Updated: July 8, 2026
The AI1 satellite — the first hardware in SpaceX's Starmind constellation — is a 20-meter-tall, 70-meter-wide orbital compute node that delivers roughly the compute payload of one NVIDIA GB300 server rack, cooled by deployable liquid radiators instead of water.
👉 Key takeaway: AI1's numbers are the most detailed specs SpaceX has released for any orbital compute hardware — but outside engineers argue the thermal and mass claims may not add up. Both sides of that debate are below.
- Compute power: 120 kW average, 150 kW peak — about the same draw as one NVIDIA GB300 rack on the ground.
- Dimensions: 20-meter deployed height, 70-meter wingspan — wider than a Boeing 747-8's fuselage.
- Cooling: 110 m² of deployable liquid radiators, redundant pumping loops, micrometeoroid shielding.
- Compute payload is modular — SpaceX says it can swap in NVIDIA GB300, Rubin, or TPU-class silicon depending on the customer.
- SpaceX also has its own radiation-tolerant orbital chip in development, internally called D3 (AI7).
💡 Context: for the full project overview see what Starmind is, and who builds the hardware in our Starmind supplier map.
What Is AI1: The First Starmind Satellite
AI1 is the name SpaceX gave the first-generation hardware for its Starmind orbital data center project, unveiled in a video presentation on June 8, 2026, ahead of the company's IPO. Elon Musk described it in strikingly simple terms: "The AI satellite is much simpler than a Starlink satellite," he said, noting it's mostly solar cells and laser links without Starlink's complex phased-array antennas, according to Digg's summary of the presentation.
Unlike a Starlink satellite, which exists purely to relay signals, AI1 is designed as a compute platform — the orbital equivalent of a server rack, running AI inference workloads on modular, interchangeable chips rather than fixed, purpose-built silicon, as Tom's Hardware reports.
AI1 Satellite: Complete Spec Sheet
Here is every confirmed AI1 specification disclosed during SpaceX's June 8, 2026 presentation and subsequent reporting.
| Category | Specification |
|---|---|
| Deployed height | 20 meters |
| Wingspan (deployed) | 70 meters — wider than a Boeing 747-8 fuselage (68.4 m) |
| Average compute power | 120 kW |
| Peak compute power | 150 kW |
| Power density | ~70 kW per metric ton (claimed) |
| Solar array output | 150 kW at roughly 250 W/m² |
| Radiator area | 110 m² of deployable liquid radiators, double-sided, "knife-edge" sun orientation |
| Heat rejection rate | ~1,400 W/m² claimed (vs. ~166 W/m² on the ISS's thermal system) |
| Thermal redundancy | Redundant pumping loops, integrated micrometeoroid shielding |
| Compute payload | Modular / interchangeable — NVIDIA GB300, Rubin, or TPU-class silicon |
| Orbital altitude | ~600 km |
| Inter-satellite links | Optical laser links, shared architecture with Starlink V3 |
| Manufacturing site | Gigasat facility, Bastrop, Texas |
These figures come primarily from SpaceX's own presentation and were compiled in reporting from Digg and Tom's Hardware.
💡 Good to know: The 150 kW peak figure lines up almost exactly with one NVIDIA GB300 rack's roughly 140 kW peak draw on the ground — that comparison is intentional on SpaceX's part.
How AI1 Handles Cooling in Space
Cooling is the single hardest problem in orbital compute. On Earth, a server rack sheds heat into moving air and circulating water. In the vacuum of space, neither exists — the only way to reject heat is by radiating it away as infrared light, as Tom's Hardware explains.
SpaceX's answer is a 110 m² deployable liquid radiator system with redundant pumping loops and micrometeoroid shielding. For scale, the International Space Station's thermal control system rejects about 70 kW of heat across 422 m² of radiator — roughly half of what AI1 needs to shed, using nearly four times the radiator area, according to figures cited by Tom's Hardware.
The Engineering Skepticism
Not everyone is convinced the numbers are physically achievable. A widely circulated Medium analysis by engineer Graham Wallington argues that AI1's claimed 150 kW from 110 m² works out to about 1,360 watts per square meter — roughly eight times the areal performance of the best thermal-control system ever flown.
The critique rests on the Stefan-Boltzmann law, which governs radiative heat loss: rejection power scales with the fourth power of a radiator's surface temperature. A radiator at a chip-friendly 30°C radiates only about 420 W/m² per face; reaching AI1's claimed rate would require surface temperatures far hotter than is typically safe for nearby electronics, per the same Medium analysis.
The same analysis also challenges the claimed 70 kW-per-ton mass density, estimating that even under the most optimistic assumptions for solar panels, radiators, and structure simultaneously, a real satellite would top out closer to 53 kW per ton — meaningfully short of SpaceX's public figure.
Musk has pushed back on thermal skepticism before, telling SpaceNews in March 2026 that it's "safe to say SpaceX knows how to do heat rejection in space," pointing to the company's existing satellite and spacecraft thermal systems, per Tom's Hardware.
⚠ Balanced view: SpaceX has a track record of shipping hardware that outside engineers doubted in advance — but AI1's thermal claims exceed anything ever flown by a wide margin, and no independent test data exists yet.
