This reads Google’s own paper, arXiv version 2, revised 17 June 2026. This publication did not read the peer-reviewed Joule version. The 1,800 launches figure is a conditional on a launch-price target, not a prerequisite, and the paper says it is not a full economic analysis.
A Google paper converts launch economics into a single power-cost equation. The 1,800 Starship figure is a conditional, not a countdown.
This publication read arXiv version 2 of Towards a future space-based, highly scalable AI infrastructure system design, last revised 17 June 2026, by Blaise Agüera y Arcas, Travis Beals, and seven co-authors, all Google. TechCrunch reported on 1 October that Google released a peer-reviewed version to be published in Joule. This publication did not read the Joule version, which returned a 403. The figures below match TechCrunch’s account of it, but the sourced version is arXiv v2.
The paper’s central argument is one unit conversion. It translates launch cost per kilogram into an annualised cost of power delivered to orbit. Using a Starlink v2 mini satellite — 575 kilograms, an estimated 28 kilowatts — and a current launch price of $3,600 per kilogram based on the reusable Falcon 9 configuration used for Starlink, the paper obtains a launched power price of $14,700 per kilowatt per year. Amortised over a five-year lifespan. That number drops to $810 if launch falls to $200 per kilogram.
The paper’s comparator, verbatim: “current power spend for terrestrial data centers in the US is reported to be roughly $570 to $3,000 per kilowatt per year, depending on regional variation in power price and operator power usage effectiveness.” Its conclusion, verbatim: “if launch costs reach $200 per kilogram, annualized cost per unit of power in space could be approximately comparable to terrestrial spend.” The word is comparable. The paper does not claim orbit would be cheaper.
The key insight: The 1,800 Starship launches is a conditional statement inside a learning-curve model — the cumulative launched mass consistent with sustaining a roughly 20 per cent learning rate to reach below $200 per kilogram by around 2035. It is not a launch count that orbital data centres must wait for, and the paper does not present it that way.
The path the paper describes has four stated points. Its starting point is the introduction of Falcon Heavy, at roughly $1,800 per kilogram and roughly 400 tonnes of cumulative mass launched. The current price it uses is $3,600 per kilogram, based on the reusable Falcon 9 configuration used for Starlink. Learning-rate estimates across its choices come out at about 18 to 24 per cent.
The conditional target is a price below $200 per kilogram by about 2035. Sustaining the learning rate to reach it would require roughly 370,000 tonnes of additional cumulative mass, which the paper says is equivalent to roughly 1,800 Starship launches at 200 tonnes each. This publication’s own arithmetic: 370,000 divided by 200 is 1,850, which the paper rounds to about 1,800.
On sensitivity, the paper says that even if the launch rate is reduced by about 70 per cent, prices could drop to $300 per kilogram in the same timeframe.
The paper calls the $200/kg result approximately comparable to terrestrial spend. It also says this is not a full economic analysis.
The Structural Read
The paper’s logic is a Product Overhang argument in aerospace clothing. The capability — power delivered cheaply to orbit — does not exist yet. But the paper maps the cost curve that would have to hold for it to materialise. That is a meaningful engineering contribution. It is not an economic proof.
The comparison the paper makes is narrow on purpose. It sets the launch cost of power-generating mass against terrestrial power spend only. It excludes compute hardware, networking, maintenance, and every other operational cost. The paper says explicitly it does not constitute a full economic analysis, but gives proof points demonstrating that there is a feasible path for launch costs to drop sufficiently to no longer be prohibitive.
The result is also highly contingent on power-to-mass ratio. Table 1 in the paper shows what happens across satellite designs at the $200 per kilogram price point. This publication’s own reading of that table: two of the four designs — Starlink v2 mini at $810 per kilowatt-year and Starlink v1 at $1,470 — fall inside the paper’s $570 to $3,000 terrestrial range. The other two — OneWeb at $7,500 and Iridium at $6,900 — sit above it.
So the approximately comparable result holds only for satellites that deliver a lot of power per kilogram. The paper does not claim the OneWeb or Iridium designs are data-centre designs. That observation is this publication’s own reading of the table.
Product Overhang Doctrine
The Cost Curve Is the Product
The paper does not describe a product that exists. It describes the launch-cost trajectory the authors say would stop launch from being prohibitive. The 1,800 launch figure is a point on that curve — one scenario, under stated assumptions, not a prerequisite or a starting gun.
