Electron Economics · Long-range Fundamentals Model
US Data Center Capacity Forecast to 2040
This is a transparent scenario model, not an extrapolation of analyst consensus. It combines a specified AI compute-growth trajectory with hardware-efficiency (Jevons-adjusted), PUE, non-AI baseline load, a grid supply ceiling, regulatory friction, a BTM bypass layer and renewal risk. Nuclear is a supply classification, not an additive driver. The formula structure and parameter names are disclosed. Outputs are directional — they test the implications of competing growth narratives rather than provide a project-level capacity forecast. The weights are published in full — FM40_WEIGHTS is a plain object in this page's source (it was base64-obfuscated until v1.9.1, which was security theatre and made the calibration un-diffable across versions). They are analytical judgement, not measured constants: treat the values as calibrated estimates, not observations. Read the flagship output as a range: ~408 GW central by 2040 within a ~303–494 GW scenario envelope (constrained–accelerated), wider still under parameter uncertainty — not a point estimate.
Capacity, defined once: every GW on this dashboard is maximum facility electrical demand (MW, facility-power basis — IT load × PUE) serving operational US data centres, including modelled behind-the-meter supply. It is not nameplate IT capacity and not generation nameplate. Where a figure uses a different basis (e.g. an external forecast), it is labelled inline.
⬤ FERC 2025 State of Markets — 50 GW primary anchor
◎ LBNL 2024 US DC Energy Usage Report
○ Goldman Sachs operational estimate (2027)
EE primary research: 5 Substack articles, May 2025 – May 2026
US 2025 (FERC)
50 GW
confirmed baseline
Base 2030
—
fundamentals model
Base 2035
—
fundamentals model
Base 2040
—
fundamentals model
Grid ex-nuclear 2040
—
grid / ex-nuclear
BTM Gas 2040
—
behind-the-meter
Nuclear 2040
—
Firm PPAs + probability-weighted
CAGR 2025–40
—
base scenario
AI share 2040
—
of IT load
2025 (FERC confirmed)
50 GW
baseline anchor
Base 2030
—
near-term signal
—
Base 2035
—
mid-term signal
—
Base 2040
—
central case · envelope ~303–494
—
CAGR 2025–40
—
base scenario
Scenario
US DC Installed Capacity Forecast · 2020–2040 · Fundamentals Model
Four scenarios driven by compute growth, hardware efficiency, grid buildout rate, and LLT friction. Calibrated weights: FERC/LBNL/Goldman historical anchors + EE primary research. Shaded band = Accelerated–Constrained range. Grid binds in selected years between 2029 and 2035 in the base scenario; demand-after-friction re-converges with deliverable capacity post-2036.
Efficiency Breakthrough reaches ~135 GW by 2040 because faster FLOP/watt gains reduce MW required per unit of compute — this is a low-MW scenario, not a high-demand scenario. Constrained reaches ~303 GW; the grid modifier applies only to annual buildout rates, not the existing installed base. Grid binds in selected years between 2029 and 2035 before demand-after-friction re-converges with deliverable capacity.
One framework for the whole model: not all demand becomes power, and not all deliverable capacity gets funded or built. Each stage applies a constraint the other tabs treat separately — LLT/grid, financeability, delivery attrition. The financeable and operational haircuts are analytical judgement, exposed here rather than buried.
Latent demand = raw net demand before friction. Energizable/deliverable = base model output after LLT friction + grid ceiling (the headline number). Financeable = less marginal BTM (BTM × 35%) that clears only under peak-demand assumptions (EE prime/marginal analysis). Operational = less ~8% EPC / permitting / community delivery attrition. The last two haircuts are illustrative.
Grid ex-nuclear vs BTM gas vs nuclear classification · Base scenario · 2025–2040
"Grid ex-nuclear" is a residual classification — total delivered capacity less BTM gas and nuclear-classified supply. Many nuclear PPAs are grid-delivered, so this is not a physical power-flow split. It is a supply-source classification: grid ex-nuclear (conventional utility supply), BTM gas (on-site generation bypassing grid interconnection, OEM-capped), and nuclear (contracted firm + probability-weighted advanced). All four scenarios shown.
Grid ex-nuclear = residual classification, not a physical power-flow split (many nuclear PPAs are grid-delivered). BTM = OEM-capped behind-the-meter gas generation bypassing the interconnection queue; cumulative annual additions capped by phased OEM ceiling. Efficiency Breakthrough shows lowest BTM because total demand is lower — the OEM cap is rarely binding when demand itself is the constraint.
The Three Narratives · Why ~303 GW · Why ~408 GW · Why ~494 GW
Bear / Constrained
303 GW
12.8% CAGR · 2025–2040
The grid builds slower than announced. LLT tightening spreads nationally. AI normalises: enterprise deployment proves slower than frontier hyperscaler demand suggested, and inference cost compression reduces per-workload power intensity faster than new workloads arrive. Renewal risk materialises at the 2030–2033 window.
Grid buildout rate −20% as OEM constraints persist and transmission permitting slows
LLT friction +25% — Dominion GS-5 model spreads; Georgia Power PLL retreats
Compute growth decelerates to mid-range near-term as inference commoditisation reduces marginal power per workload
Renewal discount 15% of AI load at 2030–2033 window vs 8% in base
Base
408 GW
15.0% CAGR · 2025–2040
AI demand is real but decelerates as the S-curve matures post-2033. The Jevons rebound mechanism means efficiency drives more compute adoption. Grid builds steadily but not faster — OEM constraints through 2027, expansion plateau post-2030. LLT friction is a national average across bimodal distribution; capital self-selects toward permissive jurisdictions. Compute transitions from frontier (1.62×/yr) to enterprise (1.15×/yr).
AI S-curve inflection already happened (2026.5 midpoint); decelerating from here
Grid builds at a steady near-term rate, decelerating mid-period — not accelerating
BTM gas adds 58.4 GW cumulative by 2040 (OEM-capped); 14.3% of total
Jevons rebound holds at the calibrated base through 2035, then softens as AI normalises to enterprise
Bull / Accelerated
494 GW
16.5% CAGR · 2025–2040
Commoditised inference at $0.001/query drives enterprise adoption that overwhelms Jevons dampening. Agentic AI and robotics create workload categories not visible in current pipeline data. Grid buildout accelerates under federal policy. LLT friction remains low as state competition limits tariff severity. Frontier scaling doesn't plateau through 2033.
Agentic AI creates always-on compute demand not in current training/inference split
Grid buildout rises to 0.30×/yr as Ratepayer Protection Pledge drives grid investment
Compute growth holds at the near-term base rate through 2033 — frontier scaling doesn't plateau
Investment Committee Questions · Direct Answers
Why 408 GW and not 250 GW?
Jevons rebound: low-end forecasts assume hardware efficiency directly reduces power demand. AI capacity grew 2.3×/yr despite 1.28×/yr chip gains — the historical data rejects the zero-rebound assumption those forecasts rely on. BTM gas layer: most sub-300 GW forecasts miss 58.4 GW of OEM-capped BTM by 2040. S-curve timing: 250 GW implicitly assumes the AI inflection is still ahead. The evidence says otherwise: hyperscaler capex doubled 2023→2025 ($200B→$400B+ combined), OpenAI token volumes grew ~4× in 2024, Google reported 8× increase in Gemini API calls Q4 2024→Q4 2025, and 451 Research confirmed US DC power demand grew 22% in 2025 alone — the fastest single-year rate since 2000. See S-Curve Evidence tab for the full dataset.
Why 408 GW and not 700 GW?
Grid physics: 53-month average interconnection queue nationally. You cannot build 700 GW without structural permitting reform or sustained BTM beyond OEM capacity. Compute deceleration: training is ~20% of AI power and grows at 4×/yr; inference is 80% and grows at 2–3×/yr. The power-weighted blend is 1.62×, not 4×. LLT friction suppresses ~14% of demand at 2040. Renewal risk embeds product-strategy optionality that credit models don't price — some discount is structural.
Which assumption is the model most sensitive to?
Near-term compute growth and the Jevons rebound. The sensitivity panel ranks them first: a ±25% move in either shifts 2040 output more than any other single parameter, because both scale the AI-load engine directly; grid buildout and LLT friction dominate the supply side. Note this is model sensitivity, not a historical error record — EE does not yet have enough preserved forecast vintages to say which parameter has been most wrong out-of-sample. That record starts accumulating now (see Calibration).
What would make you change the forecast next year?
↑ Upgrade triggers: inference below $0.001/query at scale before 2028 (agentic demand materialises faster); OEM BTM capacity expands beyond current ceiling. ↓ Downgrade triggers: PJM capacity auction clears below $200/MW-day in 2027 (grid less binding than modelled); Meta or OpenAI exercise lease renewal at significantly reduced MW at the 2030 window (raise renewal discount from 8% to 15%). v4 is triggered by whichever signal arrives first.
Model: EE Forecast · June 2026. Transparent 9-step formula (structure disclosed) · Weights disclosed in source (FM40_WEIGHTS, analytical judgement) · Calibration: 3/7 anchors within range; 4 scope-explained (LBNL/Goldman include BTM enterprise, model is grid-connected); 2025 is the calibration anchor and 2027–28 are forward benchmarks, not historical. For methodology walkthrough → . Source: electroneconomics.substack.com
Electron Economics · Model Drivers & Uncertainty
Drivers & Uncertainty
The machinery behind the headline: how gross compute, Jevons-adjusted efficiency, grid buildout and LLT friction combine; what the 2040 output decomposes into; which assumptions it is most sensitive to; and how wide the parameter uncertainty is. Moved off the landing page to keep the forecast view focused.
Grid Binding vs Demand-Led Years · Base Scenario · Why Grid Re-Converges
AI compute growth in the model decelerates from its near-term rate (1.62×) to 1.15× per year by 2040. This deceleration is an exogenous assumption of the compute-growth path — it is imposed directly in computeGrowthAt(), not derived from a separate workload model. The assumption is informed by four structural forces expected to slow AI demand growth in the 2033–2040 window (below). The training/inference split shown elsewhere is a post-hoc decomposition of the total; it does not itself drive the total.
Why we assume demand decelerates (these justify the exogenous compute-growth path; they are not separately modelled): (1) Adoption saturation. Front-loaded AI adoption implies peak demand growth in the 2026–2030 window, with growth slowing as the installed base of AI-enabled enterprises saturates. (2) Training→inference mix shift. Frontier training (4×/yr growth, highest FLOP intensity) is ~20% of AI power today and a declining share as inference serving scales — a rationale for assuming the power-weighted blend falls from 1.62× to 1.15× (the split does not mechanically produce this in the current model; that is a P1 rebuild). (3) Diminishing frontier gains. Physical limits on scaling laws (compute-optimal training, context-window saturation) suggest frontier model size growth slows post-2030. (4) Enterprise normalisation. Enterprise AI workloads (the 2033–2040 demand driver) have lower Jevons rebound than frontier AI — they are primarily workload displacement, not net-new demand creation. Green bars = demand-led years. Amber bars = grid-binding years. Blue line = compute growth rate (right axis).
Forecast Decomposition · 2040 vs 2025 · Illustrative Attribution
Approximate bridge from 2025 baseline to 2040 base forecast, decomposed by driver. Not strict additive attribution — components are directionally correct but not algebraically exhaustive. For the precise step-by-step logic, see the Methodology tab.
Illustrative bridge, not strict additive attribution. AI compute net of efficiency = gross compute growth minus hardware efficiency offset with Jevons rebound applied. PUE improvement offset (≈ −66 GW): PUE improving 1.56→1.18 saves approximately 66 GW on average across the build-out period — applied to the 2040 IT base (≈348 GW), averaged over the 2025–2040 ramp. Without PUE improvement, the 2040 forecast would be ≈474 GW. LLT friction suppresses ≈69 GW from the raw demand signal at 2040. Grid binds in selected years 2029–2035 before demand re-converges with deliverable capacity.
Driver Trajectories · Gross Compute vs Efficiency vs Net Demand
All three indexed to 2025 = 1.0×. The gap between gross compute growth and hardware efficiency is the net demand multiplier. Jevons rebound prevents efficiency from fully closing the gap.
Gross compute grows faster than hardware efficiency in all scenarios except Efficiency Breakthrough. The Jevons rebound parameter means the majority of each year's efficiency gain is consumed by new demand rather than reducing power draw — consistent with RAND RRA3572-1 finding of 2.3× installed capacity growth despite 1.28×/yr FLOP/watt gains.
Sensitivity Analysis · Impact on 2040 Forecast of ±Weight Variation
Each parameter varied independently ±15–30%, all others held at base. Bar width = total GW swing on 2040 forecast. Shows which weights matter most — and where the analytical disagreements about the future are highest-stakes.
AI compute growth rate is the dominant uncertainty. Hardware efficiency and Jevons rebound are second-tier. LLT friction matters less nationally (it redirects rather than destroys demand) and more at per-market level — see Power Intelligence Tracker for market-specific analysis. PUE trajectory is the most predictable parameter; physical limits are well-understood.
Workload Decomposition · Training vs Inference vs Non-AI · Base Scenario · 2025–2040
The industry debate that's embedded in the compute growth parameter, made explicit. Training (frontier model runs) declines from 20% of AI power to ~10% as inference serving scales. Inference becomes the dominant driver. Its lower assumed Jevons rebound is one rationale for the post-2033 deceleration we impose on the compute-growth path — it does not mechanically determine the aggregate trajectory in the current model.
