Summary
- The Information estimates that Anthropic has entered compute agreements covering at least 14.8GW of capacity since October 2025.
- Potential cloud, chip, and data centre commitments could cost as much as $517bn, largely over the coming decade.
- Contracted compute does not equal operating capacity because grid connections, construction, cooling, hardware, and financing still have to be delivered.
Anthropic has reportedly entered agreements giving it access to at least 14.8GW of additional compute capacity since October 2025, providing a measure of the physical infrastructure being assembled behind the company’s AI growth.
The estimate comes from The Information, which analysed public statements alongside its own reporting on Anthropic’s cloud, chip, and data centre agreements. It calculated that the potential value of those commitments could reach as much as $517bn, largely over the coming decade.
Anthropic has not published a single 14.8GW capacity figure or confirmed that $517bn will definitely be spent. The agreements have different durations, deployment schedules, commercial structures, and potential cancellation provisions.
The Information also cautions that Anthropic may not ultimately use all of the capacity covered by the agreements. The figures are therefore best read as an estimate of the upper scale of its contracted compute position rather than a forecast of concurrent electricity demand.
Even with that qualification, 14.8GW represents an infrastructure programme far larger than server procurement alone. Capacity at that scale has to be translated into land, grid connections, substations, generation, transformers, backup systems, cooling plant, fibre, and commissioned data halls.
Amazon and Google account for much of the reported capacity, while Anthropic has widened its supplier base through agreements involving Microsoft, Nvidia, SpaceX, Lambda, Nscale, and direct data centre leases.
Recent deals show how quickly that portfolio is expanding. Reuters reported a $35bn agreement with Nvidia-backed Lambda tied to compute capacity at a Texas data centre, while Anthropic also has a reported six-year, $45bn agreement with UK-headquartered Nscale.
DataCentral reported on the Nscale contract in August, when the agreement was linked to capacity at a West Virginia campus.
Using several infrastructure providers reduces reliance on a single cloud platform, but it creates a substantial delivery problem. A compute contract becomes useful only when the underlying site is powered, cooled, networked, fitted out, and commissioned.
Grid capacity is one of the clearest constraints. Developers can secure customers and land years before the transmission network is capable of energising the requested load, particularly in markets where several gigawatts of data centre demand are competing for the same infrastructure.
Cooling creates another physical limit. Current GPU platforms are driving rack densities beyond those for which many existing facilities were designed, requiring direct liquid cooling and more demanding electrical distribution.
Hardware delivery has to remain aligned with both. GPUs arriving before the building or substation is ready create stranded equipment; a completed data hall without the customer hardware produces the opposite problem.
Financing connects those schedules. Providers increasingly use long-term customer agreements to support debt and equity investment in campuses that may require billions of dollars before meaningful revenue begins.
Anthropic’s reported commitments are therefore creating demand for a full infrastructure supply chain rather than simply a large block of chips. Each gigawatt has to be hosted somewhere and supported continuously by energy, cooling, networking, maintenance, and physical security.
The reported $517bn estimate also extends well beyond the near term. Some agreements run beyond 2030, and capacity may ramp gradually rather than appear as one simultaneous load.
Adding every agreement together without accounting for timing would overstate immediate power demand, but the buildability requirement remains substantial even if only part of the reported capacity reaches operation.
More detail is likely to emerge as Anthropic moves towards a possible public listing and investors scrutinise its future infrastructure obligations. The central question will not simply be how much compute the company has contracted, but how much of that capacity can be physically delivered on the required timetable.
