XAI, Google, OpenAI and Meta Race to Gigawatts of AI Data Centers

by

XAI Grok 1 launched in November 2023 using rented Oracle GPUs, but xAI shifted to in-house builds due to compute shortages.

Colossus 1 in Memphis: Built in a former Electrolux factory, it reached 100,000 NVIDIA H100 GPUs in 122 days (May 3 to September 2, 2024).

It expanded to 230,000 GPUs (including H200s and GB200s) by July 2025. Challenges like power (upgraded from 15 MW to over 250 MW) were solved with 35 mobile generators (up to 420 MW), Tesla Megapacks for load smoothing and surge handling, and grid upgrades (150 MW substation by May 2025, another by fall 2025). Coherence across 230,000 GPUs was a breakthrough, enabling it to function as one giant GPU.

Colossus 2 aims for 1 million GPUs (GB200/GB300) in a 1 million sq ft warehouse that is expandable to 3 million sq ft. XAI could use multiple buildings.

Elon said Colossus 2 will be the first gigawatt AI training supercomputer.

Grok 5 will be trained using Colossus 2 and might be able to achieve AGI.

XAI will spend over $40 billion for Colossus 2 and is building a large solar farm near the site.

Phase 1 (110,000 GB200s) expected online by late July/early August 2025. A batch of 550,000 Nvidia GPUs starting mid-August 2025. The full 1 GW datacenter by January 2026 and scaling to 3 Gigawatts by mid-late 2026.

Power solutions include relocating generators from Colossus 1, imported gas plants (200-700 MW each), Megapacks, and substation expansions.

xAI outpaces competitors due to speed (3-5x faster than estimates), private structure, and motivation.

OpenAI’s Stargate is planned to have 400,000 Nvidia B200s at the end-2026. This would be 3-4x more power than Colossus 1. OpenAI has to complete all of its funding. OpenAI developed under the codename Project Ludicrous. Stargate’s Abilene plant had its phase one for two building opening earlier this year. The two buildings operating at 200 megawatts. Phase two, with six additional buildings (for a total of eight) and about 1.2 gigawatts capacity is slated for mid-2026. OpenAI Multi-datacenter training (across campuses) is their plan to compete with XAI and Google’s infrastructure. They will have gigawatt clusters connected with long-haul fiber and fault tolerance.

Meta’s Prometheus will have similar scale. Meta’s Prometheus data center is expected to be completed and come online in 2026. It will be located in New Albany, Ohio, and is engineered as a 1-gigawatt (GW) supercluster, housing approximately 500,000 GPUs (chips) and drawing about 1,020 MW of power. Prometheus is on a 740-acre industrial tract. Meta plans for a Hyperion data center. It will likely be located in Louisiana’s Richland Parish district, where Meta in December announced a $10 billion data center. The flagship plan is for it to have 2 gigawatts of capacity by 2030 and scale up from there.

Data centers-as-a-service provider ECL, a startup from Meta and Microsoft veterans, has opened phase one of an off-grid hydrogen powered facility, TerraSite-TX1, near Houston that it plans to scale up to a gigawatt capacity.

Others like Anthropic’s Project Rainier (with Amazon), Google (1 GW in Iowa/Ohio by 2025-2026, but not coherent).

xAI’s dominance could attract talent/capital, with future terawatt-scale builds requiring rethinking power (e.g., solar arrays, flex-connect agreements).

\

3 thoughts on “XAI, Google, OpenAI and Meta Race to Gigawatts of AI Data Centers”

  1. At what point will companies have to generate more revenues than CapEX? That will determine when the AI bubble pops. People are getting most, and too much, of AI for free. Most people won’t/can’t afford to pay for AI that isn’t really enhancing their own ability to make money, even if they find it personally useful, fun, or time-saving.

  2. This reminds me of Sam Altman, who said not long ago that we would have AGI very soon. After the release of ChatGPT5, he said that the concept of AGI no longer means much and that we should stop using it. We’ve managed to do some very impressive things with AI, but I don’t think it’s with LLMs and their improvement that we’re going to achieve AGI.

Comments are closed.