Editor's note. Today's lead is the freshest hardware news in months: OpenAI and Broadcom shipped a chip. The story underneath it — frontier labs vertically integrating all the way down to silicon — runs straight into the second story, where Europe just put public supercomputing behind an open, sovereign answer. The third stays with the Fable 5 saga, now in its twelfth dark day, where the mechanism for the model's return turns out to involve handing over a government ID. Regulation appears only in the quick hits, so there are zero regulation mains today.
OpenAI builds its own chip, and the full stack gets one owner
OpenAI and Broadcom on 24 June unveiled Jalapeño, OpenAI's first custom silicon — an inference-optimised ASIC the two companies say went from initial design to tape-out in nine months, which they call the fastest such cycle ever for a high-performance chip. Part of that speed came from using OpenAI's own models to accelerate the design work. It is reticle-sized, built for serving rather than training, and explicitly framed as a step away from Nvidia for high-volume inference — first deployments at Microsoft and other partners by the end of 2026, scaling toward 10 gigawatts of OpenAI-designed compute by 2029. It does not remove Nvidia from the picture; training still runs on GPUs. But the direction is unmistakable.
CNBC quoted OpenAI's own framing best: this is about building the full stack — model, software, and now the chip it runs on. For a European buyer that is the part worth sitting with. When you rent inference from a lab that owns its models, its serving software, and increasingly its own accelerators, every layer of the thing you depend on sits inside one US company's vertical, and you cannot reason about — let alone substitute — any of it independently. It is the same "who controls the silicon beneath the model" question that the China and EU infrastructure plans raised last week, now answered by a private lab with capital rather than by a state with a quota. The sovereign counterweight is unglamorous and well understood: commodity hardware you own, running open weights you hold, with the routing layer in your hands.
Europe answers with public compute behind an open frontier model
The same week, the European Commission named the winner of its Frontier AI Grand Challenge: EUROPA, a consortium led by the Italian firm Domyn, tasked with building an open-source frontier model covering all 24 official EU languages at a scale of at least 400 billion parameters. The project gets up to 2.5% of EuroHPC's total computing capacity for a year on Europe's AI-optimised supercomputers — public infrastructure, an open licence, and a multilingual remit that no US or Chinese lab is structurally incentivised to serve.
It is worth being precise about what this is and isn't. It is a funded project, not a shipped model, and the gap between a grant announcement and a competitive open-weight frontier system is exactly where Europe has stumbled before. But read alongside the lead, the contrast is the whole point: while one model lab integrates downward to own its chips, the EU is trying to build the opposite — a stack whose compute, weights, and languages are deliberately not owned by any single private vendor. For regulated European buyers, a credible sovereign open model on EU supercomputers is the supply-side precondition that makes "swap the model underneath" more than a slogan. The execution risk is real; the strategic direction is the right one.
To get Fable 5 back, you may have to show your passport
Twelve days after the 12 June US export-control order, Claude Fable 5 and Mythos 5 remain offline for every user, with no announced restoration date. What's new is the mechanism. Anthropic's updated privacy policy, effective 8 July, now enumerates government-issued ID images, photo or video of the user, and facial-geometry templates among the data it may collect. Reporting frames this plainly as a US-citizens-only path around the ban: with no way to verify nationality at API scale, Anthropic builds a compliant identity gate that lets verified US users back in without the directive being lifted. The other date on the wall is 1 August, the 60-day deadline for the executive order's covered-frontier-model framework.
Stack this on top of last week's reported 30-day mandatory retention terms and the shape of the returning product is clear. To regain access to a rented frontier model, a user may have to submit a scan of a government ID and a facial-geometry template, to a US company, under a policy that also keeps their traffic for 30 days. That is the precise inverse of data minimisation, and for an EU organisation handling regulated data it is close to a non-starter regardless of how the export politics resolve. The episode keeps teaching the same lesson in escalating detail: access to a hosted frontier model is governed by someone else's compliance machinery, and that machinery only ever tightens. Weights you run yourself never ask to see your passport.
Quick Hits
- The EU AI Act Omnibus reaches its Council vote on 29 June — after the European Parliament adopted the text on 16 June, the Council is expected to formally adopt it on 29 June, with Official Journal publication to follow. Standalone Annex III high-risk obligations move to 2 December 2027 and the watermarking grace period to 2 December 2026; until OJ publication the original text remains law. Four days out, this finally closes a thread that has run for weeks.
- Gemini 3.5 Pro is still a preview with five days left in June — Google's promised general availability (2M-token context, Deep Think) remains limited to Vertex enterprise preview, absent from the public app, AI Studio and the consumer subscription. The top of the closed-model price ladder stays unsettled until it ships.
- Open weights keep crowding the top of the leaderboards — Artificial Analysis still shows open-weight models holding several of the top ten on its Intelligence Index, led by GLM-5.2, Kimi and DeepSeek V4, all under MIT or Apache licences. The story remains the trajectory: a self-hostable near-frontier model now exists at most capability tiers — though vendor figures still run ahead of neutral reproductions, so verify on your own data before a regulated workload.
- OpenAI keeps consolidating the distribution layer — following its $4bn Deployment Company, OpenAI added a $150M Partner Network on 14 June to push enterprise rollouts through systems integrators and consultancies. As at the frontier labs generally, the model commoditises while the channel that puts it inside regulated businesses concentrates.