The bottleneck is moving from chips to power: SemiAnalysis's 30 gigawatts for 2027 meets the IEA's power-demand curve

In mid-2026, the binding constraint on the AI buildout is a fabrication plant. SemiAnalysis is explicit about the ranking: logic, the front-end chip manufacturing step, is the dominant bottleneck, ahead of memory, which is absorbing roughly 30 percent of hyperscaler capex, and ahead of power. I think that ranking has about eighteen months of shelf life left.

The arithmetic that ends it is already published. It just sits in two documents that rarely get read side by side.

Thirty gigawatts meets a doubling curve

Document one is SemiAnalysis's delivery forecast: roughly 30 gigawatts of new data-center capacity coming online in 2027, up from the 20 GW the firm counts for 2026. Document two is the IEA's 2026 Electricity report, which projects global data-center power demand roughly doubling from 485 terawatt-hours in 2025 to 950 TWh by 2030, with AI-focused data centers specifically growing threefold over the same window.

The IEA's regional split is where a developer should slow down. The United States adds 240 TWh between 2025 and 2030, a 130 percent increase. China adds 175 TWh, a 170 percent increase. Europe adds 45 TWh, a 70 percent increase. Every one of those terawatt-hours has to come out of a generation fleet and across a transmission system that exists today, or out of one that has not been permitted yet.

On the supply side, the same report projects renewables meeting roughly half of that growth, with natural gas adding roughly 175 TWh worth of generation to help cover the rest. Half committed to a technology whose projects still face their own siting fights. The remainder split between gas plants that mostly are not built and a residual the report does not assign.

Now hold the two documents together. The 2027 chip-side story is 30 GW of delivery, a 50 percent jump in a single year, produced by private supply chains spending their way out of a shortage. The 2027-through-2030 power-side story is a demand curve doubling against a generation plan that is half committed and half hopeful. One of those problems gets easier as capital arrives. The other one changes shape.

A fab problem and a grid problem do not clear the same way

This is the core of the argument, so I will make it plainly. A chip shortage is a factory problem. It clears on private timelines, driven by margins, and when the product finally exists it ships anywhere on earth in a cargo hold within days. Painful and expensive, but self-correcting.

A power shortage is a jurisdiction problem. New generation needs siting approval, and new transmission needs corridors plus years of environmental review. A new substation for a campus needs an interconnection study whose queue is measured in years, run by a utility answering to a commission, in a county where the neighbors get a microphone. Capital speeds up almost none of that. The developer who wires another billion dollars does not move up the queue.

An interconnection queue does not air-freight.

That asymmetry is why the constraint migrates. Fabs respond to price signals and are responding now; SemiAnalysis's own numbers, 20 GW to 30 GW in one year, show the delivery machine accelerating. The grid responds to dockets. By 2027-2028, on the IEA's curve, the marginal data center will be waiting on electrons, not silicon, and the marginal delay will be issued by a public process rather than a production line.

Read the queue the way you used to read allocation news

For two years, the sophisticated data-center underwriter tracked chip allocation the way a bond desk tracks the Fed. Who got the next tranche of accelerators decided which campus mattered. That habit is about to be obsolete, and the replacement habit is duller and more local: the serving utility's interconnection queue, its integrated resource plan, its pending rate cases. Those documents now tell you what allocation news used to, which is who actually gets to run workloads in 2028 and who owns an entitled shell with no feed.

The reading itself is not exotic. Rate cases are public dockets. Queue positions in most jurisdictions are published spreadsheets. The skill is not access; it is the habit of checking before the land contract gets signed instead of after, back when the answer could still change the price.

We have argued before that power is the second vote on any data-center site, and that you should order the transformer before you pick the architect. The IEA numbers turn that from house preference into arithmetic. A US developer is competing for a share of a 240 TWh expansion in a system that has to grow its data-center load 130 percent in five years. The winners of that competition are being decided now, in queue positions and signed interconnection agreements, years before any of the 2029 capacity shows up in a delivery count.

The site with power and permission is the scarce asset. The chips will find it.

Here is the falsifiable version. I expect that by the end of 2027, when a data-center project slips in public, the stated reason will be an energization date rather than chip supply in the majority of disclosed cases, and the 2028 trade press will cover utility commissions the way the 2026 trade press covers fabs. If 2028 opens with fabs still the excuse of record, I called the handoff a year early. Either way, the queue position you hold at the end of 2027 will be worth more than the GPU allocation you hold, and the teams reading rate cases today will be the ones explaining to everyone else how they got their 2029 megawatts.

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.