The Chips Inside AI1: GB300, Rubin, and SpaceX's Own Silicon
For the initial AI1 units, SpaceX plans to use existing NVIDIA GB300 and Rubin-class chips with SpaceX-designed reference hardware around them, according to reporting compiled by Digg.
Longer term, SpaceX is developing its own radiation-tolerant orbital processor, internally referred to as the D3 chip (also called AI7), part of what the company calls its "Terafab" chip architecture. Rather than using expensive traditional radiation-hardened fabrication, the D3 reportedly relies on commercial TSMC foundry nodes (likely N5 or N3) made radiation-tolerant through chip architecture rather than process, per Carthage Electronics' technical breakdown.
The chip sits in a broader SpaceX/Tesla silicon lineage: AI5 powers Tesla's Full Self-Driving stack, AI6 targets Cybercab and Optimus Gen 3, and AI7/D3 is the orbital compute variant — a detail reported by Carthage Electronics.
- If D3 works as intended, it could replace $5,000–$50,000 traditional rad-hard chips with commercial-foundry economics.
- SpaceX's Bastrop facility is targeted to reach meaningful D3 production volumes by the end of 2027.
- Until then, initial AI1 units will fly with off-the-shelf GB300/Rubin-class silicon rather than the custom chip.
Launch Math: What One Starship Flight Delivers
A single Starship launch can carry 30 to 50 AI1 satellites — the compute equivalent of dozens of GB300-class server racks delivered in one flight, with no land acquisition, no grid interconnection approval, and no ground cooling infrastructure required, as reported via Google News aggregation of the announcement.
At scale, SpaceX has told investors the constellation could add 100 gigawatts of AI compute capacity annually once Starship reaches a regular launch cadence, with a long-term vision extending toward 1 terawatt per year of deployed orbital AI capacity, per Carthage Electronics and Gadgetbond.
✔ Bottom line: The economics only close if Starship becomes fully and rapidly reusable — every scale projection SpaceX has published is downstream of that single dependency.
AI1 vs. a Terrestrial GB300 Rack: Quick Comparison
| Factor | Ground-Based GB300 Rack | AI1 Satellite (Orbit) |
|---|---|---|
| Peak power draw | ~140 kW | 150 kW |
| Cooling method | Liquid cooling + chilled water + air | Radiative liquid radiators only |
| Power source | Grid electricity (~24/7, weather-dependent) | Solar, near-constant in sun-synchronous orbit |
| Land/site needs | Data center site, zoning, grid interconnect | None — deployed in orbit |
| Deployment status | Operating today at scale | Prototype stage, launch targeted early 2027 |
Checklist: What to Watch on AI1's Specs
- Whether the two early-2027 AI1 prototype launches confirm the claimed 120 kW / 150 kW power figures in real operating conditions.
- Independent thermal telemetry showing whether the 110 m² radiator system actually rejects heat at the claimed rate.
- Progress on the D3 (AI7) radiation-tolerant chip and whether it reaches meaningful production volumes by end of 2027 as targeted.
- Whether actual satellite mass comes in near the claimed 70 kW-per-ton density or closer to independent estimates around 53 kW per ton.
- Starship's launch cadence and reusability progress, since the entire cost model depends on it.
👉 Bottom line on the checklist: AI1's numbers are the most specific SpaceX has shared for any orbital compute hardware, but every figure is still unverified by independent flight data.
Frequently Asked Questions (FAQ)
What chips does the AI1 satellite use?
Initial units will use existing NVIDIA GB300 or Rubin-class chips. SpaceX is separately developing its own radiation-tolerant orbital chip, internally called D3 or AI7, for future generations.
How much power does an AI1 satellite generate?
SpaceX states a 150 kW solar array producing 120 kW of average compute power and up to 150 kW at peak — roughly matching one NVIDIA GB300 rack's power draw on the ground.
How does AI1 stay cool without air or water?
AI1 uses 110 square meters of deployable liquid radiators with redundant pumping loops, rejecting heat as infrared radiation into space. Some independent engineers question whether the claimed heat-rejection rate is physically achievable at this scale.
How big is the AI1 satellite?
AI1 stands 20 meters tall with a 70-meter wingspan when fully deployed — wider than a Boeing 747-8's fuselage.
When will AI1 satellites actually launch?
Two AI1 prototypes are scheduled to launch in early 2027, with volume production targeted for late 2027 at SpaceX's Gigasat facility in Bastrop, Texas.
The Bottom Line
AI1 is the most technically detailed orbital AI compute satellite ever disclosed — SpaceX has published specific numbers on power, mass, radiator area, and chip strategy rather than vague concept art. At the same time, independent engineering critiques raise real questions about whether the thermal and mass figures hold up against known physics. Both sets of claims deserve tracking as actual flight data starts to arrive in 2027.
Bookmark this page and check back as AI1 prototype telemetry, D3 chip progress, and independent thermal analysis roll in over the next year.
NEVER MISS AN OPTIMUS UPDATE
We track every Tesla Optimus development — specs, deployment milestones, pricing and competitive moves — updated as news breaks.