Anthropic’s cloud bill is ballooning as rivals tighten the screws

Anthropic is struggling to secure enough cloud capacity from Amazon Web Services, Google and Microsoft as the three tech giants increasingly view the Claude developer as a competitor, not just a customer. The company’s latest filing, described in reports as part of an S-1 shared with The Wall Street Journal, also suggests its infrastructure ambitions may cost far more than previously understood and could shape the economics of any future public listing.

The disclosure arrives as Anthropic races to raise capacity, lease its own data centers and prepare for a possible initial public offering. At the same time, the company is reporting explosive revenue growth, heavy losses and a growing concentration of sales in a small number of customers — all of which underscores how expensive frontier AI has become to build and scale.

What the filing reveals about Anthropic’s infrastructure dilemma

Anthropic’s challenge is straightforward: it needs more compute than the market is currently willing to provide on easy terms. According to the leaked filing cited by the Journal, the company has encountered increasing friction in its relationships with the major cloud providers that power much of the AI industry.

That tension matters because Anthropic is one of the most compute-hungry AI companies in the world. Running large language models, training new versions of Claude and serving millions of users requires enormous clusters of GPUs, specialized networking and enough data center space to keep the system online. In practical terms, cloud access is not a commodity input for Anthropic; it is the core industrial base of the business.

The filing suggests that the company’s reliance on AWS, Google and Microsoft is becoming more complicated precisely because those firms are no longer neutral suppliers. Each has its own AI strategy, its own model ambitions and its own incentive to prioritize internal products or favored partners.

Why the cloud giants are now more than vendors

The answer is that Anthropic and the hyperscalers increasingly overlap in the same market. Anthropic sells AI models and tools to enterprises, while AWS, Google Cloud and Microsoft all offer AI platforms and assistant products of their own. As those offerings mature, the companies that once competed to host Anthropic’s workloads may now also see it as a direct rival for enterprise AI customers.

That shift can make negotiations more difficult. Cloud providers control the scarce resources Anthropic needs most — capacity, chips, networking and long-term contracts — while Anthropic needs to keep that access broad and reliable to avoid depending too heavily on any one partner.

The leaked filing indicates that Anthropic’s cloud partners are increasingly treating the company as a competitor, which appears to be complicating its ability to secure enough capacity on favorable terms.

How much is Anthropic expected to spend on cloud contracts?

Anthropic believes it could spend more than $750 billion on cloud contracts over time, including agreements still being negotiated. That figure is striking not only because of its size, but because it suggests the company expects infrastructure costs to remain massive for years, not months.

The number also helps explain why investors have been watching Anthropic’s spending trajectory so closely. A business that may eventually generate enormous software margins can still burn through huge amounts of cash if the cost of training, inference and customer delivery stays elevated. In Anthropic’s case, infrastructure is not a side expense; it is a defining line item.

The company’s plans are also evolving beyond pure hyperscaler dependence. The filing indicates that Anthropic is leasing its own data centers and working with third-party operators. That move can reduce reliance on the biggest cloud firms, but it introduces a different set of risks: execution, reliability, and the operational maturity of the companies supplying the facilities.

Why Anthropic is worried about third-party data centers

Anthropic’s concern, according to the filing, is that these operators have a limited operating history. In plain terms, the company is questioning whether newer or smaller providers can match the reliability, scale and technical discipline of Amazon, Google or Microsoft.

That caution is understandable. The most demanding AI workloads are unforgiving. A shortfall in power delivery, cooling, network throughput or maintenance can disrupt model training and model serving. For a company trying to serve enterprise customers and open the door to public investors, reliability is not optional.

So Anthropic appears to be balancing two imperfect choices: stick with hyperscalers that may be harder to work with, or diversify into less proven infrastructure partners that may be cheaper or more available but carry operational uncertainty.

How bad are Anthropic’s losses, and what does revenue growth change?

Anthropic reported an $80.7 billion net loss in the first half of the year, a staggering figure that reflects how aggressively the company is investing in model development and infrastructure. Over the same period, it posted $16.3 billion in revenue, according to the WSJ report.

That combination tells two different stories at once. On one hand, the company is clearly seeing demand for Claude and related products. On the other, its cost base is so large that revenue growth alone may not be enough to offset near-term losses.

Just last week, Anthropic told prospective investors that third-quarter revenue had reached $16.8 billion, up 45% from the previous quarter. If accurate, that would represent strong momentum in a short time frame. But for a company with infrastructure needs measured in the hundreds of billions, even rapid revenue expansion does not automatically solve the underlying capital problem.

