xAI is leasing enormous amounts of GPU capacity to its rivals, ramping to US$1.25bn a month from Anthropic for 300MW (about 220,000 GPUs) and US$920mn a month from Google for roughly 110,000 GPUs, according to a breakdown by engineer Martin Alderson.

The economics are stark. Alderson estimates power, the obvious big cost, runs at no more than about 1% of revenue; the Colossus site largely burns its own on-site gas. If the deals hold for 18 months, xAI recoups its capex and still has hundreds of megawatts of GPUs in hand. Both contracts carry 90-day cancellation clauses after an initial lock-in, and since xAI merged into SpaceX in February, the revenue lands in the entity now filing to go public.

Renting your training fleet to direct competitors is an odd look for a self-described frontier lab, and it reads as a quiet retreat from Grok's frontier ambitions. Alderson's argument is that it is three things at once: pre-IPO financial engineering, a genuine compute shortage, and a real edge in building datacentres fast. Colossus 1 went up in 122 days.