The xAI IPO and the circular compute economy

Bottom line. SpaceX prices its IPO this week, with first trade targeted for 12 Jun 2026 at roughly a $1.75 trillion valuation. The filings’ real news for IT leaders is the compute economy they expose: Anthropic pays roughly $1.25 billion a month and Google roughly $920 million a month for capacity in xAI’s data centers, about $26 billion a year flowing into an AI segment that lost $6.355 billion operationally in 2025, on leases both sides can cancel on 90 days’ notice.

Report. SpaceX filed its S-1 on 20 May 2026, amended it on 1 Jun and 3 Jun, and targets roughly $75 billion at $135 per share, around a $1.75 trillion valuation, first trade 12 Jun on Nasdaq. The filings disclose the economics of its AI segment, which combines xAI, X and the AI data centers: $3.201 billion of 2025 revenue against a $6.355 billion operating loss, and Q1 2026 revenue of $818 million against a $2.469 billion operating loss, figures reported by TechCrunch from the filing. The segment’s largest disclosed revenue lines are leases of GPU capacity to rival labs: Anthropic at roughly $1.25 billion a month for about 220,000 NVIDIA GPUs in the Colossus data centers, in the S-1 itself, and Google at roughly $920 million a month for about 110,000 GPUs, disclosed 5 Jun in a free writing prospectus.

Google’s explanation of why it is renting a competitor’s GPUs is the quote of the season:

“Google Cloud and SpaceX are long-time partners. This is a short-term, timely agreement to ensure we have bridge capacity to meet surging customer demand for our agent platform, Gemini Enterprise, which has been even higher than we expected.”

— a Google Cloud spokesperson, in a statement to TechCrunch, 5 Jun 2026.

Figures from the SEC filings as reported by TechCrunch; the ~$26B/year total is arithmetic on the two monthly figures.

Observe. The pattern is circularity. The frontier labs we described as a hyperscaler-backed top two in the Anthropic valuation read are now also each other’s landlords and tenants: Anthropic trains and serves on a rival’s data centers, and Google bridges its own enterprise agent demand on hardware owned by the company behind Grok. Capacity, not capability, is what is scarce, which is the same conclusion the energy-bill read reached from the power side and the Maia chip read from the silicon side.

Reflect. For a CIO, three implications. First, when the second-largest cloud builder needs a competitor’s GPUs because Gemini Enterprise demand “has been even higher than we expected,” capacity risk sits underneath every vendor SLA, and it belongs in your assumptions the way the enterprise AI cost evidence already prices spend. Second, the circularity correlates vendor risk: providers that look independent share physical infrastructure, a concentration pattern rhyming with the Big Four model-concentration read. Third, for anyone evaluating Grok specifically: the AI segment’s revenue base is two rivals’ leases, both cancellable on 90 days’ notice, now subject to quarterly public scrutiny.

Share thoughts. The move this week is not to trade the IPO; it is to re-read your AI vendor assumptions with the supply side in view. Ask each frontier vendor where your workload physically runs, whose data center it is, and what happens to your capacity and price at their next contract break. The filings just demonstrated that even the largest players cannot take compute for granted, and the FinOps discipline that meters your own agent spend is the demand-side half of the same posture.

Holding-up note

The primary claim of this piece (that the SpaceX filings disclose a circular frontier-compute economy, roughly $26 billion a year flowing from Anthropic and Google into xAI’s loss-making AI segment on 90-day-cancellable leases, making compute supply rather than model capability the binding constraint under enterprise AI roadmaps) is on a 45-day review cadence, short because the IPO completes this week and the Google lease’s ramp dates land within the window. Three kinds of evidence would move the verdict: either lease being cancelled, renegotiated or materially expanded; post-IPO disclosures showing the AI segment’s economics differ from the S-1 picture; or capacity loosening (falling GPU lease pricing, idle capacity disclosures) that would falsify the supply-constraint reading. The Holding-up record for AM-208 captures what changes, dated. Figures are from the SEC filings and contemporaneous reporting as of 9 Jun 2026.