“This is an ambitious target, but the required roughly 180 Starship launches a year, on average, with a ramp-up period almost certain in reality, would fall well below stated launch rate targets.”
The paper also notes that reaching the $200 per kilogram threshold is unlikely without Starship-like launch vehicles, because of Falcon payload volume limits. And it notes that $200 per kilogram is often cited — by SpaceX and others — as a threshold beyond which launch could cease to be the limiting cost factor for ambitious programmes.
TechCrunch’s framing presents the 1,800 as launches Starship has to make before orbital data centres get started. The paper presents it as the cumulative mass consistent with a price target under assumptions. Those are different framings of the same number. TechCrunch also notes, separately, that Google is a major investor in SpaceX. Not established, and therefore absent from this analysis: whether launch prices fall at all, the timing of any fall, the cost of compute hardware, any total system cost, any independent review, and anything in the Joule version that differs from arXiv version 2.
Three Implications
POWER-TO-MASS IS THE REAL VARIABLE The paper’s own Table 1 shows the approximately comparable result holds for high power-to-mass satellites only. Any orbital compute play lives or dies on how much power it can deploy per kilogram launched — not on the launch count alone. Hardware design is not a footnote to launch economics; it is co-equal with it.
THE COMPARISON IS NARROW BY DESIGN Launched power cost versus terrestrial power spend is one input into a much larger system-cost equation. The paper excludes compute, networking, and maintenance explicitly. Anyone extending this analysis to a total-cost-of-ownership claim is adding assumptions the paper does not supply. That gap is where the real business model uncertainty lives.
LEARNING CURVES ARE SENSITIVE TO STARTING CONDITIONS The paper states that the calculation of required launched mass is highly sensitive to the chosen initial conditions. An 18 to 24 per cent range in learning-rate estimates is a wide band. At the low end of the rate and a 70 per cent lower launch volume, the model still reaches $300 per kilogram. The range of outcomes is wide, and the paper is transparent about that.
None of the above is investment advice. It reports what a research paper says under its own stated assumptions, and it makes no claim about whether launch prices will fall.
Business Engineer Framework
Product Overhang Doctrine
The Google paper is a textbook Product Overhang case: a capability that does not yet exist, mapped through a cost curve that would have to hold for it to become structurally relevant. Understanding where those inflection points sit — and what assumptions they carry — is the analytical skill that separates structural reads from headline reactions. The Map of AI framework gives you the full stack context to place this kind of infrastructure bet correctly.
Google’s paper does one thing carefully: it converts launch economics into a power-cost number and holds it against the terrestrial range. At $200 per kilogram and with the right satellite design, that number lands inside the terrestrial band — approximately comparable, not cheaper, and not total cost. The 1,800 Starship launches is a unit-conversion result inside a learning-curve conditional, and the paper is explicit about its own sensitivity and limits.
What it does not do — and does not claim to do — is prove that orbital data centres work as a business. That question is open, and the paper’s own hedges say so plainly.
Every figure and quotation above comes from the arXiv version 2 of Google’s paper, Towards a future space-based, highly scalable AI infrastructure system design (arXiv 2511.19468, last revised 17 June 2026), read in full by this publication. TechCrunch reported on 1 October 2026 that a peer-reviewed version was released for publication in Joule. This publication did not read the Joule version, which returned an access error, and nothing above says whether it differs.
The figure of roughly 1,800 Starship launches is the paper’s statement of what sustaining a roughly 20 per cent learning rate to reach under $200 per kilogram by about 2035 would require, under stated assumptions. It is not a number of launches that orbital data centres must wait for, and the paper calls the target ambitious.The comparison in the paper sets the launch cost of power-generating mass against terrestrial power spend.
It is not total system cost, and the paper says it does not constitute a full economic analysis. The paper’s word is approximately comparable; nothing above says orbit would be cheaper.The figure of 1,850, and the observation that two of the four satellites in Table 1 fall inside the paper’s quoted terrestrial range at $200 per kilogram while two sit above it, are this publication’s own arithmetic and are not statements made by the paper.
TechCrunch is cited for context only; its headline frames the 1,800 differently from the paper, and nothing above says it is wrong.Nothing above predicts that launch prices will fall or that orbital compute will be built, and nothing here is investment advice.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
