Nvidia described Colossus as a 100,000-H100 cluster using Spectrum-X Ethernet, with expansion toward 200,000 Hopper GPUs. Later scale claims are useful directionally, but should be verified against filings, vendors, power/cooling evidence, and permits.
If Grok/xAI internal demand were already absorbing capacity at high margin, leasing a major block to Anthropic would be less intuitive. The lease is bullish for revenue, but may also signal surplus monetizable capacity, training/inference mismatch, or a need to convert capex into bankable revenue before IPO.
Keep Anthropic as its own probability-weighted AI compute line. Do not let it mechanically validate Grok demand or justify a full AI-platform multiple without evidence of durable utilization, margin, renewal power, and non-Anthropic customer depth.
History and configuration
- Colossus began as xAI's rapid-build Memphis supercomputer and was publicly framed around Grok training.
- The disclosed initial configuration centered on Nvidia H100/Hopper GPUs, high-speed networking, liquid cooling, and large power infrastructure.
- Later Colossus II plans push toward much larger power and accelerator footprints, but full-run-rate capability should be checked against cooling, grid, turbine, and capex evidence.
Controversies
- Memphis residents and environmental groups have challenged turbine use, emissions, air permitting, and local health impacts.
- Power availability is a financial risk as much as a permitting risk: idle GPUs, temporary generation, or constrained cooling can break ROI math.
- Grok content, data, privacy, and government-use controversies can affect enterprise adoption and utilization.
Recent development
- The SpaceX filing packet discloses an Anthropic compute agreement for COLOSSUS and COLOSSUS II capacity at a $1.25B full monthly fee through May 2029, subject to ramp and 90-day termination risk.
- May 28 update: Musk said SpaceX had agreed to lease Colossus AI training clusters to Anthropic for an initial six months, with a multi-year extension possible. The model should therefore treat the initial lease and extension period as separate confidence cases.
- The default investment estimate below is $25B: roughly triangulated from recent AI capex in the filing packet and public reporting around a $20B Colossus II package.
- The critical read is not "bad deal." It is "good revenue anchor, still unproven platform economics."