Ten trillion tokens a day, and capital wired to silicon

The read

OpenRouter announced it is joining Stripe, making official a deal first reported in July on sourcing nobody would confirm. The announcement carries the first hard operating figures the company has published — more than ten trillion tokens a day across 400-plus models — which is the number to hold onto, because until now the routing layer's scale was inferred rather than disclosed. An investor argues the acquisition was a security buy; that is an argument by a disclosed board member, not an account from either company.

Capital kept getting wired directly to silicon, in three different shapes. A filing gives Google the right to purchase up to 58,970,907 Marvell shares at US$206.58, roughly US$12.2bn, tied to purchasing targets running to Marvell's 2033 fiscal year — equity as the settlement currency for a supply commitment, on a seven-year horizon. Nvidia has been introducing companies holding its GPUs to Nordic data-centre operators with spare capacity, two sources told CNBC, and in one case sounded out an operator about potential offtakers; John's question was what Australia might learn from a chip vendor acting as a matchmaker for its own installed base. And Etched has shipped its first rack — to Jane Street, which led its US$700m round at a US$21bn valuation after testing the hardware. Customer and lead investor being the same party is worth naming plainly. Meanwhile OpenAI's CFO told an all-hands the company will be public in 2027, calling the IPO "another fundraise" rather than a finish line — and the run-rate comparisons circulating alongside it do not hold up well.

On the models, two releases and a processor. GLM-5.3 scores 60 on Artificial Analysis, eighth of 182, and pays for it in tokens — the verbosity is the cost, and it shows up in the bill rather than the benchmark. Ornith-1.5 puts the harness inside the training loop, optimising scaffold and policy jointly, which is the most direct test yet of an argument we have run for a year: that the harness moves results more than the model does. Its own numbers show parity at best, and Kimi K3 is ahead on both flagship benchmarks. And Alibaba is running its 27B model on its own RISC-V processor, at edge speeds rather than API speeds — the point is not the throughput but the absence of a GPU anywhere in the sentence.

Two items go to whether you can trust what an agent tells you it did. Two ICML papers argue the reasoning trace is not evidence of alignment: unfaithful chain-of-thought turns up on ordinary, non-adversarial prompts with no injected bias, which undercuts reading the trace as an audit record. Against that, Linear's own instrumentation reports agents now authoring just under half of issues created — a vendor-defined metric measuring work described rather than work done, but a striking one, and the adoption curves underneath it more than doubled across every job function in six months.

Three on what it means for people. Goldman finds AI job pressure real, narrow, and worst at entry level — an association with confounders named, not a causal finding, and the entry-level concentration is the part worth arguing about. Terence Tao's "Mathematics in the age of AI" deliberately declines the capability argument and conditions on it instead, which is why the controlled evidence in it will be misquoted in both directions. And a hypothesis worth testing: a stable core plus sandboxed user extensions, on the reasoning that models have collapsed the cost of authoring an extension while sandboxing has collapsed the cost of allowing one.

Finally, AI hardware may be acquiring a cargo-theft profile — two California incidents and one investigator's account, which is thin sourcing for a trend, but the physical-security question follows the value density whether or not this particular story holds.

Read all 14 signals →