Three-House Cross Research
AI / Data Centre Power Demand
GS SUSTAIN
JPM AM
MS Research
GS: Apr'24 / Oct'25 / Feb'26 / May'26 (Commodities) · JPM: GTA 1Q26 + 2Q26 · MS: "Flexible Power" Mar'26
Key Takeaways
Six messages for the Investment Committee from 7 reports across three houses
1
Demand has converged — and GS May'26 just put a hard near-term anchor on itGS + JPM + MSGS May'26 ✨
GS +220%, JPM +228%, MS 320 GW build-out — three independent teams arrive at near-identical conclusions on the 2030 picture. GS Commodities May'26 adds the operational anchor: US DC power demand goes from 31 GW (2025) to 66 GW (2027) — more than doubles in 24 months. US DC capacity reaches 94.7 GW by end-2027. DC share of US peak summer power: 4.1% → 8.5%. This is equivalent to adding a new country the size of Japan to global electricity in 24 months. US power demand CAGR of 3.2-3.8% is the highest since the 1990s.
2
Supply constraint is real but unevenly distributed — and only 50-60% of planned capacity actually landsCRITICALJPM 2Q26 ✨GS May'26 ✨
Raw generation capacity (Bottleneck 1) is solvable — JPM 2Q26 confirms 221.2 GW of US grid additions planned over 2026-28 with batteries doubling (+115%) and gas additions halving (-52%). But the new GS Commodities May'26 realisation analysis is the missing piece: historically only 72% of DCs scheduled within 4 quarters come online on time; the haircut tightens to 60% next year, 50% over two years. Raw schedule reaches 133.3 GW US DC capacity by Dec'27 but GS forecast caps at 94.7 GW (38.6 GW / 29% haircut). The true binding constraint remains grid interconnection & T&D labour — GS quantifies a 78,000 skilled worker gap requiring 3-4 years of training. The slippage is structural, not cyclical.
3
Inference load flexibility changes everythingMS NEW
As AI inference rises to 45-50% of DC workload, power demand shifts from stable/predictable (training) to volatile/spiky (inference). The system challenge shifts from "can we supply enough power" to "can we supply power flexibly enough." This creates a new investment category: energy storage systems (ESS) for millisecond-response peak shaving and frequency regulation.
4
Sector rotation follows a clear sequence — ESS is next, and JPM 2Q26 just validated itMS ROTATIONJPM 2Q26 ✨
MS maps the AI power investment rotation: Nuclear (2023-24, +4.8x) → Gas Generators (2024-25, +3.2x) → Backup Generators (2025, +2.1x) → Fuel Cells (2025-26, +4.5x) → ESS (next). Each wave lasted 6-12 months. The JPM 2Q26 data confirms the rotation is already happening in the supply mix: 26% of new 2026-28 US capacity is batteries, up from 12% in 1Q26. Direct read-through to TSLA Megapack, FLNC, CATL, BE BTM. Understanding where you are in this rotation is more important than the demand headline number.
5
GS warns we are still in the "Appraisal" phase — but transition signals are building
Using their AI Innovation Cycle framework (Shale Oil analogy), GS places AI infrastructure in the Appraisal/Hopes & Dreams phase — the best window for infrastructure equities. But reinvestment rates at 87% and CROCI declining from 31% toward the 24% historical low are early warning signals. Three transition triggers to monitor: financial inflexibility (not triggered), return erosion (deteriorating), product oversupply (not triggered).
6
Four dimensions, one complete frameworkSYNTHESIS
Goldman Sachs SUSTAIN tells you WHERE in the cycle (6P constraints, Innovation Cycle phase, BTM framework). GS Commodities May'26 ✨ adds WHEN it lands: 31→66 GW in 24 months with a 50-60% realisation haircut, plus regional reliability tiering (PJM/MISO/BPA at risk). J.P. Morgan tells you HOW BIG the gap is (219 GW global DC capacity by 2030, 221.2 GW US grid additions, >2,600 GW queue, 9-18 GW shortfall). Morgan Stanley tells you WHERE money goes next (inference flexibility, ESS rotation, Na-ion cost curve). All four dimensions now converge on: flexibility-led capacity build, ESS as the next investable wave, regional concentration of investment opportunity (TX / GA / VA = winners; PJM / MISO / BPA = stress beneficiaries; TN / NE / FL = excluded).
Goldman Sachs
WHERE in the cycle + WHEN it lands
6P constraint framework · AI Innovation Cycle / Shale analogy · BTM 14 GW · Vera Rubin server data · CO₂ social cost $145-170B · AI drug discovery value $83-412B · Green Reliability Premium · 87% reinvestment rate · 32→82 Buy-rated stocks · May'26 (Commodities) ✨ US DC 31→66 GW by 2027 · 94.7 GW capacity end-2027 · 8.5% peak summer share · 60%/50% realisation haircut · Regional reliability tiering (PJM/MISO/BPA at risk)
J.P. Morgan
HOW BIG the gap
2Q26 ✨ 82→219 GW global DC capacity (CAGR 21.7%) · AI workload 54%→71% of DC · US DC: 220→600 TWh (12% of US power) · 221.2 GW US 2026-28 additions (Batteries 26%, Gas 13%, Solar 49%) · 35.1 GW retirements · DC power mix: Tech 45% / Cooling 38% / Power conv 11% / Network 5% · >2,600 GW interconnection queue
Morgan Stanley
WHERE money goes next
Training vs inference load profiles · ESS 321 GWh (bull: 590) · Sector rotation map · 9-18 GW US shortfall (Exhibit 12) · Na-ion cost RMB 0.32→0.21/Wh · BTM ESS IRR 23%/36% · System rebalancing to flexibility · CATL, Tesla, LGES, Fluence, BYD
Detailed Headline Forecast Comparison
All data points across six publications for line-by-line comparison. GS May'26 and JPM 2Q26 are the latest releases.
