xAI leases 1.2M GPUs, earns $41.1B yearly from Microsoft, Anthropic and others
xAI, SpaceX’s AI division, is positioning itself as an AI infrastructure provider by leasing out its GPU capacity. On October 1, 2026, reports emerged that Microsoft was in early discussions with xAI about leasing compute resources, signaling a shift from xAI’s original focus on training its own models. The division’s Colossus 2 data center already houses 550,000 GPUs, with plans to expand to…
Key points
- xAI’s Colossus 2 will have **1.21 million GPUs by year-end**, up from 550,000, with **220,000 GB300s** added in October, November, and a conditional batch in December
- Annual lease revenue reached **$41.1 billion**, with clients like Anthropic ($15B/year), Google ($920M/month), and an undisclosed client paying **$1.11B/month starting December 2026**
- xAI is leasing older GPU clusters (Colossus 1) to competitors like Anthropic while using newer GPUs (Colossus 2) for its own model training
xAI’s annual lease revenue hit $41.1 billion in September 2026, per CFO Bret Johnson. Key clients include Anthropic ($15B annually), Google ($920M/month), and Reflection AI ($1.5B annually). A fourth undisclosed client is set to pay $1.11B/month starting December 2026, bringing monthly revenue to $3.4B. xAI’s ability to bypass power grid limitations—by building its own 1.2GW gas-fired plant—has accelerated its expansion. Meanwhile, Colossus 1, originally built for Grok training, is now leased to Anthropic for inference, while Colossus 2 remains dedicated to xAI’s own model training. Musk aims for 50 million H100-equivalent GPUs by 2030, hinting at future orbital data centers via SpaceX’s satellite network, though NVIDIA’s Jensen Huang dismissed this as unrealistic for now.
The story so far
3 episodes →- xAI leases 1.2M GPUs, earns $41.1B yearly from Microsoft, Anthropic and othersthis story
SpaceXAI Transforms into an 'AI Infrastructure Landlord': Cloud Business Gains Momentum with Microsoft Contract Negotiations and 1.2 Million GPU Capacity
note.com · 29 October 2026
Loading the full article…
This text was published by note.com and written by バンブー・ラボ｜BambooLab. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Hardware & Compute
All →- OpenAI CEO criticizes home RAM use for AI inference · 1 src
- Taiwan Semiconductor Manufacturing could provide AI exposure despite bubble fears · 1 src
- Nvidia CEO sees Kazakhstan as potential AI hub for Europe · 1 src
- Amazon plans to sell Trainium AI chips to third‑party data centers starting late 2027 · 3 src
- PsiQuantum unveils 8x8 optical switch cutting AI data center losses to 1.2 dB · 1 src
Comments
via GitHub Discussions