SpaceX's Compute Ambition: Starmind AI1, an Nvidia Lock-In, and 10 Gigawatts by 2027
SpaceX is making AI compute its core business. The company said this week that every data center it builds will run exclusively on Nvidia hardware, with more than 2 gigawatts of capacity targeted by the end of 2026 and close to 10 gigawatts by late 2027. The orbital half of that plan is the Starmind AI1 program, unveiled alongside SpaceX's first earnings report since its IPO, a report in which soaring AI costs outweighed a revenue beat.
The partnership, confirmed on August 4 the same day SpaceX reported results, gives Nvidia responsibility for designing the compute payload on each Starmind satellite, while the ground-side buildout uses Nvidia Vera Rubin NVL72 rack-scale systems, also known as Kyber. SpaceX confirmed the arrangement in a post on X, describing the payload as datacenter-class compute for orbit. Markets read the announcement as a reordering of the AI chip market. SpaceX shares fell about 10 percent as investors absorbed the cost of the compute push, Nvidia rose roughly 2 percent after Musk described its products as the best AI computers available, and AMD dropped about 9 percent despite a record quarter of its own, shut out of the arrangement entirely.
The framing matters as much as the hardware. SpaceX has historically sold launch services; the compute program converts the company into a capacity provider that happens to own rockets, a satellite constellation, and factories. Musk framed the 2027 capacity target as closer to ten gigawatts than five, which would place SpaceX among the world's largest AI infrastructure operators within roughly two years of starting the buildout, a category that until now has been defined by the largest cloud operators.
Inside the Starmind AI1 Satellite
The Starmind AI1 spacecraft stands 30 meters tall with a 75-meter solar wingspan, and its compute payload draws up to 250 kilowatts at peak and 175 kilowatts on average, or 75 kilowatts per ton of vehicle mass. Power comes from space-based solar panels, and cooling uses the vacuum of orbit, sidestepping the grid capacity and water limits that cap terrestrial data center growth. The design goal is localized high-performance compute, processing data where it is collected instead of beaming it back to ground stations, which cuts the bandwidth cost of shipping raw imagery and telemetry to Earth.
| Starmind AI1 at a glance | Value |
|---|---|
| Height | 30 m (98 ft) |
| Solar wingspan | 75 m (246 ft) |
| Compute power | 250 kW peak / 175 kW average |
| Power and cooling | Space-based solar; vacuum cooling |
| Payload | 36 Nvidia Vera CPUs, 72 Rubin GPUs |
| Interconnect | NVLink 6 at 260 TB/s all-to-all |
| Data path | Laser links via Starlink |
| Deployment target | Late 2027 |
Each satellite carries an NVL72-equivalent system: 36 Vera CPUs and 72 Rubin GPUs connected by NVLink 6's 260-terabyte-per-second all-to-all fabric, six times the GPU count of Nvidia's Space-1 module. In effect, one spacecraft holds the computing density of a rack-scale data center. The thermal budget is the hard constraint. A full NVL72 rack generates substantially more heat than the 150-kilowatt design envelope documented for the first AI1 prototype, so the vacuum-cooling approach must outperform anything demonstrated in orbit to date. That is precisely why the two companies are co-designing the payload: rack and spacecraft have to be engineered as a single system, and any shortfall pushes the late-2027 deployment target.
Data moves through high-bandwidth laser links into the Starlink constellation, which effectively becomes the interconnect for a distributed orbital computer. That design reuses infrastructure SpaceX already operates, and it is the part of the plan competitors would struggle to replicate: a satellite fleet that carries compute traffic between orbital nodes and back to Earth without new ground stations. The same laser network that routes Starlink internet traffic now doubles as the fabric linking orbital compute nodes. For workloads such as earth observation, processing in orbit removes the round trip to a ground station and the latency it adds.
The 10-Gigawatt Ground Buildout
The near-term numbers are a ground story. Even a fleet of a thousand Starmind AI1 satellites would contribute only about a quarter of a gigawatt at peak power, so the 2-gigawatt target for the end of 2026 and the roughly 10-gigawatt goal for late 2027 depend on terrestrial data centers built on Nvidia's Vera Rubin NVL72 racks. Growing from 2 gigawatts toward 10 in twelve months is a fivefold ramp, a pace that implies most capacity lands in conventional facilities while the orbital program matures in parallel. SpaceX has described the 10-gigawatt figure as cumulative capacity built over that period. With the end-2026 milestone less than five months away, the 2-gigawatt target is the first test of whether the buildout can move at the promised pace.
Manufacturing capacity is being assembled on the same timeline. The Gigasat Factory in Bastrop, Texas, will produce the satellites at volume, and a separate Terafab facility is planned for SpaceX's own advanced AI chips. The Starmind architecture is deliberately chip-agnostic, which makes the Nvidia exclusivity a contract rather than a permanent dependency: the deal covers today's buildout, while Terafab keeps an in-house silicon path open for later generations of the fleet. For procurement strategists, that combination couples near-term vendor lock-in with a visible path to second sourcing.
Nvidia does not have to wait for orbit to monetize the relationship. Nvidia rates its Space-1 Vera Rubin module at up to 25 times the AI performance of an H100 GPU, and commercial shipments of the platform are expected to begin later this year, giving the partnership a revenue path independent of the satellite schedule. The Space-1 line also gives Nvidia a space-ready product it can sell beyond the SpaceX relationship. A customer planning 10 gigawatts of capacity would rank among Nvidia's largest by volume, a position that carries allocation weight in a supply-constrained market. The co-design relationship also keeps Nvidia inside a payload architecture that SpaceX has deliberately kept chip-agnostic, a position that matters if Terafab silicon ever ships.
The divergent stock moves capture how investors now value SpaceX. Its first quarterly report as a public company beat revenue expectations, yet the stock sold off about 10 percent because AI spending dominates the cost side of the ledger. Nvidia's roughly 2 percent gain and AMD's 9 percent decline show capital markets treating the exclusivity agreement as a structural shift in AI supply, not a line item. The AMD drop landed even as the company reported record quarterly results, a sign of how much of the AI hardware market's future value now flows through Nvidia.
For AI buyers, the pipeline is the point. A 10-gigawatt source of capacity arriving by 2027 offers an alternative to cloud incumbents at a moment when GPU availability still constrains model development, and a provider that generates its own power and runs its own network can price differently than one renting both. FCC filings sketching constellations of up to a million AI satellites extend the ambition beyond the first unit; at that scale, orbital compute would eventually dwarf the ground plan. The near-term milestones are concrete: more than 2 gigawatts by the end of this year, Space-1 commercial shipments within months, and first Starmind AI1 deployments in late 2027. Each one changes the availability picture for teams planning large AI workloads, and any renegotiation of the exclusivity deal as Terafab matures would reset both pricing and supply.
Why This Matters
SpaceX's pivot recasts a launch company as a compute operator with its own orbital capacity, a laser-linked network fabric, and a multi-gigawatt roadmap. The Nvidia lock-in secures near-term supply while the Terafab plans signal possible vertical integration later. Whether vacuum cooling and orbital clusters can deliver on the 10-gigawatt promise is the open question, with the first Starmind AI1 deployments scheduled for late 2027.
Photo by Brecht Corbeel on Unsplash
Related Articles
- SpaceX Pentagon AI Compute Deal Could Reshape Defense Cloud Market
- SpaceX Terrestrial AI Compute Business Generates $28B from Colossus Deals
- SpaceX Targets $2 Trillion Valuation in IPO Filing Focused on AI Infrastructure Pivot
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.