| Metric | GS Apr'24 | GS Oct'25 | GS Feb'26 | GS May'26 ✨ | JPM 1Q26 | JPM 2Q26 ✨ | MS Mar'26 | Signal |
|---|---|---|---|---|---|---|---|---|
| DC Power Growth 2030 vs 2023 | +160% | +175% | +220% | US +113% to 2027 | +228% | +220% | 320 GW | CONVERGED |
| Global DC Demand 2030E (TWh) | ~1,068 | 1,131 | 1,316 | N/A (US-only) | ~1,350 | ~1,310 | N/A | tightening |
| Global DC Capacity 2030E (GW) ✨ | N/A | N/A | N/A | N/A (US-only) | N/A | 219 GW | N/A | NEW |
| Global DC Capacity 2025 base (GW) ✨ | N/A | N/A | N/A | N/A | N/A | 82 GW | N/A | CAGR 21.7% |
| US Power CAGR to 2030 | 2.4% | 2.6% | 3.2% | near-term sharp | ~3.1% | ~3.2% | 3.8% | MS highest |
| AI Share of DC 2030E | ~20% | ~39% | ~50% | N/A | ~65% | ~71% | 45-50% | JPM ↑↑ |
| DC % of US Power 2030E | 8% | 11% | ~14% | 8.5% (2027 peak) | 13.5% | ~12% | N/A | CONVERGED |
| Hyperscaler CapEx+R&D 2026E | N/A | N/A | >$1T | N/A | $664B+ | N/A | N/A | Unprecedented |
| US 2026-28 Total Additions (GW) ✨ | N/A | N/A | N/A | ~62 DC-only* | 222 | 221.2 | N/A | refined |
| US Batteries Additions 26-28 (GW) ✨ | N/A | N/A | N/A | N/A | 26.7 | 57.5 (26%) | N/A | ↑115% ✓ESS |
| US Nat Gas Additions 26-28 (GW) ✨ | N/A | N/A | 42 peaker+12 CCGT | N/A | 60 | 28.8 (13%) | 15-20 | ↓52% ⚠ |
| US DC Demand 2025-28 (GW) | N/A | N/A | N/A | 31→66 by 2027 | N/A | N/A | 74 GW | MS unique |
| Net Shortfall 2025-28 (GW) | N/A | N/A | N/A | 3 high-risk regions | N/A | N/A | 9-18 GW | MS unique |
| DC ESS Deploy 2030E (GWh) | N/A | N/A | N/A | N/A | N/A | N/A | 321 (bull:590) | MS unique |
| BTM Solutions (GW) | N/A | N/A | 14 GW | N/A | N/A | N/A | BE 5-8 GW | GS+MS |
| CO₂ Increase (mn tons) | 215-220 | 215-220 | 285-290 | N/A | N/A | N/A | N/A | ↑35% |
| Innovation Cycle / Next Rotation | N/A | N/A | Appraisal | N/A | N/A | N/A | ESS next | Complementary |
| Na-Ion Cost (RMB/Wh) | N/A | N/A | N/A | N/A | N/A | N/A | 0.32→0.21 | MS unique |
| Reinvestment Rate 2026E | N/A | N/A | 87% | N/A | N/A | N/A | N/A | GS warning |
| US DC Demand 2025 (GW) ✨ | N/A | N/A | N/A | 31 GW | N/A | ~25 GW | N/A | GS NEW |
| US DC Demand 2026E (GW) ✨ | N/A | N/A | N/A | 41 GW | N/A | ~33 GW | N/A | GS NEW |
| US DC Demand 2027E (GW) ✨ | N/A | N/A | N/A | 66 GW | N/A | ~41 GW | N/A | GS NEW |
| US DC Capacity end-2027 (GW) ✨ | N/A | N/A | N/A | 94.7 GW | N/A | N/A | N/A | vs 133.3 raw |
| DC % of US Peak Summer 2027 ✨ | N/A | N/A | N/A | 8.5% | N/A | N/A | N/A | 2x in 2yrs |
| Realisation Rate (Adjusted) ✨ | N/A | N/A | N/A | 60% / 50% | N/A | N/A | N/A | vs 72% hist |
*Note on US 2026-28 Total Additions: GS May'26's ~62 GW is DC-specific capacity, JPM's 221.2 GW is total US grid additions (all sectors). Not directly comparable. GS realisation-adjusted DC additions: 2026 ~13.7 GW + 2027 ~18 GW (50% of 36.3 raw schedule) + 2028 estimate. GS Feb'26 figure "42 peaker + 12 CCGT" refers to gas-fired additions specifically.
Global DC Power Demand Trajectory (TWh)
GS revised 3x: +160% → +175% → +220%. Now within 3% of JPM's +228%. Equivalent to adding a Top 10 power consuming country.
Forecast Evolution: 23 Months of Upward Revisions
Apr 2024 (GS)+160%(1,068 TWh)
Initial framework. Introduced Jevons Paradox / 3-constraint model (budget, demand, no constraint). AI = ~20% of DC. ChatGPT = 6-10x power per query vs Google search. 32 Buy-rated stocks.
