Starship's Orbital Compute Gateway: 1,800 Launches to Make Space AI Work

Google has run the numbers on what it would actually take to beat Earth-based AI datacenters from orbit—and the answer is roughly 1,800 Starship launches, a figure that amounts to about five missions a day for a full year, according to a peer-reviewed paper the company published October 1, 2026 in the journal Joule.

The analysis, titled "Towards a future space-based, highly scalable AI infrastructure system design" (a preprint of the Joule paper), projects that if SpaceX's reusable rocket achieves sufficient production scale, launch prices could fall below $200 per kilogram to low-Earth orbit by the mid-2030s. At that price, Google's researchers calculate that lifting the 370,000 tonnes of hardware needed for an orbital datacenter constellation would cost roughly $74 billion in launch fees alone. The figure appears in the paper's Section 2.3 on economic feasibility, which uses a sustained learning rate—meaning each doubling of cumulative launches reduces the per-kilogram cost—with the Google blog citing a 20% figure and the paper providing the detailed derivation.

The 1,800-launch figure is a derived model output: 370,000 tonnes ÷ approximately 200 tonnes per Starship mission ≈ 1,850 launches, which the paper rounds to "about 1,800." TechCrunch's headline of 1,600 uses the same underlying arithmetic but appears to round differently; both are Google's own numbers, reported with appropriate attribution.

Google's radiation testing, conducted on Trillium Tensor Processing Unit chips and described in the Joule paper's revised test section, found that the chips survive a total ionizing dose equivalent to a 5-year LEO mission without permanent failures, a meaningful result for long-duration satellite operations. However, the chips showed elevated logic error rates—characterized for bit-flip behavior under radiation—during extended large-scale training workloads. Google's researchers distinguish typical inference tasks, where chips perform reliably, from months-long training runs across thousands of accelerators, where cumulative radiation exposure becomes a limiting factor. The paper attributes these findings to the authors; they have not been independently replicated.

The 370,000-tonne figure represents the cumulative payload mass required to build out a constellation large enough to achieve cost parity with terrestrial datacenters, assuming the learning curve holds. It is not a one-time deployment requirement but the total mass launched over the build-up period. Google's own researchers treat this as a conditional projection, not an achieved milestone.

Project Suncatcher's first mission, described in Google's September 24, 2026 public briefing, is scoped as a hardware-survival and cooling experiment, with paired-satellite optical interconnection testing planned for 2027. The 1,800-launch figure is a forward-looking economic threshold, not a near-term operational target.

Google's researchers acknowledge significant unknowns: orbital dynamics for tightly-clustered satellite formations, free-space optical link performance at scale, thermal management in vacuum, and the actual pace of Starship cost reduction. The paper's analysis "works backward" from a speculative future in which most AI compute migrates to space—a vision the authors describe as a moonshot.

The paper was authored by Travis Beals, Maria Biggs, Jessica V. Bloom, Thomas Fischbacher, Konstantin Gromov, Urs Köster, Rishiraj Pravahan, James Manyika, and Blaise Agüera y Arcas. The Joule peer-reviewed version was released October 1, 2026. The earlier arXiv preprint (v1 and v2) was first posted in November 2025.

TechCrunch first reported the 1,600-launch figure on October 1, 2026, citing Google's Joule paper. The arithmetic discrepancy (1,600 vs. ~1,800) reflects rounding choices, not a factual conflict—both figures derive from the same 370,000-tonne cumulative payload assumption and ~200-tonne-per-mission estimate.