The White House Puts an Open-Weight Model on Trial
The most consequential AI story of the week isn't a benchmark — it's an accusation. Michael Kratsios, director of the White House Office of Science and Technology Policy, said on 22 July that the US has “information that Moonshot AI distilled Anthropic's Fable” to build its Kimi K3 model, describing a “sophisticated internal platform” for large-scale distillation against US models that could switch between access methods to avoid detection. He separately alleged that Moonshot obtained Nvidia GB300 servers and accessed the chips in Thailand — a possible export-control breach. The Treasury has reportedly threatened sanctions in response.
The claims are unproven, and worth reading with the caveats attached: Moonshot has not conceded, Kratsios offered no detail on how the government knows, and outside experts note that Fable has only been publicly available since 1 July — a thin window in which to distill a 2.8-trillion-parameter model. But the timing lands hard: Kimi K3's open weights are still scheduled to go free on 27 July. For a European bank or hospital weighing whether to self-host it, the practical shift is that a model's provenance now sits alongside its scores as a diligence item — and the cleanest hedge against a contested lineage is an architecture that treats any single model as swappable, hosted where you control it rather than wherever a leaderboard points.
The Enterprise Battle Moves to the Governance Layer
While labs fight over training data, the enterprise money is chasing something less glamorous: the layer that makes agents safe to deploy. Microsoft and Databricks expanded their partnership into the 2030s on 23 July, pushing Databricks' Genie co-worker and its Unity AI Gateway — which governs models, agents and cost across providers — deeper into Azure, Foundry, Purview and Microsoft 365. Databricks will also run its own core operations on Azure Databricks and lean on Azure Cobalt, Microsoft's Arm-based silicon, for agentic workloads.
The pattern is unmistakable this week. OpenAI's new Presence platform, Google's Gemini Enterprise and Meta's business-agent push are all selling the same thing: not the smartest model, but the policy, permissions, audit and evaluation scaffolding that turns a demo into a production deployment. It's a healthier competition — and a validating one for anyone who has argued that multi-model governance, not model lock-in, is where regulated buyers actually spend. The buyer's question is no longer “whose model is best” but “whose control plane fits my compliance posture,” because switching an agent platform costs far more than switching an API endpoint.
Alphabet's $205B Bet Sharpens the Sovereignty Question
Alphabet's Q2 numbers, reported 22 July, show why the AI-infrastructure arms race keeps escalating: Google Cloud revenue jumped 82% to $24.8 billion, backlog swelled to $514 billion, and the company raised full-year 2026 capital-expenditure guidance to $195–205 billion — with a warning that 2027 spending will “increase significantly.” Investors flinched, sending the stock down roughly 5% despite a revenue beat, but the message to the market was that demand still outruns supply.
That spend is overwhelmingly concentrated in US hyperscaler campuses, and it lands as OpenAI, Meta and others commit tens of billions more to gigawatt-scale sites where power, not silicon, is the binding constraint. For European enterprises, the arithmetic cuts two ways: the frontier keeps getting cheaper to rent, but the compute — and the jurisdiction it sits in — concentrates further from home. It's the backdrop against which Europe's own gigafactory ambitions and sovereign-hosting arguments will be judged over the next 18 months.
Quick Hits
- EU AI Act enforcement powers go live 2 August — In nine days the Commission's GPAI obligations become enforceable: the AI Office can demand documentation, run technical evaluations, order mitigations and levy fines up to €15M or 3% of global turnover under Article 101. The single most consequential date left on the Act's calendar.
- DeepSeek completes its V4 cutover today — As of 24 July 15:59 UTC, DeepSeek retires the legacy deepseek-chat and deepseek-reasoner names, routing all traffic to the open-weight V4 family — 1M-token context by default and roughly $0.28 per million output tokens on V4-Flash, the price floor the industry gets measured against.
- OpenAI launches Presence — The enterprise agent platform (announced 22 July) wraps voice and chat agents in shared policies, guardrails and evaluations; BBVA (Mexico) and SoftBank (Japan) are among early testers, and it already runs OpenAI's own support line at ~75% auto-resolution.
- Kimi K3's measured performance is real regardless of the dispute — The 2.8T-parameter model topped independent coding evaluations before the White House statement; the accusation is likely to slow cautious enterprise adoption more than individual-developer use, which is arguably its intended effect.