Oct 2025 (GS)+175%(1,131 TWh)
AI share doubled to 39%. Introduced 6P constraint framework. Added labour analysis (78K T&D gap). $790B grid capex. Green Reliability Premium $40/MWh. 82 GW capacity needed.
Feb 2026 (GS)+220%(1,316 TWh)
Major revision. >$300B hyperscaler capex upward revision. 87% reinvestment rate. Vera Rubin integrated. 105 GW (incl 14 GW BTM). CO₂ up to 285-290 mn tons. US CAGR 3.2%. AI Innovation Cycle in Appraisal phase.
1Q 2026 (JPM)+228%(~1,350 TWh)
Independent validation. US DC: 168→620 TWh. AI = 65% of DC capacity. Hyperscaler capex $664B+. Binding constraint: baseload mismatch and interconnection queue. 222 GW total US additions.
2Q 2026 (JPM, Apr 30)219 GW(global DC 2030)
NEW DATA: Global DC capacity 82→219 GW (2025-30, CAGR 21.7%). AI workload share of DC: 54%→71% (2025→30). US DC consumption: ~220→~600 TWh, reaching 12% of total US power. US 2026-28 build: 221.2 GW additions (revised mix: Solar 49% / Batteries 26% / Gas 13% / Wind 11%) vs 35.1 GW retirements. Material shift: Batteries doubled vs prior, gas additions cut in half — validates MS ESS rotation thesis.
Mar 2026 (MS)320 GW
New dimension: inference flexibility. ESS as next rotation. 74 GW US demand with 9-18 GW shortfall. DC ESS: 321 GWh by 2030. Sodium-ion cost disruption. BTM ESS IRR 23%/36%.
May 2026 (GS Commodities) ✨+113%(US DC 31→66 GW by 2027)
NEW DATA — Near-term operational view: US DC power demand 31 GW (2025) → 41 GW (2026) → 66 GW (2027) — more than doubles in 24 months. US DC capacity reaches 94.7 GW end-2027 at 70% utilisation. Realisation haircut: historical 72% on-time → only 60% next year → 50% over 2 years materialise (raw schedule reaches 133.3 GW Dec'27, GS adjusted 94.7 GW = 38.6 GW haircut). DC % of US peak summer power: 4.1% (2025) → 5.3% (2026) → 8.5% (2027). Regional tiering (NEW): Mid-Atlantic / Mid-Continent / Northwest = elevated reliability risk; TX / GA = marginal tightening; TN / NE / FL = critically tight already. Authors: Wei, Struyven, Dart.
GS Commodities May'26 ✨ US DC Capacity Trajectory — Raw Schedule vs Risk-Adjusted
Monthly data (May'26 → Dec'27) read directly from GS chart. Headline: 31 → 41 → 66 GW demand (capacity 44 → 59 → 94.7 GW end-Dec'27 at 70% utilisation, matches article exactly)
2025 Demand
31 GW
2026E Demand
41 GW
2027E Demand
66 GW
2027 / 2025
+113%
This is GS Commodities Research's most operational forecast to date — bottom-up from Aterio facility-level data (locations, permitting, construction status, satellite imagery). Both lines start identically at 47.1 GW (May'26) and diverge progressively: raw schedule reaches 133.3 GW by Dec'27 while GS risk-adjusted forecast caps at 94.7 GW. Capacity utilisation assumption: 70%. The Dec'27 GS forecast of 94.7 GW × 0.70 = 66.3 GW demand validates the article's "66 GW in 2027" headline to within rounding.
Realisation Haircut — Why the GS Line Diverges from Raw Schedule
YoY additions scheduled: 13.6 GW (2026), 36.3 GW (2027) vs realised: 6.4 GW (2024), 8.5 GW (2025). Historically only 72% of DCs scheduled for activation in the next 4 quarters actually come online on time. GS applies a tighter haircut: ~60% of next-year capacity, ~50% over two years. Net result by Dec'27: raw 133.3 GW vs adjusted 94.7 GW = 38.6 GW haircut (29% of raw). Drivers of slippage: developers submitting in multiple regions and choosing only the best site, supply chain, labour shortages, 18-24 month build cycle post-permit.
GS May'26 ✨ DC Share of US Peak Summer Power
Most direct measure of grid stress: how much of the summer peak DCs eat. 4.1% → 8.5% in 2 years.
This is the cleanest single metric for grid stress. Peak summer demand is when reliability margins are thinnest — adding DCs that are essentially 24/7 baseload to an already stressed peak window creates outsized grid strain. The 8.5% by 2027 is more than double the 2025 baseline in just 24 months. No comparable load category has ever scaled this fast against the US summer peak.
⚠️ Reliability Read-Through
If DC share grows to 8.5% of peak by 2027, the marginal reserve margin in the most stressed regions (PJM = Mid-Atlantic, MISO = Mid-Continent, BPA = Northwest) compresses materially. GS flags these three regions as having planned generation additions insufficient to absorb the incoming DC load — implying potential brownout risk or forced load-shedding during summer peaks. Direct beneficiaries: backup power (Generac), BTM fuel cells (BE), demand response platforms, and ESS for peak shaving (TSLA Megapack, FLNC).
