Happy Wednesday. Anthropic filed a draft S-1 that puts hard numbers on a problem every founder building on cloud infrastructure already senses: the companies supplying your compute can also be your fiercest rivals.
Elsewhere, Google moved to loosen Nvidia's hold on AI developers, and a Boston AI startup proved that winning critical government contracts is its own fundraising strategy.
FUNDING & M&A
Anthropic's S-1 Puts a Dollar Figure on the Cost of Building on a Rival's Cloud
Anthropic's draft S-1 discloses fixed commitments of $111 billion to Google, $110 billion to AWS, and $31 billion to Microsoft, with total potential cloud spending exceeding $750 billion. The company spent nearly $9 billion on compute in the first six months of 2026, and its filing names those same cloud partners as competitors that have already declined to supply as much capacity as Anthropic requested.
The filing is the clearest public accounting yet of what it costs to be infrastructure-dependent on your competition. AWS CEO Matt Garman defended the arrangement publicly, but the structural problem is plain: Anthropic's access to compute capacity from Google is explicitly conditioned on its continued commercial success, per a Broadcom regulatory filing. Founders scaling AI products should read this S-1 not for the revenue numbers but for what it reveals about how leverage accumulates inside cloud contracts as a company grows.
ARTIFICIAL INTELLIGENCE
Google's TorchTPU Lets PyTorch Developers Ditch Nvidia With One Line of Code
Google open-sourced TorchTPU on Oct. 20 at PyTorch Conference North America, giving developers a native way to run PyTorch workloads on Google TPUs. Switching requires replacing .to('cuda') with .to('tpu') and nothing else. The project was co-built with Meta's PyTorch team, replaces the now-deprecated PyTorch/XLA, and has been validated at 8,192 Ironwood TPUs with near-linear scaling.
PyTorch powers roughly 85% of AI applications, and nearly all of them were built assuming Nvidia hardware. That assumption has made Nvidia's CUDA ecosystem nearly impossible to displace, not because developers preferred it but because the cost of moving was too high. Google just lowered that cost to one line. Frameworks including vLLM, SGLang, and Ray are already integrated. Watch whether Hugging Face's device_map='tpu' shortcut drives meaningful adoption, because that is the path of least resistance for the broadest slice of AI builders.
ARTIFICIAL INTELLIGENCE
ASI Hits $3B Valuation After Beating Palantir for an $875M FAA Contract
ASI, a Boston-based AI infrastructure company, closed a $325 million Series C led by Spark Capital at a $3 billion valuation. The raise follows an $875 million, 12-year FAA contract awarded in June 2026, which ASI won over Palantir and Thales, and a $99 million U.S. Army contract signed in October. Alaska Airlines saves roughly one million gallons of fuel per year using ASI's Flyways product.
The investor list is notable: Andreessen Horowitz, Craft Ventures co-founded by former White House AI adviser David Sacks, and affiliates of Citadel Securities all participated. That group does not typically cluster around early-stage bets. What they are backing is a company that targets sectors where AI errors carry real-world consequences, which sets a different pricing dynamic than general-purpose AI vendors. Founders competing for enterprise AI budgets should note that government contract wins of this size are functioning as the new Series A proof point.
Created by the robots at The SaaS Sentinel