AI Ledger Hepburn Advisory Hepburn Advisory

Overview/Power Ledger

The questionCan the revenue curve be physically served?

The power question is physical capacity, not just dollars.

The Power Ledger tracks whether the grid, interconnection queue and named projects can support the compute demand implied by AI growth. The Overview spine shows a dollar layer; this page leads with GW capacity.

Physical GW demand vs committed supply to 2030 · dollar layer on the Overview spine

2030 capacity gap

157 GW

337 GW demand against 180 GW committed supply.

Demand by 2030

337 GW

The power AI will need by 2030, implied by the revenue growth path.

Supply by 2030

180 GW

What is committed and likely to be switched on by 2030.

Cost to fill the gap

$1.74T

What building the missing capacity would cost at today's build rates. Capital no one has committed yet.

Why this matters — the money and the chips can be lined up quickly; the electricity to run them cannot. Power stations and grid connections take years to build, so if the power arrives late, the billions already spent can't earn what they were meant to. The Compute Ledger shows how much computing has to be powered, which is what sets the electricity needed here.

AI data-centre capacity needs to grow to 337 GW by 2030 to match expected demand. Despite all the investment, we're still going to pull up short by 157 GW.

From AI revenue you can work out the computing it needs, then the chips, then the electricity those chips draw. Add up every announced project and grid-connection request and the pipeline reaches 180 GW by 2030 — but only 118 GW of that is likely to be switched on and drawing power; the rest is still being built. Even counting the whole pipeline, demand runs ahead by 157 GW (46.5%) with no committed source. The gap starts to hurt in 2029.

Energised (drawing power) Under construction Announced + queued Gap (beyond all pipeline) Demand (central)

The pale line inside the blue is energised capacity, the power drawing today. The amber bands above it are committed but not yet built (219 GW still in the pipeline by 2030). The red shading above the full pipeline is the gap: demand with no committed source at all.

Data refreshed 27/05/2026 · demand anchored on third-party 2026 baselines; supply built up from grid-connection queues and named projects.

How the gap manifests todaysix ways the shortfall shows up

The 2026 gap (~20 GW under demand) is real but not user-visible. It clears through six mechanisms, each with a leading indicator below.

1 Behind-the-meter premiums

Stargate Abilene, Susquehanna co-location, TMI restart, Lancium West Texas, xAI Memphis — only exist because grid can't deliver on the needed timeline. Hyperscalers pay 1.5–2× normal $/MWh to bypass the queue.

2 Frontier-model pace slowing

Inference (the highest-margin workload) gets prioritised. Training runs become a power-availability queueing problem. The lab count shrinks to those with secured power.

3 Geographic migration

Power-rich jurisdictions (US South, Texas, Iceland, UAE, Nordics, parts of Australia) win the build-out. Northeast US grid is saturated; Ireland and Singapore moratorium; Western Europe slow-walking. Sovereign-AI strategy is mostly a power-procurement strategy.

4 Revenue capped by power

The forward apps-revenue curve assumes the compute to serve it gets built. If only ~118 GW of supply lands by 2030, revenue cannot reach the implied $529B — it caps somewhere around $180–220B Base. The Bull case is physically infeasible at current build-pace.

5 Capital allocation distortion

GPUs purchased can't be energised. Idle compute CapEx hits hyperscaler depreciation P&L as stranded assets. The depreciation drag in 2027–29 is materially worse than current analyst models assume.

6 Political backlash on builds

Virginia residential bills are up because PJM is pricing for DC load. Texas legislators eyeing DC-specific tariffs. Ireland and Singapore moratorium. If voters see their bills rising because of AI training, jurisdictional hard-caps follow.

Capital, silicon, power — which is on track for 2030?

Three layers, each read as a % of what 2030 needs. Capital and silicon have paper gaps, but the largest hyperscalers' capex guidance and foundry expansion already fund the build that's planned. Electric power is the one that can't keep up. Its red segment is the binding constraint.

Built today Committed pipeline Remaining (on-trajectory) Remaining (gap)

Capital allocation has been the focus of the AI narrative for two years, and there is enough of it to fund the supply that's already planned. The catch is physical: even fully funded, that planned supply still falls short of 2030 demand, and the build-pace would have to roughly double to close the rest. As the next section prices out, building all the way to demand is a multi-trillion-dollar bill no one has committed to yet, so the apps-revenue path at the top of this page is physically capped until that changes.

