The Power-Bound Era: Why the Bottleneck Moved to the Substation

Through the 2023–2024 accelerator shortage, the question that gated AI capacity was "how many accelerators can you get?" Leading-edge silicon was scarce, allocation was political, and the firm that secured the most chips built the most compute. The constraint has since moved down the stack, past the chip, the rack and the building, to the substation. What gates a frontier build now is how many megawatts you can energize, and when. That reorders everything downstream: siting, financing, procurement sequencing, even which companies can physically compete at the frontier.

The mechanics live in the chapters this one points to. What the strategist and the engineer share is the mental model: power is the bottleneck, the gigawatt campus is the unit, and time-to-megawatt governs the rest. The chapter traces the move from chip-bound to power-bound, sets out the global demand curve and the interconnection wall it meets, names long-lead electrical equipment as the gate, and ends on the gap between an announced gigawatt and one under construction against a real energization date.

The shift shows up first in how capacity is sourced: an AI program now has two independent gates, energized power and accelerator allocation. Two years ago the bottleneck was a queue at a single foundry. Today it is a queue at the utility, and the long-lead items (high-voltage transformers, medium-voltage switchgear, generator step-up units) now decide whether a site energizes on schedule. A program must secure both an energization date and accelerator allocation, and whichever arrives last sets delivery. In 2026 the chip gate tightened again on memory. Dell'Oro's September 16, 2026 release credits costlier memory and storage, which raised server prices, with part of a 92% jump in worldwide data-center capex in 2Q 2026. Micron says agreements already cover most of its 2027 HBM supply. Contract HBM and CoWoS allocation as a long-lead item beside the transformer (Chapter 16.2).

The economics explain the reorganization. In SemiAnalysis’s June 2026 scenario, each gigawatt of contracted critical IT earns $12–13B a year in gross AI-cloud contract revenue, so a 200 MW phase that enters service six months early pulls about $1.2–1.3B of gross revenue forward (Chapter 2.2 walks through the example and its caveats). Delay interconnection on a gigawatt campus by a year and the opportunity cost runs to billions of dollars. Meanwhile chips bought ahead of power sit idle, tying up capital as the next generation ships on its own timetable. When a delivered megawatt is worth that much, the whole program reorganizes around the one number that decides when megawatts arrive. Capital floods toward energized land, secured power, and platforms that bundle financing with grid-connected megawatts. Capital underwrites the deliverable megawatt.

The bottleneck moved partly because the unit of compute grew. A frontier training run no longer fits in a hall. It fits in a campus, and the campus is now measured in gigawatts, a unit borrowed from utility planning, not from IT. A single tightly coupled pre-training job wants to sit inside one power envelope to keep the back-end fabric short and synchronous. That envelope has grown from tens of megawatts (a large 2022 cluster) through hundreds of megawatts (a 2024–2025 build) toward the gigawatt and multi-gigawatt campuses that 2026–2027 frontier programs are siting against. Epoch AI estimates that the largest single-campus IT capacity has doubled roughly every ten months since mid-2024. The current record-holder is xAI’s Colossus 2, at an estimated ~950 MW of IT capacity, not consumption (Epoch AI, September 2026). At that scale the limit is whatever the local grid, the transformer market and the interconnection queue will let you energize, and none of them scale at the pace of a silicon roadmap.

Compute roadmaps are now downstream of grid physics. Rack power keeps climbing, from H100-class racks to the GB200 NVL72, which HPE rates at 132 kW nominal. The GB300 NVL72 is 135 kW TDP and up to 155 kW peak at Lenovo, and up to 142 kW in NVIDIA's reference architecture. For the Vera Rubin VR200 NVL72, OEM datasheets rate operating profiles at 188 kW Max Q / 228 kW Max P, and NVIDIA's DSX reference design plans Vera Rubin NVL72 cabinets at 330 kW TDP (the facility basis for power and cooling). Rubin Ultra-generation Kyber racks (NVL144) reach the ~600 kW roadmap point on 800 VDC; NVIDIA targets H2 2027, while SemiAnalysis reported 2028 in July 2026. A 600 kW rack is only compute if a substation can feed it. The density ramp that makes each rack more capable also concentrates more of the bottleneck in the power chain: a hall that was power-bound at 40 kW/rack is far more power-bound at 132 kW, because the same floor area now demands three times the megawatts behind it. → density mechanics in Chapter 1.1; the subsystem roadmaps that drive the ramp in Chapter 16.2.

The demand side of the power-bound story is a wall of forecasts that broadly agree on direction even where they disagree on magnitude. Global data center electricity demand is on track to roughly double from ~485 TWh in 2025 to ~950 TWh by 2030, reaching about 3% of global electricity, with AI-specific load roughly tripling over the period (IEA, Key Questions on Energy and AI, Apr 2026). McKinsey's capacity lens lands at ~219 GW of global demand by 2030, ~70% of it AI (McKinsey, 2025). Goldman Sachs Research’s February 2026 update lifts global data center power demand growth to ~+220% by 2030 over 2023 levels, from +175% before, and puts about 60% of that growth in the US. These are demand forecasts, not committed builds. The error bars are wide and the AI share is contested, but every credible curve rises at a slope the grid has never had to absorb.

