Anthropic and Nscale Eye Second $45 Billion Compute Agreement in Two Months
The AI foundation model maker continues its aggressive infrastructure spending spree despite investor concerns over capital intensity and path to profitability
A Pattern of Billion-Dollar Commitments
Anthropic and AI infrastructure provider Nscale have entered discussions for a second $45 billion compute agreement, according to people with knowledge of the negotiations. The talks come roughly six weeks after the two companies announced an initial deal of identical size in early September 2026.
The negotiations remain fluid, and the final terms may shift before any formal announcement, the sources indicated. Both companies declined to comment on the discussions.
For Anthropic, the potential agreement would extend a year-long pattern of securing enormous compute capacity commitments. The San Francisco-based foundation model developer has pursued a strategy predicated on the assumption that training ever-larger models will unlock commercially viable capabilities that justify the capital outlay.
The Economics Under Scrutiny
The compute arms race has drawn scepticism from a segment of institutional investors who question whether the revenue potential of large language models can support the capital intensity required to build them. At Opentechwire, we've tracked similar concerns emerging across the sector, particularly as companies approach public markets.
Anthropic's willingness to lock in multi-decade compute commitments reflects a thesis that current spending will compound into durable competitive advantage. The company has bet that frontier model performance will translate into enterprise adoption at margins sufficient to service these obligations.
Yet the economics remain unproven at scale. Inference costs, while declining, still represent a substantial share of revenue for AI application companies. Meanwhile, the pace of model improvement has not followed a predictable trajectory, introducing execution risk into long-horizon capacity planning.
Two IPO Candidates, Different Reception
Both parties to the potential deal are pursuing public listings, though with divergent momentum. Anthropic has maintained its timeline for an initial public offering before the end of 2026, a schedule that would make it one of the highest-profile tech debuts of the year.
Nscale's path has encountered friction. The infrastructure provider is seeking a $30 billion valuation in its offering, a figure that has met resistance from prospective investors, according to people familiar with the roadshow. The pushback centres on questions about Nscale's customer concentration and the sustainability of its growth rate as hyperscale cloud providers expand their own AI-optimised infrastructure.
The timing of a second deal announcement, if it materialises, could serve both companies' listing narratives. For Anthropic, it would signal confidence in future compute needs and, by extension, model development progress. For Nscale, it would demonstrate demand visibility and the ability to secure repeat business from marquee customers.
What the Deal Structure Reveals
Multi-billion-dollar compute agreements typically span five to ten years and blend capacity reservation with usage-based pricing. The arrangements provide infrastructure operators with revenue predictability while offering AI developers guaranteed access during periods of tight supply.
The structure also shifts risk. If Anthropic's models achieve breakthrough performance and enterprise adoption accelerates, the locked-in pricing could prove advantageous relative to spot market rates. Conversely, if model efficiency improves faster than anticipated or demand growth disappoints, the company faces utilisation risk on committed capacity.
At the scale Anthropic is operating, even marginal improvements in training efficiency or inference optimisation can translate into hundreds of millions in saved costs. The company's recent architectural work has focused on reducing the compute required per unit of model capability, a priority that exists in tension with its capacity expansion strategy.
Regional Implications and Supply Chain
The geographic distribution of compute capacity has become a strategic consideration for AI developers, particularly as export controls and data sovereignty requirements shape infrastructure decisions. While neither company has disclosed where the capacity covered by these agreements will be deployed, the scale suggests a multi-region footprint.
Asia-Pacific markets represent both a demand source and a regulatory complexity for foundation model providers. Anthropic has pursued partnerships in Japan and South Korea, jurisdictions where local compute infrastructure can ease compliance with data localisation mandates. Nscale's own capacity buildout has emphasised edge locations in high-growth Asian markets, a positioning that aligns with latency-sensitive enterprise use cases.
The capital flowing into these deals also reflects a broader reordering of infrastructure investment. Across the region, we've observed sovereign wealth funds and strategic investors backing compute capacity as a strategic asset class, a shift that has implications for pricing power and long-term supply dynamics.
The Profitability Question Persists
For all the capital committed to compute, the path to durable unit economics in foundation model development remains a subject of debate. Anthropic's revenue has grown substantially over the past year, driven by API access and enterprise licensing, but the company has not disclosed whether it operates at a gross margin sufficient to cover both compute costs and research investment.
The company's willingness to continue signing deals of this magnitude suggests confidence in its revenue trajectory. Yet the broader market for large language models is evolving in ways that may challenge pricing assumptions. Open-source alternatives have narrowed the capability gap with proprietary models, while enterprises are increasingly exploring fine-tuning and model distillation strategies that reduce reliance on frontier-scale inference.
Anthropic's strategic calculus appears to prioritise maintaining a performance lead over near-term profitability. That approach has precedent in technology markets where scale and capability advantages compound into defensible positions. Whether the same dynamics hold in AI, where model weights can be copied and architectures reverse-engineered, is an open question that this compute spending will help answer.
