Anthropic Books Compute on the Ground — and Eyes the Sky

Anthropic to consider using SpaceX orbital data center satellites - SpaceNews

A 300-megawatt deal for SpaceX's Colossus 1 supercomputer comes paired with a memo of intent on multi-gigawatt orbital data centers — a concept whose physics is harder than its press releases admit.

 

By Stephen L Pendergast, Senior Life Member, IEEE  |  7 May 2026

BLUF — 

Anthropic on 6 May 2026 announced an agreement to consume the entire compute output of SpaceX's Colossus 1 facility in Memphis — about 300 MW driving more than 220,000 NVIDIA GPUs — and, separately, said it had "expressed interest" in partnering with SpaceX on multi-gigawatt orbital AI compute capacity.^1,2,3 The terrestrial half is concrete and will come online within a month; the orbital half is, today, an aspirational MoU layered atop a January 2026 SpaceX FCC filing for up to one million data-center satellites.^4,5 The deal sits at the intersection of three forces: AI demand growing at an annualized 80× year-over-year (per Anthropic CEO Dario Amodei),¹ a SpaceX-xAI merger and a planned June 2026 SpaceX IPO valuing the combined entity in the $1.75–2 trillion range,^6,7 and a Stefan-Boltzmann-bounded thermal problem that still constrains how many watts of silicon can credibly be parked in low Earth orbit.^8,9

The terrestrial deal: 220,000 GPUs, one customer

The headline transaction is straightforward. Anthropic has contracted for the full compute output of Colossus 1, the Memphis, Tennessee facility built by xAI and now owned by SpaceX following Elon Musk's all-stock acquisition of xAI in February 2026.^2,6 Anthropic's own announcement specifies "more than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs) within the month," with the GPU mix dominated by H100, H200, and next-generation GB200 accelerators according to the SpaceXAI release.^3,10

The capacity will be applied to lifting usage limits on Claude Pro, Max, Team, and Enterprise tiers — Anthropic specifically called out doubled rate limits for Claude Code on paid plans, removal of peak-hour caps, and a step increase in Claude Opus API throughput.^3,11 Amodei, speaking at Anthropic's developer conference the same day, framed the urgency in compounding terms: through Q1 2026, the company was growing at an annualized 80× the prior year's revenue and usage levels — a factor of eight above the 10× planning case that had governed its capacity procurement.¹

For context, the Colossus 1 contract joins a stack of recent Anthropic capacity agreements that together approach what would once have been considered the entire global merchant AI cloud: an up-to-5 GW Amazon agreement (with about 1 GW of new capacity by year-end 2026); a 5 GW Google/Broadcom deal coming online in 2027; a $30 billion Azure capacity arrangement with Microsoft and NVIDIA; and a $50 billion U.S. AI infrastructure investment with Fluidstack.³

Note on Colossus 1's power source

Colossus 1 became operationally controversial in 2025–2026 because xAI and its subsidiary, MZX Tech, LLC, installed dozens of natural-gas turbines on site and operated them under a temporary-use claim that bypassed federal permitting.² Local air-quality complaints and protests around the Memphis facility have continued. Anthropic's customer relationship now puts a frontier-lab AI workload on top of that infrastructure stack.

The orbital component: an MoU sitting on top of a megaconstellation filing

The space-based portion of the announcement is best read as a letter of intent rather than a procurement. Anthropic stated it "expressed interest" in working with SpaceX on multiple gigawatts of orbital compute; SpaceX's framing was nearly identical.^3,10 Neither side disclosed a price, schedule, or service-level architecture.

That MoU rests on a regulatory filing made on 30 January 2026, when SpaceX submitted an application to the Federal Communications Commission for a non-geostationary satellite system of up to one million spacecraft, branded the "SpaceX Orbital Data Center system." The filing specifies operation between 500 and 2,000 km altitude, in 30-degree and sun-synchronous inclinations, with orbital shells spaced as closely as 50 km.^4,12 The FCC's Space Bureau accepted the application for filing on 4 February 2026 (DA 26-113) and opened a public comment window.¹²

Several design choices in the filing carry engineering content. Sun-synchronous dawn-dusk orbits keep satellites in sunlight more than 99 percent of each orbit, allowing solar arrays to operate near continuous duty and minimizing battery mass.⁵ Communication is to occur primarily over optical intersatellite links, with traffic relayed through Starlink to ground stations; the filing seeks Ka-band only as a backup for telemetry, tracking, and command, and on a non-interference basis.⁴ SpaceX requested waivers from the FCC's standard milestone rules requiring half a constellation to be deployed within six years and the full system within nine — a request that on its face would be hard to satisfy at million-satellite scale even with Starship in full operation.^4,13