Training share of AI DC power: 20% in 2025 → 10% by 2040. Calibration: frontier model training grows faster (2.5× gross demand) but is a declining fraction of total power as inference serving of deployed models scales. Non-AI baseline (traditional enterprise + cloud) grows at 5.5%/yr near-term, 2.2%/yr far — stays roughly flat in absolute GW as AI displaces growth. The scale factor applied to each workload reflects the grid ceiling constraint: in grid-binding years, all three workloads are proportionally compressed.
Each run samples the six analytical-judgement parameters from independent Gaussian distributions using a seeded PRNG (reproducible across sessions). These are illustrative uncertainty bands, not calibrated probabilities: parameters are sampled independently (correlations not yet modelled) and the distributions are analytical assumptions, not empirically fitted. The median run approximates — but is not forced to equal — the base deterministic output (~408 GW at 2040). The 10th and 90th percentile outcomes span the sampled space — wider than the scenario range because scenarios hold most parameters fixed while the simulation varies all six simultaneously. These are percentiles of EE-selected parameter distributions, not fitted probabilities.
10th–90th pct band
25th–75th pct band
Median run
Base (deterministic)
Percentile
2030
2035
2040
CAGR 25–40
Running simulations…
Sampled parameters (analytical judgement only — empirical anchors held fixed), expressed as ~1σ Gaussian standard deviations relative to the base value: compute_growth_near ±14%, jevons_rebound ±12%, llt_friction_per10pp ±20%, grid_buildout_near ±16%, btm_bypass_2035 ±25%, renewal_risk_2030 σ = 40% of (1−renewal). Sampled independently — correlations between parameters are not yet modelled, which understates some joint tails. FERC 50 GW baseline (grid-connected; S&P/451 Research reports 62 GW total installed as of Mar 2026 including BTM/enterprise scope), PUE trajectory, and efficiency_per_year chip roadmap are treated as empirical anchors and held at base values. Results cached after first render within a session.
Electron Economics · Model Methodology
How the Forecast Is Built
A worked example walking through all 9 steps of the fundamentals model for a single forecast year, using base scenario weights. Scenario cases (Accelerated, Constrained, Efficiency Breakthrough) apply multipliers to compute growth, efficiency, LLT friction, and grid buildout rate — the steps are identical, only the inputs differ. Select a year to see the actual numbers the model computes at each step.
Step through year:
Nine-Step Model Pipeline · 2030
Each box shows the value the model computes at that step for the selected year. Steps 5 and 7 are the supply-side constraints — in the base scenario, Step 5 becomes binding from 2029 onward.
Green boxes = demand-side drivers. Amber boxes = supply-side constraints. Arrow values carry forward to the next step. The final box is the forecast output for that year.
Step-by-Step Calculation · 2030
Exact formula inputs and outputs at each step. Weights shown as W labels; the values behind them are disclosed in FM40_WEIGHTS in the page source, and the formula structure is fully transparent.
Step 2 is the most counterintuitive and most challenged parameter in the model. Three independent calibration anchors across different compute eras all point to the same conclusion.
The naive assumption: if chips get 1.28× more efficient per year, power demand should fall 22% per year for the same compute output. Over 15 years, DC power should shrink ~97%. This is plainly wrong — actual capacity grew 24%/yr 2020–2025.
Jevons correction: In 1865, William Stanley Jevons showed that more efficient steam engines increased coal consumption because efficiency made coal cheaper, expanding its use. Computing shows the same pattern: cheaper FLOP per watt → more total FLOPs demanded → higher total power draw.
The formula in Step 2:
Loading…
Jevons rebound calibration — three anchors:
1. AI compute era (primary anchor). RAND RRA3572-1 (the primary calibration anchor): installed AI compute capacity grew 2.3×/yr 2020–2025 despite hardware efficiency improving at 1.28×/yr. Implied rebound = ~0.44 on a simple model; higher on the compounded Jevons formula accounting for demand-price elasticity. This is the closest direct calibration to the model's forecast period.
2. Cloud computing adoption (2010–2020). AWS/Azure/GCP compute capacity grew ~35%/yr 2012–2019. Over the same period, server efficiency (FLOP/watt) improved ~20%/yr per LBNL. If efficiency were capturing demand, total power should have grown ~15%/yr. Observed US DC power growth: ~4%/yr (LBNL 2016, 2024). Lower rebound than AI era — consistent with enterprise cloud being a substitution play (moving existing workloads) rather than net new demand creation. This is why the model's calibration is higher than the cloud-era level.
3. Hyperscaler capex precedent (2021–2026). The five major hyperscalers are on track for ~$725–780B of capex in 2026 (up from 2025's record $410B). Post-Q2 2026 figures, basis-labelled: Alphabet FY26 guide $195–205B; Microsoft ~$175B calendar-2026 guide; Amazon $173B trailing-twelve-month actual (no full-year guide issued); Meta FY26 guide $130–145B; Oracle ~$50B. Basis caveat (added Aug 2026): on its Jul 29 call Microsoft extended data-centre useful life from 15 to 25 years and reclassified some leases from finance to operating, lowering reported capex by ~$15B with no change to physical buildout — headline capex now understates real spend, so cross-company comparisons must be basis-adjusted. On a comparable basis the trajectory is unchanged, and AWS (+37% YoY) and Azure (+43% YoY) confirm the spend is converting to metered revenue. Over the same period, GPU efficiency improved ~2× per generation (H100→B200). If efficiency were dampening investment, capex should have grown at half the compute growth rate. Instead it accelerated. This is consistent with the Jevons mechanism at the capital-allocation level: cheaper compute per dollar has so far drawn more compute investment, not less.
The Jevons rebound coefficient is a judgement, not an identified elasticity. The historical evidence — AI-era capacity growing 2.3×/yr against 1.28×/yr efficiency gains — strongly rejects a zero-rebound assumption, but it does not, on its own, identify a coefficient of 0.72: too many variables move at once (compute price, model capability, demand elasticity, training/inference mix, capital and GPU availability, adoption, utilisation). EE therefore applies a high-rebound central case and treats the magnitude as analytical judgement rather than an empirically identified elasticity. The open question is whether AI-era rebound persists as workloads normalise toward enterprise (possibly lower, closer to the cloud-era level) — the primary uncertainty the scenario builder lets you test directly.
Parameter Classification · Empirical Anchors vs Analytical Judgement
The most important transparency disclosure in this model. Empirical anchors are grounded in filed data, peer-reviewed studies, or directly observable market outcomes — they can be falsified by a counter-citation. Analytical judgements are informed estimates where the right value is genuinely uncertain and reasonable practitioners will disagree.
Empirical Grounded in filed data, FERC filings, peer-reviewed studies, or directly observable outcomes. Held fixed in Monte Carlo.
Judgement Informed analytical estimate. Uncertain by construction. Varied in Monte Carlo and Scenario Builder.
Forecast Changelog · v1 → v2 → v3 · What Changed and Why
Every forecast should be accountable to its prior versions. This table shows what moved, by how much, and what data drove each revision. Investors who see a forecast evolving with stated reasons trust it more than one that appears immutable.
Calibration Sources · Electron Economics Primary Research
These pieces are load-bearing: each one supplies a parameter, a constraint or a scope rule used by the model. They are named 11+ times across this file and, until v1.9.1, none of them was linked. Where a published slug exists it is linked directly; where the source is a live internal tracker rather than a published piece, the link goes to the publication root.
The four multipliers each scenario applies on top of the base weights. This table is rendered directly from FM40_SCEN_MODS — the same object computeForecast2040() reads — so a published value cannot drift from the executed one. Modifiers scale compute growth, hardware efficiency, annual grid buildout rate and LLT friction; the nine steps are identical in every scenario, only these inputs differ.
Calibrated Weights vs v1 · What Changed and Why
Calibration changes from EE primary research and two new variables added since initial build.
Nuclear capacity tier framework · Source confidence by classification
The nuclear layer is a supply-source classification within total delivered capacity. Four tiers with different source confidence levels and inclusion rules.
Expected based on current contracting pace but not yet announced
Excluded from base · upside scenario only
Nuclear is a classification layer within total delivered capacity — not additive. Total capacity is computed first (demand × friction constraints), then split into grid ex-nuclear, BTM gas, and nuclear classified supply. The 24 GW base figure is the probability-weighted sum of Tier 1 + Tier 2 only. Upside if Tier 3 materialises: 35–40 GW. Per World Nuclear News, Meta's TerraPower and Oklo deals are development support agreements — different from the firm PPAs with Vistra and Constellation.
US policy validation (May–Jul 2026): DOE UPRISE initiative targets 2.5 GW of added nuclear capacity by 2027 and 5 GW by 2029, primarily through uprates and efficiency improvements at existing plants — directly supports the Tier 1 ramp the model assumes. Separately, the National Nuclear Security Administration selected Amentum (Jul 20 2026) to negotiate a 1 GW AI data centre co-location at the Savannah River Site in South Carolina — first federal-site nuclear DC project, no hyperscaler named yet. Classified as Tier 3 (government/speculative) — excluded from base. Source: DOE May 2026 · NNSA Jul 20 2026.
Formula disclosure: all 9 steps are public, and since v1.9.1 so are the weights — FM40_WEIGHTS is a plain object literal in the page source, along with the Step-9 scenario modifiers (FM40_SCEN_MODS). Current = post-Substack primary research calibration (Jun 2026). Contact via electroneconomics.substack.com for institutional licensing or weight rationale questions.
Electron Economics · Model Validation
Calibration & External Benchmark Comparison
Seven external anchors across 2014–2028. This is a calibration and benchmark comparison, not an out-of-sample backtest: the model is anchored to the FERC 2025 print (50 GW), the pre-2025 series is mechanically back-cast, and 2027–2028 are forward benchmarks the model has not yet been tested against. Three anchors fall within the external uncertainty range; four show a scope residual explained by the FERC grid-connected vs LBNL/Goldman behind-the-meter definitional gap. A genuine backtest — preserved forecast vintages compared against subsequently realised capacity — will be added as vintages accumulate.
The two most-cited 2025–26 anchors differ by ~12 GW. That gap is scope, not disagreement. EE anchors to the FERC grid-connected figure; the middle steps below are illustrative allocations of the definitional difference (they sum to the observed 12 GW), not separately measured quantities.
FERC 2025
50 GW
grid-connected, utility-served (EE anchor)
+
BTM / captive gen
+~4 GW
on-site generation outside the FERC queue
+
Enterprise / non-utility DCs
+~6 GW
counted by S&P/451, not in FERC scope
+
Jan–Mar 2026 adds
+~2 GW
later measurement date
=
S&P/451 Mar 2026
62 GW
total installed, all scopes
Both figures are correct on their own basis. EE reports on the FERC grid-connected basis throughout; the ~10–15% scope gap to total-installed sources (S&P/451, LBNL, Goldman) is applied consistently wherever those sources are compared. Middle allocations are EE estimates pending a line-item build. Source: FERC 2025 State of Markets; S&P Global / 451 Research (Jun 2026 webinar, 62,242 MW as of Mar 2026).
Model vs Historical Anchors · 2014–2028
Fitted model output vs independent external anchors. Grey bands = anchor uncertainty range. Orange dots = point estimates. Model uses dual-rate historical CAGR: 7% (2014–18) → 18% (2018–23) → 24% (2023–25) reflecting observed acceleration phases.
Sources: LBNL 2024 US DC Energy Usage Report (TWh converted to GW using utilization factors); FERC 2025 State of Markets (50 GW end-2025, 24% CAGR 2020–2025); S&P Global / 451 Research (Jun 2026 webinar): US installed capacity 62,242 MW as of March 2026 — 12 GW above the FERC 50 GW anchor, reflecting broader scope (includes BTM + enterprise DCs not in FERC grid-connection data); US projected to reach 151,734 MW (152 GW) by 2030, broadly consistent with EE base case; Goldman Sachs ~76 GW operational estimate (2027). Residual at 2018/2020: LBNL includes BTM enterprise load not in FERC grid-connection data. Residual at 2027: Goldman includes BTM enterprise (~15% of total); adding EE BTM bypass estimate (~4 GW) closes ~60% of the gap (69 GW vs 76 GW). Scope note: EE anchors to FERC grid-connected 50 GW; S&P/451 62 GW is total installed including BTM — gap is expected, not a contradiction.
CAGR Validation Against External Ranges
Anchor-by-Anchor Backtest Detail
Electron Economics · Analytical Transparency
Known Limitations & Open Questions
A model is only as useful as its acknowledged blind spots. The following are the highest-stakes analytical gaps in this framework. This section exists because practitioners need to know what to discount.
The single most important caveat: This model forecasts total delivered data-centre capacity, split into grid ex-nuclear, BTM gas, and nuclear classification (Firm PPAs + probability-weighted advanced nuclear). In the base case, raw 2040 demand is ~477 GW. LLT friction reduces deliverable demand to ~408 GW, broadly consistent with modeled output. Grid constraints bind in selected years between 2029 and 2035 but are not the dominant suppression mechanism by 2040 — LLT friction is.
Currently Unmodeled
SMR/nuclear BTM at scale: Kairos Power, X-energy, and Microsoft-Constellation could deliver 1–3 GW/yr US BTM nuclear by 2033–2035. Currently folded into btm_bypass at a lower rate than warranted if SMR timelines compress.
Per-market grid buildout rates: The model uses three national rate parameters. PJM territory is structurally different from ERCOT, WECC, and MISO. ISO-level rates would improve 2030–2035 accuracy materially.
