Alibaba promises 20GW of cloud by 2032 and names no grid

Alibaba intends to hold more than 20 gigawatts of cloud capacity by 2032, a figure released alongside a new AI chip and one that sizes a power portfolio more than a server fleet. Gigawatts are the unit utilities and interconnection queues use, and Eddie Wu, the chief executive who gave the number, conceded that Alibaba may not bring it online as fast as customers would take it.

The chip is the Zhenwu V900, built by T-Head, Alibaba's semiconductor subsidiary, and Wu says it delivers three times the performance of its predecessor, the M890, which the company released in May. It carries more memory and higher bandwidth and can be linked into arrays of up to 500,000 chips for training and running large models; an array that size is a campus-scale machine, which is where the chip announcement and the capacity announcement begin to describe the same project. Wu expects what he called "significant growth in annual AI chip shipment volumes," though none of the performance claims has been independently verified and the part is not due for mass production and commercial release until the first quarter of 2027.

For anyone underwriting the buildings, the three-times multiplier matters less than what Wu told analysts on Alibaba's August earnings call: profit and gross margin should rise as the company deploys more of its own silicon inside its data centers and those parts replace chips it would otherwise buy. That is the lever available to a cloud operator whose largest input is priced by somebody else. Owning the accelerator design moves the biggest line on the cost sheet from an expense a supplier sets to an asset the company depreciates on a schedule it controls. It also lets Alibaba specify the hall around its own part—power density, cooling, network fabric, the arrangement of the linked chips—and under that plan the proprietary parts account for an increasing proportion of the chips in Alibaba's own data centers.

The business that silicon feeds is substantial: AI and cloud products generated $7.139bn in the second quarter of 2026, up 45 percent year on year. On the same August call, Wu said he was confident of $15bn in external cloud revenue by 2030, at a gross margin of 20 percent, though the two figures measure different things—the quarterly number covers AI and cloud products as a whole, the 2030 target covers external cloud revenue, and Alibaba has not reconciled them.

Twenty percent is a modest payoff for a vertical-integration program, which implies the proprietary chip is buying supply security and scale more than pricing power—a defensible trade for a company whose own chief executive says demand outruns what the company can build.

The demand claim itself rests on Wu's characterization, because Alibaba has offered no independent measure of its AI demand or of the utilization of the capacity it already runs, and the performance and shipment numbers for the new part are the company's own.

The most consequential sentence in the disclosure is Wu's caveat, and he has now said a version of it twice: AI customer demand is "extremely robust," but supply chain constraints are likely to affect how quickly Alibaba brings the capacity online, because "the industry's mid-to-long-term demand far outpaces our supply capabilities." In August he struck the same note for analysts, pairing confidence in the AI and cloud business with a plan to ramp proprietary silicon.

This publication has argued that the binding constraint on data center delivery has moved past capital into supplier slots and raw materials, with staffing the last bottleneck before energization; Alibaba's statement is that argument arriving from the buy side of the table, unambiguous about what the company is short of: supply, in Wu's telling, with the gap stretching years.

A target measured in gigawatts is a claim on interconnection, and in this cycle grid permission is the asset that prices first. Alibaba has attached no site list, no utility, no interconnection position, no contract and no capital figure to the 20GW figure. A 20GW program announced without a cost beside it is a program whose returns have not been underwritten in public; what Alibaba has offered instead is a growth rate, a revenue target and a margin.

For the utilities, power developers and landlords who would serve a capacity number that large, the announcement reads as an expression of demand rather than a signed commitment. Nothing in the disclosure ties Alibaba to a site, to a schedule beyond the decade, or to a counterparty. That is the gap between a target and an offtake, the same gap that opens whenever a project milestone is announced with the price column left blank.

Alibaba has published a chip date and a power decade, and the two will be judged by different audiences: the three-times claim will get independent testing or it will not, and the first quarter of 2027 settles the shipment question. The 20 gigawatts will need a utility, a site and an energization schedule, and Alibaba has named none of them. The announcement that pairs the number with a substation is the one that belongs in an underwriting file.