Anthropic Paid $3.4 Billion for 1.7 GW of Gas Generators
Anthropic paid OnePWR Solutions $3.4 billion, and the money went to gas generators rated at 1.7 GW for data centers, not to chips. According to the Financial Times, it is the largest single equipment purchase for power by an AI company. Grid connections in the US are stretching out to around 2030, so companies are building their own on-site generation. For AI projects, megawatts now matter as much as GPUs.
AI-processed from Habr AI; edited by Hamidun News
For the past couple of years, chips were considered the main bottleneck in AI: graphics cards and high-performance memory, whose supply is booked almost through 2030. Now the bottleneck is becoming the thing those chips cannot work without: electricity. A fresh example: according to the Financial Times, Anthropic paid OnePWR Solutions $3.4 billion for gas generators that will give data centers 1.7 GW of capacity. It is the largest single purchase of power equipment among AI companies.
Why there are not enough chips, and there are not enough outlets either
In the past, multi-billion-dollar sums in AI companies' reports mostly related to GPU supplies. But a data center is not just racks of accelerators. To run one, you need connected capacity, and the current energy resources of the US are limited. Grid connection timelines are also stretching to roughly 2030: you can buy the equipment, but wait several years for it to be switched on.
A social factor is added on top. New AI centers provoke protests from local residents: energy costs for them rise, because suppliers find it more profitable to sell power to data center owners. The result is a vicious circle: the more AI campuses there are, the harder it becomes to reach agreements with the grid and with neighbors.
To bypass the queues, companies are switching to on-site generation. Some go further and buy retired aircraft turbines that can be adapted to produce electricity.
Gas and nuclear instead of sun and wind
It is telling that Anthropic chose not green sources but gas and nuclear power. The logic is pragmatic: a data center needs a constant load and cannot depend on cloud cover or wind strength. Companies treat possible carbon emission fines as just another line in the budget, not as a reason to abandon the project.
Anthropic is not alone here:
- Elon Musk uses gas turbines for the Colossus series of supercomputers in Tennessee and Mississippi (the source names the company SpaceXSI);
- Meta (recognized as extremist and banned in Russia) is developing its own gas generation in Ohio;
- Microsoft, Amazon and Google are reviving US nuclear power as a stable source;
- Google invested more than $4.4 billion in Constellation Energy to upgrade existing nuclear plants and obtain an additional 890 MW of capacity over the next 20 years.
The difference in approaches is clearly visible: gas delivers speed and relatively fast commissioning, while nuclear offers a long and predictable horizon. Major players are using both.
What this means for the industry
Nikita Caplin, CEO of the cloud provider RUVDS, put it this way in a column for the WEALTH Navigator club: computing is ceasing to be a purely digital story and is running into physics: generation, networks, cooling and equipment placement. In his assessment, when an AI company buys power plants, it is effectively integrating vertically into the energy sector, and a GPU without megawatts is expensive metal, not a product.
For practitioners, several conclusions follow. First, the availability of computing will be determined not only by chip makers but also by who has guaranteed energy. This may affect prices and lead times for capacity in the cloud. Second, companies with their own generation gain a structural advantage: they depend less on connection queues. Third, startups planning to build or rent large clusters should factor energy into their business plan from the start, not after choosing hardware.
There are also limitations. Betting on gas makes AI infrastructure dependent on fuel and increases its carbon footprint, and public discontent around data centers is not going away. In addition, the source data contains no details on the commissioning timeline of the generators or the terms of the deal, so it is not yet possible to assess how quickly 1.7 GW will turn into real computing.
The main takeaway is simple: in the race for AI, energy is becoming a strategic asset on par with chips, and major labs are already spending billions on it.
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026Anthropic cuts Claude Fable 5.2 price by 20%, and on the coding test it nearly matches Opus 5.5
- Nov 21, 2026Dreamteam Raises $53M for a CRM That Fills Itself from Email and Calls
- Nov 21, 2026Claude Opus 5.5 Builds Out Dimension 9 Horror Game From the Author's 15 Rooms
Key Questions Answered
Why there are not enough chips, and there are not enough outlets either?
In the past, multi-billion-dollar sums in AI companies' reports mostly related to GPU supplies. But a data center is not just racks of accelerators. To run one, you need connected capacity, and the current energy resources of the US are limited. Grid connection timelines are also stretching to roughly 2030: you can buy the equipment, but wait several years for it to be switched on. A social factor is added on top.
What are the key details on Gas and nuclear instead of sun and wind?
It is telling that Anthropic chose not green sources but gas and nuclear power. The logic is pragmatic: a data center needs a constant load and cannot depend on cloud cover or wind strength. Companies treat possible carbon emission fines as just another line in the budget, not as a reason to abandon the project.
What this means for the industry?
Nikita Caplin, CEO of the cloud provider RUVDS, put it this way in a column for the WEALTH Navigator club: computing is ceasing to be a purely digital story and is running into physics: generation, networks, cooling and equipment placement. In his assessment, when an AI company buys power plants, it is effectively integrating vertically into the energy sector, and a GPU without megawatts is expensive metal, not a product. For practitioners, several conclusions follow.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.