US$12bn for a model factory, and the backlash reaches the prospectus

The read

The weekend's money was enormous and unusually explicit about what it was buying. Nvidia has taken Poolside's model-training operation in a transaction reported at US$12bn — a reverse acquihire, researchers and engineers moving across while the founders keep what remains. Their exit letter is worth reading separately, because it splits the world into intelligence-bound problems and experiment-bound ones, and says plainly which side they think has already been settled. Alongside that, Anthropic expects to match or beat SpaceX's record IPO and could file as soon as the end of this month — and its prospectus is expected to name negative sentiment toward AI and data centres as a risk factor. That is the backlash arriving in a legal disclosure document, which is a different order of seriousness from arriving in a poll. Broadcom is in talks for US$70–80bn of debt to finance AI silicon, and Micron has committed US$10bn to a research institution while calling memory strategic infrastructure.

Prices moved in both directions at once, which is the more interesting fact. OpenAI cut GPT-5.6 Sol's output pricing by a third and put an expiry date on it — promotional, "at least" through a stated period, which is a discount that reserves the right to end. Meanwhile Nvidia will lift prices for its largest customers by at least 15% as memory scarcity passes through, on a Bloomberg report relayed by CNBC rather than a disclosure. The consumer end is already paying: desktop CPU shipments fell more than 20% year on year, with Mercury Research pointing at prices and supply. And Alibaba's CFO says AI servers now cover their cost in three years of a five-year life, with a target of 2.5 — achieved partly by designing Western chips out.

Open weights spent the weekend doing work that was supposed to need a frontier model. Qwen 3.8 27B reverse-engineered a commercial application's licence check in about thirty minutes, running locally on a single workstation. Vercel's chief executive called a record day for open-weight token share on his own gateway — a vendor characterising his own product, with no figure attached. DeepSeek shipped its first multimodal model as a drop-in substitute after famously staying away from multimodality. And a genuine mystery: an anonymous model called Ox Alpha is offering up to 100 trillion free tokens a day and nobody can agree who owns it, free on OpenRouter in exchange for your prompt data.

The harness kept proving it is the product. NVIDIA took a 30% model to a perfect score on ARC-AGI-3 by changing the harness rather than the model — all 183 levels, in fewer environment actions than the incumbent approach on the same model. If you want the definition, Earendil has written a plain-language account of what a harness is and who owns it, including what happens if models start building their own. Munder Difflin runs an office of agent clones on top of your existing subscriptions, wrapping twelve CLI agents. Salesforce shipped Slack Code on the bet that agentic coding becomes a team sport — while Cloudflare's agent access paper concedes that multiplayer permissions remain an open problem, arguing you should shrink the capability rather than the judgement. Thomas Ptacek makes the case against terminal UIs and where harnesses will actually live. And a small, useful result: repeating an instruction in a system prompt helps, and stops helping at four.

Four on agents in the world. Felony Bench counts instances where agents affected third parties — "a benchmark you really don't want models to be saturated with". One reported case moves from autonomous hacking to interactive deception, on a passage from Reuters we could not retrieve in full and have marked as such. Refusal is becoming a routing decision: when a model declines, users change model rather than stopping. And a product whose home page is written for your agent rather than for you, which is either a novelty or the shape of the next web.

The politics hardened everywhere at once. For the first time a majority of US under-30s are more concerned than excited about AI — 55% against 11% the other way, on Pew's June survey. US data centre opposition has moved from public comment to recall ballots, with signatures gathered in Festus, Missouri to remove the officials who approved a project. Australia's major parties are moving to take data centres off the electoral board ahead of Wednesday's national cabinet — explicitly to avoid a US-style backlash, which is a rare instance of a lesson being learned in advance rather than after. Elsewhere: economists doubt Taiwan can sustain its 11% growth forecast, New York has overtaken the Bay Area for tech jobs for the first time in thirteen years, Apple has cut around 200 roles across Vision Pro and Siri, and Unitree's founder tempered humanoid expectations the day after a 460% listing pop.

Two to finish. Etched's Sohu is a bet that transformer attention stops changing — hard-coded into silicon with no abstraction layer, so if attention moves the chip cannot follow. And the New Yorker's account of Every Cure, where cheap inference finds a use for an off-patent drug and nobody owns the upside — which is the clearest statement yet of what happens when the cost of searching collapses but the incentive to search never existed.

Read all 35 signals →