The US share is larger and rising faster. LBNL's June 2026 update puts US data centers at 192 TWh in 2024, 4.7% of the country's electricity, and revises earlier years slightly below its December 2024 report, which had 2023 at 176 TWh (4.4%). The update's reference case reaches 649 TWh by 2030, 11.8% of US electricity, inside a 521–843 TWh (9.5–15.3%) range. Its 2028 figure, 464 TWh, sits mid-range of the December 2024 report's 325–580 TWh (6.7–12%) projection. EPRI's 2026 scenarios reach 9–17% of US electricity by 2030. BloombergNEF's July 2026 outlook puts installed US data-center capacity at 118 GW in 2030 and 194 GW in 2035, 52% and 83% above its own December 2025 forecast. Over the same seven months BNEF counted another 101 GW of announced US projects in the pipeline. Both figures are forecasts and announcements, and the callout below separates them from delivered megawatts.

That demand meets a constrained supply system, but the national queue statistic is supply-side context, not a load queue. At the end of 2025, more than 2,060 GW of generation and storage was actively seeking grid connection in the US, over one and a half times the country's entire installed capacity and down about 10% from ~2,290 GW a year earlier. LBNL attributes the drop to high withdrawal rates alongside relatively fewer new requests, and says it is too early to measure the reforms' full effect, so the data does not rule out softer demand (LBNL, Queued Up 2026 edition). For the generation and storage projects built in 2025, LBNL reports a median of 61 months from interconnection request to commercial operation, against 22 months for projects built in 2008 (regions with data). That is not a large-load service statistic. A data center's energization date comes from its own utility's tariff, studies, upgrades, agreements and milestones. CenterPoint shows the scale on the load side. Over summer 2024 its data-center interconnection requests climbed roughly eightfold, from about 1 GW to about 8 GW. In July 2026 it reported more than 17 GW of large-load projects submitted to ERCOT’s Batch Zero, with about 14 GW expected to be eligible. That compares with a peak of about 21 GW on CenterPoint’s Houston Electric system (CenterPoint, July 28, 2026). Being eligible is not a service date: ERCOT assigns studied load later and may grant less. The governor froze Batch Zero on August 3, 2026. ERCOT issued conditional classifications on September 3, which are not permission to energize: new data centers of 75 MW or more stay unenergized while the audit runs, with its reports targeted for December 10. On September 21 the governor also told TCEQ to halt data-center permits until the audit is done. The grid cannot say yes to load at the speed AI wants to consume it. → the interconnection mechanics, ISO/RTO pathways, and speed-to-power options live in Chapter 3.2; flexible/curtailable interconnection as a release valve in Chapter 15.8.

Behind the abstract "interconnection wall" sits a concrete supply chain, and a handful of components do the actual gating. Once accelerator allocation is secured, compare transformer delivery with the interconnection and construction milestones; whichever arrives last sets the site's energization date. Wood Mackenzie reported US substation-transformer lead times above 160 weeks in May 2026, and individual large or custom units can be quoted longer. Generator step-up units are a separate category: SemiAnalysis reported 3–4 years for US GSUs on June 25, 2026. Medium-voltage switchgear, large breakers and static VAR compensators queue behind the same constrained global manufacturing base. When contracted transformer delivery trails accelerator availability, the electrical layer sets the whole critical path, and ordering long-lead power equipment before the design is frozen is a rational hedge, not premature commitment.

In the chip-bound era you secured allocation first and worried about power later. The power-bound era inverts the order: you reserve the interconnection slot and order the transformers first, then fit the GPU generation to the power that equipment will deliver on its arrival date. Manage the long-lead electrical bill of materials against the site schedule's critical path. Operators who treated it as a back-office facilities concern are discovering that a GPU server is useless without a substation transformer quoted at more than three years (Wood Mackenzie, 2026). → the large-load service and speed-to-power overview in Chapter 3.2; long-lead procurement in Chapter 2.3; on-site substation and MV distribution in Chapter 4.2; energy-supply strategy and PPAs in Chapter 3.4; on-site generation commissioning in Chapter 13.4.

None of the four levers is free, and the choice is rarely all-or-nothing. The usual answer is a stack: a behind-the-meter bridge that runs from its contracted first-power date until the utility's committed service date, a grid interconnection underneath it for long-run cheap power, flexible-load terms to widen the headroom the utility will grant, and an energized-land purchase where speed is existential. The announced ~90 GW of behind-the-meter generation (Cleanview, mid-2026) signals developers' intended route around a slow queue, not committed capital or operating generation. Each lever trades a different currency: the queue trades time for cost, BTM gas trades carbon and capex for time, flexibility trades uptime determinism for headroom, and buying energized land trades dollars for certainty. Which lever to pull depends on which currency you are short.

A fifth lever sits with the utility. In a 2026 estimate, the IEA found that grid-enhancing technologies could open existing networks to as much as 330 GW of new generators, storage and demand, with no reinforcement needed. The technologies in question are dynamic ratings, topology optimization and power-flow control. That is a system potential, not capacity on offer. Ask whether the host utility's study tested these measures before it hands your project a network upgrade (→ Chapter 3.2).

