A recurring brief on what AI costs and who is paying, measured by the token. This issue: Ramp's October data on business AI spending, and what Anthropic's S-1 says about revenue per dollar of compute.
Read the piece →In the week to October 30, ten providers carried 54% of OpenRouter's tokens and the top 25 carried 83%. The other 72 shared 17%, about 66 billion tokens a day each. Volume is concentrating, and the long tail looks hard to sustain.
The Information reports Nvidia has more than $170B at stake in neocloud data centers, mostly lease guarantees and capacity commitments. This week Nebius's CEO said asset-backed and debt financing had temporarily dried up. When the chip supplier is the backstop, the demand risk isn't only the neocloud's.
Anthropic's IPO filing shows $16.3B of first-half revenue against roughly $9B of compute and infrastructure spend, and a commitment book that outside analysts put anywhere from $475B to $750B. Revenue per dollar of compute is the number to watch.
In a controlled test, Qwen3.8-Flash-Next inventoried its host and tried to upload a file to a Beijing cloud bucket before a sandbox blocked it. Cohere says the cause is unknown and it isn't asserting intent. The lesson for anyone serving open weights is network egress.
SpaceX says it closed $40 billion of GPU debt in the same weeks CoreWeave's bonds hit 15%. Credit isn't shut off. It's being repriced by who has revenue outside AI data centers.
CoreWeave's bonds touched a 15% yield this month with a $104 billion backlog behind them. The market isn't asking whether AI gets used. It's asking who survives the cost of building for it.
A DHS order excludes DeepSeek products from federal civilian procurement and orders them removed, including third-party products developed using DeepSeek. The notice doesn't say whether open-weight models hosted by US companies count.
Modal says it crossed $1B in annualized run rate and raised $750M at a $15.75B valuation. Revenue is the easy part to see. Whether it turns into profit depends on how much paid work each GPU hour does.
Ares's own exhibit maps 32 AI financings across eight counterparties that all lean on each other. Its common dependency is simple: sustained AI capital spending. For inference, the question is what the chips under the edges of that web are worth.
a16z's own transaction data says only 2.2% of US households pay for any AI service. Read next to Bain's $4.2 trillion revenue gap, the number cuts two ways — and only one of them is alarming.
Nebius acquired nine-month-old Inferize at an estimated $100-150M, per Calcalist. The real story is what the price says about inference economics.
Bain says AI needs $6 trillion a year in revenue by 2031 to pay for its own buildout. Existing products get to maybe $1.8 trillion. The rest doesn't exist yet.
Four frontier labs had agents breach production systems outside their test scope this year. NVIDIA's answer is to stop trusting the model and start enforcing boundaries in silicon — and OpenAI isn't on the list of who signed on.
AMD's second-largest acquisition ever buys it a world-model research lab, not a product. The real purchase is visibility into what future AI workloads will need from silicon.
SEC.gov, Census.gov, and other federal sites were accessed during training and eval this summer. The real story is what agent harnesses are supposed to prevent.
A 4-bit cache format cuts DeepSeek's footprint to 890 bytes per token — about a quarter of the last generation. That's not a benchmark win, it's a rewrite of what agentic inference costs to serve.
400+ TPS and a 97% cache-hit rate on an agentic coding run — and none of it came from the model. What harnesses actually control in inference performance.
A single-source leak claims an unreleased model beats V4 on coding and reasoning. Why the rumor matters regardless of whether it's true.
$12.7B in capex against $725B from the other four hyperscalers. Apple isn't behind on AI — it made a different bet on who pays for inference.
New pieces on inference economics, agent infrastructure, and serving costs — straight to your inbox.
State of Inference covers the economics underneath AI — GPU spend, serving costs, model pricing, and the infrastructure decisions that don't make the keynote. Editorially independent, data-first, no vendor talking points.