The thermal problem nobody can write its way around

The most persistent engineering issue with orbital AI data centers is not radiation hardening, debris, or launch cost. It is heat rejection. In a vacuum, convection and conduction to the surrounding environment are unavailable; waste heat must be radiated, and radiative power is governed by the Stefan-Boltzmann relation:

P = εσAT⁴

where σ ≈ 5.670 × 10^-8 W·m^-2·K^-4, ε is surface emissivity, A is radiator area, and T is the radiator's absolute temperature. Because P scales as T⁴, doubling allowable radiator temperature increases rejection capacity 16-fold — but AI accelerators tolerate only modest increases in junction temperature before they throttle or fail. Practically, designers are forced to grow A.^8,9

The numbers compound unfavorably. Starcloud, the most operationally advanced of the orbital-data-center startups, says its Starcloud-2 will field "the largest commercial deployable radiator sent to space" to support roughly 8 kW of compute power.¹⁴ SpaceX's per-satellite power figures have not been disclosed, but to reach the multiple gigawatts that Anthropic and SpaceX both invoke, the radiator area summed across the constellation must be measured in many square kilometers.^9,15 A consultancy review by Arthur D. Little catalogs the second-order problems: temperatures swinging from +120 °C to −250 °C through eclipse, the need for variable-emittance coatings and louvers, and the conflict between continuous sunlight (good for power) and continuous illumination of radiators (bad for cooling).⁹

Radiation is the second hard constraint. Google's recent Project Suncatcher preprint reported no hard failures in its Trillium TPUs up to a tested total ionizing dose of 15 krad(Si), suggesting some commercial silicon can survive LEO radiation environments better than once believed.¹⁶ Starcloud has flown an NVIDIA H100 on Starcloud-1 and reported successful AI inference workloads, but CEO Philip Johnston has been candid that the H100 is "probably not the best chip for space."^14,17 SpaceX's own answer to the chip problem is the D3, a radiation-hardened processor designed specifically for orbital operation and intended to be produced at the proposed Terafab facility in Texas.¹⁸

Terafab: the missing link

Terafab is the Tesla–SpaceX–xAI–Intel chip-fabrication joint venture announced by Musk on 21 March 2026 at the defunct Seaholm Power Plant in Austin.^18,19 A 6 May 2026 Grimes County, Texas hearing notice puts the project's first-phase capital cost at $55 billion, scaling to $119 billion across all phases — figures Musk publicly described as a 15-year program.^19,20 The plant targets 2-nanometer process technology, with Intel committed in April 2026 to contribute its 14-angstrom (1.4 nm) technology and packaging expertise.^18,19

Two product lines are planned. Tesla's fifth-generation AI5 inference chip is intended for vehicles and the Optimus humanoid robot. The D3 is the orbital chip. Musk has stated that 80 percent of Terafab's eventual output will be directed toward space.¹⁸ Volume production for AI5 is targeted for 2027; D3 has no public schedule. Without Terafab — or a comparable foundry partnership — there is no plausible path from a 220,000-GPU Memphis data center to a million orbital ones, since the merchant foundry capacity to fabricate space-qualified accelerators at constellation scale does not currently exist.

The competitive field

SpaceX is not alone in orbit-as-substrate, only the largest claimant. The current census of major filings and operating programs:

• SpaceX Orbital Data Center system — up to 1,000,000 satellites, 500–2,000 km, sun-synchronous and 30° inclinations. FCC application accepted DA 26-113 (4 Feb 2026).^4,12
• Blue Origin "Project Sunrise" — up to 51,600 satellites, 500–1,800 km sun-synchronous, with TeraWave broadband as backhaul. FCC filing 19 March 2026; NASA filed a formal objection on 5 May 2026 citing collision risk with crewed and science assets.^21,22,23
• Starcloud — up to 88,000 satellites; FCC application filed February 2026; $170 million Series A at $1.1 billion valuation closed 30 March 2026 (Benchmark and EQT Ventures co-led). Starcloud-1 (one H100) flew November 2025; Starcloud-2 (NVIDIA Blackwell B200, AWS server blade, 8 kW) is targeted for 2026; Starcloud-3 is a 200 kW, 3-ton bus designed to deploy from Starship.^14,17,24
• Google "Project Suncatcher" — research-grade preprint exploring TPU-based modular constellations; not a commercial filing.¹⁶
• Axiom Space — deployed an ISS-hosted compute node in September 2025; Aetherflux, Aethero, Sophia Space, Kepler Communications — earlier-stage entrants.^17,25

