Anthropic just rented Musk's supercomputer

Anthropic Just Rented Musk's Supercomputer

**The short version:** Anthropic has reportedly become the first external customer for the Colossus 100K+ GPU cluster in Memphis — the single-site training supercomputer operated by xAI, whose parent company SpaceX has filed a confidential S-1 with a reported $1.75–2T valuation ahead of a June 2026 NASDAQ listing. One frontier lab is now renting compute from a rival ecosystem. That is the actual story.

What Happened

According to signals 47c46e8d and acaa8ec5, Anthropic has signed on as the first named external tenant for Colossus — xAI's Memphis facility housing more than 100,000 NVIDIA GPUs in a single coherent-training site. The arrangement was not publicly announced; it surfaced as a reported deal, and no independent confirmation exists beyond those signals. Hold it as reported, not confirmed.

The facility itself is documented. Colossus went from groundbreaking to operational in roughly twelve months — a compressed Hopper-class buildout that would not have been feasible before 2024. NVIDIA is confirmed as the sole GPU supplier. NVIDIA's data-center segment posted approximately $115.2B in FY2025 revenue, up roughly 142% year-over-year (signal daad3d00; flagged AI-prior, not S-1-confirmed — model accordingly).

Why a Frontier Lab Rents a Competitor's Cluster

The reflexive read is "Anthropic ran out of compute." The structural read is more useful: multi-provider compute sourcing became default practice across frontier labs in 2025–2026. When your training bottleneck is GPU-hours and your own buildout has a 12–18 month lead time, the question stops being *whose* cluster and starts being *which cluster is available now*.

Colossus is available. It is the largest single-site coherent-training cluster currently operating, and until this week it had no named external tenants. Anthropic changes that calculus — not because the deal is large (no dollar figure has been confirmed), but because it establishes Colossus as merchant infrastructure rather than a captive xAI asset.

The question worth tracking: **does a second named external tenant sign before the SpaceX IPO?** That single data point does more to price Colossus as a compute business than any valuation range in a confidential S-1.

The IPO Overlay

SpaceX has filed a confidential S-1. Reported figures — $1.75–2T valuation, approximately $75B raise, a June 2026 NASDAQ listing — are sourced to signal 50da90f1 and should be held as reported until the public filing lands. Key figures remain AI-prior rather than S-1-confirmed, so any model built on them inherits that uncertainty tag explicitly.

Here is the infrastructure-desk frame: this is the first IPO underwritten in part by a 100K+ GPU single-site coherent-training asset. That asset class did not exist before the 2024–2025 Hopper buildout. Investment banks are pricing a rocket company; AI-infra desks are pricing a neocloud. Those are different spreadsheets, and right now only one of them has public comps.

The load-bearing public comp is CoreWeave — currently the only clean public neocloud comparable for Colossus-class economics. Any CoreWeave re-rate is the leading indicator for how SpaceX/xAI compute gets priced in the run-up to listing. Watch it weekly, not as a side figure.

What the Signal Density Is Actually Telling You

The open-role count at SpaceX/xAI AI-infra sits at roughly three openings over the trailing 90 days, all level-unspecified. That is not a growth curve. It reads as a cold-start cohort — the hiring posture of infrastructure that is already staffed and whose external-tenant sales motion has not yet been formalized.

That matters because the buildout narrative has been running well ahead of observable activity. A reported ~$2T-scale compute thesis with three open roles and one named external customer is a thin signal set. Do not extrapolate a supplier read to Marvell or Coherent from structural inference alone — the signal feed on both is currently empty.

Observable state as of this week:

| Metric | Current Reading | |---|---| | External Colossus tenants (named) | 1 — Anthropic (reported) | | GPU supplier | NVIDIA (confirmed sole) | | Public neocloud comp | CoreWeave | | S-1 status | Confidential; public filing pre-June 2026 | | Open AI-infra roles (90d) | ~3, level-unspecified |

Reverse-Hype Check

The "$1.75–2T SpaceX is now an AI company" framing is doing heavy lifting while the S-1 is still confidential. The rocket business is real; Starlink revenue is real; the GPU cluster is real. But AI-infra revenue from Colossus — as of this week — consists of one reported external customer, no disclosed contract value, and a job board that reads as infrastructure maintenance rather than hyperscaler sales motion.

The hype version: Musk built the world's most powerful AI supercomputer and is now selling compute to his competitors. The grounded version: a frontier lab signed a reported compute agreement with a facility that has capacity to spare, ahead of a public offering where proving external demand materially affects pricing. Both can be true simultaneously. Only one should anchor a model.

Three Things to Watch This Week

1. **Second Colossus external tenant.** The single falsifiable trigger. Does another named lab or enterprise sign before the listing? Yes → captive-to-merchant thesis advances. No → Colossus prices as a one-customer pilot, not a platform.

1. **CoreWeave multiple.** The only clean public neocloud comp. Any move in its valuation directly updates the SpaceX/xAI compute read. Treat it as a leading indicator, not a footnote.

1. **S-1 public filing.** When it drops, every figure currently tagged AI-prior gets confirmed or revised. Build your model now with explicit confirmation-status tags — it will let you update cleanly on filing day without reconstructing your sourcing.

*Signals: 47c46e8d, acaa8ec5 (Anthropic/Colossus deal, reported); 50da90f1 (SpaceX S-1 and valuation range, reported); daad3d00 (NVIDIA data-center figures, AI-prior / not S-1-confirmed).*