Anthropic has committed nearly $90 billion to secure compute infrastructure through Nscale: $44.6 billion confirmed in a September SEC filing, plus a second deal worth approximately $45 billion now in advanced negotiations, a figure reported by Bloomberg on October 19 and confirmed in secondary coverage, not carried by any Nscale or Nvidia primary surface. Both deals target the Monarch Compute Campus in West Virginia. The numbers alone are the story: model weights are becoming a commodity, and the race has moved to who controls the watts.
That shift did not happen gradually. It happened twice in one week. On October 20, DeepSeek open-sourced DeepInfer, releasing its production inference engine under Apache 2.0. The same day at PyTorch Conference, llm-d 1.0 reached general availability: a CNCF Sandbox, vendor-neutral serving layer backed by Google, IBM, Red Hat, CoreWeave, and NVIDIA, with 500-plus contributors across 25-plus organizations. IBM claims, vendor-reported, that self-hosting on llm-d delivered a 5-10x cost reduction over commercial API pricing for its GLM-5.2 deployment on 544 H100s, at roughly $0.11 per million tokens. When the engine is free and the serving stack is standardized, the highest-margin part of the stack is no longer the software layer. It is the physical layer underneath.
Anthropic appears to have drawn that conclusion early. The Monarch campus near Point Pleasant, West Virginia, spans over 2,250 acres with a planned investment of roughly $71 billion, including approximately $47 billion earmarked for Nvidia Vera Rubin chips. Each building carries 460 megawatts of power capacity. Phase one targets 1.35 gigawatts total. The initial agreement, $44.6 billion disclosed in Nscale’s IPO filing, makes Anthropic the largest customer of an Nvidia-backed infrastructure company that is itself competing for the same scarce resource: power. Google and Microsoft both evaluated the first Monarch building before Anthropic secured it. Microsoft separately holds a $43.8 billion commitment with Nscale running through 2033.
The scarcity signal is sharpest in China. ByteDance, Alibaba, and Tencent have collectively ordered over 3 million units of Nvidia’s RTX PRO 5500, the Blackwell-generation chip configured for the Chinese market, representing roughly $35 billion in committed capital. According to unverified reports circulating on social media, the current fill rate on those orders sits between 5 and 10 percent. That is not a supply delay; it is a structural bottleneck. Even with deep pockets, Chinese firms cannot secure enough silicon to maintain competitive parity in frontier model training and deployment. The hardware access question has become an industry-shaping constraint, not merely a procurement inconvenience.
The convergence of these signals tells a coherent story. On the software side, two independent releases in one week, DeepSeek’s engine and the llm-d serving layer, commoditized the entire inference stack for open-weight models. On the physical side, Anthropic is locking down gigawatt-scale power before competitors can, and Chinese firms are placing orders so large that no supply chain can absorb them. The license gates that four labs erected at $20 million in monthly revenue only make strategic sense if the weights themselves still carry premium value. If they do not, then the defensible asset is not the model: it is the campus, the chips, and the power agreement behind it.
Nvidia occupies both sides of this structural contradiction. Its hardware is the bottleneck that Anthropic is trying to buy its way out of, and that Chinese firms are trying to order their way through. But Nvidia also helped fund the open-source serving stack that commoditizes the software layer sitting on top of that hardware. The company profits whether the compute is scarce or efficient. The value is draining out of the model layer and accumulating at the infrastructure layer, exactly where Nvidia already sits.
Within a day of the events at the center of this analysis, NVIDIA proved the thesis in public. On October 21, at GTC Berlin, Jensen Huang told the audience, “Long before the AI models emerge, the infrastructure has to exist.” The program then delivered the same argument in hardware terms: Vera Rubin infrastructure reaching general availability with Google Cloud, DSX AI Factories in modular containers that cut deployment from six months to fourteen days, and a Joint Declaration of Intent with the German federal government to advance physical-AI models for European industry. NVIDIA’s own primary carries a cumulative figure just under $10 billion committed to Europe, with Paris, London, Stockholm, and Dublin named in its accounting: not the approximately $2 billion figure that circulated as media expectation before the keynote. The weights were already free. The infrastructure contract was the Berlin headline.
The question for the industry is no longer which model is best. It is who can keep the lights on at scale. As the software layer becomes a public utility, the moat is the watt.