Editor's note: yesterday's brief covered WWDC as a reported-expectations preview — the keynote ran at 17:00 UTC on 8 June, after we published. Today is the day-after, with the facts on the record, so the lead and the second story both develop the WWDC thread rather than introduce it. The lead is an Enterprise/sovereignty story and the second main is Enterprise/Models, so the one-Regulation-main cap holds: the EU AI Act countdown is today's single Regulation main, and it pairs directly with the lead. No fresh open-weight model dropped in the last 24 hours — MiniMax M3's weights are still pending (~11 June) — so Open Source sits in the quick hits today, noted here so the imbalance is visible.
Apple Ships a Gemini-Powered Siri Everywhere Except the EU — and the Buyer-Side Question Is Where the Model Runs
At his final WWDC keynote as CEO, Tim Cook unveiled a fully rebuilt "Siri AI" running on a custom 1.2-trillion-parameter Google Gemini model that Apple licenses for roughly $1 billion a year, alongside a new Extensions system that lets users pick ChatGPT, Gemini (the default), or Anthropic's Claude as the model behind Apple Intelligence — Claude's first native presence on a platform with around 2.2 billion active devices (Apple, TechCrunch). The detail most consumer coverage skipped past is the one that matters to a regulated European buyer: the new Siri will not ship on iPhones or iPads in the EU at launch, with Apple saying it needs to find "a path forward for regulatory approval," and it will not ship in China either (MacRumors). The withholding is widely read against the Digital Markets Act — the most capable consumer assistant on earth is being gated by jurisdiction, not by hardware.
Underneath sits the question Apple spent the keynote trying to pre-empt: where does the reasoning actually happen? Apple's stated architecture keeps light tasks on-device and routes heavier work to its own Private Cloud Compute with queries anonymised and tokenised so neither Apple nor Google can tie a request to a person; some reporting says the heaviest Gemini reasoning still lands on Google Cloud hardware (Business Standard). Which version holds up is precisely what regulators and security researchers will probe in the months ahead. For European organisations the takeaway is cleaner than the architecture: a model's quality and the jurisdiction it runs in are now separable decisions, and the second one can gate a launch entirely — which is the whole premise of running open models on infrastructure you control.
Even Apple Would Rather Rent the Model Than Own It
The strategic read is as striking as the privacy plumbing. The world's most valuable company, sitting on well over $100 billion in cash and investments, decided that building and maintaining a frontier model was not worth it and paid Google instead — making Gemini the default assistant across iPhone, iPad and Mac while offering Claude and ChatGPT as switchable alternatives (Bloomberg via buildfastwithai). It is the largest distribution win in consumer AI to date and the clearest signal yet that "which model" is becoming an operating-system setting rather than a fixed product decision. A week earlier Microsoft made the same move on the enterprise side, with a Foundry catalogue spanning 11,000+ models — GPT-5.5, Claude Opus 4.8, Gemini, its own MAI family — behind one Azure endpoint, and Claude now reachable inside Excel's agent mode.
When both the dominant consumer platform and the dominant enterprise platform turn the model into a user-selectable field, the durable points of difference stop being the model and move underneath it: where inference runs, who can see the data, what audit trail it leaves, and how easily a workload can be moved between providers. That is the layer European regulated buyers are forced to compete on anyway — and the case for routing per workload across infrastructure and jurisdictions you actually control, rather than betting a compliance posture on a single vendor's roadmap.
The EU AI Act's Real Deadline Is 54 Days Out — but Its High-Risk Teeth Just Moved to 2027
The bulk of the EU AI Act starts applying on 2 August 2026 — 54 days from today — with fines reaching €35 million or 7% of global turnover for the most serious breaches. The GPAI duties themselves have applied to newly placed models since August 2025 — what bites on 2 August 2026 is Commission enforcement, with the 10^25-FLOP systemic-risk tier sweeping in GPT-5.5, Claude Opus 4.8, Gemini 3.1 Pro and Grok 4.3, layering transparency, documentation and evaluation duties onto the providers behind most enterprise deployments (European Commission). What changed recently is the timing of the law's sharpest edge: under the Digital Omnibus, EU institutions reached a provisional deal on 6 May, confirmed by member states on 13 May, to postpone the high-risk obligations — standalone Annex III systems (hiring, credit scoring, education, law enforcement) slip to 2 December 2027 and AI embedded in regulated products to 2 August 2028 (Council of the EU, Gibson Dunn).
The change only takes effect once the Omnibus is adopted and published in the Official Journal, which is expected just before 2 August. The practical message for compliance teams is to separate the two clocks: the high-risk conformity regime has more runway than the original calendar implied, but the GPAI and transparency layer arrives on schedule with real penalties — and it lands the same summer Apple is withholding a flagship AI feature from EU devices over a different European rulebook. Whichever way the politics break, the evidence a deployer can produce about how a system behaves has to hold regardless of which vendor sits behind it.
Quick Hits
- Open-weight reproducibility test still pending. No fresh open-weight model shipped in the last 24 hours; MiniMax M3's full weights and technical report remain due around 11 June, the cohort's next real check on its frontier-coding and 1M-context claims (MiniMax, llm-stats).
- June's model window is open, and the variable is price. Google has confirmed Gemini 3.5 Pro "this month" (Flash already scores 55 on the Intelligence Index at 284 tokens/sec for $1.50/$9), and a mid-June Claude Sonnet 4.8 is a credible read on the March npm source-map leak — capability has plateaued at the top, so token cost and jurisdiction are deciding EU deployments (buildfastwithai).
- SpaceX prices Thursday. The IPO prices 11 June and trades 12 June on Nasdaq as SPCX, targeting ~$75B at ~$1.75T — the largest listing on record; the AI line is xAI's ~$3.2B 2025 revenue against ~$14B cash burn, and the read-through to Anthropic's reported ~$1.25B/month compute bill (CNBC via buildfastwithai).
- The Pentagon's "too safe" problem. DoD is testing OpenAI, Google and xAI models to replace Claude on classified systems, and analysts say a swap could take months; the constraints the Pentagon calls a liability — no mass surveillance, no autonomous targeting — are close to the ones a regulated European institution calls a feature (develops 8 June) (Scientific American).