Deep|Space Compute: SPCX Could Reach 1 GW of On-Orbit Compute in 2028

Executive Summary

1. A thermal design is now in place: chips run hotter, and each satellite carries one NVL72 rack. SpaceX has disclosed AI1 at 250 kW peak and 175 kW average with a 160 m² radiator, but not the operating temperature. Working backward from our June radiator table, that implies a radiator surface of ~65–75°C and chip junction temperatures of 80–90°C. Our industry checks corroborate this: the space version runs its chips at junction temperatures of 80–90°C. That is feasible from an engineering standpoint, since chips in terrestrial racks often run above 80°C at full load, and the temperature itself is not a problem for the chips; the binding constraint is memory. In orbit, running hot is worth far more for heat rejection: in a vacuum, heat can only be shed by radiation, and radiative capacity rises with the fourth power of temperature. Rejecting heat at ~40°C, as is customary on the ground, would require 240–290 m² of radiator; by running hotter, a ~2 t satellite carries exactly one NVL72, a Starship carries ~40 satellites, and no on-orbit assembly is needed.
2. The compute payload is locked: a space version of NVL72, codename Kessel. In August, SpaceX announced it is co-designing AI1’s compute payload with NVIDIA, using Vera CPUs and Rubin GPUs. the space rack architecture was finalized in October: it keeps the 72-GPU NVLink domain, compresses the 48U upright terrestrial rack into a ~9U flat enclosure, and cuts the weight to a few hundred kilograms. The first generation targets inference; large-model training stays on the ground. The chips and HBM are standard parts. Radiation is addressed through an external shield, HBM’s built-in ECC and system-level fault tolerance, and heat through a single liquid loop running straight to the external radiator, concentrating demand for advanced materials and lightweight alloys in the shielding, radiator and structural components.
3. Supply-chain orders are in: 6,000 racks for 2027, equal to 1.05 GW. SpaceX has ordered 6,000 space-specific racks for 2027 from a leading server ODM, the sole JDM (joint design manufacturer) on the program, which is due to complete NPI in May 2027, in time for the test satellite, and finish production in December. The 6,000 racks equate to 432,000 GPUs and ~$42–45 billion of hardware value at terrestrial rack pricing; at one rack per satellite, they equal 1.05 GW. Lifetime volume for the program is ~30,000 racks, matching the plan for a first phase of ~30,000 satellites, or ~2.16 million GPUs at 72 per rack, equal to 5.25 GW (7.5 GW peak); the first 6,000 racks are a fifth of that.
4. Constraints on the ground are pushing compute into orbit. Democrats won both houses of Congress in the November 3 midterms, and Texas’s incoming governor has pledged to pause new data centers until the legislature sets rules. Permitting, grid-connection and cost-allocation pressure on terrestrial data centers is rising, and the impact falls mainly on capacity added from 2028, just as Starmind scales. Orbital compute needs no land, grid connection or water, and no state or local data-center permits. As constraints on terrestrial capacity tighten, orbital compute becomes more attractive.
5. Bull case: ~1 GW on orbit by end-2028. In our bull case, Starship has ~220 launches available in 2028 (Starbase expanded to 100, plus 44 from LC-39A and 76 from SLC-37), of which ~70%, or ~154, go to compute satellites. At 40 satellites per launch, that is enough to launch all 6,000 racks, so we estimate end-2028 on-orbit compute of ~1 GW in our bull case (average compute-payload power at 175 kW per satellite).
6. Two checkpoints ahead. Flight 15, expected no earlier than late November, will attempt to catch both stages with the tower; a successful catch and reflight are what support 30–40 launches in 2027. The test satellite launching on Starship in May 2027 will deliver the first on-orbit thermal data and show whether the 175 kW average per satellite is achievable in orbit, which is what our GW estimates rest on.

Five Months After the SPCX IPO