The Energy Waitlist

The constraint everyone can see and no one can buy past

For most of the build-out, the story was about chips. Who had them, who could get them, what they cost. That story is over, and the people closest to the metal have moved on from it. The gating input is no longer the accelerator. It is the megawatt that powers it, and the megawatt runs on a clock that money does not control.

The shape of the problem is simple. GPUs are manufactured on a schedule measured in weeks and shipped in days. The electrical infrastructure that turns a building full of GPUs into a working data center -- the grid connection, the transformers, the switchgear, the generation -- runs on a schedule measured in years. When the slow input is also the one you cannot accelerate with capital, it sets the pace for everything. That is the situation the industry is in now, and it is the hinge of the fit argument: when the constraint stops being financial and becomes physical, buying more is no longer the move.

The grid queue is the first wall

The first wall is the interconnection queue, the line a project stands in to draw any utility power.

The line is long and getting longer. The US interconnection queue has grown to roughly 2,600 GW of proposed generation and storage waiting for grid access, more than double the country's entire existing operational capacity.¹ That backlog grew from about 1,400 GW in 2021 to over 2,000 GW by 2024, and the wait that comes with it now runs four to ten years for projects that need meaningful transmission work.²

For the markets that host the most capacity forecast for 2026, the numbers are grim. Grid interconnection wait times in Northern Virginia, Phoenix, and Dallas now run four to seven years.³ A campus that joins the Northern Virginia queue in mid-2026 cannot realistically expect utility power before 2030 to 2033, no matter how fast the building goes up or how much capital the operator commits, because the queue is set by the utility, not the developer.³ The demand side makes the congestion concrete: in ERCOT, 198 GW of large load applied for interconnection in the first quarter of 2026 alone, mostly data centers, which is roughly equal to the grid's entire existing peak load.⁴⁵ PJM, for its part, interconnected about 3 GW of capacity in all of 2025 while receiving 220 GW of new applications in a single 2026 cycle, a logjam, not a pipeline.⁶

This is why power, not land or capital, now dictates the pace and the location of development. A site can have land control, zoning, GPUs, and tenant demand and still fail, because the queue is the queue.⁷

The transformers are the second wall

Clearing the utility queue does not give you a working data center. It gives you the right to build the electrical plant that connects to it, and that plant has its own shortage.

The large power transformer is the chokepoint. Pre-pandemic lead times of seven to fourteen months have stretched well beyond 24 months in most major markets, with specialized units running 36 to 48 months.⁸ Wood Mackenzie's order-based survey puts standard power transformers at about 128 weeks and generator step-up transformers at roughly 144 weeks, with some orders extending to four years.⁹ The demand that caused it is structural, not a blip: generator step-up transformer demand grew 274 percent between 2019 and 2025, and roughly 80 percent of large power transformers used in the US are imported, which exposes the most critical component to global supply chains the buyer does not control.⁹¹⁰

Transformers are not the only item on the critical path. Switchgear, breakers, busway, and grid-tie batteries are all competing for the same constrained capacity, and a site can clear the utility queue and still slip if any one of them is out of sequence.³¹¹ The framing that lands hardest: electrical equipment is under 10 percent of total data center cost and 100 percent of the bottleneck.¹²

The generation is the third wall

Faced with multi-year queues, the largest operators have tried to go around the grid entirely, building their own generation on site. The escape route has the same problem.

Large-frame gas turbines from the three makers that supply them are effectively booked through 2028, and a wave of 2030 delivery slots sold through in early 2026.¹³¹⁴ GE Vernova described roughly a three-year lead time as of spring 2026, with about 10 GW of combined 2029 and 2030 capacity remaining; the backlog now covers four to five years of revenue.¹⁴¹⁵ Building a new gas plant has gotten both slower and more expensive: combined-cycle construction costs jumped 66 percent from under $1,500 per kilowatt in 2023 to about $2,157 in 2025, with build times running roughly a quarter longer, both driven by data center demand.¹⁶ The self-generation move is real -- xAI assembling a 1.9 GW standalone microgrid from heavy-duty turbines is the benchmark case -- but it trades one multi-year queue for another.¹³

So the three walls compound. Get through the interconnect queue in four to seven years, or build your own generation booked out to 2029, and you still wait two to four years for the transformers either path requires. There is no order of operations that makes the megawatt fast.

The tell: powered shells with nothing running

The cleanest evidence that power has become the binding constraint is the gap between what has been announced and what is actually energized.

Sightline Climate tracked about 12 GW of 2026 US data center capacity announced across roughly 140 projects. Only about 5 GW is under construction. Roughly 11 GW sits in the announced stage with no physical progress, despite typical build times of twelve to eighteen months, and a quarter of those projects have disclosed no power strategy at all.¹² NVIDIA is shipping. The chips are available. The gating constraint is the electrical plant, which is why the announced pipeline converts into a late-decade energization curve rather than a 2026 event.³¹²

The same pressure shows up inside the hyperscalers' own numbers. Microsoft has reported that its data center crunch will persist, with new Azure subscriptions restricted in crucial hubs including Northern Virginia and Texas, because in those regions it lacks the power to bring capacity online, not the demand to fill it.¹⁷ When the company with the deepest capital position in the industry is gating its own cloud on power, the constraint is not financial.

Why this is the fit argument, not a detour

Lay the walls end to end and a single conclusion falls out. The next megawatt of new capacity is, realistically, two to seven years away, and no amount of spending compresses that. Capital can win an auction for an existing slot or a turbine reservation. It cannot manufacture an energization date.

