AI News Daily 2026-07-28
- Semiconductor and AI-related shares fell sharply around the world on 28 July, with South Korea’s KOSPI dropping far enough to trigger a temporary trading halt and both Samsung Electronics and SK hynix losing more than 9 percent.
- SK Group and NVIDIA agreed a wide-ranging partnership spanning AI factory construction through the joint development and supply of next-generation AI memory (HBM), with SK Telecom planning a two-gigawatt-class DSX AI factory built on NVIDIA Vera Rubin.
- OpenAI started the US rollout of Health in ChatGPT, a feature that connects to Apple Health and supported medical records, with the company stating the data is not used for model training or advertising.
- Anthropic announced Claude Opus 5, available the same day across all platforms, priced at the same level as the previous model and set as the new default for Claude Max.
- The through-line for the period: markets are re-testing the economics of AI infrastructure spending at the same moment that frontier labs push both cheaper models and deeper real-world integrations.
01 AI-linked chip stocks sell off worldwide on funding and China concerns
Published: 2026-07-28 · Category: Corporate developments · Source tier: Tier 2
Facts
On 28 July 2026 the sell-off in semiconductor shares deepened across global markets. In South Korea the KOSPI fell far enough that trading was temporarily halted. Two of the region’s largest memory makers were at the centre of the move: Samsung Electronics and SK hynix each fell by more than 9 percent.
The reported drivers were twofold: growing doubt about whether the returns on AI infrastructure investment justify its cost, and the rise of Chinese-made semiconductors and memory as competition.
Background
Memory suppliers such as Samsung Electronics and SK hynix sit close to the centre of the AI hardware supply chain, so their share prices function as a proxy for how investors price the AI build-out as a whole. A move large enough to interrupt trading on a national index is therefore not a company-specific story — it is the market repricing an entire spending cycle at once.
The second driver named in the reporting, the emergence of Chinese semiconductor and memory suppliers, matters because it points at the revenue side rather than the demand side: even if AI demand holds up, incumbent suppliers may not capture the same share of it.
Implications
The market has begun to question the sustainability of very large AI infrastructure investments. For companies that use AI rather than build it, this is the moment to re-examine cost structure and procurement risk — what a given workload actually costs to run, and how exposed the organisation is to a single hardware or memory supplier whose pricing and availability may move with this cycle.
02 SK Group and NVIDIA strike a broad AI factory and memory partnership
Published: 2026-07-25 · Category: Corporate developments · Source tier: Tier 1 (official web)
Facts
SK Group and NVIDIA announced that they had agreed a large-scale comprehensive partnership. Its scope runs from the construction of AI factories through to the joint development and supply of next-generation AI memory (HBM).
As part of the agreement, SK Telecom plans to build a two-gigawatt-class DSX AI factory equipped with NVIDIA Vera Rubin.
Background
The announcement links the two halves of an AI system that are usually negotiated separately: the compute platform and the high-bandwidth memory that feeds it. By committing to joint development rather than a purchase agreement alone, the two groups tie a memory roadmap to a compute roadmap over a multi-year horizon.
The scale marker to note is the two-gigawatt class figure for the planned DSX AI factory — a power-denominated unit, which is how capacity in this generation of AI build-outs is increasingly described.
Implications
A long-term partnership with a major memory supplier that sits at the core of the AI semiconductor supply chain further strengthens NVIDIA’s AI platform supply network. Read alongside the market move in the previous chapter, it is a useful reminder that supply-side commitments are being made on a longer clock than the one the equity market is currently trading on.
Source: SK hynix Newsroom
03 OpenAI begins rolling out Health in ChatGPT to US users
Published: 2026-07-23 · Category: Model releases · Source tier: Tier 1 (official X post and official web)
Facts
OpenAI announced that it had begun the US rollout of Health in ChatGPT, a feature that can securely connect with Apple Health and supported medical records.
The feature is designed to let the model understand context drawn from information such as medication details and test results. OpenAI states that this data is not used for model training or for advertising purposes.
Background
What distinguishes this from a general-purpose assistant answering health questions is the connection itself: the model is reading structured personal records rather than only what a user types into a conversation. That changes the nature of the data an AI provider holds, which is why the stated boundary on training and advertising use is part of the announcement rather than a footnote to it.
Implications
Cases in which generative AI links directly to an individual’s medical data are becoming more common, and the central question ahead is how to make healthcare utility and data governance work together rather than trade off against each other. For any organisation considering a comparable integration, the governance commitments — what the data may and may not be used for — are as much a part of the product as the capability.
Sources: OpenAI (official web) · @OpenAI (official X post)
04 Anthropic announces its new flagship, Claude Opus 5
Published: 2026-07-24 · Category: Model releases · Source tier: Tier 1 (official web)
Facts
Anthropic announced a new model, Claude Opus 5, and made it available the same day across all platforms.
Pricing is held at the same level as the previous model. Claude Opus 5 becomes the new default model for Claude Max, and it carries a control that lets users switch the amount of processing applied between low, medium and high.
Background
Two of the three changes are commercial rather than architectural: holding price flat across a flagship generation, and making the new model the default for the top subscription tier. The third — a low / medium / high processing setting — moves the cost-versus-quality decision to the point of use, so the same model can be run cheaply for routine work and expensively where the answer justifies it.
Implications
By delivering performance approaching that of higher-tier models at lower cost, the release widens the range of cost-effectiveness options available to enterprises adopting AI. In practice this favours organisations that have already worked out which of their workloads need maximum capability and which do not, because the processing switch only pays off when that distinction is made deliberately.
Source: Anthropic
05 Trend overview
- Concern about the profitability of AI infrastructure investment is driving a broad global sell-off in semiconductor and AI-related shares.
- The EU is easing the operational burden of its AI regulation while keeping key obligations, such as transparency requirements, on their original schedule.
- Frontier models are advancing on two fronts at once: lower prices with higher efficiency, and expanded practical features such as healthcare integration.
The EU regulatory item above appears in the source material only as a trend observation. No individual story with its own sourcing was adopted for it in this edition, so it is reported here without further detail.
06 Editor’s note
The four stories in this edition pull in opposite directions, and that tension is the story. Chapter 01 is the market marking down the AI build-out: a trading halt on the KOSPI and two memory makers each off more than 9 percent, driven by doubts about the returns on AI infrastructure spending and by Chinese competition. Chapter 02, dated three days earlier, is the same supply chain committing to a multi-year partnership that runs from AI factories to next-generation HBM. Capital markets and capital expenditure are working on different clocks.
Chapters 03 and 04 describe what that infrastructure is being asked to carry. OpenAI is connecting ChatGPT to Apple Health and supported medical records in the US, and pairing the capability with an explicit limit on how the data may be used. Anthropic is holding flagship pricing flat while adding a low / medium / high processing control and making Claude Opus 5 the default for Claude Max. One release moves AI deeper into a regulated, high-sensitivity domain; the other moves the cost of using it down and puts the dial in the user’s hand.
For decision-makers the practical read is this: the supply side is not slowing down, but its pricing is now visibly contested, and the capability side is getting cheaper per unit of work. That combination rewards organisations that can say precisely which workloads justify the expensive setting — and, where personal or regulated data is involved, can state as clearly as the vendors now do what that data will and will not be used for.
This is a retrospective edition. As the source notes record, only two items published within the preceding 48 hours met the adoption criteria, so the edition was widened to cover the recent days rather than padded with weaker material.