Anthropic Spends $3.4 Billion on Gas Generators to Power AI Data Centers

Anthropic is spending $3.4 billion on gas-fired generators to supply its AI data centers, according to a Financial Times report published on November 8, 2026. Reuters carried the FT account the same day.

The equipment is a set of Caterpillar generators capable of producing up to 1.7 gigawatts of electricity, or 1.7 billion watts. Equipment supplier OnePWR had ordered the same machines a couple of years earlier but could not find financing to complete the purchase. Anthropic did not take over that order as originally planned. It bought the generators directly from Caterpillar, according to TechCrunch, which reported the change on November 9, 2026.

The direct purchase is what allows Anthropic to move ahead of the queue for this class of equipment. Tom's Hardware reports that the deal lets the company skip a six-year waitlist for this equipment.

Laid out in order, the sequence ran as follows. A large, financed order sat for about two years without the capital to close it. The machines were then bought by a buyer with the balance sheet to take delivery, and the turbine queue did the rest.

For anyone who has spent time on hyperscale capacity planning, the constraint is no longer only accelerators and interconnect. Firm, dispatchable power, meaning electricity a plant can deliver on demand, and the equipment slots that deliver it now sit on the critical path. Whoever holds a financed position in that queue sets the pace for everyone downstream.

In my view, the more important detail is the financing gap, not the headline number. A generator order is a promise of future power, and a promise without capital behind it does not energize a single rack. The OnePWR order shows that demand for capacity alone did not carry the project to completion. That is a useful signal for operators, developers and equipment makers weighing how much of the AI buildout depends on who can underwrite long-lead hardware.

The broader context here is that AI compute is now a power-procurement problem as much as a silicon one. Gas-fired generation at this scale raises engineering and capital questions, along with permitting, fuel and emissions issues, and none of those are settled by a single purchase. What the reporting supports is narrower: a frontier AI developer has taken direct ownership of a large block of on-site generation capacity, bypassing a supplier's stalled order to do it. Whether that approach becomes a template for others will depend on how the next wave of data-center projects is financed and sited.

In my view, the practical lesson for readers building or buying capacity is that a turbine order is only as credible as the money behind it. The deal also suggests that companies able to commit capital early will shape which projects get powered on schedule, and that the path to more compute runs as much through equipment queues and financing as through model training runs.