Time-to-Energized: The Real AI Bottleneck Nobody Quotes

Everyone has a favorite AI bottleneck. For two years it was chips. Then it was megawatts — "warm shells," "it's the bottleneck," the quotes we opened the first piece in this series with. Both are real. Neither is the binding one. We source both ends of this market — the turbines and the GPUs and increasingly the harder middle, connecting a site to power — and from where we sit the constraint that actually decides who ships compute in 2027 isn't any noun. It's a clock. The thing that's scarce is time-to-energized: the number of months between deciding to build and having real load running. Chips you can buy. Megawatts, in the abstract, exist. What you cannot buy at any price is speed.

That's the whole post. The rest is showing you the two places the clock runs out — the grid and the turbine line — and why both roads dead-end at the same answer.

The grid can't hear you

Start with the obvious path: ask the utility for power. The honest version of what happens next is in one document, the Berkeley Lab Queued Up report, and it is brutal. As of the end of 2024 there was on the order of 2.3 terawatts of active generation and storage capacity sitting in U.S. grid interconnection queues — nearly twice the entire installed U.S. fleet, waiting in line to plug in (Berkeley Lab, Queued Up: 2025 Edition). That is not a backlog. That is a fleet's worth of power that has raised its hand and been told to wait.

And the wait is the number that matters. A typical large project that reached commercial operation recently spent somewhere north of four to five years in the queue — for projects over 200 MW, a median past 55 months (Berkeley Lab). Worse, getting in line is not the same as getting built: historically only about one in five projects that entered these queues ever reached commercial operation. So the real expected time-to-power through the front door isn't five years — it's five years times a coin flip you usually lose. You are not buying a delivery date. You are buying a lottery ticket with a five-year scratch-off.

For a data center, that's disqualifying. An AI buildout needs load in 2026 to 2028, not 2031. The demand is not theoretical: EPRI's latest read has U.S. data centers climbing from roughly 4–5% of national electricity today to somewhere between 9% and 17% by 2030 — and they raised that projection 60% over their prior estimate in a single year (EPRI, Powering Intelligence 2026). The load is coming whether or not the grid is ready, and the grid, on the queue evidence, is half a decade behind. The first thing time-to-energized tells you is that the utility is a slow lane, and for AI timelines, a closed one.

Even if it could, you can't buy the iron

So skip the utility. Build your own generation on-site, behind the meter, and energize on your own schedule. That's the move a quarter of new data-center capacity is expected to make by 2030 (per Bain, summarized here), and it's the right instinct. But it runs straight into the second clock, and this one's just as unforgiving.

The default answer for fast, firm, on-site power is a gas turbine. There is no fast gas turbine to be had. Three companies — GE Vernova, Siemens Energy, and Mitsubishi Power — make essentially all the large heavy-duty machines, and all three are sold out for years:

• GE Vernova expects to close 2025 with roughly an 80-gigawatt backlog that stretches into 2029, and its CEO has said publicly he expects the company to be sold out through 2030 by the end of next year (Utility Dive, 2025).
• Siemens Energy is sold out through 2028 even after announced capacity expansions, and is now booking slots for 2029 and beyond — a record order book past €150 billion (Enlit, 2025).
• Mitsubishi Power is telling customers that turbines ordered today won't arrive until 2028 to 2030 (Bloomberg, 2025).

The composite figure is the one to remember: the lead time for a new combined-cycle gas plant roughly jumped from about 3.5 years in 2023 to five years in 2025, with full builds running five to seven years end to end (BloombergNEF, via Bloomberg). Slot-reservation pricing now runs well above existing backlog levels — you pay a premium for a place in a line that still doesn't get you a machine before 2028. The only machines that skip that line are the ones already built — zero-hour surplus units in the mid-size class, like the ~62 MW Siemens SGT-800.

Notice what just happened. The "fast" alternative to a five-year interconnection queue is a five-year turbine wait. The two clocks converge on the same number. There is no front door and no obvious back door — both of the default paths to large, firm power put your energized date in the back half of the decade. That convergence is the actual story of this buildout, and it's the one the chip headlines and the megawatt headlines both miss, because neither of them is denominated in time.

The only fast lane left

If the queue is five years and the turbine is five years, the binding question becomes: what can you put steel on the ground behind in twelve to twenty-four months? That's a short list. It's behind-the-meter generation that isn't gas-turbine-gated — and it's geography-specific. You don't get to pick it from a catalog; the site picks it for you. On-site solar-plus-storage where the land and the sun cooperate. Reciprocating-engine packages where you can stomach the fuel logistics and the emissions profile. And — the one we spent the first post on — self-sustained geothermal in the right geology, where in places like Nevada's Great Basin you can put 10–20 MW behind the meter, bypass the utility under existing law, and never touch the queue at all.

We are not here to tell you geothermal beats gas everywhere — it doesn't, and pretending otherwise is how you lose credibility. Gas is denser, more dispatchable, and works anywhere you can get a pipeline; geothermal is cheaper to run and cleaner but only where the rock cooperates and only at a size that fits the resource. The honest framing is narrower and more useful: the winning generation type on any given site is whichever one gets you energized fastest for the megawatts you need — and right now, for a lot of 10-to-100-MW projects, the fastest option is whatever you can build behind your own meter without asking permission from a queue or a turbine OEM. Time-to-energized is the scoreboard. Everything else is a tiebreaker.

Why power and compute are one decision

Here's where this series closes the loop. The compute post argued that the GPU market's apparent glut is a generational rotation, not a collapse — that a perfectly-priced used-H100 fleet is sitting there for the buyer who reads the deal flow instead of the headlines. True. But re-read that sentence with this post's clock running: a perfectly-priced fleet of GPUs is dead weight without megawatts behind it to run it. You can win the compute decision and still lose the project, because the GPUs land in eight weeks and the power lands in 2030.

That's the trap we watch operators walk into. They run the two decisions on two different desks — someone sources the compute, someone else "handles power" — and the compute desk wins fast while the power desk discovers it's in a five-year queue. By the time the mismatch surfaces, the GPU contract is signed and the racks are stranded. Power and compute have the same answer key and almost never the same author. You do not own your compute roadmap unless you own your power timeline. That is the single sentence this whole series exists to land.

What to do now

If you're sizing a real project, three moves follow directly from the clock:

• Price your build in time-to-energized, not in dollars per megawatt. The cheapest power you can't energize until 2030 is more expensive than slightly pricier power running in 2027. Put a date on every option before you put a price on it.
• Assume the queue and the turbine line are both closed to you, and design from there. If a behind-the-meter path exists for your site — solar-plus-storage, recip engines, geothermal where the geology allows — that path is your real schedule. The grid is upside, not the plan.
• Solve power and compute on one desk, in one window. Don't let the fast half of the decision (GPUs) get signed before the slow half (megawatts) is sized. They're one procurement, not two.

We source both ends of this and we watch the deal flow that the index charts only summarize — but we don't set the market, we don't sell turbines off a shelf, and we don't own megawatts; nobody hands those out. What we can do is connect a serious project to the people who can put generation behind your meter and run power, cooling, and compute as one stack instead of three disconnected scrambles. If your grid answer is "2030," that's not the end of the conversation. It might be the beginning of a better one. Browse the stack or talk to us — and if you came in through the front of this series, the power and compute halves are where the specifics live.

---

Pantheon Research is our series on the infrastructure behind AI: power, turbines, cooling, compute, and the procurement reality that decides who actually ships. Field notes from the deal flow, not the keynote.