First Five AI Data Center With Over One Gigawatt of Power Arriving in 2026-2027

In the USA, AI data centers will need around 20 to 30 gigawatts (GW) of combined power by late 2027. 30 GW is about 5% of US average current power generation capacity and 2.5% of China’s. 30 GW would be around 25% of power generation capacity in Japan, 50% in France, and 90% in the UK. Companies connect data centers to the grid and build new power plants to support them. They are currently mostly using natural gas power and then have diesel powered backup.

Data centers are growing fast enough that we don’t need decentralized training — at least for the next two years. Decentralized training across different locations makes it easier to soak up excess power capacity.

The largest 2026 facility (xAI Colossus 2) will have the compute equivalent of 1.4M H100 GPUs. This will get further expansions that are not described by Epoch.AI.

There is concentration of AI compute in the U.S. The 13 large U.S. data centers represent around 2.5 million H100-equivalents. This is ~15% of the roughly 15 million H100 equivalents have been delivered globally as of late 2025.

Even larger data centers are coming: Meta Hyperion and Microsoft Fairwater will each have five million H100e when they reach full capacity in late 2027 to early 2028.

AI data centers require massive capital investment, typically $29B per gigawatt of total facility power.

The most expensive data center we know about is Microsoft Fairwater, projected to exceed $100 billion in total capital cost upon completion.

Several data centers will soon demand 1 GW of power, starting early next year:

– Anthropic–Amazon New Carlisle (January 2026)
– xAI Colossus 2 (February 2026)
– Microsoft Fayetteville (March 2026, borderline 1GW)
– Meta Prometheus (May 2026)
– OpenAI Stargate Abilene (July 2026)

XAI Colossus 2. Installation began July 2025 with initial 110,000 Nvidia GB200 GPUs. (330k-550k H100). ~550,000 total nearing completion by October/November 2025 [1.5 to 2.7 Million H100]. Expansion to 1 million GPUs (crossing 1.6 GW power draw) expected by end-2025, with further scaling into 2026 [5 million H100].

Anthropic–Amazon New Carlisle will have power available in Jan 2026. Facility operational since October 2025 with ~500,000 AWS Trainium 2 chips already installed. Additional ~500,000 chips slated for installation by year-end 2025 to double capacity, ahead of the power ramp-up.

1 Million Tranium 2 chips are 500 MW of chips and equal to about 300,000 H100s.

B200 specs

The NVIDIA HGX B200 propels the data center into a new era of accelerating computing and generative AI, integrating NVIDIA Blackwell GPUs with a high-speed interconnect to accelerate AI performance at scale. As a premier accelerated scale-up x86 platform with up to 15x faster real-time inference performance, 12x lower cost, and 12x less energy us. The NVIDIA GB200 NVL72 system, a rack-scale AI supercomputer with 72 Blackwell GPUs, weighs approximately 3,000 pounds (1.36 metric tons or 1,360 kg). Some configurations, like the HPE version fully loaded, may reach up to 3,245 pounds.