What it would cost to close the gap 2B

The capacity gap is a capital question as much as a grid question. At the Ledger's realised build intensity — about $11B of capital per gigawatt of AI data-centre capacity, closing it is a multi-trillion-dollar commitment that no announced pipeline yet covers.

Cost to fill the gap

$1.74T

The capacity gap priced at the realised build intensity. Range $1.63T–$2.79T across the realised-capex bases.

Built + committed

$766B + $530B

Cumulative AI capex 2023–25 plus committed 2026–30, building toward the planned supply, not the gap.

Missing capital

$1.74T

Capital with no committed source, for the capacity beyond the planned supply.

Capital required to 2030 — serving full 2030 AI power demand needs about $3.74T of capital at this intensity. $766B is built and $530B committed; a further $704B is required just to finish the planned supply, and roughly $1.74T more to build the gap. Total uncommitted: about $2.44T.

Basis: capex-anchored intensity (realised Capital-Ledger spend and the planned forward trajectory), Tier 2B, published as a range, not a point. Priced instead at the marginal cost of newest-generation silicon, the bill rises several-fold; that basis is excluded from the headline because it rides the Ledger's published demand-method variance.

Where the AI is landing — and what it costs to power 2B

Two-thirds of identified AI data-centre power sits in the United States. The rest of the world is small but uneven, and the price of industrial electricity explains most of the build-pace gap. Cheap power attracts scale (US South, UAE, Iceland); expensive grids and political constraint cap it (Ireland, Singapore, UK).

Country	AI DC (GW, 2026)	+ Pipeline (GW)	Industrial ¢/kWh	Status	Source

GW figures are 2026 estimates of identified AI-attributable data-centre IT-load. They sum to ~78 GW, within the ±5 GW range on the page's 2026 anchor. Electricity prices are 2024–2025 large-industrial tariffs (Eurostat band IS5000 for EU; EIA industrial for US; comparable national series elsewhere) in US¢ equivalent. Confidence is Tier 2B for most rows: country totals are triangulated across IEA, BloombergNEF, SemiAnalysis and operator disclosures, but no single primary filing covers all markets.

What to watch

Closing the gap means committed supply catching the demand the revenue curve implies. Three levers decide whether capacity arrives in time. Each is shown against where it stands now.

1Grid build pace

Needs~2x rate

→

Now~1x rate

Build-pace must roughly double from today's trajectory to track the demand curve through 2030.

Watch: interconnection-queue clearance and named-project energisation dates each quarter

2Committed supply

Needs337 GW

→

Now180 GW

Committed and likely-energised supply must climb toward the 2030 demand line, not stall at today's pipeline.

Watch: under-construction capacity converting to energised each quarter

3Lead-time components

NeedsIn window

→

NowSlipping

Transformers, switchgear and turbines must ship inside the window the 2030 demand curve assumes.

Watch: transformer, switchgear and turbine lead times in operator disclosures

Detail & references

Supply pipeline detail queue · lead times · named projects — expand

Three views of the supply stack: where the grid queue sits today, how long each component takes to ship, and which builds are publicly named.

Interconnection queue by ISO 1B

Across seven US ISOs the interconnection queue totals ~1,084 GW (as of 2026-Q1), of which ~95 GW is data-centre-attributable. PJM leads. Full table loads below.

ISO	Total queue (GW)	Data-centre attributable (GW)	As of	Source

Lead times by component 2A

Long-lead components run 18–36 months: HV transformers 18–36, substation permitting 12–24, large-frame gas turbines 24–36. Full table loads below.

Component	Min (months)	Median	Max	Source

Announced projects 2A

Project	Operator	Planned (GW)	Route	Status	Online target	Tier

How the demand chain is computedtriangulation & anchor sources

Demand is triangulated across three independent methods — an anchored projection from third-party 2026 baselines (~76 GW), a silicon-shipments bottom-up, and operational tokens × FLOPs, with variance between methods published as a health metric. Supply is built up from interconnection-queue data and named projects, each row carrying its own provenance.

Full methodology, including the demand formula, anchor sources, efficiency-curve inputs, and triangulation health, lives on the methodology page:

Read the Power Ledger methodology →

Sourcesprimary filings, operator disclosures & measured estimates

Every number on this page traces back to a primary filing, operator disclosure, or third-party measured estimate.

Methodology caveat

Projections are directional arithmetic, not forecasts. Benchmarks are dated and sourced. Demand defaults to the anchored projection, cross-checked against two independent methods; every assumption is editable. Where a number appears on a Ledger page, that page is the canonical source. Full method on the Methodology page.