GS May'26 ✨ Regional Power Reliability Tiering
2027 annual DC additions in each of Mid-Atlantic / Texas / Mid-Continent are individually scheduled to exceed the entire nation's total 2025 additions.
| Region | RTO / Grid | Reliability Risk | Driver |
|---|---|---|---|
| Mid-Atlantic | PJM (VA, MD, PA) | 🔴 ELEVATED | Largest DC cluster (Loudoun). Planned generation lags incoming demand. |
| Mid-Continent | MISO (IL, IN, MI, IA) | 🔴 ELEVATED | Coal retirements + new DC siting. Planned gen insufficient. |
| Northwest | BPA / WECC (WA, OR) | 🔴 ELEVATED | Cheap hydro draws DCs, but limited new generation. May refuse some projects. |
| Texas | ERCOT | 🟡 MARGINAL | Heavy DC additions but solid generation pipeline (gas + solar + storage). |
| Georgia | Southern Co. / SERC | 🟡 MARGINAL | Vogtle nuclear + new gas. Major DC growth absorbed. |
| Tennessee | TVA | 🟣 CRITICALLY TIGHT | Constrained DC additions — already at capacity ceiling. |
| New England | ISO-NE | 🟣 CRITICALLY TIGHT | Pipeline + gas constraints. Won't see major DC builds. |
| Florida | FRCC | 🟣 CRITICALLY TIGHT | Constrained — DC growth bottlenecked by generation. |
GS May'26 ✨ Investment Implications by Region
Where the regional tiering points capital in 2026-27
🔴 Elevated Risk Regions (PJM / MISO / BPA): Backup power becomes mandatory — Generac portable+stationary, fuel cells (BE BloomEnergy, FCEL), and demand response. ESS for peak shaving (TSLA Megapack, FLNC) sees pricing power. Co-location at owned generation becomes critical (ETR, CEG, VST own-and-supply model).
🟡 Marginal Tightening (TX / GA): These remain the highest-growth DC markets in absolute GW terms. Hyperscalers will pivot toward TX in particular — ERCOT's faster interconnection process is a major edge. Read-through to solar+storage developers focused on ERCOT (NextEra residential is less relevant; utility-scale developers are the play). Nuclear in GA (Vogtle units 3&4) is a unique national asset.
🟣 Critically Tight (TN / NE / FL): These regions are excluded from the next wave of DC capex. Watch for spillover demand to neighbouring regions (TN → MISO/PJM border zones, NE → upstate NY ISO).
Cross-House Read
GS May'26 regional tiering complements MS Mar'26's ESS rotation thesis — the elevated-risk regions are precisely where ESS economics will pay first (peak shaving + frequency regulation at scale). The JPM 2Q26 mix shift (batteries +115%, gas -52%) is the macro confirmation. The three houses now converge on: flexibility-led capacity in stressed regions = the highest IRR sub-segment in AI power.
JPM 2Q26 ✨ Global Data Centre Capacity (GW) — AI vs Non-AI
82 GW (2025) → 219 GW (2030) — CAGR 21.7%. AI workload share rising from 54% → 71%.
2025 Base
82 GW
2030 Forecast
219 GW
AI Workload 2030
~155 GW
CAGR 5yr
+21.7%
JPM's 2Q26 release (April 30, 2026) introduces a GW-denominated capacity forecast that complements the existing TWh trajectory. The 2.67x build-out by 2030 is consistent with GS Feb'26 (105 GW US + 72 GW RoW = 177 GW DC-specific) and MS (320 GW total power including non-DC). AI workload rises from ~54% of capacity in 2025 to ~71% by 2030 — at the high end of all houses, exceeding GS Feb'26's 50% and MS's 45-50%.
Source: JPM AM Guide to Alternatives, 2Q26 (Apr 30, 2026), p.39 — citing IEEE Communications and McKinsey & Company forecasts. Implication for IC: A 137 GW capacity build over 5 years requires ~$4.1tn cumulative capex at $30bn/GW DC capex (excl. compute), tracking the GS hyperscaler reinvestment rate of 87%.
JPM 2Q26 ✨ Data Centre Power Consumption Breakdown
Where the GW actually goes — informs which equipment vendors capture the spend
| Component | % Power | 2030 GW Equivalent | Key Beneficiaries (our mapping) |
|---|---|---|---|
| Tech equipment (GPU/CPU/Memory) | 45% | ~99 GW | NVDA, AMD, MU, SNDK, TSM |
| Cooling | 38% | ~83 GW | Vertiv (VRT), liquid cooling players |
| Power conversion & regulation | 11% | ~24 GW | Eaton, Schneider, Vertiv (VRT) |
| Networking (optics, switches) | 5% | ~11 GW | COHR, LITE, ALAB, CRDO, GLW |
| Lighting | 1% | ~2 GW | (de minimis) |
Key Investment Insight
Cooling at 38% of DC power is the surprise number. This is up from ~25-30% historically because liquid cooling at high-density AI racks (e.g. Vera Rubin NVL8 at 24kW/server) requires far more pumping power than air cooling. Read-through: Vertiv (VRT) and other thermal management vendors capture both cooling (38%) and power conversion (11%) = ~49% of DC power spend — a structurally larger wallet share than the headline narrative implies.
JPM 2Q26 ✨ U.S. DC Power Consumption + % of Total US Power
McKinsey-sourced forecast: U.S. DC consumption 220 → 600 TWh by 2030, reaching ~12% of total US power
Critical anchor for the entire "AI power" investment thesis: data centres rise from ~5% of US power demand in 2025 to ~12% by 2030. This is the single largest concentrated source of incremental US power demand growth — the entire residential sector for comparison contributes only ~0.6pp to the 3.2% CAGR per GS. The 2.7x growth in 5 years is the structural tailwind behind every name in the AI power supply chain — generators (GEV), fuel cells (BE), ESS (TSLA/FLNC), and optical networking (COHR/LITE/ALAB).
