The AI build-out is quoted in announcements — “a 10-gigawatt campus,” “$750 billion of capex.” But almost none of that is compute yet. Every input a data center needs arrives on its own clock, and the slowest one sets the real date. The gap between announced and deployable is the whole story — and most coverage misses it. Hover or tap any underlined term.
When a hyperscaler announces a “5-gigawatt AI campus,” the number does a magic trick: it turns a construction project with a decade of dependencies into something that sounds like it’s already humming. It isn’t. A gigawatt announced is a gigawatt of intent — permits to file, equipment to order, a grid connection to wait in line for. The compute that actually earns revenue is years behind the press release.
The 2026 numbers make the gap concrete. Hyperscalers have guided to more than $750 billion in data-center capital spending this year. Yet of the roughly 12 GW of new US capacity planned to energize in 2026, only about a third is under active construction — and industry trackers estimate that 30–50% of announced builds are delayed or cancelled outright, choked by shortages of power equipment and parts. The announcements are running years ahead of the concrete.
Picture cooking dinner for 500 people, where every dish has to reach the table at the same minute. Your oven, your delivery truck, and your prep cook each run on a different schedule. It doesn’t matter how fast the salad is ready — if the main course is two hours out, nobody eats until then. The latest item sets dinner.
A data center is that dinner party. A world-class chef — the newest, fastest GPUs — standing in a kitchen with no power hookup serves exactly nothing. Engineers call the longest chain of must-happen-in-order steps the critical path. In the AI build-out, the loudest, most-tracked input — the chips — is almost never on it. The critical path runs through the quiet, unglamorous, long-lead items nobody live-tweets.
Here is the same build-out, sorted the way it actually behaves — by lead time and by how visible each input is. The pattern is the point: the things markets watch move fastest; the things that set the date move slowest and get the least attention.
| Input | Realistic lead time (2026) | How visible | Sets the date? |
|---|---|---|---|
| GPUs / accelerators | Weeks to months | Very loud — everyone tracks it | Rarely |
| HBM memory | Booked out — 2026 supply sold out | Semi-quiet | Sometimes |
| Advanced packaging | Quarters | Quiet | Sometimes |
| Liquid cooling / CDUs | Months to quarters | Quiet | Sometimes |
| Hyperscale construction | 1–2 years | Semi-quiet | Often |
| Switchgear | 1–2 years | Quiet | Often |
| Large power transformers | ~2.5 years average; up to 4–5 | Very quiet | Frequently |
| Grid interconnection | ~5 years typical; up to 7 in PJM | Very quiet | Usually THE gate |
| Permits / water / community | Variable — can halt a project | Quiet | Sometimes |
Read the bottom of that table slowly. A large transformer now averages roughly 128 weeks — about two and a half years — and high-capacity units can run four to five years, versus two before 2020. The interconnection queue holds more than 2,000 GW of projects nationwide, with typical waits near five years and the busiest grid regions pushing seven. And every AI accelerator needs its HBM, whose entire 2026 output was sold out — spoken for by the biggest buyers — before the year even began.
So the chip you read about is real, but it’s the salad. The dinner is set by a transformer being wound in a factory backlog and a utility study that started years ago.
Here’s the cruel arithmetic: because the inputs must all arrive together, a delay in any one of them delays the whole site. Speed up the GPUs and you’ve sped up nothing — they weren’t the gate. Slip the transformer by six months and the entire campus slips six months, GPUs sitting in a warehouse the whole time, capital already spent.
That is the synchronization tax, and it’s why the money and the reality fall out of step. The capital is committed up front — land, chips, contracts — but the revenue can’t start until the last, slowest piece lands. Spending happens on the announcement’s clock; earnings happen on the interconnection queue’s clock. Those two clocks are years apart.
