Wood Mackenzie puts the orbital premium at three times the cost of a ground facility and says reaching parity would need a 70% cost cut—but the real driver is a seven year US grid queue that makes the calculus different than it looks.
A 1-gigawatt orbital data centre would cost about US$170 billion to build, more than three times the price of an equivalent ground facility, according to energy researcher Wood Mackenzie. Yet the largest AI companies are publicly entertaining the idea anyway, because the bottleneck they would be escaping, a seven-year wait for a US grid connection, may be worse than the bill.
The trade-off lands in a new Wood Mackenzie analysis, distributed via GlobeNewswire, that tries to price orbital compute as if it were a serious answer to AI's power problem rather than a vision pitch. The firm's base case: a 1 GW facility in low-Earth orbit runs $170 billion, with launch and satellite hardware eating roughly 60% of that cost. A comparable terrestrial facility costs about a third as much.
For the orbital math to ever work, Wood Mackenzie says launch and satellite hardware costs would have to fall another 70%. That is a steeper drop than the 90% reduction reusable rockets have already delivered over the past decade, and it depends on a launch-cost trend that has held but is not guaranteed.
Why bother, then? Because the ground is not getting easier. Wood Mackenzie projects global data centre power demand at 460 terawatt-hours (TWh) in 2026, rising to 1,280 TWh by 2030 and 3,700 TWh by 2040, a 16% compound annual growth rate and a 703% increase from today's level. For scale, the 2026 figure alone is roughly half of Japan's entire annual electricity generation. The 2040 figure is roughly three times that.
The United States and China together account for 78% of the global planned data centre pipeline, according to the firm. In the US, grid-connection queues can run up to seven years, gas-turbine equipment is on waitlists stretching past 2030, and dry-region cooling has become a binding constraint. In that environment, the orbital pitch starts to look less like a vision statement and more like triage.
The demand curve is being pulled upward by what AI applications themselves are expected to do. Wood Mackenzie projects that next-generation AI agents will consume 10,000 to 40,000 times more compute per task than today's chatbots, a step-change that turns today's pilot-scale data centres into a rounding error on the 2040 number.
That math has pushed the largest technology companies to explore space-based data centres, per the firm's reporting. The most aggressive announced aspiration comes from SpaceX and xAI, which have floated "100 GW of orbital compute annually," a figure roughly ten times the combined announced pipeline of every other operator. Wood Mackenzie treats it as an aspiration, not a plan, and the gap between that headline and the rest of the industry's commitments is itself a credibility signal worth watching.
For now, the orbital thesis survives on two conditions: continued launch-cost reduction toward the 70%-further threshold Wood Mackenzie names, and engineering progress on the problems that do not appear on the launch line item, including beaming gigawatts of power to orbit without losing most of it on the way down and rejecting waste heat in vacuum where there is no air to blow across a heatsink.
The terrestrial pipeline still has to be built regardless. Orbital compute, if it ever pencils out, would add capacity on top of a grid expansion that is already behind.