Good morning,
This week we cover the impact of AI on the chip industry (aside from Nvidia) and what it’s starting to mean for consumers, plus OpenAI’s new GPT-5.6 model, blocked by the U.S. government—and what this might hide. Plus the usual top news, including Micron’s extraordinary Q3 results, Anthropic’s Claude Tag, and the new 2026 AI in Insurance Index.
Back to the chip bottleneck
For two years the AI-hardware story had one name: Nvidia—and one question: could you get their GPUs? Since the beginning of the year, this bottleneck has spread outward: first to rival chipmakers AMD (up +133% YtD), then to memory makers such as SK Hynix (+280%) or Micron (+296%). However, supply is tightening, and it’s now affecting everyone’s prices, including retail customers.
Just three firms—Micron, SK Hynix, and Samsung—make more than 95% of the world’s DRAM, and all three are racing to convert their factories to high-bandwidth memory (HBM), the pricey, high-margin chips that sit inside AI servers. Every chip redirected to AI is a chip pulled away from the ordinary memory in phones, laptops, and consoles. The result: contract memory prices jumped roughly 90% in the first quarter of 2026 and another 60% in the second, per TrendForce.
The AI companies saw this coming and have started trying to own their supply rather than queue for it. OpenAI unveiled its own inference chip ("Jalapeño," with Broadcom); and Anthropic signed a multi-year memory deal with Micron and took it on as an investor. But everyone downstream is exposed, including us. On 25 June, Apple raised prices mid-cycle across nearly its whole line-up — the entry MacBook from $599 to $699, with similar bumps to iPads, Apple TV, and the Vision Pro — blaming a memory shortage it called an "unprecedented challenge." Apple shares fell about 6%. Hours later, Microsoft raised Xbox prices, warning that console memory costs had already more than doubled.
Why it matters: The bill for the AI build-out is starting to land on people who never bought an AI product. Memory has become the new strategic commodity, concentrated in three suppliers, sold to the highest-margin buyer, and short until at least 2028 on the makers' own forecasts. Two practical consequences: budget for hardware inflation across any device or data-centre refresh through 2027, and start treating memory and compute supply as a genuine procurement risk, not a line item that only ever falls.
Who gets to use the best AI?
Among the many predictions from AI 2027 — a much-discussed 2025 scenario from the AI Futures Project — one stands out particularly today: once models become capable enough to accelerate their own development, and to confer real economic and military advantage, whoever leads will have every incentive to keep that lead. Not just by building faster, but by controlling who else gets to use the technology.
Two weeks ago, Anthropic's most capable models, Fable 5 and Mythos, were switched off worldwide after a US export-control order. This week it was OpenAI's turn: the White House asked it to restrict its new GPT-5.6 family (Sol, Terra and Luna) to roughly 20 government-vetted partners, approved customer by customer, with no public release. The legal hook is a 2 June executive order that lets Washington benchmark and gate "frontier" models with strong cyber capabilities. For the first time, the US government — not the company — is deciding who may use America's best AI, at home and abroad. OpenAI complied, while making clear it doesn't think the arrangement should become permanent.
On one hand, the protective instinct isn't paranoia. Chinese labs — DeepSeek, MiniMax, and Moonshot among them — have been caught reverse-engineering US models through "distillation": quietly querying them at industrial scale to train cheaper copies. In February, Anthropic says it traced more than 16 million such exchanges across some 24,000 fraudulent accounts, and recently accused Alibaba of unlawfully extracting Claude capabilities.
On the other hand, it seems questionable for one government to appoint itself sole gatekeeper (and sole beneficiary) of a technology that was never purely American. Of the eight authors of the 2017 "Attention is all you need" paper that allowed the AI revolution, only one was American. Google DeepMind, the most decorated lab of the lot, is run by a Briton. The talent behind frontier AI is profoundly European and Asian; the control over it is becoming narrowly national. That mismatch may prove a more consequential story than any single model launch.