Cross-house reconciliation: JPM 2Q26's 12% (~600 TWh) sits between GS Feb'26 (~14% = ~790 TWh) and the prior consensus. The difference is largely how aggressive each house is on AI inference scaling — JPM 2Q26 implies more efficient AI compute per query than GS Feb'26.
JPM 2Q26 ✨ U.S. 2026-28 Generation Mix — Major Revision
Total 221.2 GW additions (vs 1Q26 estimate of 222 GW). But the mix has shifted dramatically.
| Source | JPM 1Q26 | JPM 2Q26 | Change | % |
|---|---|---|---|---|
| Solar | 111.1 GW | ~108.4 GW | -2.4% | 49% |
| Batteries / ESS | 26.7 GW | ~57.5 GW | +115% | 26% |
| Natural Gas | 60 GW | ~28.8 GW | -52% | 13% |
| Wind | 22.2 GW | ~24.3 GW | +9% | 11% |
| Other | 2.2 GW | ~4.4 GW | +100% | 2% |
| TOTAL ADDITIONS | 222.2 GW | 221.2 GW | -0.5% | 100% |
| Total Retirements (Coal 69% / Gas 30%) | 34 GW | 35.1 GW | +3.2% | — |
| NET ADDITIONS | 188 GW | 186.1 GW | -1.0% | — |
⚡ Material Mix Shift — Validates MS ESS Rotation Thesis
Batteries doubled (+115%, 26.7 → 57.5 GW) while gas additions halved (-52%, 60 → 28.8 GW). This is exactly the rotation MS flagged in March: as inference workload makes DC power demand spiky and unpredictable, the marginal capacity addition shifts from baseload gas to flexible ESS for peak shaving and frequency regulation. The 2026-28 grid is being built more for flexibility than for raw capacity. Direct read-through to ESS platforms (TSLA Megapack, FLNC, CATL) and BTM fuel-cell flexibility providers (BE).
AI Share of DC Power (%) — All Houses
JPM 2Q26 most bullish at 71% (revised up from 1Q26's 65%). GS Feb'26 + MS converge at ~50%. Gap widened.
The AI share divergence is the single biggest remaining disagreement across houses. JPM 2Q26 revised up to 71% (from 65% in 1Q26) implies even faster inference scaling and broader AI adoption. MS aligns with GS at 45-50%, framing inference as the dominant workload driver by 2030.
US Power Demand CAGR — Five Estimates
Levels not seen since the 1990s. MS highest at 3.8% (incl non-DC drivers).
GS Feb'26 breakdown of 3.2% CAGR: Front-of-meter DC = 1.5pp, BTM DC = 0.5pp, residential = 0.6pp, commercial (ex-DC) = 0.4pp, industrial = 0.5pp, transport = 0.2pp, other = -0.4pp. Data centres alone contribute 2.0pp.
GS: Hyperscaler Investment Surge
CapEx + R&D exceeds $1T in 2026E. Reinvestment rate 87% of OCF.
| Metric | Prior Est. | Feb'26 Rev. | Change |
|---|---|---|---|
| 2026-27 CapEx+R&D | ~$700B | >$1,000B | +$300B |
| Reinvestment Rate '26 | 79% | 87% | +8pp |
| Net Debt/EBITDA | ~0.3x | ~0.3x | Stable |
| CROCI 2027E | ~31% | 28.7% | -2.3pp ⚠ |
| Hyperscaler EBITDA '27E | $972B | $1,079B | +$107B |
Over the last two months, GS analyst forecasts for 2026-27 hyperscaler CapEx + R&D rose by over $300 billion. Balance sheets remain healthy at 0.3x ND/EBITDA. However, CROCI declining from 31% toward the 24% historical low-end is an early warning signal for AI Innovation Cycle phase transition.
GS: AI Innovation Cycle — Shale Analogy
Currently in Appraisal / Hopes & Dreams Phase — best window for infrastructure equities
ExplorationComplete
Public/private companies pursue unlocking new opportunities. Higher risk. (GPU development, early LLM research)
Appraisal / Hopes & DreamsCURRENT ← We are here
Street most bullish. Multiple expansion for infrastructure. Thematic investment dominates stock-picking. Power supply chain +196pp since 2025.
Execution / EfficiencyApproaching — monitor signals
Stock-picking replaces thematic buying. Focus shifts to corporate returns, balance sheets, market share. Multiples compress for laggards.
Technology Extension / LegacyFuture
Potential second leg via optimisation or wider applications. Mature technologies.
Three Phase Transition Signals:
1. Financial inflexibility — reinvestment rate at 87%, but balance sheets healthy (0.3x). NOT triggered.
2. Corporate return erosion — CROCI declining from 31% toward 24% low. DETERIORATING.
3. Product oversupply — compute/token demand still voracious. NOT triggered.
2. Corporate return erosion — CROCI declining from 31% toward 24% low. DETERIORATING.
3. Product oversupply — compute/token demand still voracious. NOT triggered.
MS: Training vs Inference Load Profiles
Core MS insight: as inference share rises, power demand characteristics fundamentally change
Training
Hyperscalers, academia, startups
Months-long continuous GPU runs
Stable, predictable load curve
Location-flexible (remote OK)
Baseload power sufficient
Workload: batch processing
Power demand: STABLE & PLANNABLE
Inference
Software, auto-drive, search, ads
Real-time user-triggered requests
Spiky loads with sudden large ramps
Latency-sensitive → must be urban
Millisecond-response flexibility needed
Workload: streaming + real-time
Power demand: VOLATILE & SPIKY
Inference rising to 45-50% of DC load by 2030 means power systems must shift from "can we supply enough" → "can we supply flexibly enough"
GS: Jevons Paradox & Emissions Impact
Budget constraint operating — efficiency gains absorbed by rising demand
Jevons Paradox: Three Scenarios
Budget Constrained ← CURRENT
Same budget, 2x compute speed, 1/5 servers. Max power +20%. Efficiency gains drive increased demand at same spend.
