Anthropic in Talks with Nscale for $45 Billion Compute Deal, a High-Stakes Bet Ahead of IPO

The AI infrastructure arms race is still escalating. According to Bloomberg, citing people familiar with the matter, AI safety and model company Anthropic is in talks with compute provider Nscale over a compute agreement valued at $45 billion. Negotiations are ongoing and terms may change at any time.

This potential deal continues Anthropic's aggressive expansion pace over the past year, and once again draws attention to a core tension: when AI companies routinely sign compute contracts worth tens of billions of dollars, can these investments translate into verifiable commercial returns?

Multiple Mega-Deals in a Single Year: Anthropic's Compute Anxiety

According to the report, this is not Anthropic's first large-scale compute purchase. Over the past 12 months, the company has signed multiple high-value compute agreements to support its model training and inference needs. The original article notes that the $45 billion negotiations follow a similar deal in August 2026, suggesting the two parties may already have an established relationship.

This successive escalation reflects a reality: the cost of training frontier models is rising at an astonishing rate, and compute supply has become one of the most critical bottlenecks for model companies. For Anthropic, locking in long-term compute resources is in some sense more important than short-term profit—especially when competing with rivals like OpenAI and Google, where compute reserves directly determine the speed of model iteration.

The Structural Sources of Frontier Model Training Costs

Frontier large-model training costs consist mainly of three components: compute (GPU/TPU hours), data (collection and cleaning of high-quality corpora), and human capital (research and engineering teams). Compute's share has risen sharply in recent years because model scale follows an empirical Scaling Law—when parameter count, data volume, and computation scale up together, model capability improves predictably. This means that to remain competitive, model companies must continuously procure larger clusters, and the cost of training a frontier model once has risen from millions of dollars in the early days to hundreds of millions. More critically, training is not a one-time expense: failed reruns, ablation experiments, post-training and alignment, and inference serving after launch all continuously consume compute. Therefore, locking in long-term compute resources is essentially locking in the pace and certainty of model iteration—which explains why model companies are willing to sign multi-year, tens-of-billions-of-dollars contracts.

Nscale's Dual Identity: Supplier and Soon-to-Be Public Company

The other party to the deal, Nscale, also merits attention. As a compute provider, Nscale is both a beneficiary of this deal and under its own capital pressure. The report notes that Nscale is seeking to go public, but its proposed $30 billion valuation is being questioned by investors.

A $45 billion compute agreement is undoubtedly a shot in the arm for Nscale's IPO narrative—it means long-term revenue visibility. But conversely, if investors' confidence in the valuation logic of AI infrastructure wavers, such mega-deals could also be interpreted as "circular deals" or overly optimistic capacity expansion.

What Is the "Circular Deal" Concern?

In the AI infrastructure space, a "circular deal" refers to a situation where a supplier and customer develop financial or revenue interdependencies that make reported revenue appear stronger than actual market demand. Typical forms include: chip or cloud vendors investing in model companies, which then use that money to buy back their compute; or compute providers financing capacity expansion against future revenue expectations, where that revenue comes from a small number of yet-unprofitable customers. Critics argue that such arrangements disconnect revenue recognition from real end demand, and once financing conditions tighten or customer growth falls short of expectations, valuations along the chain will be revised downward in tandem. For Nscale, a $45 billion long-term agreement is proof of revenue visibility in the IPO narrative, but investors may also press: how concentrated is the customer base, do payment schedules match capacity buildout, and does the contract include adjustable or exit clauses?

What Investors Are Worried About

The original article highlights a key backdrop: some investors are concerned about the ratio between the high cost of AI investment and the profits it generates. This concern is not unfounded. When compute contracts reach the hundreds-of-billions-of-dollars level while AI companies' revenue growth has yet to keep pace, capital markets' patience will be tested.

Both companies are advancing IPOs almost simultaneously—Anthropic is expected to go public in 2026, and Nscale is also preparing—making the above tension even more sensitive. Going public means financial data will be publicly scrutinized, and the market will ask: how long is the payback period for these massive compute investments?

One Deal, Multiple Signals

From an industry perspective, this news conveys at least three layers of information. First, demand for compute among leading AI companies is still growing rapidly, with no sign of slowing in the short term. Second, compute providers are racing to capital markets on the AI boom, but the reasonableness of their valuations will be tested. Third, investors' scrutiny of the "input-return" balance is intensifying, which could affect the negotiating terms and market sentiment of similar future deals.

It should be noted that negotiation details have not yet been finalized, and whether the deal closes and whether the amount is adjusted remain uncertain. But regardless of the outcome, this potential deal has already become a typical sample for observing the AI infrastructure investment cycle.

For readers following the AI industry, the nodes worth tracking next include: formal confirmation of the deal, financial disclosures in the two companies' IPO processes, and the market's continued pricing of compute-intensive business models.

One-line Summary

Anthropic is negotiating a $45 billion compute agreement, reflecting valuation anxiety over AI investment and returns.

Paragraph Summary

According to Bloomberg, Anthropic is in talks with compute provider Nscale over a $45 billion compute agreement; negotiations are ongoing and terms may change. This is the latest in a series of large-scale compute purchases by Anthropic over the past 12 months, reflecting how locking in long-term compute has become a core competitive strategy for model companies as frontier model training costs climb. The other party, Nscale, is seeking to go public, but its $30 billion valuation is being questioned by investors; the mega-deal is both a boost to its IPO narrative and potentially interpretable as a circular deal or excessive capacity expansion. The article notes that investors are increasingly focused on the ratio between high AI investment and profit returns, and that both companies are advancing IPOs almost simultaneously (Anthropic is expected to go public in 2026), which will make this tension even more sensitive. The deal itself conveys three signals: compute demand is still growing rapidly, compute providers are racing to capital markets on the AI boom, and market scrutiny of investment returns is intensifying.