Editor's note. Today's thread is the terms of access to intelligence — who sets them, and what it costs to stop renting. The lead is the clearest signal: Anthropic just absorbed an entire data centre's worth of GPUs to raise its limits, a reminder that how much of a hosted model you can use is a capacity decision made elsewhere. The second story is the other lever — price — as OpenAI opens a front against Claude's mid-tier. The third is the structural counter that ties the brief together: if capacity and price are both the vendor's to set, the only supply you fully control is the model you hold yourself. Regulation stays in the quick hits, where two milestones landed at once: the EU Omnibus became law on Monday, and Colorado's landmark AI law was due to take effect today and instead was gutted and pushed to 2027.
Access becomes a capacity question
Anthropic has been buying headroom. In early May the company signed a deal to use all of the compute at SpaceX's Colossus 1 data centre in Memphis — more than 300 megawatts and over 220,000 Nvidia GPUs, due online within weeks of the announcement — a facility that until this spring was associated with Elon Musk's xAI. With that capacity now coming onstream, Anthropic raised usage limits sharply: Claude Code five-hour limits doubled, the peak-hours reduction was removed, and top-end Opus API limits rose by as much as ~16×.
The detail worth holding onto is what it reveals about the binding constraint. The labs are no longer competing only on model quality; they are racing to lock up gigawatts — Anthropic alone is stacking a 5GW Amazon agreement, a 5GW Google–Broadcom deal for 2027, $30 billion of Azure capacity and a $50 billion Fluidstack build on top of the Memphis site. For a European buyer the implication is concrete: how much of a hosted frontier model you can use, and when, is downstream of someone else's data-centre deal in Tennessee. Capacity you operate against a known number of your own GPUs is a different kind of guarantee than capacity you queue for at limits another company raises and lowers as its supply allows.
OpenAI opens a price front
The other lever is price, and it is moving too. GPT-5.6's mid-tier — Terra — is priced at $2.50 input and $15 output per million tokens, undercutting Claude Sonnet 4.6's $3/$15 on input by roughly 17% at comparable capability, a deliberate shot at the model that serves as the default workhorse in many enterprise stacks. The Wall Street Journal has separately reported OpenAI is weighing deeper token cuts ahead of its IPO. Read alongside the lead, the picture is a supply side responding hard to the efficiency squeeze that surfaced last week — when enterprises started capping budgets and routing to cheaper models, the labs answered on both throughput and price.
That is good news for buyers in the near term, with one caveat that keeps recurring: the savings are captured by the routing, not the rate card. The technique enterprises are converging on — send the bulk of traffic to a cheap model and escalate to a frontier model only for the subset that needs it — is what actually drops effective cost by the reported 70–90%, and it works only if the model underneath is something you can swap on your terms. It is also worth noting the price is being set on a model you cannot yet buy: GPT-5.6 remains gated to roughly twenty government-vetted partners, with general availability still only "in coming weeks."
Europe's answer is to own the supply
When both capacity and price are the vendor's lever, the structural counter is to stop renting access and hold the model. That is the explicit European pitch now. In a mid-June post that has aged into a position statement, Mistral's Arthur Mensch argued that AI is "a commodity technology that every organisation needs a secured and affordable supply of," and that "states and organizations need to have sovereignty over [it] — they should own and control the systems that are specifically embedding their IP and tacit knowledge, and will end up running their most critical processes." The supply that makes the argument real already exists: open-weight models such as Z.ai's GLM-5.2 now land within a point of Claude Opus 4.8 on agentic benchmarks at roughly a fifth of the price, and the frontier-tier open field is several models deep. (Mistral is a European competitor; noted via third-party coverage.)
The honest caveats stand. Chinese-origin open leaders carry provenance scrutiny that a regulated buyer has to resolve, a vendor's benchmark is not a production workload, and Europe's own bet on a sovereign open frontier model on public compute is funded rather than shipped. But the principle survives all of them. A model you can pull down and serve on hardware you operate is the one supply in this market that no capacity deal, no price change and no export directive issued by another government can re-term. The week's news on access, price and gating is, read together, the strongest argument yet for treating the model as something you hold rather than something you queue for.
Quick Hits
- The EU AI Act Omnibus is now law — the Council gave the package final adoption on 29 June, closing a months-long thread. Standalone high-risk systems under Annex III now face full compliance from 2 December 2027, the watermarking grace period for transparency obligations was cut from six months to three (2 December 2026), the sandbox deadline moved to December 2027, and a ban on AI-generated non-consensual intimate imagery is written into the Act.
- Colorado's landmark AI law was due today — and isn't — SB 24-205, the most detailed US state AI law, was set to take effect 30 June, but Colorado passed SB 26-189 substantially reducing obligations on deployers and employers and pushed the effective date to 1 January 2027. The most-watched US enforcement date quietly became a non-event.
- Anthropic's Economic Index finds AI moving from helper to doer — in a survey of ~9,700 users linked to actual usage, over 35% expect AI to perform most or nearly all of their work within twelve months, and automation has overtaken augmentation as the primary mode in professional (API) use — production deployments are already more automated than most internal AI policies acknowledge.
- Gemini 3.5 Pro slips to July — Google missed its own June deadline for the second consecutive I/O; the 2M-token flagship stays in Vertex preview, with Deep Think reasoning expected to be gated to Ultra subscribers at $250/month.
- Mythos 5 partially restored, Fable 5 still banned — a 26 June Commerce letter exempts roughly 100 US critical-infrastructure organisations and federal agencies from the export-control directive on Anthropic's cybersecurity model; everyone else still needs an export license, and Fable 5 remains suspended for general use. The next markers are 8 July (ID-verification policy) and 1 August (the executive-order framework deadline).
Update (added 19 July, at publication of the June backfill): Commerce lifted the export controls later on 30 June, and Fable 5 access was restored on 1 July.