Interconnection queue depth, and the Texas pause: PJM has 4+ year average queue times; ERCOT was 18–24 months, which is why the near-term grid buildout parameter leans on Texas absorbing load the eastern RTOs cannot. That assumption is now exposed. On Aug 3 2026 Governor Abbott directed the PUCT and ERCOT to audit every data centre in the interconnection queue before further approvals, and ERCOT paused the Batch Zero study process in response, covering a total request queue of roughly 474 GW of which about 90% is attributed to data centres, so the data-centre share is nearer 425 GW; applications are still accepted and what halted is advancement through the study. ERCOT expects the verification to run under nine months and the Apr 9 2027 Batch Zero moved to conditional classification (PUCT, 20 Aug 2026) and the 9 Apr 2027 deadline, while not formally extended, has been conceded unachievable by ERCOT with no replacement date. The pause therefore has NO known length, so the earlier reasoning for not re-running a scenario (that a sub-nine-month pause sat inside the model's timing tolerance) no longer holds and is withdrawn. A re-run is now warranted for 2030 and beyond. It becomes a downgrade trigger if the audit runs past the Batch Zero deadline or if a material share of queued Texas capacity is denied rather than delayed. grid_buildout_near is the parameter that would move.
Inference-as-a-service commoditization: If inference costs drop 10× by 2028 (consistent with Jevons), demand for inference compute could grow faster than compute_growth_mid captures. This is the primary upside risk to the accelerated scenario.
EPC delivery risk (EE CoVolt/EPC analysis, May 2026): The model assumes grid buildout rate reflects OEM turbine availability. It does not model EPC contractor delivery risk: fixed-price project overruns of 20–30% are a realistic central scenario in 2026 given steel/copper/labour cost moves, transformer lead times of 3–5 years, and switchgear at 18 months minimum. GE Vernova Gas Power backlog rose from 55 GW to 62 GW in Q3 2025 and, on the current print (Q2 2026, reported Jul 22), stands at 116 GW of equipment backlog plus slot reservation agreements — against a year-end 2026 target of ≥125 GW. Siemens Energy is at 87 GW on the comparable basis (Q3 FY2026), with Gas Services book-to-bill at 2.65 — demand running at 2.65× what it can produce. Two OEMs therefore carry roughly 200 GW of forward obligation against a combined production ramp measured in tens of GW/yr. This creates a second-order grid delivery friction layer that grid_buildout_near does not capture — effective deliverable grid capacity in 2029–2032 may be 5–10% lower than the model implies. External validation from two independent sources: Goldman Sachs (Aterio, May 2026) applies a 40–50% haircut to announced pipeline (~60% of next-year capacity materialises on time; ~50% of two-year-out). Sightline Climate (Jun 2026) tracked 12 GW of 2026 US DC capacity announced across 140 projects — only 5 GW is actually under construction, with 11 GW at announced stage and 25% of projects having disclosed no power strategy. Both sources converge on roughly the same 40–60% slippage rate, independently validating the model's grid buildout discount.
Community opposition (emerging Jun 2026): As of mid-2025, more than 36 data centre projects representing $162B in investment were blocked or significantly delayed by local opposition (Data Center Watch / Ropes & Gray 2026). State-level moratoria are a real risk in some markets. Air permitting for gas BTM projects is a major gating issue — local and state opposition to gas turbine installations is growing. This sits between LLT friction and EPC delivery risk in the constraint chain and is not yet parameterised in the model. Practical effect: could reduce effective BTM capacity in high-opposition markets (Virginia, California) and add 6–18 months to grid-connected DC project timelines.
PJM BYONG escape valve (EE PJM CIFP analysis Jan 2026): PJM's Bring Your Own New Generation pathway allows large loads to move faster through the interconnection queue by bringing incremental generation with them. This creates a partial bypass of the LLT friction the model applies to PJM territory. The national LLT friction coefficient does not distinguish BYONG-eligible developments. Practical effect: model may slightly overestimate LLT suppression in PJM territory post-mid-2026, when BYONG expedited pathways become operational — partially offsetting the grid constraint in the 2029–2033 grid-binding period. ERCOT counterpoint (PUCT Docket 59220, Jul 2026): the Commission affirmed full curtailability of co-located load uncapped by the BTM generator — grid-synchronous colocation is curtailable, reinforcing fully off-grid BTM as the firmer path.
Calibration Limitations
Jevons rebound: RAND confirms high rebound for AI generally. But enterprise AI (the 2030–2035 demand driver) may have lower rebound than frontier AI (the 2025–2029 driver). A two-speed Jevons would be analytically defensible.
Renewal risk quantification: The 8–12% probability weights are analytical judgement, not market-implied. If options pricing for DC lease renewals becomes available, this parameter should be re-anchored.
LLT friction is modelled as a uniform national coefficient (a single disclosed value, llt_friction_per10pp, applied per 10pp of coverage). EE's Tariff Lottery (May 2026) shows the actual distribution is bimodal: Georgia Power PLL (near-zero friction, 11 GW already contracted) vs Dominion GS-5 ($1.5M/MW collateral, 14-yr minimum-take). Capital self-selects toward low-friction jurisdictions — the effective national friction rate likely rises more slowly than raw LLT coverage expansion implies. Per-jurisdiction coefficients would diverge significantly from the national average in both directions.
Unconstrained Demand Signal vs Grid-Deliverable Capacity · 2025–2040
The raw demand signal at 2040 is ~477 GW. After LLT friction (~14%), demand-after-friction is ~408 GW ≈ the base model output. LLT friction is the dominant suppression mechanism at 2040; the grid ceiling binds in selected years 2029–2035 before demand re-converges with deliverable capacity post-2036.
Demand signal = computeNetDemandMW(yr) — raw demand before friction or grid. Demand-after-friction = demand signal × lltFrictionAt(yr). Grid-deliverable = actual model output after gridCeilingMW(). Grid modifier in constrained/accelerated scenarios applies to the annual buildout rate, not the existing stock. High Efficiency scenario: efficiency_per_year scaled 1.30×, which reduces net power demand per compute unit via Jevons — the output of ~135 GW is correct: fewer MW are needed for the same compute workload.
Prime vs Marginal Capacity · Quality Split of Base Forecast
Not all forecast MW are equally financeable. EE's OpenAI Cost of Capital (May 2026) and Who Wears the Risk (Apr 2026) identify a structural split: prime capacity (grid-connected, IG offtake, $0.06–0.08/kWh) attracts infrastructure capital at scale; marginal capacity (BTM gas, sub-IG, weak queue position) only clears under peak-demand assumptions. The model does not currently distinguish — this chart estimates the split using BTM bypass rates and queue conversion data. Q2 2026 sharpened this: the market repriced hyperscalers on spend-to-revenue legibility rather than spend size — Alphabet's first negative-FCF quarter since its 2004 IPO (−$5.9B) and Meta's ~91% YoY FCF drop (to $784M) show financeability, not demand, is the emerging constraint this prime/marginal split is meant to capture.
Prime MW — grid-connected, IG offtake, clean power profile
BTM gas MW — behind-the-meter gas, marginal financeability
Estimated unfinanceable at trough (marginal × cycle discount)
Prime MW = base forecast − BTM additions. BTM MW = phased OEM-capped BTM additions (EE Gas Turbine Reservations Dec 2025: GE booked through 2027, post-Greenville expansion to 20 GW/yr from 2027). Trough-unfinanceable = BTM MW × 35% (estimated share that only clears under peak-demand assumptions, per EE OpenAI analysis). This is illustrative — the model does not carry quality tiers. A production version would track IG-backed grid-connected MW separately from BTM gas MW. Source: EE primary research.
Model version: v3 (post-Substack primary research calibration, June 2026). For questions on methodology, weight rationale, or calibration: electroneconomics.substack.com. For institutional licensing, contact directly.
Electron Economics · Bottom-Up Cross-Check
Queue Signal vs ForecastQUARANTINED — DO NOT CITE
⚠ Methodological caveat — this tab is quarantined. The regional "queue" figures (935 GW raw) are drawn from FERC / Berkeley Lab interconnection-queue data. That data — Berkeley Lab's Queued Up series — is a dataset of generation and storage projects seeking transmission interconnection. It explicitly excludes load interconnection requests, distribution-connected projects, and behind-the-meter projects. Applying generation-project completion/attrition rates to it and reading the result as data-centre demand is a category error: these are supply-side (generator) queues, not a data-centre load pipeline. The tab is retained only as an illustration of grid supply-side congestion and is not a validated demand cross-check. A legitimate load-pipeline module requires a large-load dataset (customer, utility, requested load, study stage, signed ESA, deposits, energisation date, BTM vs grid, LLT status) with stage-specific probabilities — not generation-queue attrition. Do not cite these GW figures as data-centre demand. The legitimate dataset exists and points the same direction: ERCOT logged ~198 GW of large-load interconnection applications in Q1 2026 alone (≈ ERCOT's peak load), and PJM data-centre-zone waits run ~40 months (EE, Jul 2026). A rebuilt load-pipeline module should be built on that large-load data, not the generation queue.
The queue pipeline and the forecast model answer different questions. The queue figures here are a stock of generation/storage interconnection projects in-flight today (935 GW raw) — NOT a data-centre load pipeline (see caveat above). The forecast is a trajectory of annual installed data-centre capacity to 2040. Read this tab as grid supply-side congestion context, not as bottom-up demand validation. PJM CIFP note (Jan 2026): PJM's BYONG pathway allows large loads bringing their own generation to move faster through the queue — a partial escape valve from the LLT friction the model applies nationally. PJM also formalised curtailment as a contractual reality for connect-and-manage loads, meaning reliability is no longer binary for new large-load additions.
LLT friction: model parameter applied per ISO risk tier
Queue scenario
Queue scenario only — forecast model is always base
Forecast Trajectory vs Queue Pipeline · 2025–2040
Green line: EE base forecast (annual installed capacity). Amber band: queue realistic demand range (pessimistic → optimistic). The queue is a pipeline stock — it represents current demand intent, not a point-in-time annual figure. Where the forecast line crosses the queue base, the model implies installed capacity has caught up with the current pipeline.
Queue realistic demand = raw queue GW × ISO-specific conversion rate (base: 26% avg). The gap between raw queue (935 GW) and realistic demand (~239 GW base) reflects attrition — not disinterest. PJM alone cancelled 38 GW in 2025. The queue band does not grow over time in this chart — it represents the current pipeline stock at three conversion scenarios.
Queue Realistic Demand · Before and After LLT Friction
Each ISO's queue realistic GW (grey) vs after applying the model's LLT friction parameter by ISO risk tier. High-LLT ISOs (PJM, ISO-NE) see the largest discount. Low-LLT ISOs (ERCOT, WECC) see minimal drag.
LLT friction applied: High-risk ISO → lltFrictionAt(2035). Medium → avg(2030,2035). Low → lltFrictionAt(2030). The national LLT parameter in the forecast model averages across this distribution — ERCOT is structurally under-discounted; PJM may be under-discounted at the national level.
Processing Time vs Attrition Rate · Bubble = Realistic GW
Slower ISOs have higher attrition. Bubble size = realistic GW in current scenario. ERCOT (bottom-left) is the structural outlier: fastest processing, lowest attrition. PJM (top-right) is the constraint case: most demand, highest friction.
The negative correlation between processing time and conversion probability is the empirical basis for the model's ISO environment scoring dimension. The national grid buildout rate parameter (0.26/yr near-term) implicitly assumes a weighted average of these ISOs — ERCOT headroom offsets PJM constraint at the national level.
ISO-Level Queue Signal vs Forecast Parameters
Queue realistic demand by ISO cross-referenced with the forecast model's LLT friction and grid constraints. "% of 2030 forecast" shows each ISO's realistic demand as a share of the model's 2030 base output — a sense-check on whether the queue is consistent with the trajectory.
Interpretation: queue realistic GW is a pipeline stock (projects currently in process), not an annual flow. "% of 2030 forecast" uses the total realistic GW divided by the model's 2030 base output — it will exceed 100% because the queue represents multi-year pipeline, not just next-year additions. A ratio above 200% suggests the pipeline comfortably supports the trajectory; below 100% is a potential demand signal concern.
Key Analytical Tensions
Queue data: FERC interconnection queue filings, Berkeley Lab Queued Up 2024, FERC State of Markets 2026. ISO conversion rates: EE Queue Intelligence model (base). LLT friction: forecast model parameter applied per ISO risk tier (High/Med/Low). The model's national parameters (grid_buildout_near, llt_friction_per10pp) flatten the per-ISO distribution — this tab exists to surface that flattening. PJM CIFP (Jan 2026): PJM defined large load as ≥50 MW/POI; endorsed BYONG pathway; formalised curtailment for connect-and-manage loads. FERC June 18, 2026 — six tailored orders issued today: FERC ordered all regional grid operators to (1) justify within 60 days why existing tariff structures can accommodate large loads or immediately reform them, and (2) submit a mandatory reliability report within 30 days on generation capacity adequacy. Action taken under Section 206 of the Federal Power Act — targeted, region-specific orders maximising legal durability. If utilities reform tariffs under the 60-day mandate, the model's LLT friction parameter may loosen for 2027–2030 faster than the base case assumes. Source: AAF June 18, 2026; FERC Docket RM26-4-000. For the full queue scoring tool: electroneconomics.substack.com
Electron Economics · External Calibration
EE Forecast vs Analyst Consensus
How does the EE model compare to the major public forecasts? Key differences, where EE is higher or lower, and the methodological reasons behind each divergence.
The question every institutional reader asks first: "How different is your number, and why?" EE base 2030: 115 GW. EE base 2040: 408 GW. Where we diverge from consensus, it is because we model different mechanisms — not different assumptions about AI's importance.
