Tech

Ars Technica analysis reveals steep technical and economic hurdles for SpaceX orbital data centres

While physics challenges like radiation tolerance appear manageable, the true barriers to SpaceX’s plan for one million orbital data centre satellites are heat dissipation, latency, and the logistical demand for up to 42 Starship launches daily.

Author
Owen Mercer
Markets and Finance Editor
Published
Draft
Source: Ars Technica · original
How hard is it to build orbital data centers, actually?
Feasibility of AI1 constellation hinges on unprecedented launch cadence and heat management solutions

Ars Technica has published the second instalment of a three-part series examining the technical and economic viability of SpaceX’s proposed "AI1" orbital data centre constellation. The analysis details the engineering hurdles involved, particularly heat dissipation, latency, and radiation tolerance, alongside the logistical demands of deploying one million satellites. SpaceX founder Elon Musk and satellite engineering director Ian Dahl recently unveiled specifications for the AI1 satellite, which is designed to host frontier-class GPUs using Nvidia Rubin chips. While the physics are considered manageable, the project requires unprecedented manufacturing scale and launch cadence, potentially necessitating 10 to 42 Starship launches daily. The report estimates that ground systems to handle global data traffic could cost approximately $100 billion, with heat management identified as the most critical engineering challenge.

The proposed AI1 satellite design includes solar panels covering approximately 600 square meters, generating 150 kW of peak power and 120 kW of average power for computing. The solar panels alone are estimated to weigh 1 to 2 metric tons, with radiators adding another 1 to 2 metric tons, bringing the total satellite weight to between 3.5 and 7.5 metric tons. To support a constellation of one million satellites, SpaceX would need to increase its launch cadence significantly, potentially requiring 10 launches per day in an optimistic scenario and up to 42 launches per day in a pessimistic one. SpaceX is currently planning a Starship V4 rocket with a payload capacity of 200 metric tons to low-Earth orbit, up from the V3’s estimated 100 metric tons.

The analytics firm Quilty Space estimates Starlink V3 satellites cost around $1 million; the report suggests a similar best-case cost for AI1 satellites, though orbital data centres will require more expensive computer hardware. Ground systems to handle global data traffic are estimated to cost roughly $100 billion. Starcloud, a startup, is launching the Starcloud-2 mission in October, featuring a 450 kg satellite with 8 kW of power generation, to demonstrate low-cost, low-mass radiator technology. Google’s experiments with V6e Trillium TPU compute trays indicated that ionizing radiation can cause device failures over time, but devices can operate reliably for about five years.

Iridium Communications chief executive Matt Desch noted that while the concept is a "hot" area of discussion, it may be driven more by valuation opportunities than immediate technical necessity. He described orbital data centres as a "really, really long-term opportunity" and emphasised his company’s pragmatic focus on cash flow and growth. Desch suggested that the enthusiasm surrounding the technology might not be solely about solving an immediate problem, but rather about positioning for future market shifts.

Heat dissipation remains the most significant engineering hurdle, as convective cooling is impossible in the vacuum of space. The International Space Station’s radiators weigh over 6 metric tons to dissipate 70 kW of heat, a mass that is considered too heavy for scalable commercial operations. Starcloud’s Philip Johnston stated that two-thirds of his engineering team is focused on developing low-cost, low-mass radiator technology to make the concept viable. Latency issues also pose challenges for distributed workloads, though inference tasks may be more adaptable to satellite constellations than large-scale training requiring tight synchronisation between GPUs.

Continue reading

More from Tech

Read next: The Walrus warns of collapsing digital memory as AI erodes search reliability
Read next: Open-source tool claims 97 per cent token savings for AI agents
Read next: Valvoline Unveils August 2026 Promotional Offers for Service and Retail Buyers