Supply is the cap, and the money went to the primaries
The board
▲ CRM 22.6% ▲ CRWD 20.5% ▲ FIG 13.3% ▲ PANW 12.8% ▲ NOW 10.0% ▲ NVDA 8.7%
- Markets — Thursday's session was the largest single-day move we have recorded on this board: Salesforce +22.6%, CrowdStrike +20.5%, Figma +13.3%, Palo Alto +12.8%, Atlassian +10.2%, ServiceNow +10.0%, UiPath +9.4% and Nvidia +8.7%. Every name in the slice finished higher.
- Open weights — Qwen3.8-Flash-Next is climbing, with GLM-5.3-Flash now trending alongside it and Qwen3.8-27B still there.
- Models — GPT-5.6 Sol (max) holds the value pick at $8/M blended; Opus 5 leads intelligence at 63.1. OpenAI's provider status read degraded at capture.
The read
Nvidia had the kind of day that reorganises a sector's assumptions. It guided to 70% revenue growth in FY28 and said the constraint is supply, not demand — which is a different claim from a bullish forecast, because it locates the ceiling outside the company's control. It is also reported to be acquiring Hugging Face at around US$13bn, roughly 80 times ARR, on reporting that contradicted itself within a day: one outlet said talks had not concluded while another called it confirmed, so treat the deal as reported rather than done. It is optimising for DeepSeek and Qwen while warning that Washington may stop it, which is an awkward position stated in public. And AI money has arrived in US primaries, with an Nvidia PAC and US$500,000 for Senator Markey's challenger.
The supply story ran underneath all of it. Memory remains the constraint: SK Hynix broke ground in Indiana — packaging, with a cleanroom due October 2028 — while Android gets a RAM diet. Chip tariffs may reach servers and consoles at the worst possible moment, though they remain under consideration and the White House called the reporting baseless speculation. And Z.ai says 100,000 China-made chips serve all GLM-5.3-Flash traffic — the company's own claim, and if it holds it is the most concrete evidence yet that a frontier-adjacent service can run without Western silicon.
Two items on what agents are actually doing to organisations, and they point opposite ways. Meta built a plan to cut teams by 60% using agents, then cancelled it. Coding agents installed unowned packages inside Fortune 500 networks — a supply-chain exposure created by the tools themselves, which is the kind of failure that arrives before the governance does. Alongside them, a distinctive vocabulary cluster now reaches 45% of a GitHub pull-request corpus, which is a measurement of penetration rather than of quality.
The local-model argument reached a marker. Mark calls the watershed passed: a near-frontier model on tap locally, his own reading of Artificial Analysis figures rather than a published claim. A working developer argues the harness, not the model, is now the product. Segment's co-founder makes the token-spewer case for small models. And against the enthusiasm, a sobering denominator: two per cent of US households pay for a generative AI subscription.
Four more. Anthropic has previewed a Model Hardware Standard for lab and factory machines — a preview, not a shipped specification. MIT's committee says every subject may need rebuilding around AI, a recommendation rather than policy. Irish data centre water use nearly doubled through a hosepipe ban, which is the sort of fact that decides local politics. Wagtail shipped an agent-facing API while Microsoft renamed a roadmap. And OpenExecutive packages a C-suite as eight agents behind a single persona, which is either a joke or a preview, and we are not yet sure which.