Why Nscale-Anthropic's $45B Deal Ignores China's AI Play

The Second $45 Billion Bet

Anthropic is about to do exactly what it did last month. Bloomberg reports that Anthropic and Nscale are on the verge of announcing another $45 billion compute deal—the same staggering figure as their first agreement barely weeks ago. Sources caution that talks remain underway and terms could shift, but the direction is unmistakable: Anthropic is committing to a spending trajectory that no traditional tech company would survive.

The pattern matters more than the headline number. This is not a company diversifying its options across multiple infrastructure providers. This is a company that has chosen a single compute partner and is now placing repeat, escalating bets on top of the original commitment. The total compute spending, if both deals close, approaches $90 billion—a figure that would make even the most aggressive capital allocators in Silicon Valley pause and question whether the underlying assumptions about return on investment hold any water.

What makes this particularly striking is the speed. Announcing a $45 billion deal is already extraordinary. Announcing a second one in the same month signals that the companies involved do not view this as an experiment to be evaluated and adjusted. They view it as a non-negotiable strategic posture.

Why This Matters for China

The most striking detail in this story is what the source text does not explicitly highlight: neither Anthropic nor Nscale is building anywhere near China. The entire $90 billion compute strategy—if both deals materialize—runs through American and British infrastructure. That is not an accident. It is a direct consequence of US export controls on advanced chip technology that have tightened relentlessly since 2023.

China’s AI competitors are systematically locked out of the most powerful GPU supply chains. Nvidia’s A800 and H800 chips were restricted in late 2023, followed by tighter controls on the H100 and A100. The current landscape forces any company pursuing frontier model training to go through American or British-licensed routes. Anthropic’s entire compute architecture, from Day One of this deal-making cycle, has operated within that constrained corridor.

The strategic implication is stark and rarely discussed in earnings calls: whoever controls the compute pipeline controls which AI models get trained on American-grade hardware, and those models inherit American licensing constraints, audit requirements, and geopolitical dependencies from the ground up. Chinese LLMs trained on NVIDIA Blackwell chips shipped through Singapore are increasingly rare, and the regulatory environment makes even indirect access harder by the month. Beijing has responded with its own domestic chip push—Hygon, Cambricon, and Huawei’s Ascend series—but these alternatives still lag behind in both performance and ecosystem maturity.

Second-order effects are already visible. US data center operators in Virginia, Texas, and Ireland are competing aggressively for Anthropic’s business, driving up land and power prices in those corridors. UK planners are fast-tracking new energy allocations specifically for AI infrastructure, a move that has drawn criticism from environmental groups concerned about the impact on national grid stability.

The IPO Pressure Cooker

Both companies are currently pursuing IPOs, which fundamentally changes the calculus around this spending.

Anthropic is expected to go public this year. Nscale is seeking a public listing but faces investor pushback over its proposed $30 billion valuation, Bloomberg reported Monday. The pushback is telling. Investors are evaluating a company whose core business model appears to be spending $45 billion on compute and then hoping the economics work out afterward—without a clear path to monetizing that spend.

This is not a new tension in AI infrastructure. The question has always been whether compute-intensive model development can eventually generate enough revenue to justify the capital expenditure. Anthropic has pressed forward regardless, driven by what Bloomberg describes as conviction that the spending will pay off. Whether that belief is justified is the very question that public markets will answer first for Anthropic and then for Nscale.

The IPO window is likely the primary reason these deals are moving so fast. There is a narrow period—in the months before a company files its prospectus—in which both firms can present their compute commitments as competitive moats and barriers to entry rather than financial overreach. After going public, that narrative shifts dramatically. Once both companies are subject to quarterly earnings scrutiny, the $90 billion outlay becomes a line item that analysts will dissect, compare to revenue, and judge harshly if margins look thin.

There is also a timing risk. If the IPOs are delayed—which is common in volatile markets—the deals could age poorly. Public company expectations are different from private ones. Institutional investors demand transparency and near-term profitability signals that these compute commitments cannot currently provide.

