Pacing, day four: who actually pays

Pacing, day four: who actually pays

The people who buy and sell compute have now answered the pacing essay, and their answer is that it changes nothing they can see. Broadcom's Hock Tan has not revisited his fiscal 2027 and 2028 forecasts "in the least", and says Anthropic — the company proposing the slowdown — is on track to become his largest custom-chip customer next year; pacing slows training, and his guidance runs on inference. OpenAI's CFO said "absolutely" to pacing in the same breath as calling it a business decision she would manage, while Musk proposed rivals grade each other's homework and conceded nobody has agreed. Geoffrey Hinton backed the essay and made the cleanest case against the cynical reading: warnings are bad for business, so a pause "may decrease the value" of Anthropic — though a pause that applies to everyone still favours whoever already holds the capital. The Register argued the labs want to be "too big to fail"; John's challenge to its premise that they are "years from profitability" stands, given yesterday's report that Anthropic expects a second straight quarter of operating profit. Rob Manson's essay supplies the sharpest exhibit against a distillation crackdown: when Hugging Face investigated the July break-in, the reconstruction was reportedly done with Z.ai's open GLM 5.2 after closed models refused — "open models were the defence, not the threat." An ex-DeepMind researcher wants compute treated as fissile material; Mark's response, that damage scaling with capability is "provably wrong" on fifty years of evidence, is recorded as his claim, not established. And in Washington, Sanders shared a stage with Bannon under the Future of Life banner while the House prepares to leave until after 3 November: a coalition without a vehicle.

Australia asks for the safety side and is offered the bill

OpenAI told Canberra there will be no Australian training centre unless copyright loosens, and it will not discuss the Prime Minister's renewable-generation condition until that "gating problem" clears. Leaked Attorney-General's slides show the concession already drafted: two routes, both opt-out by default, one paying a central body and one paying only the rights-holders large enough to negotiate a quota. ASD's chief asked for an AI "early warning system" modelled on cyber incident reporting — the one measure every camp accepts — and confirmed Australian agencies ran Mythos, the model US export control had withdrawn from foreign nationals, to find vulnerabilities in hours. If the warning system is formalised before the copyright question settles, Canberra will have taken the safety side of the bargain without yet paying the investment side.

Rubin, and the shortage it lands in

SemiAnalysis' first agentic-inference results on Vera Rubin show 67 times GB300's throughput per dollar at one fast point on the latency curve and 1.4 to 3 times in the band where providers actually serve — a strong generational step measured by a sympathetic tester on early software, not a write-off of Blackwell. Mark's "every data centre built before that is basically just trash" and John's "whatever our thesis was for capacity constraint is going to be out by an order of magnitude" met Mark's own reply: in which direction? The constraint sits upstream regardless: Moody's has US data-centre load doubling to 426 TWh by 2030 against grid delays of up to seven years, and the RTX 5090 has vanished from US online retail with resellers asking up to US$9,500 — a consumer card priced as an inference server. IEEE Spectrum's survey of inference silicon finds the field diverging on every axis as the CPU once did, and Samsung led a US$230m round into Euclyd, a non-GPU inference system due in 2028 — an option on a memory-bound future, priced by the date rather than the claim. Trump phoned Jensen Huang on stage to call data-centre opposition "a hoax" on the day a UMass poll found 65 per cent of Americans oppose one in their community. AWS gave up on Bahrain: region data unrecoverable six months after the strikes, and the UAE weighing whether to disperse and harden Stargate.

The open map, and what sits above it

Mozilla's State of Open Source AI puts the open-weights lag at about 4.4 months — tasks under eight hours can go to either, and closed models earn their premium on the long tail; and its census of the money finds the value already priced one layer up — Nvidia bought the hub, Stripe bought the router, US$20bn between them in a fortnight — with the open ecosystem "largely funded by Chinese capital right now". MediaTek's 2nm Dimensity 9600 Pro claims 30-billion-parameter models fully offline — the tier Mozilla says handles routine work — with no tokens-per-second figure yet. Periodic claims a trillion-parameter model trained on its own lab data beats Astra and Fable on X-ray diffraction: proprietary instrument data no frontier lab can scrape, on a single task, with no published evaluation. TypeSafe's Jev sells typed decisions at four cents per million tokens in, output free, on self-built benchmarks. And Nvidia's OpenShell team proposes proving agent permissions with Z3 rather than paying a second model to watch the first — a supervision cost that scales with agent count, not agent intelligence, and covers only what was modelled.

Marks, credentials and cheap bodies

Bending Spoons bought Miro at a US$1.355bn enterprise value, about 90 per cent below its 2022 equity mark and 2.3 times ARR — the second billion-dollar collaboration tool in five weeks, valued as if its growth has ended. Daniel Litt's positive half of "The End of Mathematics" would award the PhD on a rigorous defence and make the provenance of the mathematics irrelevant — separating evidence of progress from evidence of expertise, the same move Goedecke made about software yesterday. And cheap physical AI arrived twice: Unitree's US$4,900 humanoid, built by a founder who approves every expense over 100 yuan, and OpenArm, a US$6,500 open-source bimanual arm under a licence that lets a lab own its hardware.

Read all 26 signals →