The Kessler-syndrome math is unforgiving. Harvard astrophysicist Jonathan McDowell catalogued roughly 14,500 active payloads in Earth orbit at the end of January 2026, of which Starlink accounted for about 9,500.¹³ SpaceX's million-satellite proposal alone would multiply that by a factor of nearly 70. Amazon has petitioned the FCC to deny SpaceX's filing; SpaceX has incorporated Amazon's arguments by reference into its objections to Blue Origin's Project Sunrise — a litigation pattern that will define the regulatory schedule for years.^22,23

The IPO and the politics

The commercial timing is not coincidental. SpaceX filed confidentially with the SEC on 1 April 2026 for an IPO targeting a $1.75–2 trillion valuation, with a public S-1 expected by late May and a roadshow targeting the week of 8 June 2026.⁷ The Anthropic deal arrives in the window where SpaceX must demonstrate AI-infrastructure revenue capacity to underwrite that valuation. Likewise, Terafab's $55 billion Grimes County filing hit the public record on 6 May 2026 — the same day as the Anthropic announcement.^19,20

The political backdrop is more complex. In March 2026 the Pentagon designated Anthropic a supply-chain risk and excluded it from a 10-vendor Department of Defense AI contract; Anthropic responded with parallel suits in San Francisco and Washington that remain in litigation.² The Defense Department has, in the same period, embraced xAI's Grok and competing models. Musk himself was a vocal Anthropic critic as recently as February 2026; his pivot — announcing on X that meetings with Anthropic leadership had not "set off [his] evil detector" — preceded the deal by days.^2,11 On the same day as the Colossus announcement, Musk posted that xAI would be "dissolved as a separate company" and renamed SpaceXAI.²

What is actually being procured

For an engineering audience, the practical questions are these:

On the ground: This is a bulk-capacity rental of an existing, gas-turbine-fed GPU farm. Anthropic gets useful compute within 30 days. The technical risks are operational, not architectural — power continuity, cooling-water availability in the Memphis basin, and whatever future regulatory action emerges around the unpermitted turbines.

In orbit: Anthropic has bought an option, not capacity. Multi-gigawatt orbital compute requires (1) Starship operating at a launch cadence not yet demonstrated, with the first 50-Starcloud-3-class commercial payload deployments not expected until mid-to-late 2028 even by the most aggressive operator; (2) radiation-hardened, space-qualified silicon at constellation volumes, which Terafab is built to supply but cannot before 2027 at earliest; (3) deployable radiator structures with characteristic dimensions in the tens of meters per spacecraft; (4) intersatellite optical-link reliability sufficient to synchronize training workloads — the hardest case — across a moving platform; and (5) FCC orbital authorizations not yet granted, with NASA, Amazon, DarkSky International, and the Center for Space Environmentalism all on record opposing the proposals.^9,14,17,23

Industry analysts speaking to TechCrunch and others have pegged commercial orbital data centers as a 2030s phenomenon, not the "two to three years" Musk has publicly claimed.^5,25 Microsoft President Brad Smith summarized the dominant view among hyperscale incumbents at a recent event: the company is "keeping [its] feet on the ground."¹⁷

The bottom of the matter

The Anthropic–SpaceX agreement does two distinct things and conflates them in the announcement. It locks down a 300-megawatt, 220,000-GPU bridge to keep Claude Pro and Max usage limits from constraining Anthropic's revenue trajectory in 2026. And it inserts Anthropic into the public narrative of orbital AI compute at the exact moment SpaceX needs that narrative for an IPO. Both halves serve both parties. Whether either party ever runs a single Claude inference on hardware in low Earth orbit will depend less on press releases than on whether the Stefan-Boltzmann constant, the Kessler envelope, and the FCC milestone clock turn out to be negotiable. They are not.