What the revenue numbers suggest about demand

The revenue figures point to strong market appetite for Anthropic’s enterprise AI products and developer tools. A 45% quarter-over-quarter increase is not a routine improvement; it suggests major customer adoption, larger usage volumes or both.

Still, revenue growth in AI often comes with heavy usage costs. The more customers rely on model calls, the more the company must spend on compute to deliver them. That means top-line momentum can sometimes amplify infrastructure pressure rather than ease it.

What are the risks tied to harmful AI behavior?

The filing also warns that Anthropic could face liability for damages caused by harmful AI behavior that could exceed its insurance coverage. That disclosure extends the company’s risk profile beyond finance and infrastructure into the legal and operational consequences of deploying frontier models at scale.

This kind of warning is increasingly important as AI systems are embedded into customer support, coding, search, business workflows and other high-stakes settings. If a model produces harmful outputs, misleads a user or contributes to a damaging decision, companies could face claims that are expensive to defend and potentially costly to settle.

For Anthropic, that risk matters even more because the company positions itself as a safety-focused AI developer. The warning does not mean liability is imminent, but it does indicate that the company believes the downside case is significant enough to disclose to investors.

According to the filing described in the report, Anthropic is warning investors that certain AI-related harms could trigger damages beyond what its insurance coverage would pay.

Why does this matter for a possible IPO?

This matters because public-market investors tend to scrutinize AI companies far more closely than private backers do. A business approaching a reported November IPO would be expected to explain not just its growth prospects, but also its dependence on outside infrastructure, its customer concentration and the durability of its margins.

Anthropic’s cloud situation could become a major topic in any public offering discussion because it touches the heart of the company’s business model. If hyperscaler partnerships are strained, the company must demonstrate that it can still obtain enough compute to meet demand. If it turns to alternative operators, it must prove those relationships can scale safely.

Either way, investors will likely ask how sustainable Anthropic’s economics are in an industry where the cost of keeping models online can be enormous and volatile.

What investors will likely focus on

• Whether Anthropic can secure enough compute without sacrificing margins or control.
• How much of its revenue depends on a small number of large customers.
• Whether future growth can outpace infrastructure spending.
• How the company manages legal and safety-related exposure from model behavior.
• Whether its supply chain for chips, data centers and cloud services is diversified enough.

How Anthropic fits into the broader AI infrastructure race

Anthropic’s predicament is not isolated. It reflects a wider industry pattern in which leading AI labs are colliding with the companies that supply their compute. The same cloud providers that once eagerly courted model developers now face the possibility of those developers becoming strategic threats in their own right.

This is one reason the AI boom has become as much an infrastructure story as a software story. The winners are not simply the companies with the best models; they are also the ones that can reliably secure chips, power, data center space and long-term financing at scale.

Anthropic’s reported attempt to diversify through leased facilities and third-party operators shows that the market is moving toward a more complex supply chain. In the early phase of the AI boom, cloud partnerships were often enough. In the current phase, leading firms are beginning to think more like industrial operators, building or renting physical capacity to reduce bottlenecks.

Who benefits when AI firms build their own capacity?

The answer is that several players can benefit, but not equally. Specialized data center operators can win new business. Chip suppliers can gain more demand. Cloud providers may still retain important workloads, even if the customer diversifies. And AI companies can gain leverage by no longer depending on a single channel for compute.

But building and leasing infrastructure also adds complexity. It requires long-term capital planning, technical operations, redundancy, maintenance and power procurement. For a company like Anthropic, that means the AI race is increasingly about managing a physical industrial system, not just launching a better model update.

Timeline of the reported developments

The sequence of recent disclosures helps explain why Anthropic’s infrastructure story is drawing so much attention.

What happens next?

In the near term, the key question is whether Anthropic can secure enough infrastructure to support growth without leaning too hard on any one provider. The answer will shape its cost structure, its negotiating power and potentially its market valuation.

The company’s response will also matter to the wider AI ecosystem. If one of the fastest-growing model developers is struggling to obtain enough cloud capacity, it suggests the industry’s infrastructure bottlenecks are still real, despite all the investment pouring into data centers and chips.

Anthropic declined to comment when approached by TechCrunch, according to the report. That leaves investors, customers and competitors to interpret the filing on their own — but the message is already clear: the era of easy access to AI compute is ending, and the companies building frontier models are being forced to compete for physical capacity as aggressively as they compete on product quality.

For Anthropic, the stakes are especially high. It is trying to scale a business defined by huge demand, huge losses and huge infrastructure needs while convincing investors that the path to public markets is still viable. The cloud challenge is now part of that pitch.