Why it matters: Access to frontier AI is now a geopolitical variable, not a commercial one. If your roadmap assumes you can simply buy the best model on the market, build in a hedge: the very best may be reserved for a government's approved list — or switched off — with little notice, and the approved list may not include you.
More top news
Micron just had a frankly ridiculous quarter. Its Q3 results (reported 24 June) show: revenue of $41.46bn, up 346% YoY, net income of $28.24bn, and guidance of roughly $50bn for the current quarter. The stock jumped ~15% after hours and Micron's market value has passed $1tn. Sanjay Mehrotra, Micron’s CEO, expects the memory shortage to persist until at least 2028.
Anthropic launched Claude Tag, its new Slack agent. Claude Tag puts a single shared @Claude inside a channel that anyone can delegate tasks to, and that runs on its own for hours. Anthropic says an internal version already produces 65% of its product team's code.
Labs put numbers on the switch to agents. OpenAI's new study of its Codex agent found non-developer users up 137-fold since last August, with nearly a quarter of requests now representing over an hour of human work. It dovetails with Anthropic's Economic Index, which tracks the slow drift from AI that helps you ("augmentation") to AI that just does it ("automation"). Frontier labs are increasingly focusing on producing agent economic research to prove the benefits of their model.
Qualcomm goes shopping to escape Nvidia. It is reportedly in talks to buy AI-chip startup Tenstorrent for $8–10bn, and confirmed a separate software deal with Modular on 24 June.
AI takes the bar. Garfield AI, a British law firm relying on AI to prepare claims, won its first AI-prepared case. The AI conducted all the legal work prior to the trial, including preparing four witness statements.
AI is learning to imagine catastrophes. FT’s Lee Harris and Aditi Bhandari report risk modellers are increasingly using diffusion models to manufacture tens of thousands of plausible-but-never-happened extreme-weather events, filling the gaps where real history is too thin to price.
Allianz overtook AXA as #1 in the AI Index ranking. Evident published its 2026 index on AI maturity in the insurance sector, showing a rise in maturity among all players. The index takes into account Talent, Innovation, Leadership, and Transparency. Top 5: #1 Allianz (+1 spot to 2025), #2 AXA (-1), #3 Manulife (+2), #4 Zurich (+8), #5 Liberty Mutual (+2).
The DeepMind brain drain. Google's lab lost two big names: Nobel laureate John Jumper, creator of AlphaFold (to Anthropic), and Gemini co-lead Noam Shazeer (to OpenAI) — as its Gemini 3.5 Pro stayed stuck in limited preview past its June target. Noam was the latest researcher behind the Transformer technology still at Google.
Midjourney opens a Spa. The independent lab—known for its image generation model—surprised everyone by announcing Midjourney Medical. This new division will focus on building a new type of (faster and cheaper) scanner. They aim to deploy 50,000 scanners worldwide by 2031, an ambitious task they plan on supporting by opening Spas where customers will be able to come, relax, and get scanned.
Tools to try: Google Flow
Part of the wave of enthusiasm about AI in 2023/24 was about image and video generation models (often “diffusion” models). However, recently it seems the vibe has strongly shifted toward productivity tools: agents that code, analyse documents, run workflows, search the web, update spreadsheets, or sit inside Slack and email. Image and video generation have not disappeared; they have become less surprising. The frontier is moving from “look what AI can create” to “look what AI can do.”
Google Flow is a good example of why the creative side still matters. Built around Veo, Imagen, and Gemini Omni, it lets users generate short cinematic clips from text prompts, images, or scene descriptions, then iterate on shots, characters, and camera movements. It is not just a toy for making surreal clips. It is closer to a lightweight production studio: storyboard an idea, test a campaign concept, create a product mock-up, or translate a written scene into something visual in minutes.
Workflow is more controlled. Flow tries to make generation feel more like editing: keep a character consistent, extend a shot, change the framing, or build a sequence. Still, do not expect a full flawless experience—video models still require a lot of tweaking and correcting.