Editor's note: today's brief carries a declared Regulation main. The EU AI Act's enforcement powers over general-purpose AI go live Monday, 2 August — a fixed statutory deadline now two days out and the single dominant regulatory event of the cycle, which is the carve-out under which we run more than the usual one regulation story. The lead stays on the freshest 24-hour fact — Amazon's after-close print.
Amazon closes the hyperscaler sweep with AWS's fastest growth in over four years
Amazon reported Q2 2026 results after the 30 July close, and the number that mattered was AWS: cloud revenue grew 37% year over year to $42.2 billion, its fastest rate in 18 quarters, lifting the division to a $169 billion annualised run rate at a 39.4% operating margin (CNBC). Group net sales reached $200.6 billion (up 20%) and operating income rose 43% to $27.5 billion, and the shares jumped roughly 9% after hours (Yahoo Finance). CEO Andy Jassy said Amazon now expects capital spending of about $220 billion this year — raised from around $200 billion on higher memory costs — and the company's AI and custom-chip lines are each past a $25 billion run rate, growing triple digits.
That closes the sweep. After Microsoft's Azure-crossing-$100-billion quarter and Meta's punished spend earlier in the week, AWS reaccelerating — rather than merely holding — hands the bulls the one data point the market had been waiting on, and every one of the big three is now guiding capital expenditure higher into 2027. For European buyers the takeaway is the same as it was on Wednesday, only firmer: the compute that the AI economy runs on is being built faster than ever, and it is being built almost entirely inside US-headquartered, US-jurisdiction cloud estates. Reaccelerating growth is not the same as diversifying supply — it concentrates it.
What actually changes on Monday when the AI Act's enforcement powers switch on
On 2 August 2026 the European Commission's supervision and enforcement powers over general-purpose AI models formally take effect, one year after GPAI obligations first applied (European Commission). The obligations themselves are not new; the machinery to compel them is. From Monday the AI Office can require a provider to hand over technical documentation, demand model access to run its own evaluations, order risk-mitigation, market-restriction, recall or withdrawal measures, and levy fines of up to €15 million or 3% of global annual turnover, whichever is higher, under Article 101 (European Commission — Article 101). Models placed on the market before 2 August 2025 have until 2 August 2027 to come into line; those released since are already exposed, and the Commission can act on conduct dating back to last August.
The practical shift for anyone deploying a frontier model in Europe is that “who can be compelled, in which jurisdiction” stops being a compliance abstraction and becomes an operational one. A regulator that can demand model access and documentation needs a party it can actually reach — which is exactly why the sovereign-AI vendor scramble covered here over the past month reads less like marketing and more like the market pricing in Monday. The durable question for a buyer is not which model tops a leaderboard this week, but whether the data and the accountable operator sit somewhere a European authority — and no other — has reach.
Open at the frontier, still lagging in the serving rack
Artificial Analysis's Intelligence Index is the clearest read yet on where the models actually stand: on the live leaderboard Claude Opus 5 leads overall at 61 and tops the Agentic Index at 55.3, ahead of Fable 5 (60 / 52.8) and GPT-5.6 Sol (59 / 54.0), with Moonshot's open-weight Kimi K3 the top open model at 57 (Artificial Analysis). The v4.1 methodology matters as much as the ranking: it retired IFBench for saturation and rebuilt the composite around agentic and coding-heavy evaluations with per-task cost, time and token metrics (Artificial Analysis v4.1) — a leaderboard that now reads like a unit-economics sheet rather than a trophy cabinet.
The gap the benchmark doesn't capture is the one that decides whether an open model is usable in production: the serving stack. Kimi K3's weights have been downloadable since 26 July, but at roughly a 594GB MXFP4 download and about 1.4TB in memory it needs an eight-GPU H100-class node just to load, and vLLM's support for its bespoke Kimi Delta Attention is still maturing — the project's own preview frames production-stable local inference as a Q4-2026 task for most teams (vLLM; Northflank). The lesson repeats: at the frontier, capability is no longer the bottleneck. The stack, the licence and the jurisdiction are.
Quick Hits
- Snowflake ships an MCP governance gateway — At Black Hat 2026, Snowflake introduced Cortex AI Gateway, governing how first- and third-party agents reach models, data and MCP servers, with agent-identity controls and data-exfiltration prevention — the same agent-identity problem Hush raised its $30M on this week, now a platform feature (Snowflake).
- Mistral's open-weight MoE is still a promise — The “fat but sparse” model Arthur Mensch teased remains in partner early access with no parameter count, benchmarks or licence disclosed; a broader release is still expected “later this summer” (TechTimes).
- GLM-5.5 reportedly lines up for August — Z.ai's next open-weight flagship is said to be close, per a JPMorgan note not confirmed by Zhipu; GLM-5.2 scored 51 on the AA Index, June's open leader before Kimi K3's 57 overtook it (Technosports).
- Gemini 3.5 Pro slips again — Google's next Pro model is still short of internal coding goals with no fresh timeline, even as Gemini 4 pre-training is reported to have begun (9to5Google).
- Enterprise AI money keeps flowing to workflows — Encore AI raised $30M (Team8, Planven) for “interaction mining” that trains agents from calls, chats and CRM signals, one of several late-July rounds landing where investors see fast enterprise revenue rather than model ambition (Tech Startups).