Editor's note: WWDC day. We lead on Apple's keynote (Enterprise/Models) — not the consumer features, but the one decision that matters to a European buyer: where the new Siri's reasoning is actually processed. The keynote runs at 10:00 PT / 17:00 UTC, after this brief publishes, so the specifics below are reported expectations, and the data-residency posture is what to watch rather than what is confirmed. We carry one Enterprise main (the platform layer going deliberately multi-vendor) and exactly one Regulation main — the genuinely fresh and widely-misreported story that Colorado just gutted and delayed its AI law, sharpening the transatlantic divergence 55 days before the EU AI Act's high-risk obligations bite. Models and Open Source are handled in the quick hits, where the frontier-release queue and MiniMax M3's pending weights sit.
Apple Buys a Brain It Doesn't Own — and the Question Is Where It Thinks
Apple's WWDC keynote today is expected to confirm what has been reported for months: a completely rebuilt, system-wide Siri running not on Apple's own models but on a custom ~1.2-trillion-parameter Google Gemini model, licensed for roughly $1 billion a year. Alongside it, Apple is expected to open Apple Intelligence to multiple providers through an Extensions system — letting users pick ChatGPT, Gemini or, for the first time as a native option, Anthropic's Claude. The world's most valuable company has decided that owning a frontier model is not worth the capital and talent it demands, and has chosen to rent one instead while keeping the surface, the distribution and — crucially — the compute layer under its own control.
That last point is the whole story for a regulated European buyer, and it is the detail most coverage underplays. Apple's reported architecture runs the heavy Gemini reasoning on its own Private Cloud Compute servers rather than Google's infrastructure, so that a third party's model can power the assistant without that third party seeing the data. Whether that distinction survives scrutiny — and whether Apple commits to a specific data-processing location for European users — is exactly the question an EU compliance officer would ask before trusting any of this with regulated information. It is the same question the EU's Cloud and AI Development Act now turns into a scored procurement criterion: not only which model, but whose infrastructure, in which jurisdiction, processing the inference data. Apple is demonstrating, at consumer scale, that you can separate the model from the place it runs. That separation is the entire premise of running open models on infrastructure you control.
The Platform Layer Goes Deliberately Multi-Vendor
Apple's multi-provider Extensions move is not happening in isolation. A week after Build, Microsoft has been consolidating the same pattern from the enterprise side: its Foundry catalogue now spans thousands of models — OpenAI's GPT-5.5, Anthropic's Claude Opus, Sonnet and Haiku, Google's Gemini, Microsoft's own MAI family and a long tail of open-weight and specialised models — all reachable through one Azure endpoint with one billing relationship. The concrete new step is that Claude now arrives in Foundry with MACC-eligible Azure billing and lands inside Excel's Agent Mode, putting a frontier model directly into the spreadsheet used by hundreds of millions of knowledge workers. Microsoft has also made the agent harness itself a flex point rather than a lock-in, with the GitHub Copilot SDK and Claude Agent SDK now in stable release.
When the two largest platform owners in consumer and enterprise computing both make the model a user-selectable setting, the competitive question quietly shifts. "Which model is best" stops being the decision a platform makes on your behalf and becomes a routing choice — and once it is a routing choice, the durable differentiators are the ones underneath the model: where it runs, who can see the data, what the audit trail looks like, and whether you can move between providers without re-architecting. That is the same logic that makes a sovereign router, rather than a single-vendor bet, the defensible position for a European deployer.
Colorado Blinks as the EU AI Act Clock Runs Down
The cleanest illustration of the transatlantic divergence arrived quietly and has been widely misreported since. Colorado's AI Act — the most-cited piece of US state AI legislation, and long flagged for a 30 June 2026 effective date — is no longer taking effect then. On 14 May, Governor Polis signed SB 189, pushing the effective date to 1 January 2027 and scaling the law back sharply — stripping the duty of care against algorithmic discrimination, the risk-management-programme obligation, and the impact-assessment requirements that made it look like a state-level analogue of the EU framework. Several national trackers still describe it as a live June deadline; it is not.
The contrast is the point. As the leading US state AI law retreats toward disclosure-only, and a federal proposal (the Great American AI Act) floats a three-year preemption of state rules entirely, the EU AI Act's high-risk and general-purpose obligations take full effect on 2 August — 55 days out — carrying fines of up to 7% of global turnover. The asymmetry that results is worth naming plainly: for any company selling into both markets, the binding high-risk baseline is now European, not American, and it is the one that arrives first with real teeth. A US deployer can wait; an EU deployer, or anyone serving EU users, is building the documentation, human-oversight and robustness evidence now — and that evidence has to hold regardless of which vendor's model sits behind the API.
Quick Hits
- The frontier release window is open, and the variable is price. Gemini 3.5 Pro was confirmed "this month" at I/O and has not yet shipped; a Claude Sonnet 4.8 is widely anticipated on the strength of a source-map leak that correctly called Opus 4.7's timing — high-confidence rumour, not a dated commitment. With five frontier models now shipping commercially and capability largely plateaued at the top, the live competition is on token cost, which is why cost and jurisdiction increasingly decide European deployments.
- Open-weight cost counter holds; MiniMax M3 weights imminent. The full open weights and technical report for MiniMax M3 (59% SWE-Bench Pro, 1M-token context) are due around 11 June — the cohort's next real reproducibility test — with Alibaba's Qwen 3.7 still drawing developers at roughly half the input and a quarter of the output cost of comparable frontier models.
- The Pentagon is shopping for a less-restrictive Claude. The US military is testing OpenAI, Google and xAI models to replace Anthropic's Claude on classified systems, having labelled the company a supply-chain risk over its refusal to relax guardrails on surveillance and autonomous-weapons use, with a six-month wind-down underway. The inversion is worth noting: the constraints the Pentagon calls a liability are close to the ones a regulated European institution would call a feature.
- SpaceX prices Thursday. The SpaceX IPO prices 11 June with Nasdaq trading under SPCX from 12 June, targeting up to $75B at a ~$1.75T valuation — the largest listing on record. The AI line investors are scrutinising is the xAI/Colossus compute draw, and the read-through is Anthropic's reported ~$1.25B/month single-vendor compute dependency, which its own prospectus will have to explain.
- Claude's consumer traffic is compounding. Momentic's June chatbot share report (via Similarweb) puts ChatGPT at 54.7% of AI-chatbot web visits and Gemini second at 27.4%, but the standout is Claude growing ~306% in a quarter — enterprise developer mindshare converting into consumer brand awareness just as Apple prepares to add a two-billion-device distribution channel.