Anthropic Just Took All of SpaceX's Memphis Compute

Anthropic announced on May 6, 2026 that it has reserved all of the compute capacity at SpaceX's Colossus 1 data center in Memphis, Tennessee, securing more than 300 megawatts and signaling interest in working with SpaceX on multiple gigawatts of orbital compute. The deal is the loudest commercial validation yet of SpaceX's pivot from rockets-and-Starlink into AI infrastructure.

Most of the headlines this week are chasing the orbital piece. They have it backwards. The 300 MW lockup in Memphis is the real story. The orbit pitch is a 2030s option that SpaceX itself has started to publicly hedge.

What happened

Anthropic confirmed to CNBC that it will use the entirety of SpaceX's Colossus 1 site in Memphis, a facility SpaceX has been quietly building out as a dedicated AI training campus. The contract covers more than 300 MW of capacity and includes joint exploration of "multiple gigawatts" of in-orbit compute, per reporting from NPR.

The orbital piece is downstream of an FCC filing SpaceX submitted in January 2026 requesting authorization to deploy up to one million orbital data center satellites. SpaceX has not disclosed unit economics or a launch timeline for the program.

Elon Musk's framing on the orbital plan, delivered at an X all-hands and recapped by Datacenter Dynamics, was that "the cost of deploying AI in space will drop below the cost of terrestrial AI much sooner than most people expect" and that the crossover "may be only two or three years" away. SpaceX is also confirming a 2026 IPO partly to fund this buildout, per FNEX coverage of the S-1 timing.

There is a complication. In late April, SpaceX engineers walked back several core assumptions in the orbital architecture, as reported by Futurism, citing thermal dissipation, radiation hardening, and bandwidth-to-Earth constraints that the public timeline does not yet account for.

Anthropic has not disclosed pricing terms for the Memphis tenancy.

What it actually means

Why does the Memphis deal matter more than the orbit pitch?

Because Memphis is real, dated, contracted compute, and orbit is a press cycle. Anthropic gets 300 MW of training capacity it can route to Claude post-training and inference within months. The orbital plan is a regulatory filing, a partner statement of interest, and an unproven thermal envelope. One ships models, the other ships slides.

That gap matters because Anthropic's binding constraint right now is not strategy. It is power. Frontier labs are queueing for grid interconnects in markets where utilities are quoting 24-to-36-month wait times for new substations. A signed tenancy at a campus that already has power on contract collapses that timeline. SpaceX, sitting on Starlink ground-station rights and high-density site engineering, can move faster than a typical hyperscaler co-lo.

What does 300 MW actually buy Anthropic?

Roughly enough power for a 100,000-to-150,000 H200-equivalent GPU cluster running steady state, depending on the rack design. That is in the same order of magnitude as the largest publicly known training clusters, and meaningfully larger than what Anthropic is believed to operate today. It is not a "next Claude" cluster on its own. It is a credible piece of one.

For context, Brandon Lucia, a Carnegie Mellon professor who works on space-based computing, told MIT Technology Review that Musk's two-to-three-year orbital timeline is "an optimistic interpretation," noting it "may be feasible, but not next year and certainly not in three years." Translation: if Anthropic is planning capacity for the next two training runs, all of that capacity has to be terrestrial. Memphis is where it lives.

This is the same Anthropic that has spent the last 18 months trading compute access for distribution: cloud commitments with AWS, then Google, now an asset-heavy operator. It mirrors the capex logic Meta is running at a different scale, and it lines up with the strategic posture Anthropic has been telegraphing for a year. The company is willing to lock in long-dated capacity at premium rates because the alternative, queueing for utility power, is fatal at the frontier.

What does SpaceX get out of this?

A signed tenant that lets SpaceX call Colossus 1 "Anthropic-anchored" in its IPO marketing. Anchor tenants for a new infrastructure asset class lower cost of capital, validate the engineering, and make the orbital pitch look like a credible roadmap rather than a vision deck. Even if the orbital piece never ships at scale, the Memphis revenue is recurring and contracted.

