The gigawatt is the new unit of account

2026-10-21

SpaceX reported third-quarter revenue of roughly fourteen and a half billion dollars, up 161% on the year and 87% on the quarter. The launch business was under two billion of that. Connectivity, which is Starlink, was just over five billion. The compute business, which the company now calls Super Intelligence, was seven and a half billion and growing at over 800%.

Three years ago this was a rocket company. The numbers now describe a power and compute utility with a rocket division attached.

Capacity as the headline

The company ended the quarter with about 1.7 gigawatts of nameplate compute and expects to be near three gigawatts by year-end, adding around two hundred thousand GPUs a month. For 2027 the stated target is a further ten gigawatts, almost all on Nvidia's next platform, with power and cooling infrastructure being built toward twenty. Fourth-quarter capital spending alone is guided at thirty-five to forty billion dollars.

The constraint, management says, is not chips. It is power, cooling and the parts that go into gas turbines, where lead times for blades and vanes run a year or more behind the rest of the machine. The company's answer is to cast those parts itself, using the foundry it built for rocket engines, and to prefabricate entire data centres as modules that can be dropped anywhere there is a grid connection.

Pricing per watt

Compute is now sold by the watt per year. The company says short-term contracts are pricing above fifty dollars a watt, that it expects a premium for next-generation capacity, and that it could rent out its capacity several times over if it accepted every request. Its own build cost is quoted in the mid thirties of billions per gigawatt, against fifty billion or more for others, with the saving coming from the building rather than the chips.

Put those two numbers together and the stated payback on new capacity is under a year. That claim, more than any revenue figure, is what the capital markets are being asked to underwrite.

Who pays

Ten gigawatts at the company's own cost is around three hundred and fifty billion dollars of capital spending in a single year. Management says it has several pools of debt available and is not planning to raise equity. Operating cash flow will carry more of the load as monetisation rises. The target the company has set itself is annualised revenue of five hundred billion to a trillion dollars by the end of 2027.

Whatever one thinks of the target, the financing need is not in dispute. It is the largest single-company capital programme we are aware of, and it will be funded in the debt markets, with lock-up expiries releasing more float into the equity market through December.

The rest of the company

Starlink added 1.6 million consumer subscribers to reach 13.6 million, with monthly revenue per user rising to around sixty-nine dollars as the mix shifted toward developed markets. Mobile service in the US is now gated on satellite deployment rather than spectrum, with launch expected in early 2028. The next Starship flight will attempt to catch both the booster and the upper stage, which would be the first full recovery.

Our read

We do not take views on listed securities. We do take views on how capital gets deployed, and this is the cleanest case study available. Capacity is measured in gigawatts, revenue in dollars per watt, and returns in months of payback. Every neocloud, every data-centre developer and every power project in Asia will be priced against those three numbers from now on. The companies that can build a gigawatt cheaply and fill it quickly will set the cost of capital for everyone else.