Demand Constrained
AI solutions well-defined. Efficiency → 1/10 servers, -50% money, -40% power. Would signal Execution Phase.
No Constraint
Same servers, 4x budget, 9x compute, 5x power. Pure scale-up. Unlikely but highest demand scenario.
Emissions Impact (GS Feb'26)
DC CO₂ emissions revised to 285-290 million tons (2030 vs 2023), up 35% from 215-220 mn. Driven by higher demand + BTM simple-cycle nat gas + softer renewables PPAs. Social cost PV: $145-170 billion at $190/ton. Partially offset by AI drug discovery value ($83-412B).
US Data Centre Capacity Build-Out
GS: US DC capacity 32 GW (2025) → 95 GW (2030). RoW: 42 GW → 72 GW.
| Metric | 2023 | 2025E | 2027E | 2030E | CAGR |
|---|---|---|---|---|---|
| Global DC Power (TWh) | 411 | 607 | 870 | 1,316 | +26% |
| US DC Power (TWh) | ~165 | ~305 | ~520 | ~790 | +25% |
| US DC Capacity (GW) | ~25 | 32 | 55 | 95 | +21% |
| RoW DC Capacity (GW) | ~30 | 42 | 55 | 72 | +13% |
| DC as % of US Power | ~4% | ~5.5% | ~9% | ~14% | — |
Of the 905 TWh growth by 2030, GS sees ~60% in the US (up from ~50% previously). Hyperscale + cloud workloads growing at 14% CAGR; AI workloads at 98% CAGR over 2023-26.
AI Drug Discovery Value (GS Healthcare)
First quantification of AI's "Pervasiveness" benefit
| Metric | Without AI | With AI | Delta |
|---|---|---|---|
| Drug discovery success rate | 6.4% | 10.3% | +370 bps |
| Additional discoveries/yr | — | +28 drugs | +28 |
| Pre-clinical + testing time | 13 years | ~10 years | -3 years |
10-Year Pipeline PV Uplift:
Discount Rate 21%
$83B
Discount Rate 12%
$236B
Discount Rate 8%
$412B
AI boosting drug discovery success rates by 370bps and cutting development timelines by 3 years. First concrete quantification answering "what are the goods AI is delivering?"
MS: US DC Power Shortfall 2025-28 (Exhibit 12)
Total demand 74 GW. After all solutions: 9-18 GW net shortfall remains.
| Solution | Low | Mid | High | Probability |
|---|---|---|---|---|
| Nat Gas Turbines | 15 GW | 18 GW | 20 GW | 90% |
| Bloom Energy Fuel Cells | 5 GW | 7 GW | 8 GW | 90% |
| Nuclear Co-location | 5 GW | 10 GW | 15 GW | 75% |
| Bitcoin Site Conversions | 10 GW | 13 GW | 15 GW | 90% |
Net Shortfall After All Solutions:9 GW (mid) to 18 GW (low)
MS: "We believe the most likely outcome skews towards the low end of our range" — i.e. closer to 18 GW shortfall
Three Bottleneck Layers (Cross-House Framework)
Disaggregated supply constraint analysis from our research
1. Raw Generation Capacity
OVERSTATED as constraintJPM: 222 GW gross US additions 2026-28 (net 188 GW after 34 GW retirements). GS Feb'26: 105 GW DC-specific capacity including 14 GW BTM. Green Reliability Premium $40/MWh = only 3.4% of hyperscaler 2027E EBITDA ($1,079B). Hyperscalers will pay — this is solvable with capital and time.
Sources: GS + JPM
2. Grid Interconnection & T&D Labour
UNDERSTATED — true binding constraintGS: 78,000 skilled T&D worker gap requiring 3-4 year apprenticeship training. Current: ~45K apprentices/yr, need ~65K from 2027. JPM: >2,600 GW in interconnection queue, 5+ year wait. Transmission: 7-10 year lead time. Transformer lead times 128-144 weeks (2.8 years) at 4-6x cost. This is a human capital problem money cannot solve quickly.
Sources: GS + JPM + IOU Article
3. Load Flexibility (Inference)
NEW DIMENSION — MS uniqueAs inference rises to 45-50% of DC workload, power demand shifts from stable baseload (training) to volatile, spiky, unpredictable load curves. System needs millisecond-response peak shaving and frequency regulation. ESS provides this — not replacing generation, but complementing it. DC ESS deployment: 321 GWh by 2030 (bull case: 590 GWh).
Sources: MS
GS Feb'26: DC-Specific Capacity Additions (105 GW)
Up from 82 GW (Oct'25) and 72 GW (Apr'24). Includes 14 GW BTM (new).
Gas dominates near/medium-term: peakers (40%) + CCGT (12%) = 52% gas total. BTM (13%) is entirely new in Feb'26 — reflects hyperscaler onsite simple-cycle nat gas to bypass 5+ year grid queues. US DC nat gas demand projected >7 Bcf/d by 2030. Grid capex: >$600B in 2026-2030.