EE EE base scenario · Jun 2026
Goldman Sachs · May 2026 · Aterio facility-level data
LBNL 2024 US DC Energy Usage Report
McKinsey / EPRI / IEA / BloombergNEF · 2025–2026
EE vs External Forecasts · US Data Center Capacity · 2025–2040
All figures converted to GW total installed capacity (IT load × PUE where source gives IT load only). EE base shown as solid green. External forecasts as coloured dots/lines. Where ranges are published, shown as error bars. Key divergence points annotated.
Sources and scope notes in table below. LBNL figures include BTM enterprise not in FERC grid-connection data — adds ~15% to comparable grid-connected figures. Goldman 2027 estimate (95 GW) is total installed; EE model implies ~79 GW grid-connected + BTM gas at same date (scope gap explained by BTM enterprise). McKinsey $6.7T investment figure converted to GW using the EE Capex Stack all-in rate of $32.75 per critical IT watt: $6.7T ÷ $32.75/W = ~205 GW of IT load, which on this model's facility basis is ~205 × PUE(2030) 1.35 ≈ 276 GW global, or ~124 GW at a US ~45% share. Superseded (v1.9 and earlier): ~185 GW global / ~83 GW US, which applied a mixed all-in rate and omitted the IT→facility PUE step, understating the capex-implied figure. This remains the crudest conversion on the chart — cumulative investment is not installed capacity, and McKinsey's $6.7T includes silicon and non-US spend.
All external forecasts converted to comparable basis where possible. Scope differences (BTM enterprise, international, crypto mining inclusion) noted per source. EE model is total delivered capacity (grid ex-nuclear + BTM gas + nuclear classification) — excludes enterprise BTM (generators, UPS not connected to utility grid). This produces a systematic ~10–15% gap vs LBNL/Goldman total-market figures. Scope note: Goldman Sachs (May 2026) forecasts power demand of 41 GW in 2026 and 66 GW in 2027, implying ~95 GW installed capacity by end-2027 at 70% utilisation — directionally consistent with EE trajectory. Goldman measures consumed power; EE measures installed capacity. Goldman also applies a 40–50% delivery-risk haircut to the forward pipeline. Sources: Goldman Sachs Aterio analysis May 2026; LBNL 2024 US DC Energy Usage Report; McKinsey Global Institute 2025; EPRI Powering Intelligence 2024; IEA Electricity 2025; BloombergNEF Dec 2025.
Electron Economics · Infrastructure Translation
Implied Generation Requirements
Translating the capacity forecast into what must actually be built. Gas, solar, storage, and transmission requirements by scenario — the strategic value layer above the headline GW number.
The forecast number alone has limited capital allocation value.408 GW of US DC capacity by 2040 (base case, read live from the model — this line previously read a stale 411 GW) implies specific generation, transmission and storage requirements — and specific bottlenecks. This tab makes those explicit.
Generation mix: EE analysis · current ISO dispatch data
Capex: EE Capex Stack v20 · $32.75/W per critical IT watt · 2026 dollars
Transmission: FERC Order 1000 · typical 40% of DC load
Scenario
Capex basis. $32.75/W is per critical IT watt (EE Capex Stack v20 — IT $17.50/W + powered shell $4.50 + physical shell $1.25 + gap/overhead $9.50). Every GW figure in this forecast is facility capacity: the demand chain ends in × PUE, and the 50 GW FERC 2025 anchor is facility load. Facility capacity is therefore converted to IT load at the model's own PUE before costing (408.5 GW facility → 346.2 GW IT at the 2040 PUE of 1.18). Figures are 2026 dollars — no construction escalation applied; the Capex Stack's observed build-cost escalation is 5.5%/yr (7.0% CAGR 2020–25), so a nominal-dollar 2040 stock would be materially larger. Correction, v1.9.1: prior versions multiplied facility MW by the IT-watt rate directly, overstating 2040 base-case DC capex by the PUE factor — $13,377B superseded by $11,336B.
Required Generation by Fuel Type · Base · 2025–2040
Incremental generation needed to serve DC load growth. Mix based on current ISO dispatch trends and announced clean energy commitments from hyperscalers. Gas is firm capacity (always-on); solar + storage is contracted clean energy.
Generation requirements = DC capacity × 1.6 — a simplified planning multiplier combining load factor and reserve margin. It is NOT resource-specific accredited capacity (ELCC); gas, solar, wind, storage and nuclear are not interchangeable MW-for-MW. A resource-adequacy build is noted in Limitations. Mix assumptions: natural gas 38% (firm baseload + peaking), solar 32% (contracted hyperscaler PPAs), storage 12% (co-located with solar for 24/7 matching), nuclear 8% (existing + announced SMR pipeline), wind 7%, other 3%. These ratios reflect announced hyperscaler procurement mix 2024–2026, not current grid average.
Capex & Infrastructure Requirements · Base · Cumulative to 2040
Gross value of DC + generation + transmission infrastructure implied by the forecast (full 2040 stock, not incremental spend). DC capex at $32.75/W of critical IT load (EE Capex Stack v20), applied to facility GW ÷ PUE(yr) — ≈ $27.8M per facility MW at the 2040 PUE of 1.18. Generation at blended ≈ $2.31M/MW (fuel-mix weighted). Transmission at $1.8M/MW-equivalent.
DC capex = EE Capex Stack v20 all-in rate of $32.75 per critical IT watt, converted to facility basis by dividing by the model's own PUE for the year (1.18 at 2040) → ≈ $27.8M per facility MW. Generation capex = blended across the fuel mix ($2.5M/MW gas, $1.2M/MW solar, $1.6M/MW wind, $1.4M/MW storage, $8M/MW nuclear) ≈ $2.31M/MW of generation; at 1.6 MW of generation per MW of DC capacity that is ≈ $3.7M/MW of DC. Transmission = FERC Order 1000 cost-sharing at 40% of DC load × $1.8M/MW ≈ $0.7M/MW. All-in per MW of facility DC capacity ≈ $27.8M + $3.7M + $0.7M ≈ $32M/MW at 2040. (Superseded: v1.9 and earlier stated ≈ $37M/MW, which mis-applied the $32.75 IT-watt rate directly to facility MW; an even earlier draft cited $80–95M/MW, not produced by this cost stack. Both retained here per the versioning policy. KPI totals below are gross value of the full 2040 stock, not incremental forecast-period investment.)
Annual Build Rate Required · Base · GW per Year by Type
Annual installation requirements to stay on trajectory. Peak build years 2028–2032 require ~30–40 GW/yr of combined generation — roughly 3–4× current US annual solar installation rates. Gas turbine OEM constraints (GE Vernova, Siemens) are binding through 2027.
Annual requirements derived from year-over-year capacity additions in the forecast. OEM bottleneck annotation: GE Vernova and Siemens effectively sold out through 2027 at current production rates. Solar installation: US installed ~50 GW in 2025; DC-serving solar requirements peak at ~18 GW/yr in 2029–2031 — significant but within 35% of total US solar industry capacity. Storage is the most under-supplied component relative to announced hyperscaler requirements.
Workload Mix · Training vs Inference vs Non-AI · Base
Where the GW is actually going. Training (frontier model runs) declines from 20% → 10% of AI power as inference serving scales. Inference becomes the structural driver. Non-AI stays roughly flat in absolute GW — it's being displaced as AI absorbs all capacity growth.
Training share: 20% of AI DC power in 2025, declining to ~10% by 2040 as inference serving of deployed models scales. Non-AI (traditional enterprise + cloud) grows at 5.5%/yr to 2030, 2.2%/yr thereafter — roughly flat in absolute GW as AI absorbs all growth. The split does not change the model's total GW output — it decomposes the AI portion of the existing forecast. Training power grows in absolute terms despite falling share because total AI power grows faster.
Generation requirements are derived mechanically from the capacity forecast — they represent what must be built to serve the implied DC load, not a separate generation forecast. Mix ratios are based on announced hyperscaler procurement patterns (2024–2026), which skew cleaner than current grid average. Actual mix will depend on policy, geography, and tariff evolution. For ISO-level generation requirements, see Queue Signal tab.
Electron Economics · Analytical Position
Why We Are Above Consensus
A cumulative bridge from comparable-scope consensus (~80 GW, 2030) to EE base (~115 GW), turning on one EE assumption at a time in a stated order: base/scope/LLT differences, then the Jevons rebound, then the BTM layer. Each step is computed from the model. The S-curve is not a bridge step — it is evidence for the front-loaded compute-growth path, not a forecast input. Because the parameters interact, the split depends on ordering.
The single question every reader asks: if I believe a respected forecaster at 80 GW and you at 114 GW, which assumption explains most of the difference? Jevons rebound explains roughly half. S-curve timing and BTM account for most of the rest. This page shows the arithmetic.
Starting from ~80 GW consensus (IEA/BloombergNEF grid-connected, comparable scope), each bar shows the GW contribution of one analytical difference, computed from the model.
Consensus baseline: ~80 GW 2030 (IEA Electricity 2025 midpoint 80 GW; BloombergNEF ~85–90 GW implied). This is a CUMULATIVE bridge computed by re-running the EE model and turning on one assumption at a time in a stated order: (1) base/scope/LLT differences — the gap between the consensus print and the EE model configured with consensus-like assumptions (low Jevons, no BTM); (2) + Jevons rebound (0.72 vs consensus ~0.30); (3) + the BTM layer (grid-connected-only consensus counts none). The steps are exactly additive to the EE base because each is a sequential re-run. The S-curve is not a step — it is evidence for the front-loaded compute-growth path, not a forecast input. Because the parameters interact non-linearly, the split depends on the ordering shown.
Assumption Comparison · EE vs Consensus
If EE Were Wrong On One Assumption
Gap decomposition: each driver isolated by substituting consensus-equivalent parameter value into EE model. Consensus values from IEA Electricity 2025, BloombergNEF, LBNL 2024 (grid-connected scope). Interaction terms arise because parameters compound multiplicatively.
Electron Economics · Empirical Evidence
Why 2026.5? · S-Curve Evidence
The S-curve midpoint is the single most consequential timing assumption. Six independent datasets all point to the same answer: the AI compute demand inflection has already happened.
The analytical question: has AI adoption already reached its inflection point — where growth transitions from accelerating to decelerating? Six convergent signals say yes, and it happened in 2025–2026.
Hyperscaler capex: public earnings filings 2022–2026
DC capacity: FERC State of Markets 2026 · 451 Research
Hyperscaler Capex Trajectory · 2020–2026 · The Capital Signal
Capex is a leading indicator of compute deployment. 2020→2025 tripling at this scale is the strongest signal that S-curve inflection has passed — capital doesn't commit ahead of inflection at $400B+ scale.
Combined capex: Meta + Microsoft + Google + Amazon + Oracle. 2020: ~$100B. 2023: ~$200B. 2025: ~$410B. 2026: ~$725B reported (Alphabet $195–205B; MSFT ~$175B CY26; Amazon $173B TTM; Meta $130–145B; Oracle ~$50B). Basis note: the Microsoft figure is ~$15B lower than physical spend after a Jul 29 accounting change (DC useful life 15→25yr; some leases finance→operating), so comparable-basis 2026 is ~$740–780B. The acceleration itself is the signal — at ~75%+ growth off a $410B base, the market is not pre-inflection.
The 2028 case: enterprise AI is still experimental, agent workflows nascent, productivity gains unrealised. The S-curve hasn't peaked yet.
Why EE disagrees: the S-curve tracks compute demand, not enterprise productivity. Compute demand accelerated visibly 2023–2025 through hyperscaler capex and DC additions — regardless of whether end-users see productivity gains. The inflection in compute investment rate precedes enterprise productivity inflection by years.
Model robustness: moving the midpoint from 2026.5 to 2029 changes the 2030 output by ~8 GW (114→106 GW) and 2040 by ~15 GW. The directional conclusion is unchanged. Test it in the Scenario Builder.
S-curve calibration: midpoint set to when second derivative of DC capacity additions turned negative — i.e. when growth rate peaked. Six datasets all show acceleration peaking 2024–2026. Sources: FERC State of Markets 2026, Meta/MSFT/GOOG/AMZN earnings, 451 Research, SemiAnalysis GPU tracker 2025, OpenAI usage benchmarks.
Electron Economics · BTM Supply Chain
BTM Gas Supply Chain Model
Behind-the-Meter & Nuclear Power Layer
The 58.4 GW BTM gas + 24.2 GW nuclear model by 2040 represents two distinct pathways beyond the conventional grid. Gas (RICE-led) is deployable now. Nuclear (restarts + first SMRs) is contracted and building. Together they account for 82.6 GW, or 20.2% of the 408 GW base forecast.
Two paths, very different timelines. Gas BTM: 8–10 month RICE delivery, 14.9 GW firm (20.5 GW total incl. framework, 21 tracked projects, EE Modular Gas Tracker Aug 2026), developer-willingness constrained from 2028. Nuclear: 835 MW Microsoft/Constellation TMI PPA (announced 2024; restart ~2028), ~2 GW contracted and ramping by 2026, Kairos SMR fleet 2030–2035, ~24 GW cumulative by 2040 (T1 contracted + T2 prob-weighted; upside 35–40 GW if full Meta delivery). The carbon profiles are opposite — gas at ~400g CO₂/kWh, nuclear at ~12g — which makes the nuclear layer the ESG-defensible version of the BTM thesis.
Microsoft/Constellation Crane Clean Energy (TMI U1): 835 MW · PPA Sep 2024 · Restart targeted late 2027 (accelerated from 2028) · NOT yet online
Signed Nuclear Contracts for AI Data Centre Power · Jun 2026
Contracted, permitted, or under construction only. Sorted by expected online date.
