The Claude maker is scaling its infrastructure at a pace that makes most utilities look sleepy, backed by hundreds of billions in commitments
Anthropic expects to have more than 3.5 gigawatts of computing capacity online by the end of 2026. That is the kind of number usually reserved for power plant announcements, not software companies.
The company behind the Claude models is building a power-hungry machine. Whether the revenue can keep pace with the electricity bill is the question hanging over the whole AI industry.
The scale of the build
Start with the trajectory. Anthropic’s capacity was projected at an estimated 1.4 GW by the end of 2025, so the 2026 target represents more than doubling in roughly a year.
The plan does not stop in 2026. Anthropic is planning to roughly double its footprint again in 2027, targeting about 9.5 to 10 GW by the close of that year.
Getting there involves a patchwork of cloud, chip and data center agreements. The headline partnerships are substantial on their own. A deal involving Google and Broadcom covers approximately 3.5 GW of TPU capacity starting in 2027.
Amazon Web Services, meanwhile, has an arrangement for up to 5 GW of its Trainium chips.
Following the money
Capacity this large does not come cheap. In an amended S-1 filing dated November 2, 2026, Anthropic disclosed $895.4 billion in non-cancelable purchase commitments and lease agreements, mostly tied to compute and infrastructure projects.
That total reflects commitments as of late September 2026. On top of it sits an additional estimated $150-175 billion in further arrangements.
“Non-cancelable” is the key phrase. These are not wish lists or memoranda of understanding; they are obligations the company has signed up to pay regardless of how demand evolves.
The revenue side is growing quickly, though it remains a fraction of those commitments. Anthropic reported Q3 2026 revenue of nearly $17 billion, up from $11.56 billion in the previous quarter.
Profitability, however, is a different story: the company posted a net loss of $23.3 billion in the same period.
Beyond Nvidia
One of the more interesting threads in Anthropic’s strategy is chip diversification. The company is moving past a reliance on Nvidia GPUs and embracing custom silicon like Google’s TPUs and Amazon’s Trainium.
Anthropic is also seeking smaller data center sites alongside its mega-deals. Smaller sites can often come online faster than sprawling campuses, which matters when every quarter of delay is a quarter of capacity your competitors might grab.
Background: the AI capacity arms race
Anthropic has drawn investment from some of the biggest names in technology, including Amazon, Google and Microsoft. Those relationships double as supply chains, since several of the backers are also the cloud providers selling it compute.
The S-1 filing itself signals how far the company has come. Public filings bring a level of disclosure that private companies typically avoid, which is why figures like the $895.4 billion commitment are now on the record.
What this means
For investors and the broader technology sector, the most immediate read is demand. A company does not sign close to $900 billion in binding commitments unless it believes customers will keep showing up, and the near-$17 billion quarterly revenue figure suggests that belief is not purely aspirational.
The ripple effects reach well beyond Anthropic. Cloud providers like AWS and Google, plus chip designers like Broadcom, are direct beneficiaries of this spending, and the push toward custom silicon could reshape competitive dynamics in a market long dominated by Nvidia.
The risk is equally visible. A $23.3 billion quarterly loss against commitments of this size leaves little margin for error if AI demand cools, pricing compresses, or a rival delivers comparable models more cheaply.
There is also an execution question. Bringing more than 3.5 GW online by end-2026, then targeting roughly 9.5 to 10 GW a year later, depends on chips being delivered, data centers being built and power being available on schedule.