JPM 2Q26: Total US Grid Additions 2026-28 (221.2 GW gross)
2Q26 revision (Apr 30): Net 186.1 GW after 35.1 GW retirements. Material mix shift toward batteries.
Material 2Q26 revision: Solar still dominates at 49% (108 GW) but at ~25% capacity factor = only ~237 TWh effective generation over 3 years vs nameplate of ~950 TWh. Batteries now 26% (57.5 GW) — doubled from 1Q26's 26.7 GW. Natural gas cut in half to 13% (28.8 GW) from 60 GW. Retirements: 35.1 GW (69% coal, 30% gas). The grid is being rebuilt for flexibility, not baseload.
GS: Power Generation Timeline
Renewables + BTM GasNear Term
Key constraint: IRA safe harbour, land/supply
Solar, battery storage, simple-cycle nat gas. BTM = 14 GW of onsite generation bypassing grid. Hyperscalers deploy in months vs years for grid connection.
Nat Gas CCGT2029
Key constraint: Turbine availability
Combined cycle more efficient than peakers. GE Vernova, Siemens Energy key suppliers. Turbine lead times 3-4 years. >7 Bcf/d US DC nat gas demand by 2030.
Nuclear2030-35+
Key constraint: Permitting, build, uranium
Large-scale + SMR. Meta signed 2,600 MW with Vistra (20yr PPA). 50 GW nuclear needed to fully offset DC emissions. Capacity factor 90%+ vs solar 25%.
NVIDIA Server Evolution — Power vs Compute
Efficiency +650% over 4 gens, but absolute power per server +269%
| Generation | Max Power | Compute | Intensity (kW/pF) | vs A100 |
|---|---|---|---|---|
| DGX A100 | 6.5 kW | 5 pF | 1.30 | Baseline |
| DGX H100 | 10.2 kW | 32 pF | 0.32 | -75% |
| DGX B200 | 14.3 kW | 72 pF | 0.20 | -85% |
| NVL8 (Rubin) | 24 kW | 140 pF | 0.17 | -87% |
GS 6P Constraint Framework (Feb'26) + Cross-House Overlay
Six constraints governing the pace and shape of AI data centre power buildout
Pervasiveness
MediumGS: Inference power intensity rising. AI drug discovery success rates +370bps (6.4%→10.3%). Still in Appraisal phase — not yet demand-constrained.
MS: Inference reaches 45-50% of DC load by 2030. Drives fundamental shift from capacity to flexibility requirements.
Productivity
MediumGS: Vera Rubin NVL8: 0.17 kW/pFLOPS (vs A100: 1.30). +650% efficiency over 4 gens. But max power per server +269%. Pent-up demand absorbs gains.
MS: Efficiency gains per unit confirmed, but offset by higher absolute power per inference server. Jevons Paradox operating.
Parts
HighGS: 105 GW total (incl 14 BTM). Nat gas >7 Bcf/d by 2030. Turbine availability key constraint for CCGT.
MS: Bloom Energy fuel cells: 5-8 GW at 90% probability. Transformer lead times 128-144 weeks at 4-6x cost. Brownfield sites with grid connections = premium assets.
People
CriticalGS: 78,000 T&D skilled labour gap. 3-4 year apprenticeship training. Current: ~45K/yr, need ~65K from 2027. Wage inflation may help supply.
MS: Confirms labour as most severe bottleneck. Drives accelerated BTM adoption (less T&D workers needed). Grid automation and contractor premium persist.
Price
LowGS: Green Reliability Premium $40/MWh = 3.4% of hyperscaler 2027E EBITDA ($1,079B). CROCI impact: -0.8pp. Not a meaningful constraint.
MS: Solar+ESS LCOE $74-100/MWh approaching CCGT $67/MWh. Na-ion at RMB 0.21/Wh further reduces ESS costs. BTM ESS: 23% unlevered IRR.
Policy
High ↑GS: Rising public concerns about DC impact on electricity affordability. Push for ring-fencing costs. IRA sunset modest near-term impact.
MS: White House Ratepayer Protection Pledge (Mar'26) = BYOP era. PJM BTM rules may increase co-located DC fees. Texas SB-6 'kill switch' bill.
MS: Global DC ESS Deployment Forecast (GWh/yr)
From ~15 GWh (2025) to 321 GWh (2030). Bull case: 590 GWh. US = 53% share. DC ESS CAGR ~85% vs utility-scale ~45%.
DC-specific ESS is a fraction of total utility-scale ESS (total market reaching 1,200+ GWh by 2030 per Rystad). But growth rate is faster because inference volatility requires onsite ms-response capability that utility-scale grid storage cannot provide due to transmission latency.
MS: Power System Rebalancing (Exhibit 19)
Traditional generation's share declining as ESS/flexibility rises to ~22% by 2030. Structural shift.
MS frames this not as ESS replacing generation, but as the marginal investment dollar shifting toward flexibility. Traditional baseload remains essential (gas CCGT, nuclear), but "demand shorts" — the gap between what generation provides and what inference loads need in real-time — are absorbed via ESS. By 2030, ~22% of system contribution from flexibility/storage.
MS: Sector Rotation Sequence (Exhibit 4)
AI Power basket accreted >$1.5T market value since Jan 2023. Capital rotated sequentially. ESS is next.