Status definitions: Operational = generating power to DC campus. Under construction = NRC licence issued, civil works started. Contracted = binding PPA or power purchase agreement signed, pre-construction. SMR timelines carry FOAK (first-of-a-kind) risk — Kairos, Oklo, and TerraPower are all first commercial deployments. Model applies a conservative ramp that assumes 12–18 month delays on first SMR units relative to stated timelines.
Gas BTM OEM Production Capacity · Annual GW · 2025–2032
Published capacity, reservation backlog where disclosed, and estimated US DC-serving fraction. RICE section: not supply-constrained. Heavy-duty GT section: supply-constrained through 2027–2028.
GE Vernova: Greenville SC expanding ~50→70–80 units/yr by 2026, targeting 20 GW/yr from 2027; total backlog + slot reservation agreements 116 GW at Q2 2026 (up from 100 GW at Q1 end — 44 GW firm + 56 GW SRA — against a ≥125 GW year-end target). Siemens Energy: 87 GW gas turbine backlog on the comparable basis (Q3 FY2026), of which 12 GW is explicitly data-centre linked (Q2 2026 call). Lead times 40+ months. Caterpillar/reciprocating: smaller units (1–50 MW), faster lead times, edge BTM. US DC-serving fraction: ~35–45% of total global output based on disclosed regional allocation.
OEM Cap vs BTM Model Additions · Annual GW · 2026–2040
Annual OEM cap available to US DC BTM vs annual model additions. Where model is below cap, demand intent (not OEM supply) is the constraint.
RICE vs aero split: Wärtsilä 50SG/34SG (8–10 month delivery, stable, not constrained) is the volume leader — 2.4+ GW across 5 US DC orders by Q2 2026. INNIO/Rehlko 1.25 GW framework (Apr 2026) adds ~417 MW/yr RICE 2026–28. Heavy-duty GTs (GE/Siemens): supply-constrained, 2028+ delivery. Grey market PE6000/CF6-80C2: hard ceiling — fixed pool of ~1,000 retiring engines competed by aviation MRO + defence + data centres. Model additions run well below the combined cap after 2027 — developer adoption (not OEM supply) is the binding constraint from 2028. 14.9 GW firm + 5.6 GW framework = 20.5 GW total across 21 tracked projects (EE Modular Gas Tracker, Aug 2026). This supersedes "11.9 GW firm + 3.8 GW framework = 15.6 GW total (Jun 2026)" — the 3.8 GW figure was the tracker's INNIO/VoltaGrid single-OEM total, not a market framework total. OEM cap held at 5.0→8.0→10→14 GW/yr: it is a forward manufacturing-capacity judgement drawn from the OEM capacity table above, not a fit to the order book, so the pipeline restatement does not change it.
Cumulative BTM Stock · Model vs OEM Maximum
58.4 GW by 2040 is 32% of the OEM-constrained theoretical maximum — well within the supply envelope. Working: the maximum is the cumulative sum of the phased OEM cap plotted above (5.0 GW/yr to 2026 + 8.0 to 2027 + 10.0 to 2030 + 14.0 from 2031, summed 2026–2040) = 183 GW; 58.4 ÷ 183 = 32%. The constraint is developer adoption rate, not turbine availability.
Key Assumptions & Upside/Downside
BTM pipeline validation (S&P Global Jun 2026 webinar): Pipeline of projects planning to co-locate power supply with data centres has reached 180 GW maximum capacity, with ~50 GW expected near-term over the next five years. EE model has 58.4 GW BTM by 2040 — 32% of the 180 GW maximum, and 117% of the near-term 50 GW expectation. The 180 GW figure is maximum developer-stated ambition; applying the Goldman/Sightline 40–60% delivery haircut yields ~72–108 GW realised. That range does not bracket the EE base — it sits entirely above it. At 58.4 GW the model is ~19% below the low end of the haircut band, i.e. the EE BTM layer is more conservative than a haircut applied to the S&P pipeline, not a midpoint of it. The framing is retained as a supply-envelope check (the pipeline can comfortably supply the model), not as a two-sided validation. GE Vernova × Blue Energy (Jul 2026): GE Vernova formed a strategic collaboration with Blue Energy to advance the world's first gas-plus-nuclear hybrid plant using BWRX-300 SMRs and GE Vernova gas turbines — targeted at AI and advanced manufacturing demand. Hybrid gas-nuclear is a new supply chain category not currently in the OEM table; added as a note pending commercial contracts. Source: ANS Nuclear Newswire Jul 2026.
Nuclear sources: Microsoft/Constellation Crane Clean Energy (TMI Unit 1, 835 MW, PPA signed 2024, restart ~2028); Amazon/Talen Susquehanna (960 MW campus, confirmed); Google/Kairos Power SMR fleet (500 MW, Hermes 2 TVA 50 MW first reactor 2030, Aug 2025 announcement); Meta/Constellation Clinton Clean Energy Center (Illinois, ~250 MW, 2027+); Oracle SMR campus (~1 GW, 3 reactors); Equinix/Oklo 500 MW PPA; TerraPower Natrium 345 MW Wyoming (2030 est.); Helion fusion 50+ MW Microsoft commitment (2028 target, excluded from base). Gas sources: Wärtsilä press releases Q1–Q2 2026 — 790 MW Texas off-grid (Q2 2026), 429 MW IOU plant (Q1 2026), 507 MW Ohio 50SG (Q4 2025), 412 MW Ohio 34SG (Q2 2026); INNIO/Rehlko 1.25 GW framework Apr 2026; ProEnergy PE6000 portfolio >1 GW + 650 MW Crusoe Apr 2026; GE Vernova Q4 2024–Q2 2026 earnings; Siemens Energy Q2 2026; Meta Socrates South OPSB Jun 2025; EE Modular Gas Tracker Jun 2026; EE Gas Turbine Reservations Dec 2025.
Electron Economics · US AI Data Centre Forecast 2040 · Last updated: Aug 2026
Changelog — What Changed and Why
Every model update, data revision, and source addition since the first version. All changes are sourced from public filings, named analyst reports, or EE primary research. No version removes data — superseded figures are noted with their replacement.
New First time this data appears
Update Existing figure replaced with newer source
Fix Error corrected
Structure UI / tab / layout change
v1.12Sep 15 2026Corrections
Installed capacity could fall. The supply constraints do not bind in the base case. And the one channel through which financing conditions reach this forecast was published but never applied.
Capacity now has a floor at the existing stock. The line read min(demand, ceiling) with nothing beneath it, so when modelled demand dipped below what was already built, the installed capacity series followed it down. The constrained case lost 50.0 → 43.9 GW in 2026, an 11 per cent contraction of the installed base in a single year, and the efficiency case shed capacity across 2039–40. Nothing in this model retires a data centre — it says so itself in the behind-the-meter prose — so a falling stock was arithmetic, not a forecast.
Changed series: constrained 2026 50.0 GW (was 43.9), efficiency 2039 and 2040 both 141.5 GW (were 139.3 and 134.7). Base and accelerated are untouched at every year.
The base case is demand-bound, and that is now stated on the model. The deliverable ceiling binds in 3 of 15 years in base and 0 of 15 in accelerated; it sits 2 to 56 GW slack. A reader looking at a grid buildout rate, a bypass share and an OEM ceiling reasonably assumes they are doing work. In the base case they are not.
The consequence, measured. A ten-fold swing in the OEM delivery ceiling moves 2040 by 0.0 GW. Deleting behind-the-meter entirely moves it by 1.1 per cent. No supply-side or capital-side input can move the base forecast, because everything they would relax is already slack.
Scenario weights have an owner, and the lever is live. They were published in the sync block and applied nowhere — so the single channel through which cheaper capital, residual value guarantees and rated execution reach this forecast could not even be observed. Shifting 10 points from constrained to accelerated is worth +5.1 GW at 2030 and +19.1 GW at 2040, roughly twenty times the entire supply-side machinery.
The verifier grew from 519 to 607 assertions, and the declared binding regime is recomputed rather than trusted — if the declaration ever stops matching the model, the gate fails.
Why the weights were reviewed and held rather than moved
Every input that should ease the constrained case moved in the right direction between August and September: the triple-A ceiling broke at a fourth agency, residual value guarantees appeared in S&P’s own hyperscaler commentary, and AI-related debt issuance ran US$489bn year to date against US$322bn for all of 2025. Every input that should tighten it also moved: AAA seniors priced 30 to 50bp wide of guidance, Vantage could not place seven-year paper, DataBank faces an October repayment, and the median large-load contract term went from five years to twelve. Both lists moved. Netting them is a scenario decision taken deliberately, not a side effect of a maintenance release — so the weights carry a lastReviewed date of 15 Sep with lastChanged still null, which makes “held” distinguishable from “never looked at”.
FM40_SCENARIO_PROBS_PROV · two-sided checklist of what moves the weight
What the floor asserts, and what it does not
Zero net retirement to 2040. That is defensible over this horizon — the 50 GW installed base is young, built mostly since 2018 against facility lives of 15 to 25 years — and it is not a claim that no facility ever closes. If retirements belong in this model they belong as a schedule, the way the nuclear layer is a schedule, rather than arriving as an accident of demand falling below the stock. Every floor activation is logged, and the verifier distinguishes a legitimate flat year from a silent clamp by requiring one to explain the other.
FM40_FLOOR_HITS · 3 activations: constrained 2026, efficiency 2039 and 2040
v1.11Sep 15 2026Corrections
This model had no verifier, and one of its inputs had drifted 48 per cent from the model that owns it without anything noticing.
The behind-the-meter OEM ceiling is now a read from EE_SYNC.gas.oemCapacityGWyr, not a hardcoded copy. It had been returning 5.0 / 8.0 / 10.0 / 14.0 GW per year while carrying a comment saying “canonical values now live in EE_SYNC.gas” — citing the owning model and then ignoring it. By 2030 the two had drifted to 10.0 against 19.2 GW/yr, a 48 per cent gap.
The forecast series does not move. 2030, 2035 and 2040 are unchanged at 114.7, 257.1 and 408.5 GW, because the ceiling is not the binding constraint — a ten-fold swing in it moves the 2040 total by nothing. The bypass-share term binds first, by roughly three times. That non-binding fact is now asserted, so the day it stops being true the gate says so.
The BTM bypass share has an owner for the first time. It is the term that actually governs the behind-the-meter path, it is a judgement, and it is now labelled one — with the reason it cannot be derived, a cross-check against the Gas Tracker carrying its basis warning, and two falsification conditions.
A new verifier: 519 assertions. This was the only one of the nine models with no verifier, no Python source and no data file — 26,000 characters of JavaScript inside one page. The verifier extracts the model from the page and runs it, so the assertions hit the same code the page does.
The published series is now checked against what the model computes. The previous check compared the page to the sync block, which is self-consistency, not currency. A 0.6 per cent drift in any scenario at any year now fails the gate.
An open defect, recorded rather than quietly fixed
This model seeds behind-the-meter stock at zero for every year to 2025 and reaches 0.34 GW in 2026. The Gas Tracker observes roughly 2 GW already operating, and 2.8–3.2 GW by year-end 2026. Even after an assumed 1.2–1.4× reserve margin to convert generation nameplate into facility load, the model understates 2026 behind-the-meter by six to eight times. It is not fixed here. Seeding the stock changes the published series, which is a forecast revision rather than a maintenance fix, and it should be signed off rather than slipped into a release as a side effect of wiring up a verifier. The defect record is asserted to exist and to state its own magnitude, so it cannot be forgotten — and the verifier fails if anyone marks it closed while the seed is still zero.
Why the drift survived so long, and what changed structurally
Two duplicated constants remain — the large-load friction rate owned by the Tariff Tracker, and PUE. Both still agree with their owners, but they agree by maintenance rather than by construction, which is exactly the state the OEM ceiling was in before it drifted. They are now gated for agreement, so the next divergence fails the build instead of accumulating. The model also declares which of its inputs are owned elsewhere, in FM40_SYNC_INPUTS, so the coupling is legible rather than buried in a comment.
spec/verify_forecast.js · 519 assertions, 10 of 10 injected defects caught
v1.10Sep 07 2026FixUpdate
The 50 GW anchor asserts a measurement basis that FERC never stated, and treats a floor as a point.
FERC’s words are “data centers with a collective capacity of more than 50 GW were in service at the end of 2025”. It does not say critical IT, facility, nameplate or peak. This model reads it as facility MW post-PUE, and that reading is the model’s interpretation rather than FERC’s claim.
The figure is not FERC’s own measurement. It is taken from Yes Energy’s Load Center Project Database, accessed 3 March 2026, which publishes no capacity definition either. The basis is unstated at both links in the chain.
“More than 50 GW” is a lower bound. This model anchors on it as a point estimate.
Nothing has been changed, because there is no better anchor. No FERC, EIA, NERC, DOE or LBNL figure for US installed data centre capacity in GW was published between 20 August and 7 September 2026, and none of them states a basis either. The one in-window authority, CRS R49326 of 1 September 2026, works entirely in terawatt-hours of annual consumption and does not address installed capacity at all.
FERC’s regional split of 2025 in-service capacity is added, with the same basis caveat attached: ERCOT 31 per cent, MISO 20, Southeast 18, and SPP, PJM, West and Other at 8 each.
No PJM, ERCOT, MISO or SPP load forecast revision was published in the window. The PJM revisions in circulation are from January 2026.
No PUE benchmark was published in the window. Uptime’s 2026 survey reports only “minor improvements” and publishes no number.
An anchor is only as good as its stated basis
A capacity figure with no stated basis can differ by 20 to 45 per cent depending on what was meant. Building a sixteen-year forecast on one, and describing it as post-PUE facility megawatts, imports a precision the source does not carry.