Who this affects

Enterprise buyers already on Claude. Capacity pressure is the single biggest reason Anthropic has been rate-limiting enterprise tiers and slowing model releases relative to OpenAI. A 300 MW unlock changes that math. Expect the rate-limit story to ease through Q3, and expect Anthropic to be more aggressive on enterprise discounting where they were defensive six months ago. The same compute pressure is what made OpenAI's recent product moves look reactive to Anthropic's enterprise momentum.

Solo developers and indie builders. Less direct impact. Anthropic's API tier is unlikely to see immediate price relief from a single training-class capacity add. The bigger signal is that Claude availability and model-cycle cadence should improve on a 6-to-12 month lag.

OpenAI and Google. OpenAI is locked into Microsoft and CoreWeave; Google has its own TPU stack. Neither is structurally hurt by a single Anthropic-SpaceX deal. What is new is the precedent: a non-hyperscaler, vertically-integrated infrastructure operator just won a frontier-lab anchor tenancy. Expect Oracle, CoreWeave, and possibly xAI's own infra arm to pitch similar exclusive lockups in the next two quarters.

SpaceX's IPO investors. The Memphis deal is now a line item in the S-1 narrative. The orbital plan, after the recent engineering walk-back, should be priced as optionality, not as a near-term revenue line. Read the prospectus carefully when it lands.

What to watch for next

Three signals over the next 90 days.

First, the FCC orbital filing. Watch for a public comment period response, especially from astronomers and other satellite operators. A friction-heavy ruling would push the orbital timeline out by years and shrink the "gigawatt" interest letter into a footnote.

Second, the SpaceX S-1. The filing will be the first time SpaceX has to publicly disclose data center revenue, capex commitments, and counterparty risk concentration. If Anthropic is more than 30 percent of the AI infra book, that is a flag.

Third, Anthropic's next Claude release. If Memphis is online for training by late 2026, the model that ships from it will be the public test of whether the deal moved the frontier. The competitive context for that release, especially given the open questions in Claude Opus 4.7 vs GPT-5.5, will tell you more about Anthropic's actual capacity unlock than any press release.

A useful tell on Anthropic's own posture is whether the lab keeps optioning capacity at premium prices or starts to walk back. The Project Glasswing playbook around Claude Mythos is the same operating pattern at a different layer: build the capability, then choose deployment deliberately. Memphis is that instinct applied to compute. Anthropic does not lock in 300 MW it does not plan to use.

The bottom line

The orbital data center pitch will get the headlines. The Memphis deal is what actually changes the AI infrastructure map this year. Anthropic just locked down 300 MW of training capacity from a non-hyperscaler operator, validated SpaceX as an AI infra category entrant ahead of an IPO, and bought itself meaningful capacity at a moment when grid power is the binding constraint at the frontier. If you are an enterprise buyer on Claude, plan around easier capacity through Q3 and Q4. If you are watching SpaceX, value the Memphis revenue and price the orbit pitch as optionality.

FAQ

Did Anthropic buy SpaceX's data center, or just lease capacity?

Anthropic took all of the compute capacity at Colossus 1, which is a contracted tenancy, not an asset purchase. SpaceX continues to own and operate the site. Pricing terms have not been disclosed publicly as of May 7, 2026.

Is the orbital data center plan real, or marketing?

Both. SpaceX has filed with the FCC for orbital data center authorizations and Anthropic has expressed interest in multiple gigawatts of orbital compute. SpaceX engineers have also publicly walked back parts of the architecture. Treat it as a 5-to-10-year option, not a 2027 product.

How much compute is 300 MW?

Roughly enough power for a 100,000-to-150,000 GPU cluster at H200-class density, give or take rack design. That puts Colossus 1 in the same league as the largest publicly known training clusters, though smaller than the multi-gigawatt sites Microsoft and Meta are building.

Will this lower Claude's API prices?

Probably not in the short term. The deal is most directly relevant to enterprise capacity availability and model-release cadence. API pricing tends to move on competitive pressure, not raw capacity adds.

When does Memphis come online?

SpaceX has not disclosed a phased timeline publicly. Anthropic's framing suggests partial capacity through 2026 with full ramp into 2027.