Each rotation wave lasted 6-12 months before the next segment takes leadership. Nuclear peaked in 2024 (Constellation, Cameco). Gas generators peaked 2024-25 (GE Vernova +196pp). Fuel cells are the current wave (Bloom Energy 10x in 12 months). MS identifies ESS as the next wave — the question is timing, not direction.
ESS Use Cases for AI Data Centres
ESS in DCs is not about providing baseload — it's about managing what baseload cannot handle
Peak Shaving
Absorb sudden inference load spikes (100s of MW in seconds) that would otherwise trigger demand charges or destabilise local grid. Saves $2-5/MWh in demand charge avoidance.
Frequency Regulation
AI GPU loads create high-frequency power quality noise. BESS provides millisecond-level voltage and frequency stabilisation, preventing damage to $2M+ per rack GPU hardware.
Time-Shifting / Arbitrage
Charge during low-cost overnight hours, discharge during peak windows. BTM ESS unlevered IRR 23%, levered 36% — purely from peak/off-peak price differential. Self-funding economics.
Backup / UPS Replacement
Replace diesel UPS with battery. Faster response (<10ms vs >10s for diesel), lower emissions, lower maintenance. Critical for 99.999% uptime requirement.
Renewable Integration
Buffer intermittent solar/wind to provide firm power. Addresses fundamental mismatch: 50% of new US generation is solar at 25% capacity factor, but DCs need 24/7 supply.
Grid Deferral
ESS defers ~10% of infrastructure capex. Net deferral value ~$1.15M per $10M investment delayed 5 years. Reduces upfront grid upgrade burden for both utilities and hyperscalers.
MS: ESS Levelised Cost of Energy (LCOE)
Solar+ESS approaching CCGT parity in US. Already cheaper than coal in China.
| Configuration | 2025 | 2027E | 2030E | Benchmark | Status |
|---|---|---|---|---|---|
| US Solar+ESS (20%,3H) | $90 | $82 | $74 | CCGT $67 | Approaching |
| US Solar+ESS (35%,4H) | $95 | $87 | $78 | CCGT $67 | Converging |
| US Standalone ESS (20%,3H) | $140 | $120 | $100 | Peaker $93 | 2030 parity |
| China Solar+ESS | $48 | $46 | $44 | Coal $50 | Already cheaper |
| China Standalone ESS | $65 | $55 | $45 | Gas $70 | Already cheaper |
BTM ESS Returns (DC-specific, peak/off-peak arbitrage only)
US Unlevered IRR
23%
US Levered (50%) IRR
36%
China Unlevered IRR
13%
China Levered IRR
22%
Sodium-Ion: The Next Cost Curve Disruption
CATL commercialised. At 100 GWh scale: >30% cheaper than LFP. Changes ESS economics.
| Parameter | LFP (Baseline) | Na-Ion Current | Na-Ion at Scale | Advantage |
|---|---|---|---|---|
| Cell cost (RMB/Wh) | 0.35-0.40 | 0.32 | 0.21 | -34% vs current |
| Energy density (Wh/kg) | 160-180 | 160 | 180 | Approaching LFP |
| Low-temp performance | Degraded | Superior | Superior | Cold climate key |
| Cycle life / decay | Good | Slower | Slower | Lower replace cost |
| Low SOC full power | No | Yes | Yes | Better grid response |
| Key material | Lithium | Sodium | Sodium | No supply constraint |
At RMB 0.21/Wh (from 0.32), sodium-ion undercuts LFP by 30%+ at 100 GWh scale. CATL already in commercial production. Key DC advantages: better low-temperature charging, slower degradation, full power at low SOC. Combined with ESS LCOE approaching gas parity, Na-ion could accelerate crossover by 1-2 years.
MS High-Conviction ESS Beneficiaries: CATL · Tesla · LGES · Fluence · BYD
MS: ESS Drives Lithium Demand
New ESS installations could push lithium market into deficit — unless Na-ion scales faster
| Year | ESS Install (GWh) | ESS Shipments | Li Demand (kt LCE) | Bull Case Li |
|---|---|---|---|---|
| 2024 | ~200 | ~300 | ~200 | ~200 |
| 2026E | ~500 | ~550 | ~380 | ~420 |
| 2028E | ~950 | ~1,100 | ~700 | ~850 |
| 2030E | ~1,800 | ~2,200 | ~1,100 | ~1,400 |
If ESS shipments follow the bull case curve, lithium demand from ESS alone reaches 1,400 kt LCE by 2030 — potentially pushing the market into deficit, depending on supply response and sodium-ion substitution pace.
ESS Investment Vehicles — Tiered Risk/Reward
Three tiers from lowest to highest risk
Tier 1: Scaled Platforms (Lowest Risk)
Tesla (TSLA) · BYD · CATL
Megapack is the dominant utility-scale ESS platform. Tesla also has Autobidder AI software + UK Ofgem retail licence = full vertical integration. CATL is the Na-ion leader. These companies have proven manufacturing scale, positive margins, and diversified revenue. ESS = upside optionality.
Tier 2: Pure-Play ESS (Medium Risk)
Fluence (FLNC) · LGES
Fluence: GS Buy-rated, MS high-conviction. Leading software-defined ESS platform. But dropped 55% ($33→$15) in one month. LGES: Korean battery giant with strong ESS credentials but less direct DC exposure.
Tier 3: Emerging / Pre-Profit (Highest Risk)
EOSE · FCEL · Plug Power
EOSE: zinc-based ESS with interesting chemistry but -126% gross margin, $969.6M net loss, securities investigations. Technology thesis is right but unit economics unproven at scale.