FERC 2025 State of the Markets, 19 Mar 2026, citing Yes Energy Load Center Project Database accessed 3 Mar 2026
v1.9.2Aug 23 2026Update
The Texas pause is recorded as a downgrade trigger rather than a scenario change, and a fresher combined-cycle cost print is added.
Texas stopped data centers advancing through the interconnection process on August 3 2026. The affected queue totals roughly 474 GW, about 90 percent of it attributed to data centers. The near-term buildout assumption leans on Texas absorbing load the eastern grid operators cannot, so that assumption is now exposed.
No scenario has been re-run. A pause expected to last under nine months sits inside the model's own timing tolerance for 2030 and beyond. It becomes a downgrade trigger if the audit runs past the April 2027 deadline, or if queued Texas capacity is denied rather than delayed.
NRG disclosed US$3.2 billion for a 1.2 GW combined-cycle first phase in Texas, about $2,670 per kW, against the $2,157 per kW the model carries. The gap is 24 percent arithmetically, but the two figures have not been reconciled for capacity basis, cost scope or cost year, so it is shown as unconfirmed rather than as an escalation read.
Update
Texas pause added to Known Limitations as a named downgrade trigger. The interconnection-queue-depth row previously read PJM 4+ years against ERCOT 18–24 months, which is the implicit basis for grid_buildout_near leaning on Texas. It now records the Aug 3 2026 directive and Batch Zero pause, states explicitly that no scenario has yet been re-run, that the sub-nine-month justification is WITHDRAWN because the pause now has no end date, and names grid_buildout_near as the parameter that moves if the audit overruns Apr 9 2027 or ends in denials rather than delays.
Gov. Abbott letter Aug 3 2026 · ERCOT Batch Zero
Update
Combined-cycle cost print refreshed alongside the existing figure, not replacing it. The grid_buildout reason string carried +66% to $2,157/kW. NRG's Aug 2026 disclosure of US$3.2B for a 1.2 GW first phase implies ~$2,670/kW, 24% higher arithmetically, though the two have NOT been reconciled for capacity basis, cost scope or cost year, so the gap is carried as unconfirmed. Both are shown; the older print is dated and the newer one is a single project rather than an index.
NRG disclosure Aug 2026 · Utility Dive
v1.9.1Aug 18 2026FixUpdateStructure
Data center capex to 2040 cut from US$13.4 trillion to US$11.3 trillion, with several stale figures restated.
The 2040 build cost falls from US$13.4 trillion to US$11.3 trillion. Capacity was being costed at the facility level using a rate that measures critical IT load, which double-counted cooling and distribution overhead. Figures are in 2026 dollars with no construction escalation added.
Behind-the-meter gas capacity in 2040 is restated from 70 GW to 58.4 GW, which is 14.3 percent of the 408.5 GW base case rather than the 17 percent previously claimed.
The constrained 2040 case is corrected from 318 GW to 303.3 GW, matching what the model has actually produced since the friction correction two releases earlier.
The behind-the-meter order pipeline now reads 14.9 GW firm plus 5.6 GW of framework agreements, 20.5 GW across 21 tracked projects, replacing two conflicting earlier counts.
None of the four scenarios moved. The base case still runs 114.7 GW in 2030, 257.1 GW in 2035 and 408.5 GW in 2040.
Fix
CORRECTION — capex basis error. Headline DC capex 2040 falls from $13,377B to $11,336B (~$13.4T → ~$11.3T). The Generation tab costed data centre capacity as dc2040 × $32.75M/MW. That is wrong on units. $32.75/W is the EE Capex Stack v20 all-in rate per critical IT watt (IT $17.50/W + powered shell $4.50 + physical shell $1.25 + gap/overhead $9.50), but every GW figure this model produces is facility capacity — the demand chain ends in × PUE(yr) and the 50 GW FERC 2025 anchor is facility load. Multiplying facility MW by an IT-watt unit cost overstated total capex by the PUE factor. All DC-capex calculations now route through eeSyncCapexB(), which divides facility GW by eeSyncPue(yr) before costing: 408.5 GW facility → 346.2 GW IT at the 2040 PUE of 1.18 → $11,336B. The cumulative capex chart, the KPI tile and the all-in per-MW figure move with it (all-in ≈ $37M/MW → ≈ $32M per facility MW). Superseded figures retained on the tab per the versioning policy. A visible basis line was added stating the IT-watt basis, the PUE conversion, and that figures are 2026 dollars with no construction escalation applied — the Capex Stack's observed build-cost escalation is 5.5%/yr (7.0% CAGR 2020–25). The McKinsey $6.7T → GW conversion carried the same error in reverse and is corrected from ~185 GW global / ~83 GW US to ~276 GW global / ~124 GW US on a facility basis. No model weight or scenario modifier changed; all four scenario series are byte-identical.
EE Capex Stack v20 · EE_SYNC.capex.allInPerW · internal reconciliation Aug 18 2026
Fix
CORRECTION — changelog scenario drift. Constrained 2040 restated 318 GW → 303 GW. The v1.0 entry still carried 318 GW, the pre-v1.3 value from before the LLT-friction fix. The live model returns 303 GW. A 15 GW gap is three times this file's own "≥5 GW = major version" threshold, and it survived four releases. Accelerated (493 → 494) and Efficiency Breakthrough (135 → 135) were within rounding but were stale by the same mechanism. All four scenario figures on that line, plus the Generation tab's headline (previously a stale 411 GW) and the vs-Analysts vintage strip, now render from computeForecast2040().
Model output vs published text · Aug 18 2026
Fix
vs-Analysts EE self-rows now read live from the engine. 2027 corrected 72 GW → 65 GW (a 10.8% mismatch, not rounding).ANALYST_DATA hardcoded the EE rows at 72 / 115 / 408 GW while the model returned 65.0 / 114.7 / 408.5. The Calibration table already read base.results[yr] live; the comparison chart and table now do the same, including the 2040 scenario range (constrained–accelerated). A model's own output should never be typed into its comparison chart by hand.
computeForecast2040() vs ANALYST_DATA · Aug 18 2026
Fix
CORRECTION — BTM pipeline citations reconciled to the Gas Tracker project array. Canonical Aug 2026: 14.9 GW firm + 5.6 GW framework = 20.5 GW across 21 tracked projects. The file carried two mutually inconsistent citations, neither of which reconciled with the tracker: "11.9 GW firm + 3.8 GW framework = 15.6 GW total (Jun 2026)" and, twice, "confirmed firm orders now ~17.4 GW (was 15.6 GW Jun 2026)". The "3.8 GW framework" figure was in fact the tracker's INNIO/VoltaGrid single-OEM total, not a market framework total. Firm is now defined as Operational + Ordered + Permitted; framework as Framework + Framework signed + Testing. Every live occurrence reads from EE_SYNC.gas.pipelineGW. The phased BTM OEM cap is unchanged at 5.0 / 8.0 / 10 / 14 GW/yr. It is a forward OEM-manufacturing-capacity judgement taken from the OEM capacity table (nameplate plus stated ramp), not from the order book — it was not reverse-fitted to the pipeline, which is why restating the pipeline upward by 4.9 GW does not move it. Base forecast unchanged.
EE Modular Gas Tracker PROJ array (Aug 2026) via EE_SYNC.gas.pipelineGW
Fix
CORRECTION — BTM 2040 stock restated 70 GW → 58.4 GW. Every derived framing recomputed. The prose carried a hardcoded 70 GW in eight live places — the base scenario card, the "Why 408 GW and not 250 GW" IC answer, two BTM Q&A titles and one Q&A body, the BTM tab deck, the BTM cumulative chart note and the S&P validation note-strip — plus two changelog entries, while EE_SYNC.forecast.supply.btm[2040] — the executed output of computeForecast2040() — returned 58.4 GW. A 20% overstatement in the layer the whole BTM thesis rests on, and in one Q&A row the two were rendered in the same sentence: "The 70 GW by 2040 (14% of 408 GW total base)". Cause: 70 GW predates the v1.3 OEM-cap revision and the v1.4 friction fix and was never updated; because the model applies no BTM retirement, the 2040 stock is exactly cumulative additions, so there is no basis on which 70 and 58.4 could both be right. Derived figures, before → after: share of the 408.5 GW base 17% → 14.3%; BTM + nuclear ~18% of base → 82.6 GW = 20.2% (58.4 + 24.2, the share rises even though BTM falls, because the old 18% was itself understated); share of the OEM-constrained theoretical maximum ~35% → 32% against a maximum of 183 GW (cumulative sum of btmOEMCapMW() 2026–2040, the same series the BTM cumulative chart plots); share of the S&P 180 GW co-location maximum 14% → 32% (the old 14% was not 70/180 = 39% — it was the base-share number mis-filed against the S&P denominator); multiple of the S&P near-term 50 GW 140% → 117%. One framing withdrawn: the Goldman/Sightline 40–60% haircut on 180 GW yields 72–108 GW, which was described as "bracketing the EE base". It never did — 70 sat below 72 — and at 58.4 GW the model is ~19% below the low end. That line is now stated as a one-sided supply-envelope check, not a validation. Also corrected: the v1.0 entry's supply split 314 / 70 / 24 GW → live 325.9 / 58.4 / 24.2. All occurrences now read from EE_SYNC.forecast.supply via eeBtmFrames(). No model weight or scenario modifier changed; all four scenario series are byte-identical (base 114.7 / 257.1 / 408.5 at 2030 / 2035 / 2040).
EE_SYNC.forecast.supply.btm[2040] vs published text · internal reconciliation Aug 18 2026
Fix
CORRECTION — stale BTM pipeline figure in the BTM Q&A. "6.8 GW announced (Jun 2026), of which ~5 GW delivers by 2028" → 14.9 GW firm + 5.6 GW framework = 20.5 GW across 21 projects (Aug 2026). The Q&A row survived the v1.9.1 pipeline reconciliation and contradicted the canonical figures rendered elsewhere on the same tab. It now reads from EE_SYNC.gas.pipelineGW. The "~5 GW delivers by 2028" clause is dropped, not restated: EE_SYNC.gas carries no delivery-year split, and inventing a replacement number would reproduce the failure being fixed.
EE Modular Gas Tracker PROJ array (Aug 2026) via EE_SYNC.gas.pipelineGW
Update
GE Vernova backlog advanced a quarter; Siemens Energy comparator added. The Limitations tab still cited the Q3 2025 print (55 → 62 GW). Current: 116 GW of equipment backlog plus slot reservation agreements at Q2 2026 (up from 100 GW at Q1 end — 44 GW firm + 56 GW SRA), against a ≥125 GW year-end 2026 target. Siemens Energy is at 87 GW on the comparable basis (Q3 FY2026), book-to-bill 2.65. Both figures now read from EE_SYNC.gas.oemBacklogGW rather than being restated inline.
GE Vernova Q2 2026 (Jul 22) · Siemens Energy Q3 FY2026 (Aug 5) · EE_SYNC.gas.oemBacklogGW
Structure
Weights de-obfuscated; disclosure language corrected.FM40_WEIGHTS was stored as JSON.parse(atob('…')) — reversible in one line, so it protected nothing, while making the calibration un-diffable across versions in a file that simultaneously claimed parameter names and calibration were disclosed. It is now a plain object literal with values byte-identical to the decoded payload (verified: all sixteen years in all four scenarios unchanged). The "weights are proprietary" claims have been replaced with what the file actually does. Historical changelog entries describing the base64 encoding are retained as record.
Internal code review · Aug 18 2026
Fix
Step-9 scenario modifiers now published from the executed object. A new Scenario Modifiers table in Methodology renders directly from FM40_SCEN_MODS — the same object computeForecast2040() reads. Prior hand-written disclosure listed accelerated.efficiency 1.08 and constrained.efficiency 0.95; the engine has always executed 1.00 in both cases. Three of twelve published modifiers were wrong in a view whose entire pitch is full disclosure. Rendering from the object removes the class of error. Values themselves unchanged.
FM40_SCEN_MODS vs published table · Aug 18 2026
Structure
Dead code and dead parameters removed; sensitivity loops made exception-safe; EE research linked. (1) base_cagr_mid (0.022) was declared but never read — all three consumers branch y<=2030 ? base_cagr_near : base_cagr_far. Deleted rather than wired in, because wiring it would move forecast output. (2) s_midpoint / s_steepness retained but labelled evidence only — not a live driver, so the sensitivity and parameter tables no longer imply they move the forecast. (3) The orphan <div id="tt"> was removed — it had no CSS rule and no JS reference; the live tooltip is #global-chart-tt, created lazily by getTooltip(). (4) renderForecast40Sensitivity() and runMonteCarlo() mutate global weights and restore them; both restores are now in finally blocks, so an exception mid-loop cannot leave the globals holding a perturbed value for the rest of the session. (5) The EE pieces this model is calibrated against — Tariff Lottery, Who Wears the Risk, OpenAI Cost of Capital, Gas Turbine Reservations, Modular Gas Tracker — are named 11+ times as load-bearing sources and were previously linked zero times. Published slugs are now linked inline and collected in a Calibration Sources block on the Methodology tab; sources with no published slug link to the publication root.
Internal code review · Aug 18 2026
v1.9Aug 17 2026NewUpdate
A roughly 250 MW Baker Hughes turbine order for Twenty20 Energy added, and all second-quarter supplier figures verified current.
New
Baker Hughes / Twenty20 Energy added to BTM order ledger. 10 Frame 5 gas turbines (~250 MW) for data centre projects in Georgia and Texas, deliveries from 2027 — initial award under a broader multi-gigawatt strategic collaboration (Baker Hughes press release Feb 11 2026). Added to the BTM firm-order list; too small to move the phased cap, which is held at 5.0/8.0/10/14 GW/yr. Base forecast unchanged.
