xAI is looking more like a datacentre REIT than a frontier lab
Points and comments are a snapshot, not live.
xAI's datacentre rental deals with Anthropic and Google suggest it functions more as a REIT than a frontier AI lab.
xAI announced partnerships in May and June 2026 to lease GPU capacity: $1.25 billion per month for 300MW to Anthropic (220k GPUs) and $920 million per month for 110k GPUs to Google. Both contracts include 90-day cancellation clauses after initial lock-in. At these rates, xAI recoups its ~$40 billion datacentre capex within 18 months, excluding operating costs. The author notes three factors at play: potential financial engineering ahead of SpaceX's IPO, genuine GPU scarcity that makes even older H100s valuable, and xAI's competitive advantage in rapid datacentre construction (Colossus 1 built in 122 days). The deals leave Grok in a difficult position, with inference capacity apparently overspecified relative to demand. Motivation for the deals may include pressure on OpenAI and Google's shareholding in SpaceX, but massive GPU shortages make the commercial basis plausible regardless.
What commenters are saying
Commenters split between accepting the deals as commercially sensible versus viewing them as circular financial engineering. Top-ranked comments note that Grok underperformed as a frontier model and that overspecifying capacity followed by leasing excess compute is a rational pivot. Users who test multiple LLMs report Grok performs well on current events (X integration), less sycophantic responses, and handling sensitive professional topics others refuse. One commenter flags Google's 5-6% SpaceX stake ($88.5B-$106.2B at the proposed $1.77T valuation) as motivation to inflate the IPO. Discussion of bailout cycles and monetary inflation appears in lower-ranked threads but is less substantive than the capacity and product quality debate.