That leaves exactly one source of capacity that is available now, on the hardware already powered, cooled, and racked: the capacity that is sitting idle inside it. Average GPU utilization across tens of thousands of production clusters sits near 5 percent, which means organizations are running roughly twentyfold over-provisioned on assets they have already paid to power.¹⁸¹⁹ In a world where the next megawatt is years out, recovered utilization is not an optimization. It is the only capacity that arrives on demand. Every point of utilization recovered from already-powered hardware is a megawatt the buyer does not have to wait four years to energize.

This is the precise place the fit thesis and the physical constraint meet. The build-out has hit a wall that money cannot climb. The lever that remains is to extract more useful work from the capacity that is already on the floor. That is what fit is, and it is why fit, not capacity, is the constraint that now decides who gets ahead.

A note on the numbers

The lead-time and queue figures here are reported with their source and current to Q1 to Q2 2026, a fast-moving area where the direction matters more than any single point. Where a range exists, it is reported as a range rather than collapsed to the most dramatic end. Interconnect waits, transformer lead times, and turbine backlogs are three different clocks measured by three different bodies and are kept distinct rather than summed. The announced-versus-constructed capacity gap is a single tracker's dataset and is labeled as such. As with the execution gap, the argument does not rest on any one number. It rests on the shape of all of them together: every path to a new megawatt runs in years, and only recovered utilization runs in weeks.

References

1. US interconnection queue at roughly 2,600 GW of proposed generation and storage, more than double existing operational capacity. EnkiAI summary of national queue data; Novogradac. EnkiAI Novogradac
2. Queue growth from ~1,400 GW (2021) to over 2,000 GW (2024) to ~2,600 GW (2025+); waits of four to ten years for transmission-dependent projects. EnkiAI; Lawrence Berkeley National Laboratory "Queued Up" dataset. EnkiAI LBNL
3. Grid interconnection waits of four to seven years in Northern Virginia, Phoenix, and Dallas; queue set by the utility, not the operator; energization curves slipping to 2030–2033. Sightline Climate data cited by Bloomberg (May 2026), via Tech-Insider. Tech-Insider
4. ERCOT large-load interconnection: 198 GW applied in Q1 2026, roughly equal to existing peak load; PJM facing a capacity shortfall that could reach 15 GW by 2030. Ascend Analytics (May 2026). Ascend Analytics
5. ERCOT large-load queue dominated by data centers (cited at ~87% of demand under review). EnkiAI. EnkiAI
6. PJM interconnected ~3 GW in all of 2025 against 220 GW of new applications in a single 2026 cycle; 38 GW of projects cancelled in 2025. EnkiAI summary of PJM/Avanza data. EnkiAI
7. Transformer and switchgear procurement as front-end go/no-go feasibility; power, not land or capital, dictating pace and location. Build.inc (2026). build.inc
8. Large power transformer lead times beyond 24 months in most markets, 36 to 48 months for specialized units, against a pre-pandemic 7 to 14 months; IEA noting lead times roughly doubling since 2021. PTR Inc. via CWIEME Berlin (Mar 2026). CWIEME Berlin
9. Standard power transformers averaging ~128 weeks and generator step-up transformers ~144 weeks, some orders to four years; GSU demand +274% (2019–2025); ~80% of US large power transformers imported. Wood Mackenzie Q2 2025 survey via IndustrialSage. IndustrialSage
10. Transformer lead times extending to four years on high-capacity units; demand outpacing manufacturing capacity. PwC and Wood Mackenzie via pv magazine USA (May 2026); Power Magazine. pv magazine USA Power Magazine
11. Switchgear, breakers, busway, and battery backup competing for the same constrained capacity; a site can clear the queue and still slip on internal distribution. Energy News Beat (Apr 2026); ResearchAndMarkets switchgear forecast. Energy News Beat
12. ~12 GW announced across ~140 projects, only ~5 GW under construction, ~11 GW with no physical progress; "electrical equipment is under 10% of cost and 100% of the bottleneck." Sightline Climate data via Tech Fund. Tech Fund
13. Large-frame turbines booked through 2028 with 2030 slots selling through; off-grid self-build as the default workaround, including xAI's 1.9 GW microgrid. Enverus (Apr 2026); Longbridge. Enverus Longbridge
14. GE Vernova at roughly three-year turbine lead time as of spring 2026, ~10 GW of 2029–2030 capacity remaining; gas capacity under construction topped 29 GW with 159 GW pre-construction. Tech Fund; OilPrice. Tech Fund OilPrice
15. Turbine OEM backlog coverage extending to 4.5–5.0+ years, slots filled into 2029–2030; OEM expansion plans constrained by upstream blade and casting capacity. Longbridge (Apr 2026). Longbridge
16. Combined-cycle gas plant construction cost up 66% (from <$1,500/kW in 2023 to ~$2,157/kW in 2025), build times ~23% longer, driven by data center demand. BloombergNEF via DCD (May 2026). DCD
17. Microsoft data center crunch persisting; new Azure subscriptions restricted in Northern Virginia and Texas for lack of power, not demand; off-grid self-generation as the near-term answer. Bloomberg (Oct 2025); Windows Forum analysis (Apr 2026). Bloomberg Windows Forum
18. Cast AI, 2026 State of Kubernetes Optimization Report. Average GPU utilization of 5% across tens of thousands of production clusters; roughly 20x over-allocation. Cast AI
19. Independent reporting on the Cast AI findings: ~5% average across ~23,000 clusters, roughly 20x over-allocation. ITBrief. ITBrief