A Data Centre in Orbit Is One Rack: Audit the Nouns Before You Sign
AI infrastructure is being sold in words that carry orders of magnitude more implied capacity than the hardware behind them. The cheapest control in procurement is forcing every claim back into kilowatts.
A satellite that peaks at 150 kilowatts of compute is not a data centre. It is closer to a rack. SpaceX's disclosed AI1 specification puts peak compute at 150kW with 120kW sustained, roughly what one densely populated accelerator rack draws on the floor of a terrestrial facility that holds hundreds or thousands of them. The engineering is real, and so is the launch capability. The noun is doing work the hardware cannot do.
Be precise about the precision available here. Converting kilowatts to racks is an order-of-magnitude comparison, not an audited equivalence: rack density varies by design, generation and cooling method, and a 40kW air-cooled rack and a 130kW liquid-cooled one are both racks. The looseness is the point. Even the crudest conversion puts an orbital compute node and a hyperscale campus three or four orders of magnitude apart on installed load, and the same two words are being used for both.
How is data centre capacity measured, and why does the unit matter?
Four units survive contact with an engineer. Installed IT load in megawatts, which is what the utility contract and the cooling plant are sized against. Rack count at a stated density: physical footprint, and failure domains. Accelerator count and interconnect topology, which decides what size of model can be trained rather than merely served. Then delivered work, expressed as tokens per second at a stated latency and context length. Only that last unit can be fed straight into a business case.
Everything else is packaging. "AI factory" is a marketing term with no settled engineering definition; it can describe a 300MW campus or a shipping container. "Fleet" implies fungible reserved capacity when it often means a queue position. "Cluster" is the worst offender, because it sounds like a topology commitment while promising nothing about the network fabric that decides whether the machines can work as one system or only as many small ones. A buyer who evaluates these words in the vendor's vocabulary has already lost the negotiation, because the vendor picked the vocabulary.
So run the audit before the pricing conversation, not after.
It takes one email: restate every capacity claim in your proposal as installed IT load, rack count at density, accelerator model and count, and sustained tokens per second at our target latency. Vendors who can answer in an afternoon are selling infrastructure. Vendors who need three weeks and a call with their strategy team are selling a story. The signal is in the response time as much as in the numbers, and it costs nothing to collect. It is the same discipline as assessing readiness before anyone commits to a build: define the object, then price it.
Why has vendor pricing stopped carrying information?
Price used to be a proxy for cost, and cost a proxy for feasibility. Both links are weakening, for a reason familiar to anyone who watched the conglomerate era: cross-subsidy.
Two mechanisms are in play. A profitable incumbent can price inference far below what it costs to serve, funded from unrelated earnings, because the strategic aim is distribution rather than margin. And a company that bought accelerators for its own roadmap can resell idle capacity to rivals, booking the proceeds as AI infrastructure revenue. Expect more of the second. Compute is becoming something large firms trade sideways to each other rather than only a capital asset they consume, and a reseller's appetite to keep selling lasts exactly as long as the capacity stays idle.
For a buyer, the consequence is uncomfortable. A quoted price no longer tells you whether the supplier can serve you profitably, which means it no longer tells you whether the price will survive a strategy change upstream. So contracts need explicit repricing terms, defined exit paths and portability of workloads. The cheapest quote in the market is frequently the one most exposed to a decision nobody in your procurement team will be consulted on.
Narrative collapse is a delivery risk, not a shareholder problem
The audit earns its keep a second time here. When a supplier's funding depends on a valuation that cannot be reconstructed from its disclosed revenue lines, the gap is not a bull case you happen to disagree with. It is an unpriced dependency in your supply chain.
Scale does not settle the question either way. The Associated Press, citing SpaceX's filing, reports $18.7bn of total 2025 revenue and $4.4bn of Starlink operating income, and the company's SEC-hosted offering document specifies 555,555,555 shares at $135 each. Those are substantial, disclosed, checkable numbers. The question is not whether the business exists. It is what fraction of the market value rests on lines that do not yet appear in any filing, and whether your five-year contract depends on that fraction holding.
Founder rhetoric belongs in the same file. The Motley Fool records Musk saying on 17 July 2026: "I said SpaceX will be worth more than Earth if we achieve our goals. Obviously true." Read that as a statement about goals, which is what it says, and it is unremarkable. Read it as a valuation input and you have imported someone else's conditional into your risk register. The behaviour around the rhetoric matters more than the rhetoric: when a firm answers scrutiny with hostility rather than disclosure, that is a governance datum about the channel you will depend on when something breaks at three in the morning. Sentiment is not diligence. Units are.
Three questions hype fuses into one
Separate them and most of the confusion drains out. Is the technology real? Is the described deployment physically achievable at the stated scale? Is the vendor's business model solvent long enough to honour a multi-year contract? The first can be comfortably true while the second and third are false, and orbital compute is the cleanest current example: genuine hardware, real launches, and a capacity claim that survives only if you accept the noun without converting it.
Procurement teams keep discovering this after signature because they evaluate in the seller's language. Fixing that needs no new tooling and no consultants. Add a units column to the evaluation matrix, bar any capacity claim from the business case until it has been restated in kilowatts and delivered work, and walk when the restatement cannot be produced. The same discipline works upstream, in a strategy function that owns the architecture rather than inheriting it, and downstream, in a governance model that keeps humans in control of the decisions the system makes on your behalf.
Define the noun or the vendor will define it for you.
Questions people ask
What questions should I put in an AI data centre due diligence checklist?
Ask for installed IT load in megawatts, rack count at a stated power density, accelerator model and quantity, interconnect topology, and sustained tokens per second at your target latency and context length. Then ask what happens to each number under contention, and how quickly the vendor can answer. A supplier that cannot restate its own proposal in engineering units within a few days is describing an ambition rather than a facility.
How many servers are in a data centre rack, and why is that a better unit than 'cluster'?
It depends entirely on density: a rack might hold dozens of conventional servers or a handful of accelerator nodes drawing 40kW to 130kW or more, so the server count alone tells you little without the power figure attached. Rack count at a stated density is still a better unit than 'cluster' because it is physical and checkable, whereas 'cluster' promises nothing about the network fabric that determines whether the machines can act as one system.
How do I assess supplier risk when a vendor's valuation depends on future revenue lines?
Treat the unexplained portion of the valuation as an unpriced dependency and write contract terms that survive its repricing: defined exit paths, workload portability, escrowed configurations, and repricing clauses that trigger on ownership or strategy change. You are not making a call on the share price. You are ensuring that a funding-round outcome does not become an outage in your production environment.
Related
- The Memory Oligopoly Behind AI's Cost Inflation
- Least privilege for AI agents is a 1975 idea your pilot is skipping
- The AI Subsidy Cliff: Your Vendor's Investors Have Been Paying Your Bill
- AI & Automation
Written by an AI editorial persona of Abyshire's proprietary editorial system and reviewed by our team.