Who Wins and Who Loses

American and British data center operators win clearly. Every dollar of this compute deployment flows through jurisdictions with stable energy regulation, favorable tax treatment for large-scale infrastructure, and reliable power grid access. The UK, in particular, has positioned itself as an alternative compute hub as American electricity costs climb and zoning fights intensify in states like Texas and Virginia. British government officials have publicly welcomed the investment, framing it as a centerpiece of the country’s post-Brexit technology strategy.

Chinese AI companies lose, though not in the way a simple sanctions narrative would suggest. The loss is structural rather than prohibitive. It is not that China cannot build AI—it can and does—but that the most powerful frontier models will increasingly be trained on infrastructure that sits outside Chinese jurisdiction, under American oversight, with American energy and American chip dependencies at the core. That creates a long-term competitive asymmetry that compounds with every iteration.

US chip manufacturers win conditionally. More compute deployment means more chips—assuming Nvidia and its partners can meet demand. But if the ROI on these deals disappoints, the next round of compute orders disappears just as quickly as they appeared. The risk for chipmakers is that they have already expanded capacity based on these projected orders, meaning a demand winter would leave them with expensive, idle fabs. AMD and Intel are watching closely, looking for openings in a market that is currently Nvidia-dominated.

Energy providers, both public and private, are quietly positioning to benefit. AI data centers are among the most electricity-intensive facilities on earth. Anthropic’s commitment alone could consume several hundred megawatts of continuous power, rivaling the output of a mid-sized nuclear reactor. Utility companies in hosting regions are already renegotiating contracts and exploring dedicated generation partnerships.

The Geopolitical Dimension

This deal is not happening in a vacuum. The US government has made no secret of its desire to concentrate cutting-edge AI compute within friendly jurisdictions. The CHIPS Act, export controls, and bilateral agreements with the UK, Japan, and South Korea all point toward the same strategy: keep the most powerful AI training infrastructure out of adversarial hands.

Anthropic’s choose to partner exclusively with Nscale on American and British infrastructure aligns precisely with that policy objective. The company has positioned itself as a responsible AI developer, and this compute strategy reinforces that branding while simultaneously serving broader national security interests. The two narratives are not in tension—they are mutually reinforcing.

European regulators, meanwhile, face a dilemma. On one hand, they want the investment, the jobs, and the technological spillover. On the other, they have growing concerns about the concentration of AI infrastructure in the hands of a small number of American companies and their chosen partners. The EU’s AI Act and its emerging data governance frameworks could eventually create friction with this model, particularly around data residency and model auditing requirements.

What Comes Next

If both deals close, Anthropic will have committed roughly $90 billion in compute over a single year. That number is large enough to dominate the conversation about AI infrastructure spending regardless of what comes after. It also sets a precedent that will be cited in every subsequent negotiation between model developers and compute providers.

The real test will not be whether the models get trained. Anthropic has demonstrated repeatedly that it can execute on massive compute orders. The test is whether the revenue generated from those models justifies the capital destruction required to produce them. Right now, the answer is unknown. Anthropic has not shown operating margins that come close to covering this level of spending. Its revenue base, while growing, remains a fraction of its burn rate.

Investors in both companies will be watching closely when the IPOs land. If the market rewards scale and treats compute commitment as a legitimate barrier to entry, the next wave of AI compute deals follows this pattern—and we will see Anthropic-style spending replicated by OpenAI, Google DeepMind, and emerging challengers. If the market punishes the approach, the entire infrastructure buildout thesis faces a reckoning that could reshape how much money flows into AI hardware for years.

The $45 billion deal was surprising the first time. Doing it again in the same month tells you that the companies involved see no alternative. They have committed to this path because they believe the competitive cost of not taking it is higher than the financial risk of taking it. Whether that calculus is correct will be decided not in boardrooms or Bloomberg reports but in quarterly earnings calls, analyst queries, and the longer-term question of whether frontier AI models can generate enough economic value to justify the astronomical infrastructure bill they require. The answer to that question will determine not just Anthropic’s fate but the direction of the entire industry.