Anthropic Faces Capacity Crunch as Cloud Giants Turn Rival

The Capacity Squeeze

Anthropic is running into a problem familiar to anyone who has ever grown too fast: its suppliers are becoming its competitors. The San Francisco-based AI firm, which builds the Claude family of large language models, is struggling to secure sufficient cloud capacity from the very hyperscalers that once courted it as a marquee customer. AWS, Google Cloud, and Microsoft Azure - all of which have invested in or partnered with Anthropic over the past two years - now view the company less as a client and more as a threat to their own AI ambitions, according to details in a confidential S-1 filing disclosed in recent reporting.

The filing, which Anthropic submitted ahead of a planned November initial public offering, lays bare the scale of the compute challenge. The company estimates it could spend more than $750 billion on cloud contracts, a figure that includes deals still under negotiation. Even that staggering sum may not be enough. Anthropic is now leasing its own data centres and working with third-party operators, a move that reflects both the tightness of the market and the cooling enthusiasm of the hyperscalers.

A Relationship Built on Sand

The shift in dynamics is striking. AWS and Google have collectively poured billions into Anthropic, with Amazon committing up to $4 billion in 2023 and Google investing $300 million around the same time. Those investments came with strings: Anthropic agreed to run its workloads primarily on the investors' clouds and to collaborate on custom silicon and AI services. For a time, the arrangement worked. Anthropic gained access to vast compute resources and the credibility that comes with backing from two of the world's largest technology companies. The hyperscalers, in turn, gained a showcase customer for their AI infrastructure and a hedge against OpenAI's dominance.

But as Anthropic's models have grown more capable - and as the company has begun selling Claude directly to enterprises, often in competition with AWS Bedrock, Google Vertex AI, and Azure OpenAI Service - the relationship has soured. The hyperscalers are no longer eager to provision unlimited capacity for a firm that is undercutting their own AI product lines. Anthropic's filing acknowledges this tension implicitly, noting the difficulty in securing compute and the need to diversify supply.

The Third-Party Gamble

Anthropic's response has been to look beyond the Big Three. The company is now leasing capacity from independent data centre operators and, in some cases, building out its own facilities. This is a significant departure from the industry norm, where AI labs typically rely on hyperscalers for both infrastructure and operational expertise. Running your own data centres requires capital, technical know-how, and a tolerance for risk - all of which Anthropic appears willing to deploy.

The filing, however, expresses concern about the "limited operating history" of some third-party providers. Translation: these operators are unproven at the scale and reliability Anthropic requires. A multi-hour outage at a critical data centre could delay model training, disrupt customer inference workloads, and damage the company's reputation. The hyperscalers, for all their faults, have decades of experience running mission-critical infrastructure. Newer entrants do not.

This is not an abstract worry. At Opentechwire, we've tracked the growth of independent GPU-as-a-service providers across Asia and North America - companies like CoreWeave, Lambda Labs, and a clutch of Singapore- and Seoul-based startups. Many are well-capitalised and technically competent. But few have faced the operational stress of supporting a hyperscale AI lab through a multi-month training run on tens of thousands of accelerators. Anthropic is betting that at least some of them can.

The Financials: Revenue Growth, Catastrophic Losses

The S-1 also sheds light on Anthropic's financial position, and the picture is mixed. The company reported $16.3 billion in revenue during the first half of 2026, a figure that speaks to the rapid adoption of Claude in the enterprise. Third-quarter revenue alone reached $16.8 billion, up 45 per cent from the prior quarter, according to figures Anthropic shared with prospective investors last week.

But revenue growth has come at a brutal cost. Anthropic recorded an $80.7 billion net loss in the first half of the year, driven largely by compute expenses, headcount expansion, and research and development. The loss dwarfs the company's revenue and raises questions about the path to profitability. Even if Anthropic can continue growing revenue at 40-plus per cent quarterly, it is unclear when - or whether - margins will turn positive.

Customer concentration is another risk. Two clients accounted for roughly a fifth of Anthropic's first-half sales. Losing either would be material. The filing does not name the customers, but industry observers have speculated that they include large technology or financial services firms that have embedded Claude into internal workflows or customer-facing products. Anthropic will need to broaden its base if it hopes to reduce this dependency.

Liability, Insurance, and the Long Tail of AI Risk

Beyond the immediate challenges of compute and cash flow, the S-1 includes a disclosure that has received less attention but may prove more consequential in the long run. Anthropic warns that it could be held liable for damages caused by harmful AI behaviour - jailbreaks, misinformation, bias, or other failures - to an extent that would exceed its available insurance coverage.

This is not boilerplate. It is a frank admission that the company's models, despite extensive safety work, could produce outputs that cause real-world harm, and that Anthropic may not have the financial cushion to absorb the resulting claims. The disclosure builds on earlier "catastrophic risk" language in the company's Responsible Scaling Policy, which outlines scenarios in which AI systems could pose existential or society-level threats.

For prospective investors, this is a material risk. It is one thing to invest in a company that might lose money on its cloud bills. It is another to invest in a company that might face multi-billion-dollar judgements for harms its products enable. Anthropic's insurance coverage, while substantial, is capped. A single large-scale incident - a Claude-generated deepfake that swings an election, a faulty medical diagnosis that harms patients, a jailbroken model used in a cyberattack - could expose the company to liabilities that dwarf its balance sheet.

The IPO Calculus

All of this comes as Anthropic prepares to go public, reportedly targeting a November listing. The timing is aggressive. The company is loss-making, operationally complex, and facing supply constraints in a market where compute is the ultimate bottleneck. Its relationships with its largest suppliers are fraying. Its customer base is concentrated. Its liability exposure is open-ended.

And yet, the IPO may succeed. Investor appetite for AI exposure remains strong, and Anthropic's revenue trajectory is undeniably impressive. The company has built a model family that enterprises trust, a brand associated with safety and reliability, and a technical team that rivals any in the industry. If it can navigate the compute crunch, diversify its customer base, and manage its cost structure, Anthropic could emerge as one of the defining AI companies of the decade.

But the path is narrow. The leaked S-1 reveals a company under immense operational and financial pressure, making bets on unproven infrastructure partners and racing to go public before the window closes. For investors, the question is not whether Anthropic has built something valuable - it clearly has - but whether it can sustain that value in a market where the rules are being rewritten in real time.