Baker Hughes / Twenty20 Energy press release Feb 11 2026
Update
Full data refresh verified against live sources (Aug 17). All Q2 2026 OEM figures re-checked and confirmed current — GE Vernova 116 GW, Baker Hughes $7.1B / $37.1B RPO, Siemens €162B backlog / 2.65 book-to-bill, INNIO 1.1 GW order and margin compression, Wärtsilä Liberty €292M. No revisions required. Sync date advanced to Aug 17 2026 across all three EE dashboards (Gas Tracker · Forecast 2040 · Capex Stack).
EE verification pass · Aug 17 2026
v1.8Aug 16 2026UpdateNew
Second-quarter 2026 supplier results added: GE Vernova backlog at 116 GW, Siemens demand running at 2.65 times production.
Update
Synced with EE Modular Gas Tracker — Q2 2026 OEM earnings pass. All OEM notes in the BTM Supply Chain tab updated to Q2/Q3 2026 actuals. No parameter change — phased cap 5.0/8.0/10/14 GW/yr held; new orders are within the existing ceiling.
GE Vernova Q2 2026 (Jul 22): backlog + SRAs = 116 GW (up from 100 GW at Q1 end); year-end target ≥125 GW. Power segment orders $16.7B (+134% organic). DC electrification orders >$5B YTD — more than double full-year 2025. Production roadmap: 20 GW annualised H2 2026 → 24 GW 2028 → 30 GW 2030. >50% of 2031 production already sold. New entrant: Doosan Enerbility (South Korea) — 12 US DC units under contract (xAI Colossus 5 × 380 MW confirmed, 7 more unnamed), first non-incumbent OEM to break into US DC gas turbine market.
Baker Hughes Q2 2026 (Jul 26): IET orders record $7.1B (doubling YoY, 4th consecutive record quarter); IET RPO record $37.1B; Power Systems orders $2.6B incl. 2.7 GW DC and mobile power. Chart Industries acquisition closed. IET Horizon 2 order target raised to >$45B (2026–28). Power Systems guided to ~$5B annually by 2029 (3–4× 2025). New named deal (Jul 29): Dynamis Power Solutions — 76 × NovaLT16 (~1.3 GW) for hypermobile DC + O&G power, packaged in Dynamis DT17 platform; turbines booked Q2, generators booked Q3 2026. Establishes mobile/hypermobile as a distinct BTM sub-category.
Siemens Energy Q3 FY2026 (Aug 5): record orders €17.9B; Gas Services record €9.97B (+61.9%), book-to-bill 2.65 — demand running at 2.65× production capacity. Total backlog record €162B (+€26B QoQ, +20% YoY). Medium-sized GT production capacity lifted from ~50 units FY2025 to ~80 units FY2026. Large GT additional capacity from FY2027. DC + Middle East ~50% of gas turbine orders. Structural bottleneck confirmed: demand is 2.65× what can be produced.
INNIO Q2 2026 (Jul 28): equipment order intake $2.3B (+316% YoY); backlog $6.6B (+279% YoY, record) — visibility into at least 2030. Revenue $937.7M (+42%). New: 1.1 GW order (200+ J624 engines) for unnamed US mega-scale DC campus (Jul 28). EBITDA margin compressed 22% → 18% — cost of scaling. Share price −36% post-earnings: market separating order intake from execution confidence. Full-year 2026 guidance: revenue $3.8–$3.9B; EBITDA $720–$740M.
Wärtsilä Liberty Energy (Jun 29): €292M / ~$332M 34SG order, Q3 2026 booking, delivery 2029–30. Fifth named US DC transaction. Total confirmed DC capacity approaching 2.4 GW across 50SG and 34SG platforms.
New named projects added to tracker: (1) Joule Capital Partners / Caterpillar G3520K Utah — 4 GW campus, Phase 1: 1.5 GW via 636 G3520K units, early 2028 ramp, Kern River Gas supply (NGI Jun 2026 / Evercore ISI). (2) Dynamis Power Solutions / Baker Hughes NovaLT16 — 76 units (~1.3 GW), mobile power, DC + O&G, North America. Evercore ISI (Jun 2026): new large recip engine orders now slotting 2028 — engine supply as constrained as large-frame turbines.
BTM pipeline total (confirmed firm orders, Aug 2026 tracker):[Superseded in v1.9.1 — this roll-up did not reconcile with the tracker PROJ array; canonical is 14.9 GW firm + 5.6 GW framework = 20.5 GW across 21 projects. Retained as the v1.8 record.] ~17.4 GW across named transactions (INNIO 1.1 GW Jul 28 + Dynamis/BKR 1.3 GW Jul 29 + Joule/Cat 1.5 GW Phase 1 + Wärtsilä Liberty €292M Jun 29 + prior 11.9 GW base). Goldman 50–60% delivery haircut applied. Developer willingness (not OEM supply) remains the binding constraint from 2028.
GE Vernova Q2 2026 earnings Jul 22 · Baker Hughes Q2 2026 earnings Jul 26 · Siemens Energy Q3 FY2026 Aug 5 · INNIO Q2 2026 + 1.1 GW order Jul 28 · Wärtsilä Liberty Energy €292M order Jun 29 · Baker Hughes / Dynamis Jul 29 · NGI May 29 / Evercore ISI Jun 2026 · EE Modular Gas Tracker Aug 2026
v1.7Aug 11 2026StructureFix
Headline restated as a 303 to 494 GW range around 408 GW, and navigation cut from eleven views to six.
Structure
v4 pass — hierarchy, reconciliation, discipline (external product review). (1) Headline reframed as an envelope — ~408 GW central within a ~303–494 GW scenario range, not a point estimate. (2) Jevons de-overclaimed — the 0.72 rebound is now stated as analytical judgement that rejects zero-rebound, not an empirically identified elasticity ("the data is clear" language removed). (3) Training→inference causal claim corrected — the mix shift is a rationale for the assumed deceleration, not the mechanism that produces it. (4) Capacity defined once — a persistent definition (facility-power basis incl. modelled BTM) sits on the landing page and governs every GW figure. (5) FERC 50 → S&P/451 62 GW reconciliation bridge added to Calibration (scope allocation, not disagreement). (6) Monte Carlo relabelled — "Illustrative Parameter Uncertainty," 10th / median / 90th percentile outcomes (P10/P50/P90 shorthand removed). (7) Capacity funnel added — latent demand 477 → energizable 408 → financeable 388 → operational 357 GW, unifying LLT/grid, financeability and delivery attrition in one frame. (8) Navigation collapsed 11 → 6 (Forecast · Drivers · Why EE Differs · Supply · Methodology · Model Notes) with a secondary sub-nav; the six heavy analytical charts moved off the landing page into a new Drivers & Uncertainty view (~40% shorter homepage). (9) Colder prose — removed "largest single-cycle investment in history," "inference explosion is underpriced," and the unanswerable "most wrong parameter" framing. (10) Stale scenario numbers fixed — constrained shown as ~303 GW / 12.8% CAGR (post-LLT-fix) wherever it had been left at 318 / 13.1%. Weights kept proprietary but flagged as illustrative pending a server-side release. Base forecast unchanged (408 GW); reconciliation error 0; 80-call runtime sweep clean.
External analytical-product review · Aug 2026
v1.6Aug 11 2026FixUpdate
Texas regulators confirmed co-located load can be fully curtailed on 30 minutes notice; Bloom Energy and CoreWeave evidence refreshed.
Fix
Reconciled dashboard against current Electron Economics writeups. (1) Bloom Energy FEOC caveat — the "~95% US share" claim now flagged as final-assembly only; added the scandium-oxide (~13–15 kg/100 kW server) China-supply-chain / Foreign-Entity-of-Concern underwriting risk from "Beijing holds the off-switch on Bloom's $70B" (Jul 2026). Prior card read as a clean-domestic story that contradicted the published piece. (2) ERCOT curtailment precedent added — PUCT Docket 59220 (order Jul 23–24 2026) affirmed ERCOT can fully curtail co-located large load on 30-min notice, uncapped by the BTM generator (525.5 MW load vs 265.5 MW Goodnight Wind). Added to BYONG Q&A, Wärtsilä/ERCOT trigger, and Limitations — it reprices grid-synchronous colocation and reinforces the model's off-grid gas assumption. (3) CoreWeave refreshed from end-2025 vintage to Q1 2026: contracted backlog $99.4B (31 Mar 2026), plus the "customers becoming competitors" renewal thesis (Microsoft declined the next contract; Meta building owned capacity). (4) Renewal-window reconciliation — noted the neocloud take-or-pay renewal cliff at 2027–2028 (per CoreWeave/Jevons pieces) is earlier than the model's 2030–33 hyperscaler lease window; flagged as a candidate near-term window rather than silently moving the calibrated base. (5) Queue tab — quarantine banner now points to the legitimate large-load dataset (ERCOT ~198 GW applications Q1 2026; PJM ~40-month waits) as the correct basis for a rebuilt load-pipeline module. Base forecast unchanged (408 GW) — all evidence-layer.
EE writeups Jul 2026 (Jevons; Bloom FEOC; CoreWeave; Crusoe/ERCOT) · White & Case (PUCT Docket 59220)
v1.5Aug 3 2026UpdateFix
Big Five 2026 capex refreshed to US$725 to 755 billion; financing capacity, not demand, is now the binding constraint.
Fix
Microsoft accounting change — capex-as-evidence caveat added. On its Jul 29 2026 call (FY26 Q4) Microsoft extended data-centre useful life from 15 to 25 years and reclassified some leases from finance to operating, trimming ~$15B off its reported calendar-2026 capex (~$175B guided) with no change to physical buildout. All capex evidence cards, the S-curve capex chart, and the s_midpoint evidence note now carry a basis caveat: headline hyperscaler capex understates real spend, and cross-company comparisons must be basis-adjusted. Base forecast unchanged — this is an evidence-layer correction, not a demand revision.
Microsoft FY26 Q4 earnings Jul 29 2026 · Digital Applied capex scorecard Jul 31 2026
Update
Big 5 capex refreshed to Q2 2026 actuals/guides (basis-labelled). Alphabet FY26 $195–205B; Microsoft ~$175B CY26 guide; Amazon $173B TTM actual (no full-year guide issued); Meta FY26 raised to $130–145B; Oracle ~$50B. Reported Big 5 ~$725–755B; comparable-basis ~$740–780B. Prior cards citing "MSFT ~$190B" and "Amazon ~$200B" corrected. Demand signal held or strengthened — AWS +37% YoY (fastest in 18 quarters), Azure +43% YoY, TTM past $100B — so the 408 GW base is supported, not challenged.
Financeability is now the binding constraint, not demand. Alphabet posted its first negative-FCF quarter (−$5.9B) since its 2004 IPO; Meta FCF fell ~91% YoY to $784M. The market rewarded metered spend (MSFT/Amazon ~+8–9%) and punished unmetered spend (Meta ~−9–10%) — a spend-to-revenue legibility test. Added to the prime/marginal financeability note and the renewal-risk rationale. Microsoft's 15→25yr useful-life extension noted as a countervailing datapoint suggesting the 8–12% renewal discount may be conservative at the top end.
EE Who Wears the Risk (Apr 2026) · Q2 2026 earnings
Update
Supply side — incremental, no cap change. Meta's Greenlight behind-the-meter power plant took a positive FID in early July 2026; cumulative hyperscaler nuclear commitments now ~10 GW across Microsoft/Google/Amazon/Meta. Both are consistent with the model's existing BTM OEM cap and nuclear tier classification — noted for the trackers, no parameter change.
Data Center Frontier Jun 2026 · World Nuclear News · press Jul 2026
v1.4Jul 30 2026FixStructure
A friction error fix moves the constrained 2040 case from 318 GW to 303 GW; base holds at 408 GW.
Fix
P0 integrity pass — external technical review. (1) S-curve removed as a live driver. The AI-adoption S-curve was computed but never fed the demand calc (dead variable); it is removed from the engine and the sensitivity/waterfall panels, and retained only as evidence framing on the S-Curve tab. Base case unchanged (the variable was never used). (2) LLT friction bug fixed. The constrained scenario's exponent (mod.llt−0.5) reversed the intended drag, giving the constrained case LESS friction than base. Replaced with a transparent multiplier: friction = 1 − (1−baseFriction)×scenarioMult. Constrained 2040 moves from ~318 GW to ~303 GW (12.8% CAGR); accelerated ~494 GW; base 408 GW unchanged. (3) BTM/grid stock-flow reconciled. Grid and BTM are now explicit stocks with a shared additions basis and one OEM cap; total ≡ grid + BTM (reconciliation error 0). Nuclear is an explicit classification subset of grid, not additive. (4) "Why Higher" waterfall rebuilt as a sequential cumulative bridge (base/scope/LLT → Jevons → BTM) with stated ordering; the broken S-curve term is gone. (5) Backtest → Calibration. Tab relabelled; corrected "4 pass" to 3-within-range / 4 scope-explained; flagged 2025 as the calibration anchor and 2027–28 as forward benchmarks; removed a NaN detail row. (6) Capex claim corrected from "$80–95M/MW all-in" to ~$37M/MW (the figure the stated cost stack actually produces); wind capex note aligned to code ($1.6M/MW); totals labelled gross-stock not incremental. (7) Deceleration relabelled as an exogenous assumption of the compute-growth path, not a mechanical output of the training/inference split. (8) Monte Carlo seeded (reproducible); distribution text aligned to code (~1σ); relabelled illustrative uncertainty bands, not calibrated probabilities; independence caveat added. (9) Queue Signal tab quarantined — the 935 GW "queue" is generation+storage interconnection data (Berkeley Lab Queued Up excludes load/BTM) and must not be read as data-centre demand. (10) Publication language softened from "first-principles / fully disclosed" to "transparent scenario model."
