Editor's note. Today's lead is the brief's one regulation main, and it earns the lead slot under the editorial carve-out for "a major EU AI Act deadline": the European Parliament's plenary adoption of the Digital Omnibus on 16 June is the most consequential, concrete AI Act event in weeks for this audience, and it resolves the "still awaits a vote" thread the brief has carried — yesterday's quick hit is now superseded by a recorded 423–57 vote. The other two mains are non-regulation by design: the SpaceX–Cursor deal hitting the antitrust machine (Enterprise/markets) and the open-weight stack (Open Source). Two briefs running touch the European Parliament, but the substance is entirely distinct — Monday's sovereignty debate versus today's binding law. No fresh frontier model shipped in the last 24 hours; Models runs through the quick hits.
The EU AI Act Omnibus clears Parliament — the deadlines move, the structure does not
The European Parliament adopted the Digital Omnibus on AI on 16 June by 423 votes to 57 with 174 abstentions, turning May's trilogue political agreement into a text the institution has now formally signed off. The package is a deadline-relief measure: high-risk obligations for stand-alone Annex III systems slip to 2 December 2027, and for AI embedded as safety components under EU product law to 2 August 2028. It also trims the grace period for labelling AI-generated content from six months to three (new deadline 2 December 2026), and writes a fresh Article 5 prohibition on non-consensual intimate imagery and AI-generated CSAM into the Act. Council formal adoption and Official Journal publication are the remaining steps before the changes enter into force.
For a compliance officer the headline is tempting but partial. The high-risk clock has genuinely moved out by a year or more, but the content-labelling deadline of 2 December 2026 still stands, the watermarking obligation actually tightened, and the GPAI duties that were never part of this postponement remain roughly six weeks away on 2 August. Civil-society groups read the same vote as deregulation that weakens digital rights, which is a reminder that "relief" is contested rather than settled. None of it changes the structural requirement underneath the dates: a deployer still has to know which jurisdiction runs its inference, keep provenance and content-labelling machine-readable, and be able to show its working. A moving deadline buys time to build that; it does not remove the obligation to.
SpaceX–Cursor meets the antitrust machine
Yesterday's lead has acquired a second act. The $60 billion all-stock agreement for Cursor's maker, Anysphere, carries a $10 billion general break fee and — more tellingly — a separate $4 billion "regulatory termination fee" payable if the deal is blocked on antitrust grounds, a number SpaceX's own lawyers would only write down if they judged review a real risk. The Hart-Scott-Rodino pre-merger process is mandatory at this size, a Second Request from the DOJ or FTC could stretch the timeline past a year, and the deal is expected to draw scrutiny in the US and EU alike. Cursor is not a small target: roughly $4 billion in total ARR (about $2.6 billion of it enterprise B2B), deployed inside an estimated more than half of the Fortune 500. Microsoft examined a bid and passed; OpenAI was rebuffed twice and bought Windsurf instead.
The buyer lesson sharpens the one from Monday. The dependency problem with a coding agent embedded in your developers' workflow was always stickiness; now there is an enforcement overlay on top of it. A tool sitting inside a regulated organisation's codebase is about to spend months — possibly more than a year — in antitrust limbo while its ownership, terms and data posture are unsettled by a process the customer has no part in. That uncertainty is itself a continuity risk, and it points the same direction as everything else this fortnight: keep the model layer underneath the assistant portable, so a contested merger upstream is a vendor's problem and not yours.
The open-weight stack is the part that does not move
Neither the law nor the deal touches the option that has been quietly maturing on the open side, and on a day with no fresh frontier release that is the story worth restating. MiniMax M3 — open weights, top of the open field at 55 on Artificial Analysis's Intelligence Index — sits alongside Moonshot's Kimi K2.7-Code under a Modified MIT licence, both runnable today on vLLM, SGLang and llama.cpp. The capability gap to the closed flagships keeps narrowing from below even as the convenient commercial tools concentrate into a handful of US balance sheets and the premium tier stays partly out of reach (Gemini 3.5 Pro is still a preview; Fable 5 and Mythos 5 are still switched off).
For a regulated deployer those are not interchangeable paths. A coding or reasoning model you self-host can be evaluated on your own data, kept running regardless of an export directive or a year-long merger review, and addressed by whatever editor your developers already use. The honest caveats stand: vendor-run benchmark figures continue to run ahead of neutral-board reproductions, and model origin remains a sovereignty variable in its own right — Chinese-built weights carry scrutiny in European finance, healthcare and government that a permissive licence does not erase. But the through-line of the week is hard to miss: the deadline can move, the owner can change, the model can be revoked — and the only layer none of that reaches is the one you hold yourself.
Quick Hits
- Gemini 3.5 Pro is still in the waiting room — Google's promised June GA (2M-token context, Deep Think) remains a limited preview, with prediction markets pricing the "shipped by 30 June" odds at roughly 50–55%. Twelve days left in the month; the top of the price ladder stays unsettled.
- SpaceX's acquisition currency is a two-week-old stock — the Cursor deal's all-stock structure was paid for by SPCX's post-IPO surge in hours of trading, pushing SpaceX past roughly $2.9 trillion. A purchase priced in a freshly public, volatile equity is a reminder that the supply structure behind these tools is now tied to public-market sentiment.
- Fable 5 and Mythos 5 are still dark — six days after the 12 June US export-control directive, Anthropic maintains it is working to restore access but there is still no announced return date and no published legal justification for the order. The arc is the backdrop to the whole week's sovereignty framing.
- The capex concentration keeps widening — the four largest hyperscalers have guided roughly $635–665 billion of 2026 capital spending (Amazon ~$200bn, Alphabet ~$175–185bn, Microsoft ~$145bn, Meta ~$115–135bn), the vast majority on AI data centres and GPUs, as compute-as-a-service contracts overtake equity rounds as the primary way the build-out is funded.