xAI leases Colossus 1 supercomputer to Anthropic

According to Forbes and CryptoBriefing, on May 6, 2026 xAI (reported by CryptoBriefing as recently rebranded to SpaceXAI) signed an agreement to provide Anthropic full and exclusive access to Colossus 1, the Memphis data centre originally built for training Grok (Forbes; CryptoBriefing). Forbes and multiple outlets report the cluster contains more than 220,000 NVIDIA GPUs including H100, H200, and GB200 accelerators and offers roughly 300 megawatts of compute capacity (Forbes; CryptoBriefing). CryptoBriefing projects the lease will generate $5 to $6 billion annually; Fortune quoted New Street Research estimating about $3 billion to $4 billion in annual revenue for SpaceX/SpaceXAI from the deal (CryptoBriefing; Fortune).

Per Forbes and Data Center Dynamics reporting, the capacity being transferred includes accelerators used for both training and inference workloads and comes online within weeks, giving Anthropic immediate scale while other contracted capacity ramps over the next six to twelve months (Forbes; Data Center Dynamics). Reporting also states xAI has migrated its main training operations to the newer Colossus 2 cluster, leaving Colossus 1 available to monetise (Forbes; CryptoBriefing).

According to CryptoBriefing, the reported $5-6 billion annual lease would convert otherwise idle capital into recurring revenue ahead of public-market events; Fortune reported an alternate analyst estimate of $3-4 billion annually (CryptoBriefing; Fortune). CryptoBriefing and Fortune both link the timing of monetising Colossus 1 to xAI/SpaceXAI's wider financing and public-market positioning, though those sources present this as observer interpretation rather than a company-declared rationale (CryptoBriefing; Fortune). Forbes includes direct public remarks from Elon Musk posted on social channels about being "ok leasing Colossus 1 to Anthropic," which the outlet reproduces in context (Forbes).

Observers should track three measurable indicators: reported utilisation and uptime figures for Colossus 1 once Anthropic begins workloads; any subsequent compute-leasing arrangements between frontier labs or with hyperscalers; and how reported revenues show up in SpaceX/SpaceXAI financial disclosures if the parent company includes those lines. Coverage also highlights whether future data-centre builds are structured with leasing or resale in mind from day one (Forbes; CryptoBriefing; Data Center Dynamics).

Frontier AI labs have continued to face a structural mismatch between large sunk capital investments in specialized accelerators and fluctuating internal demand. Companies converting built-out data centres into external compute products is an increasingly visible pattern in 2026 coverage, as publicly reported by multiple outlets in this deal (Forbes; CryptoBriefing; Fortune). Leasing idle capacity to peer labs shortens the timeline for realising returns on capital-intensive builds and changes the dynamics between hyperscalers, cloud providers, and model labs.

For practitioners, the Anthropic-xAI arrangement demonstrates how access to hundreds of thousands of accelerators can be purchased outside the hyperscaler market, altering cost and deployment trade-offs for teams that scale models to frontier sizes. It also illustrates a growing secondary market for frontier compute where availability, not just price, constrains large-model development.

The deal is a clear example of monetising sunk infrastructure to address a sector-wide compute shortage. For engineering teams and platform planners, it signals that large-scale external compute capacity can now be sourced directly from peer labs as an alternative to spot hyperscaler capacity or waiting for vendor shipments.