Google's Case for Orbital Data Centres Rests on One Launch-Cost Curve. Check the Arithmetic.

Google researchers have published an analysis arguing that orbital data centres could make economic sense if launch prices fall to about $200 per kilogram by the mid-2030s, and that getting there would take on the order of 1,800 Starship launches over ten years. The authors say the work is not a complete economic feasibility study. That caveat is the right starting point, because the paper is best read as a statement of what has to be true about launch cadence, not as a forecast that it will be. This piece checks the arithmetic, which several outlets got wrong, and explains why the central variable is how often rockets fly rather than how good the chips are.

What the paper says, and what the coverage got wrong

Secondary coverage, including a TechCrunch article summarised by Universe Magazine and a longer account by Remio, reports the same core numbers: about 1,800 Starship launches in a decade, which is roughly 180 a year; about 200 metric tons per launch; and a launch price of about $200 per kilogram by around 2035. Some outlets describe the paper as forthcoming in the journal Joule. Remio links a PDF hosted on a Google domain and names no journal. We could not read the paper itself, so every figure below is as reported.

The reporting contains errors worth flagging. A KuCoin item's headline and URL gave 1,600 launches, and Remio notes that the source headline and URL originally cited 1,600 before a published correction gave 1,800. A KuCoin item also gave total mass as 3.7 million tons, while other outlets say about 370,000 metric tons. Only one of those can be consistent with the launch count. Multiplying 1,800 launches by 200 tons gives 360,000 tons, which is close to the 370,000 figure and about a tenth of 3.7 million. At 200 tons per flight, 3.7 million tons would require 18,500 launches. The sources' own text therefore supports roughly 360,000 to 370,000 tons, with the gap probably reflecting rounding or an unreported margin.

Why cadence is the binding variable

The paper's logic, as Remio describes it, is a learning curve: the price per kilogram falls about 20% each time the cumulative mass launched doubles, extended from SpaceX's history on Falcon 9 to Starship. Universe Magazine describes the assumption differently, as SpaceX having cut costs about 20% a year since Falcon 1. These are not the same assumption. A per-year decline needs only time; a per-doubling decline needs volume, and volume needs flights.

The per-doubling version has a simple consequence that the coverage did not draw out. Under a 20%-per-doubling rule, halving the price takes about 3.1 doublings, since 0.8 raised to that power is 0.5, which means about 8.6 times the cumulative mass launched. Each further halving requires another factor of roughly 8.6 in volume. That is our arithmetic on the assumed curve, not a figure from the paper. It shows why the authors frame the conclusion as a cadence requirement: price follows mass flown, so a fleet that does not fly does not get cheaper.

The same account says that if payload growth falls about 70% short of the central estimate, the price might still reach roughly $300 per kilogram. It also lists the things that must hold: high reuse, sustained cumulative volume, technical execution, competition and favourable regulation. The $200 figure is a conditional outcome, not a quote.

Setting 180 launches a year against the record

Universe Magazine, relaying TechCrunch, reports that Starship has never flown more than five times in a single year, while Falcon 9 flew 165 times in 2025. Against those two numbers, 180 Starship launches a year is 36 times the best single-year Starship total and about 1.1 times Falcon 9's 2025 count, from a vehicle that has not yet demonstrated that tempo. Remio notes that early years would fall below the average, which means later years would need to exceed 180. TechCrunch also treats with scepticism Elon Musk's suggestion that Starship could fly hourly by 2029.

Two further points bear on the number. Lower actual payloads would require more launches than 1,800, since the figure assumes 200 tons each. And Remio points out a circularity: space data centres need cheap launches, while Starship needs large payload demand to lower its costs. The project could itself supply some of that demand, but only if it exists at scale before the prices fall. Google is also an investor in SpaceX, a relationship both outlets note, and Google neither designs nor controls the rocket its economics depend on.

How big would the compute be?

The reported design is an 81-satellite cluster at about 650 kilometres altitude, with 100 to 200 metre separations and redundant provisioning for failures. Remio says Google's prototype satellite receives about 1 kilowatt of solar power and carries four TPUs. If each production satellite were similar, an 81-satellite cluster would draw on the order of 81 kilowatts. For scale, a 100-megawatt data-centre hall would need roughly 1,200 such clusters, or around 100,000 satellites. This is an illustration built on the prototype's power figure; the paper does not claim it, and a production satellite could differ. It does show why the launch count and the satellite count are different problems: 1,800 launches buys a very large tonnage, and what matters next is how much usable compute each ton delivers.

Remio reports that the prototype, built with Planet, launched on 1 October aboard a Falcon 9 on SpaceX's Transporter-18 rideshare mission and can run its processors for about 15 minutes before pausing to shed heat. That detail comes from a single account and we have not confirmed it elsewhere. It is consistent with Google's own update, as Remio relays it, calling cooling a crucial research challenge. Heat rejection in vacuum, where the only way to cool is by radiation, is a different engineering problem from cooling a building on Earth.

What the paper leaves out

• Economics beyond launch. Remio reports that manufacturing, insurance, operations, connectivity, repair and replacement costs are not fully captured. The authors say the study is not a complete economic feasibility analysis.
• Hardware lifetime. The design assumes five years of radiation exposure, so a ten-year launch campaign implies at least one full replacement cycle of the fleet, and every replaced ton counts toward the cumulative mass.
• Workload fit. Nothing in the material we reviewed says which jobs the orbital compute would take on or at what price relative to ground capacity.

The case for taking it seriously, and against

The argument for is that SpaceX has cut launch costs before, that Falcon 9's 2025 cadence shows high flight rates are achievable for a mature reusable vehicle, and that the Earth-side constraints on new capacity are real. Amazon Web Services' chief executive said on 2 October that more than 100 data-centre moratoriums are being considered across the United States, which is the kind of pressure that makes alternative siting attractive. The argument against is that every one of those favourable facts is a precondition rather than a result: the cadence has not been demonstrated for Starship, the learning curve is borrowed from a different vehicle, the economics are excluded by the authors' own caveat, and the heat problem is unsolved at scale.

The honest summary is that the paper converts a vague aspiration into a testable condition. If Starship flies in the low hundreds each year within a decade, launch cost stops being the dominant objection. If it does not, nothing else in the design matters. Readers can track the condition directly, by counting Starship flights per year, long before any orbital rack carries a paying workload.