SemiAnalysis says 20 gigawatts of data-center capacity came online in 2026, and model capability is still outrunning it

Twenty gigawatts. That is how much new data-center capacity came online in 2026, by the count of SemiAnalysis, the independent semiconductor and AI-infrastructure research firm led by Dylan Patel, and the same shop expects roughly 30 GW more in 2027. Not long ago a single gigawatt campus was a press-release event, the kind of announcement that put a governor behind a podium. The industry just energized twenty of those inside twelve months. And by SemiAnalysis's own read, it was not enough, because AI model capability is still expanding faster than the infrastructure buildout underneath it can keep pace.

That last clause is the whole story.

Everything below is the supporting evidence, and the single best piece of it is not a gigawatt figure at all. It is the price of an old chip.

The tell is the price of the old chips

Here is the data point I would put in front of an investment committee before any of the capacity numbers. SemiAnalysis observes that on-demand GPU rental prices keep rising, and that they are rising for Hopper-generation chips, silicon that is now nearly two generations old.

Capital equipment is supposed to depreciate. When a new generation ships, the old generation gets discounted, the way a three-year-old airframe or a last-cycle excavator gets discounted, because buyers with options pay up for the frontier and everyone else collects the remainder at a haircut. That is the normal physics of every equipment market I can name, and the GPU market is currently refusing to obey it.

Old chips do not usually get more expensive.

Translate this into real estate terms and the strangeness sharpens. Picture an aging Class B office tower raising rents in the same week two new Class A towers open across the street. In any balanced market that is impossible. It happens only when vacancy across the entire market is effectively zero, when every tenant who loses the bid for the new product immediately backfills the old product at whatever the ask happens to be that morning. Rising Hopper rental prices are exactly that signal for compute. There is no slack anywhere in the fleet. None at the frontier, none in the middle, none at the trailing edge that should by now be the bargain bin.

A market where the worst product on the shelf reprices upward is a market that has absorbed everything, including its own depreciation curve.

This is why "20 GW came online" reads as a huge number and still is not one. Twenty gigawatts landed and got absorbed, and the price of the least desirable compute in the market went up anyway. Delivery is not catching demand. Delivery is not even slowing the spread.

There is a second reading of the same signal, and it matters for anyone holding data-center dirt. A rising price on old compute is the least speculative demand evidence this market produces. Announcements can be vaporware and letters of intent can evaporate, but a customer paying more this month than last month to rent a two-generation-old chip is demand with a credit card attached. If your pro forma's lease-up assumptions ever get challenged in committee, this is the exhibit to bring: the market is currently paying a premium for the worst inventory it has. Tenancy risk on well-located, powered capacity is, for now, close to a theoretical concern. The risk lives everywhere upstream of the tenant.

Where SemiAnalysis says the constraint actually sits

The firm breaks the shortage into three layers. Logic, meaning front-end chip manufacturing capacity, is now the dominant constraint; the fabs cannot cut wafers as fast as the models want them. Memory is second, and it is no rounding error: roughly 30 percent of hyperscaler capex is now going to memory, a line item that reprices campus economics upstream of anything a developer controls. Power is the third.

For a development team, the ranking matters less than the clearing mechanism behind each layer. Logic and memory shortages clear inside private supply chains, on capital-investment timelines that are painful but self-correcting, because a fab owner staring at these prices has every incentive to build. Power clears through utility commissions and county hearing rooms, on public calendars with veto points that no amount of capex can compress. Chip prices are somebody else's problem to solve. The third constraint is the one that lands in your diligence file.

Thirty will not be enough either

The comfortable read on 2027 is that another 30 GW closes the gap. The Hopper price signal argues against comfort. If 20 GW landed in 2026 and the trailing edge of the compute market got more expensive while it happened, then absorption is running so far ahead of delivery that a 50 percent larger delivery year gets soaked up the same way, by workloads that were already queued and models that had already outgrown their training budgets before the concrete cured.

My bet, stated so it can be scored: on-demand Hopper rental pricing in mid-2027 sits at or above its mid-2026 level, even with roughly 30 gigawatts of new capacity landing in between. If Hopper pricing cracks before then, the market found slack somewhere I cannot currently see, and I will write that correction myself.

If you underwrite data centers, track the rental price of the oldest chip in the fleet, not the newest. The newest chip tells you what the frontier costs. The oldest one tells you whether the market has any give in it at all. Right now it has none, and 30 gigawatts of 2027 delivery is already spoken for by a demand curve that treats every fresh megawatt as back rent.

This analysis is a source-cited research summary drawn from public records, not legal advice. It can contain errors and should be verified independently before any investment decision.