External model review · Jul 2026
v1.3.5Jul 28 2026UpdateNew
Alphabet's 2026 capex guidance raised to US$195 to 205 billion, and federal nuclear uprate targets of 2.5 GW and 5 GW added.
Update
Alphabet Q2 2026 capex raised to $195–205B — Reported Jul 22 2026. Up from $180–190B guidance set at Q1. Google Cloud grew 82% YoY to $24.8B; backlog jumped $50B in one quarter to $514B. Big 4 total now ~$740–760B; Big 5 ~$790–810B. S-curve evidence card stat updated from ~$775B to ~$800B. Capex chart 2026 bar updated from $725B to $760B (Big 4 midpoint). Microsoft, Meta, Amazon Q2 results due Jul 29–30 — further revision likely.
DOE UPRISE initiative added to nuclear methodology note — DOE (May 2026) targets 2.5 GW of added nuclear capacity by 2027 and 5 GW by 2029, primarily through uprates at existing plants. Directly validates Tier 1 uprate ramp (Vistra 433 MW). Amentum selected Jul 20 2026 by NNSA to negotiate 1 GW AI DC co-location at Savannah River Site, South Carolina — first federal-site nuclear DC project. Classified Tier 3 (no hyperscaler named); excluded from base.
DOE UPRISE May 2026 · NNSA / Amentum Jul 20 2026
v1.3.4Jul 10 2026NewUpdate
S&P Global's 62 GW US installed capacity benchmark added, alongside a 180 GW on-site power pipeline.
New
S&P Global / 451 Research — 62 GW US installed capacity (Mar 2026) — S&P/451 Research webinar (Jun 2026): US installed DC capacity 62,242 MW as of March 2026, projected to reach 151,734 MW by 2030. 12 GW above FERC 50 GW anchor due to scope difference (S&P includes BTM + enterprise DCs not in FERC grid-connection data). Added to Backtest chart note and sensitivity note as complementary anchor. EE anchors to FERC 50 GW (grid-connected); both figures are consistent once scope is adjusted.
S&P Global / 451 Research Jun 2026 webinar
New
BTM pipeline — 180 GW maximum / 50 GW near-term (S&P Global) — Pipeline of projects planning to co-locate power with data centres has reached 180 GW maximum capacity, ~50 GW expected near-term over next five years. Added to BTM Supply note-strip as independent pipeline validation. Corrected v1.9.2: this entry originally read "EE model 70 GW BTM by 2040 represents 14% of 180 GW maximum … 72–108 GW realised — brackets EE 70 GW base". Both the 70 GW and the derived framings were wrong. The live BTM stock is 58.4 GW, which is 32% of the 180 GW maximum (the "14%" was never 70/180 either — it was the base-share figure mis-filed against the S&P denominator), and the 72–108 GW haircut band brackets neither 70 nor 58.4 — it sits above both.
S&P Global Jun 2026 webinar · Goldman Sachs / Sightline Climate
New
GE Vernova × Blue Energy gas-plus-nuclear hybrid (Jul 2026) — GE Vernova formed collaboration with Blue Energy to advance world's first gas-plus-nuclear plant using BWRX-300 SMRs + GE gas turbines, targeting AI and advanced manufacturing. New supply chain category — hybrid gas-nuclear — not previously in OEM table. Added as note to GE Vernova row pending commercial DC contracts.
ANS Nuclear Newswire Jul 7 2026
v1.3.3Jul 7 2026NewUpdate
Microsoft brought 1 GW online in a single quarter, while only 5 of 12 GW announced for 2026 is under construction.
New
Microsoft 1 GW commissioned in one quarter — Q2 FY2026: Microsoft stood up 1 GW of data centre capacity in a single quarter, vs 2 GW in the whole of FY2025. Added as S-curve evidence card — commissioned capacity is harder evidence than capex spend. Amazon added 3.9 GW over past 12 months, doubling its 2022 base, and expects to double again by 2027. Five 1 GW+ campuses expected across industry in 2026 per Epoch AI.
Data Center Dynamics Jun 2026 · Epoch AI May 2026
New
Sightline Climate delivery data added to Limitations — Tracked 12 GW of 2026 US DC capacity announced across 140 projects; only 5 GW under construction. 11 GW at "announced" stage with no physical progress; 25% of projects have disclosed no power strategy. Independent confirmation of Goldman 40–50% delivery haircut. Two sources now converge on same slippage rate.
Sightline Climate Jun 2026
Update
Goldman $5.3T 2025–2030 cumulative capex added — Goldman raised combined hyperscaler capex estimate from $4.5T to $5.3T for FY2025–FY2030 following Q1 2026 earnings. An $800B upward revision in a single quarter of analyst updates. Added to $775B S-curve evidence card context.
Goldman Sachs Q1 2026 earnings note
v1.3Jun 21 2026UpdateNew
On-site gas ceiling raised to 5, 8, 10 and 14 GW per year; the 2040 base case still holds at 408 GW.
EE Modular Gas Tracker Jun 2026 · Goldman Sachs Aterio May 2026
New
FERC June 18 2026 orders — Six tailored orders compelling all regional grid operators to justify tariff structures within 60 days or reform them; 30-day mandatory reliability report. Model implication: LLT friction parameter may loosen faster than base case assumes for 2027–2030. Added to Queue Signal note-strip.
FERC Docket RM26-4-000 Jun 18 2026
Update
Hyperscaler capex updated to ~$775B Big 5 — Big 4 (Google/Amazon/Microsoft/Meta) $725B per FT Q1 2026 earnings; Oracle $50B FY2026 guidance. Individual: Amazon ~$200B, Alphabet $180–190B, Microsoft ~$190B, Meta $125–145B, Oracle ~$50B. S-curve capex chart 2026 bar updated from prior estimate to $725B.
Financial Times Q1 2026 earnings · Oracle FY2026 guidance
Update
CoreWeave updated — 850 MW active + 3.1 GW contracted (end-2025) — Was: 2.9 GW contracted Q3 2025 at 5:1 ratio. New: 850 MW active across 43 DCs, 3.1 GW contracted, virtually all expected online by 2027. New S-curve evidence card added.
CoreWeave Q4 FY2025 earnings Mar 2026 · DCD Mar 2026
New
Goldman scope note added to Analyst Comparison — Goldman forecasts power demand (41 GW in 2026, 66 GW in 2027) vs EE installed capacity. Goldman applies 40–50% delivery haircut to forward pipeline. Scope difference now explicitly disclosed in note-strip.
Goldman Sachs Aterio analysis May 2026
v1.2Jun 19 2026UpdateNewFixStructure
Nuclear contribution in 2040 raised from about 15 GW to 24 GW, and Bloom Energy added as a fourth on-site platform.
Update
Nuclear base raised to 24 GW — Meta Jan 2026 deals added — Vistra operating (2,176 MW firm) and uprates (433 MW firm) now fully in base per Vistra IR. Meta/TerraPower and Meta/Oklo reclassified Tier 2 (development agreements per World Nuclear News, not firm PPAs). Total: 13 deals, 9.7+ GW committed. Nuclear ramp: 0 → 5.3 GW (2027) → 12.2 GW (2030) → 24.2 GW (2040). Was: ~15 GW.
Vistra Corp IR Jan 9 2026 · World Nuclear News Jan 2026
Fix
TMI/Crane date corrected — was "online late 2024" (wrong) — Confirmed: PPA signed Sep 2024; restart targeted late 2027 per Constellation CEO Jun 2025 (accelerated from 2028); PJM grid ready ~2031 (execution risk). Nuclear 2024 MW set to 0; 835 MW lands in 2027 ramp. Model had this wrong since v1.1.
Reuters/Fox Business Jun 2025 · Fox43 Mar 2026
New
Bloom Energy added as fourth BTM platform — Non-combustion process qualifies as EPA minor source, bypassing full permitting. 8–12 week deployment. Proven max: 73 MW (AWS Ohio). Project Jupiter (NM): 2.45 GW announced primary power, under construction. Added to OEM table and BTM assumptions panel.
EE Bloom Energy analysis May 2026
New
PJM BYONG pathway named — Jan 2026 CIFP: large load defined as ≥50 MW/POI; BYONG expedited pathway operational mid-2026. Model implication: national LLT coefficient may slightly overestimate LLT suppression in PJM territory. Added to Queue Signal note-strip and Limitations tab.
EE PJM CIFP analysis Jan 2026
New
EPC delivery risk added to Limitations — 20–30% overruns realistic central scenario in 2026. Transformer lead times 3–5 yr, switchgear 18 mo. GE Vernova backlog 55 → 62 GW in Q3 2025. Goldman 40–50% delivery haircut as named anchor.
EE CoVolt/EPC analysis May 2026 · Goldman Sachs Aterio May 2026
Structure
Nuclear tier table added to Methodology — Four-row source confidence framework: Firm operating/contracted (green), Restart (amber), Advanced nuclear development (amber), Optional/speculative (gray). World Nuclear News citation for Meta/TerraPower and Oklo as development agreements rather than firm PPAs.
Internal
Fix
FTM chart ID mismatch fixed — HTML had id="chart-ftm-btm"; JS expected "chart-f40-ftmbtm". Chart was silently not rendering on the Forecast 2040 tab.
Renewal risk variable added — 8% haircut at 2030–2033 lease renewal window. Meta/Blue Owl Hyperion JV and OpenAI lease terms as calibration anchors. Source: EE Who Wears the Risk Apr 2026, Crusoe-Oracle-OpenAI capital stack.
EE Who Wears the Risk Apr 2026
New
Monte Carlo added — 2,000 runs — P10/P25/P50/P75/P90 bands. Six analytical-judgement parameters varied simultaneously. P10 ~270 GW, P50 408 GW, P90 ~550 GW at 2040.
Internal
Update
Jevons rebound calibrated to RAND data — RAND RRA3572-1: installed AI compute 2.3×/yr vs chip efficiency 1.28×/yr (2020–2025). Rebound coefficient set to 0.72. Accounts for ~18 GW of gap between EE 2030 (115 GW) and IEA/BloombergNEF comparable-scope consensus (~80 GW).
RAND Corporation RRA3572-1 2025
Update
FERC 50 GW baseline confirmed — FERC State of Markets 2026: 50 GW installed capacity, 24% CAGR 2020–2025. Set as primary model anchor replacing prior 45 GW estimate.
FERC State of Markets 2026
Update
LLT friction parameter added — 77 utility tariff regimes across 36 states. Georgia Power PLL vs Dominion GS-5 as bimodal exemplars. National average held at 0.018/10pp coverage. Half of regimes approved since Jan 2025.
EE Tariff Lottery May 2026
Fix
Grid constraint bug fixed — Modifier previously applied to full stock level; corrected to apply to buildout rate only. No material effect on final 2040 output but improved trajectory accuracy 2028–2033.
Internal QA
Structure
Scenario Builder added — Six sliders: Jevons, grid buildout, LLT friction, compute growth, BTM bypass, renewal risk. Shows % delta from base, not absolute values. All six sliders verified effective after override-range fix (full parameter families overridden, not single-point anchors).
Internal
Structure
Ten tabs added — Forecast 2040, Methodology, Backtest, Limitations, Queue Signal, vs Analysts, Generation, Why Higher, S-Curve, BTM Supply. Changelog added as eleventh tab in v1.2.
Internal
v1.0Dec 2025New
First release: 408 GW of US data center capacity by 2040, growing 15.0 percent a year from 50 GW.
Base output: 408 GW by 2040 at 15.0% CAGR — From 50 GW baseline. Grid ex-nuclear 326 GW + BTM gas 58.4 GW + Nuclear 24.2 GW. Accelerated 494 GW · Constrained 303 GW · Efficiency Breakthrough 135 GW. Corrected v1.9.1: the constrained figure on this line read 318 GW against a live model output of 303 GW. A 15 GW gap is three times the file's own "≥5 GW = major version" threshold and should not have survived the v1.3 LLT-friction fix that moved the constrained case. All four scenario figures on this line now render from computeForecast2040() and cannot drift again. Accelerated (493→494) and Efficiency Breakthrough (135→135) were within rounding; the constrained figure was not. Corrected v1.9.2: the supply decomposition on this line also read 314 / 70 / 24 GW (grid ex-nuclear / BTM / nuclear) against a live 325.9 / 58.4 / 24.2. The BTM figure was the pre-v1.3 value; the grid figure had been back-solved from it to make the three sum to 408. All three now render from EE_SYNC.forecast.supply.
EE primary research · FERC State of Markets 2026
Versioning: major = headline GW changes ≥5 GW; minor = new parameters, data sources, or tabs; patch = copy corrections and citation updates. No version removes data — superseded figures retained with replacement noted. Model weights, parameter names and calibration sources are all disclosed — FM40_WEIGHTS is a plain object literal in the page source (de-obfuscated in v1.9.1). · Source questions: electroneconomics.substack.com
Electron Economics · Track record
Accuracy Ledger
Every figure this platform has published and later changed, across all seven models, with the size and direction of each change and what caused it. Research products ask you to trust their numbers. This is the evidence for how often those numbers have moved and by how much.
Revisions by model
Split between our own errors and figures that moved because new information arrived. A model with more revisions is not a worse model; it is usually a more heavily used one.
Our errorNew information
What actually goes wrong
Cause of our own errors only. New-information updates are excluded, because they are not mistakes.
Every revision
Rows marked with a green edge belong to the model you are looking at. Sorted by size of change.