If that pattern — heavy capital laid down long before the cash flow shows up — sounds familiar, it should. It’s the exact mechanism behind the “is this a bubble?” debate we covered in the Austrian lens on the build-out. Synchronization risk is how a boom can build real, useful things and still strand the capital that built them: the demand may be genuine, but if the compute lands three years late into a changed world, the investment can still fail to earn its keep.
Once you see the build-out as a critical path, the investable question stops being “who sells the bottleneck?” and becomes sharper: who shortens the longest pole? The value doesn’t just accrue to whoever makes the scarce part — it accrues to whoever makes the date arrive sooner. A few ways that happens:
That reframing is the difference between “power is the hot theme” and knowing which exposure actually gets paid. A company that sells a scarce component into a project that’s still four years from energizing has a great story and a long wait. A company that shortens the wait has pricing power now.
xAI’s Colossus supercomputer is the synchronization problem playing out live. Rather than wait years in the grid queue, xAI powered up by running roughly 69 mobile gas turbines at a temporary site in Southaven, Mississippi — behind-the-meter generation to get compute online now. It worked: the facility became valuable enough that Anthropic reportedly rents Colossus 1 for about $1.25 billion a month, and Google struck a multi-year deal worth roughly $30 billion for capacity there.
But the workaround hit the exact risk on the scorecard’s last row. The NAACP and the Southern Environmental Law Center sued in April 2026 over unpermitted turbine operation and air pollution; under the settlement the temporary turbines must come out by July 2027, replaced by a permanent 1.2 GW gas plant (41 turbines) running on a Clean Air Act permit. That’s the whole lesson in one site: behind-the-meter power can compress the critical path dramatically — and community and regulatory pushback can just as quickly threaten it.
Here’s the part you can use on the next headline you see. When a project is announced, don’t ask “how many gigawatts?” Ask which inputs are locked, which are pending, and which are unknown. The binding constraint — the true earliest date — is the longest-lead item that isn’t locked yet.
Count the pending and unknown rows, find the one with the longest lead time, and that — not the gigawatt number — is roughly when the site becomes real. Most splashy announcements have their power and transformers in the “pending” column. Now you know what the headline left out.
Because the real gates are quiet, the useful signals are quiet too. These are the gauges that tell you whether the announced-vs-deployable gap is closing or widening:
This lens cuts both ways, and that’s the point — it’s a framework, not a cheerleader. Synchronization risk is exactly why some of this capex will strand: build three years late into a world that changed, and even real demand may not save the return. It’s also why the firms that genuinely compress the critical path — who make the date arrive sooner — can capture outsized, durable value while everyone else waits in line.
What it is not is a reason to take any “5 GW announced” headline at face value. Nobody ships a data center. They assemble one, slowly, from a dozen supply chains that have to arrive together — and the slowest one, the one nobody’s tweeting about, decides when the lights actually turn on.
Dragonfly Lens covers the AI build-out as one connected system: power, memory, cooling, silicon — and the synchronization between them. Every claim sourced, every number dated.
Join the Lens →Sources & further reading (figures as of mid-2026 — verify current): Grid interconnection queue — Lawrence Berkeley National Laboratory, Queued Up (2025 edition), emp.lbl.gov/queues (2,000+ GW queued; ~5-year typical wait, up to ~7 in PJM). Power transformer lead times — Wood Mackenzie transformer survey via POWER Magazine and pv magazine USA (~128 weeks average; up to 4–5 years for high-capacity units). Data-center capex, construction share & delays — industry analyses incl. Build.inc ($750B+ 2026 capex; ~12 GW planned, ~1/3 under construction; 30–50% of sites at risk of delay/cancellation). HBM supply — TrendForce and supplier statements (2026 HBM capacity sold out across SK Hynix, Samsung, and Micron). Figures are point-in-time and move quickly; treat them as orders of magnitude, not fixed constants.
Educational research, not personalized investment advice. Dragonfly Lens is not a registered investment advisor. Nothing here is a recommendation to buy or sell any security. Past performance does not guarantee future results.