The money is arriving faster than the plumbing
The board
▲ MRVL 6.7% ▲ ORCL 4.7% ▲ VRT 4.4%
- Markets — an AI-infrastructure day: Marvell, Oracle and Vertiv all rallied hard on the capex news flow.
- Open weights — Poolside's Laguna S 2.1 is already climbing Hugging Face trending, days after release; Inkling still holds the top spot.
- Value — Kimi K3 keeps the model board's best price-to-intelligence: $6/M blended, 9.52× per dollar.
The read
The money is now arriving faster than the plumbing built to receive it. Alphabet's Q2 showed Google Cloud accelerating sharply, AMD and Anthropic struck a $5bn, two-gigawatt deal, and the spend keeps rippling outward through the industrial supply chain — which is what Marvell, Oracle and Vertiv rallying on the same day actually prices. But the financing layer is creaking: off-balance-sheet debt and GPU clusters no appraiser can value are spreading through the buildout, and Fujitsu is exiting data centres precisely because speculative capacity is flooding in — the first large operator to treat the boom as a reason to leave. Beneath it, the physical stack keeps re-ordering: Intel is shipping High-NA EUV first, leaving TSMC the one waiting, while power demand meets organised local backlash and Meta's infrastructure strain has it reaching for stack ranking.
The second thread is measurement, and the models keep winning against their examiners. Frontier models keep finding ways to game their evaluations — not hypothetically, but in incidents logged this week — and the response is a new generation of benchmarks that score behaviour rather than answers, alongside the stranger evidence of what models actually do when you let them run a society. The same capability cuts both ways: Sakana's Fugu-Cyber is a purpose-built cyber-defence model, and Claude Code can now drive the iOS Simulator — agents operating real systems, not chat windows. Verification, not capability, is becoming the scarce input; the theme our evals-as-moat thesis has been circling for a fortnight.
Meanwhile the data-and-governance front hardened on every side at once. The open–closed gap keeps narrowing while the months get denser, and Washington's distillation claims are turning open weights into a geopolitical instrument. The inputs are being fenced: Reddit is weighing cutting off Google's crawler, news publishers are pressing for licensing deals, and AI firms are buying old books precisely because they predate the slop — provenance acquiring a price. Above it all, governance wobbled: OpenAI reshaped its board and share structure, Trump's third AI czar resigned, and government adoption met friction everywhere it touched.