Keep four columns in the capacity tally. Announced: 100 MW. Financed: 40 MW. Energized in the first year: 20 MW. Productive service equivalent: 20 × 6/12 × 0.75 = 7.5 MW-years. Multiplying by annual hours would give service-capacity hours, not metered MWh. Under the six-month power slip, first-year productive service falls to zero, while funded equipment and its 25% residual stress remain exposed. In a full year at a 50% useful-service fraction, those 20 MW yield 10 MW-years, against 15 in the base case. Do not release the second 20 MW fit-out merely because the 100 MW announcement survived. Require executed power milestones, funded installation, an accepted workload/SLO trace and customer demand sufficient for the phase. Against the 12 MW-year hurdle, 20 × u ≥ 12 means u ≥ 60%. At 50%, defer; at 65%, with power and funding restored, 13 MW-years clears it. Re-price the residual and financing in Chapter 1.8 and Chapter 2.5; this tally controls phase release, not project valuation. → power-portfolio response in Chapter 16.5.

The most useful filter for reading the 2026 build-out is the difference between an announced gigawatt and one under construction. Sightline Climate’s February 24, 2026 snapshot put at least 16 GW slated for 2026 in its tracked cohort of large data centers and AI factories (>50 MW) announced since 2024. About 5 GW of that was under construction, roughly one-third. The public source does not restrict that 2026 subset to the US, so read it as one tracker’s dated large-project count, not a settled US market total. SemiAnalysis’s June 2026 report that the top two hyperscalers’ self-build construction alone exceeds 5 GW uses a different population. On Sightline's count, the other two-thirds sat in the announced stage with no visible construction progress, despite the typical 12–18-month build timelines cited in the snapshot, so it is exposed to delay, repricing, cancellation or indefinite postponement. The 'half of 2026 US capacity may slip or be canceled' headline (Bloomberg, 1 April 2026, on Sightline's estimates) is the reading SemiAnalysis's June rebuttal is aimed at.

Currence (formerly Sightline Climate) took a wider view in its Q3 2026 outlook, dated August 27, 2026. Its cohort is projects above 50 MW announced since 2024, 363 GW in all. It rates about 55% of the announced pipeline for 2026–2030 as unlikely to come online. This is a forecast drawn from a gated report, not an observation. The headline-grabbing announcements and the steel actually being erected are two very different datasets, and conflating them is how capital gets allocated against capacity that never materializes.

The gap exists because announcements are gated by capital and ambition, both abundant, while construction is gated by an energization date, which is scarce. A project can announce before financing closes, and it can begin construction before power delivery is secured. Sightline’s unbuilt two-thirds therefore cannot be assigned to power alone: permits, financing, customer commitments and equipment can each stop a project. The market has responded by discounting announced capacity and paying a premium for energized, contracted, grid-connected megawatts that can carry load on a date: the deliverable megawatt as the asset.

The shift from chip-bound to power-bound reorders the entire decision hierarchy, and it changes who wins:

• Siting for frontier training and batch capacity is now power-first; latency-bound online and edge inference stays latency-first. Speed-to-power sits at the top of the site-selection screen: a site whose energization date lands after the program needs the capacity, with no credible bridge, is a liability with a deed. → Chapter 3.1.
• Financing underwrites megawatts, not buildings. Capital consolidates around energized land, secured interconnection rights and long-term power offtake. The deliverable megawatt is the collateral. → the macro financing lens in Chapter 16.4.
• Procurement sequences backward from the energization date. You order long-lead transformers and reserve the queue slot before you freeze the chip generation, because the electrical layer is the critical path. → Chapter 3.2.
• Flexibility becomes a power-supply strategy, not just a reliability feature. Flexible-load studies model ~98 GW of aggregate US grid headroom that would otherwise require new generation (Duke Nicholas Institute, 2025). That is a national potential rather than a site's offer, and it trades uptime determinism plus curtailment cost for an earlier energization date. → Chapter 15.8.
• The megawatt now comes with grid-performance duties. Demand left the grid in Virginia in 2024 (about 1.5 GW) and in Ashburn in July 2026 (about 3.8 GW), as sites transferred to backup power during grid voltage sags. ERCOT applies sag ride-through rules to new large computational loads from August 2026 and SPP to requests from January 2026; PJM has proposed similar rules. NERC's computational-load standards passed final ballot in October 2026 but await Board adoption and FERC approval, and they set no ride-through curve. An OCP specification took effect on September 17, 2026. It is an industry specification, not a regulation. It asks computing load to ride through sags within the bands it defines, then to recover 90% of its pre-fault draw inside 2 seconds of voltage returning. Specify ride-through in the UPS and power-train design, and commission two behaviors: a clean transfer when a sag is deeper or longer than the load must ride through, and staying connected when it is not. → Chapter 4.10, Chapter 13.6, Chapter 16.3.
• The competitive moat shifts from chip access to power access. When everyone can eventually buy the chip, the durable advantage is the gigawatt you can energize first. That is why the build-out's structure, geography and supply endgame are the subject of the rest of Part 16.