it is supported through official distribution or how accurately the software predicts an individual's response to a meal.

APPI News cannot verify Taiwan's promised medical AI payment decision

APPI News Editorial — Mon, 03 Aug 2026 00:00:00 GMT

Taiwan's National Health Insurance Administration said in August 2024 that it aimed to finish a cost-effectiveness review of AI diagnostic tools by the end of that year. The agency cited software that could assist clinicians in detecting intracranial bleeding on computed tomography (CT) scans as a possible next candidate for payment, but did not name a product or proposed rate.

APPI News could not find a published payment decision for that CT application as of August 3, 2026. That search gap does not show that the proposal was rejected or that no review took place; it means the publicly available record located for this report does not resolve the 2024 promise.

The proposal concerns Taiwan's national insurance system

Taiwan's National Health Insurance is a single-payer program that pays contracted providers for covered care. A decision to add an AI tool would determine whether the program pays for its use under specified conditions, not whether the software is accurate in every hospital or authorized in another country.

The August 2024 announcement said the agency was considering temporary payment for AI diagnostic tools. It also pointed to an AI-assisted system used for high-risk surgical patients as the program's first paid example, while describing CT support for intracranial bleeding as the next type under evaluation.

The earlier payment case has a narrower purpose

The earlier case involved an Edwards Lifesciences sensor using the company's Hypotension Prediction Index, which analyzes an arterial pressure waveform to warn of possible low blood pressure during surgery. A 2024 report from National Cheng Kung University says Taiwan began paying for the sensor in July 2023 after reviewing six randomized trials and setting its payment at 8,593 points.

That record shows how evidence supported one defined use, patient group and payment. It cannot be carried across to brain imaging, where the model, workflow, clinical consequence of a false result and staff requirements are different.

The blood-pressure evidence also needs a cautious reading. A 2025 systematic review of 11 randomized trials found that Hypotension Prediction Index-guided care reduced the incidence and duration of low blood pressure during non-cardiac surgery, but did not change acute kidney injury or hospital length of stay. A device can therefore improve an intermediate measure without yet showing better patient outcomes or lower overall costs.

A payment review must measure more than model accuracy

Accuracy measured on a test dataset is only one input to a coverage decision. The World Health Organization defines health technology assessment as a multidisciplinary review of clinical, economic, ethical and social effects used to inform adoption and reimbursement. The comparison must address the technology as used in care, not the algorithm in isolation.

For CT triage, relevant outcomes could include how quickly a scan reaches a clinician, whether the alert changes treatment, the number and consequences of false alerts, the staff time required for review and the effect on other imaging work. The review must also count software integration, training, monitoring and updates rather than treating the purchase price as the full cost.

Published economic research offers limited shortcuts. A 2022 scoping review found only 13 cost-effectiveness studies of medical AI among 4,820 screened records, and only five of the 13 stated the assumed payment mechanism. The authors concluded that the available studies did not establish whether the assessed systems were clinically, technically and economically viable.

Taiwan built a new evidence route after the deadline

Taiwan's health ministry described a three-part evaluation structure in March 2025. Its AI Impact Research Center was designed to compare outcomes from clinicians using AI with outcomes from clinicians working without it, so that clinical value could inform a payment. Separate centers address responsible implementation and clinical validation.

The timeline became clearer in November 2025, almost a year after the original target. Taiwan launched AI impact research centers at five medical centers, and health ministry information chief Lee Chien-chang said randomized trials already in progress could take at least three years to produce results. The centers use local data and are intended to support later health technology assessment and payment deliberations.

The new structure does not by itself answer what happened to the CT review announced in 2024. It does show that Taiwan was still building the machinery for local clinical and economic evaluation after the promised deadline had passed.

Clinical adoption remains part of the calculation

An AI alert has value only if it fits the clinical pathway and changes a decision that matters. Hospitals need rules for who reviews an alert, how quickly they respond, what happens when the clinician disagrees and how performance is monitored after deployment.

Those questions affect both benefit and cost. Extra false alerts can consume staff time, while a useful triage signal may move an urgent scan forward in the queue. A defensible payment decision needs evidence on both effects in the hospitals where the tool would be used.