---

Sources

1. Foust, J. "Anthropic to consider using SpaceX orbital data center satellites." SpaceNews, 6 May 2026. https://spacenews.com/anthropic-to-consider-using-spacex-orbital-data-center-satellites/
2. Field, H. "Anthropic, SpaceX announce compute deal that includes space development." CNBC, 6 May 2026. https://www.cnbc.com/2026/05/06/anthropic-spacex-data-center-capacity.html
3. Anthropic. "Higher usage limits for Claude and a compute deal with SpaceX." Anthropic news release, 6 May 2026. https://www.anthropic.com/news/higher-limits-spacex
4. Foust, J. "SpaceX files plans for million-satellite orbital data center constellation." SpaceNews, 31 January 2026. https://spacenews.com/spacex-files-plans-for-million-satellite-orbital-data-center-constellation/
5. Moss, S. "SpaceX files for million satellite orbital AI data center megaconstellation." Data Center Dynamics, 31 January 2026. https://www.datacenterdynamics.com/en/news/spacex-files-for-million-satellite-orbital-ai-data-center-megaconstellation/
6. Foust, J. "SpaceX acquires xAI in bid to develop orbital data centers." SpaceNews, February 2026. https://spacenews.com/spacex-acquires-xai-in-bid-to-develop-orbital-data-centers/
7. "Anthropic signs Elon Musk's SpaceX for Colossus 1 compute ahead of June IPO." CoinDesk, 6 May 2026. https://www.coindesk.com/tech/2026/05/06/anthropic-signs-elon-musk-s-spacex-for-colossus-1-compute-ahead-of-june-ipo
8. "The 'Physics Wall': Orbiting Data Centers Face a Massive Cooling Challenge." SatNews, 17 March 2026. https://satnews.com/2026/03/17/the-physics-wall-orbiting-data-centers-face-a-massive-cooling-challenge/
9. Arthur D. Little. "Data centers go orbital." Viewpoint report, 31 March 2026. https://www.adlittle.com/en/insights/viewpoints/data-centers-go-orbital
10. SpaceXAI. "New Compute Partnership with Anthropic." 6 May 2026. https://x.ai/news/anthropic-compute-partnership
11. "SpaceX backs Anthropic with data centre deal amidst Musk's OpenAI lawsuit." Al Jazeera, 6 May 2026. https://www.aljazeera.com/economy/2026/5/6/spacex-backs-anthropic-with-data-centre-deal-amidst-musks-openai-lawsuit
12. U.S. Federal Communications Commission, Space Bureau. "DA 26-113: Space Bureau Accepts for Filing the SpaceX Application to Modify Its Authorization for the SpaceX Orbital Data Center System." 4 February 2026. https://docs.fcc.gov/public/attachments/DA-26-113A1.pdf
13. Quach, K. "FCC opens Musk's 1M-satellite DC plan for public comment." The Register, 5 February 2026. https://www.theregister.com/2026/02/05/spacex_1m_satellite_datacenter/
14. Henry, C. "Starcloud achieves unicorn status with $170 million raise for orbital data centers." SpaceNews, 30 March 2026. https://spacenews.com/starcloud-achieves-unicorn-status-with-170-million-raise-for-orbital-data-centers/
15. Liu, Y. et al. "Thermal Management Technologies for Space Data Centers: Current Status and Prospects." Journal of Refrigeration, 16 February 2026. https://www.sciopen.com/article/10.12465/issn.0253-4339.20251030004
16. Google Research. "Exploring a space-based, scalable AI infrastructure system design (Project Suncatcher)." 2025 preprint. https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/
17. Bishop, T. "Orbital AI: Seattle-area startup Starcloud hits $1.1B valuation to build space-based data centers." GeekWire, 1 April 2026. https://www.geekwire.com/2026/orbital-ai-seattle-area-startup-starcloud-hits-1-1b-valuation-to-build-space-based-data-centers/
18. "Terafab." Wikipedia, retrieved 7 May 2026. https://en.wikipedia.org/wiki/Terafab
19. Field, H. "Elon Musk's Terafab chip factory in Texas could cost up to $119 billion, filing shows." CNBC, 6 May 2026. https://www.cnbc.com/2026/05/06/elon-musks-spacex-chip-fab-in-texas-to-cost-up-to-119-billion.html
20. Wiggers, K. "SpaceX may spend up to $119B on 'Terafab' chip factory in Texas." TechCrunch, 6 May 2026. https://techcrunch.com/2026/05/06/spacex-may-spend-up-to-119-billion-on-terafab-chip-factory-in-texas/
21. Foust, J. "Blue Origin joins the orbital data center race." SpaceNews, 19 March 2026. https://spacenews.com/blue-origin-joins-the-orbital-data-center-race/
22. Boyle, A. "51,600 more satellites? Blue Origin adds another twist to the data center space race with Project Sunrise." GeekWire, 21 March 2026. https://www.geekwire.com/2026/blue-origin-data-center-space-race-project-sunrise/
23. "NASA Objects to Blue Origin's 'Project Sunrise.'" SatNews, 5 May 2026. https://satnews.com/2026/05/05/nasa-objects-to-blue-origins-project-sunrise/
24. Werner, D. "Starcloud Raises $170M to Fund Orbital Data Center Plans." Via Satellite, 30 March 2026. https://www.satellitetoday.com/finance/2026/03/30/starcloud-raises-170m-to-fund-orbital-data-center-plans/
25. Fernholz, T. "Jeff Bezos' Blue Origin enters the space data center game." TechCrunch, 20 March 2026. https://techcrunch.com/2026/03/20/jeff-bezos-blue-origin-enters-the-space-data-center-game/