This dashboard is password protected.
Where value has accrued in the AI capex cycle, and how to position for the migration up the stack.
Through 2025, value has overwhelmingly accrued to Layer 2 (Chips): NVIDIA Data Center revenue compounded from $47.5B (FY24) to $193.7B (FY26), a ~4.1× expansion. The next $500B-$1T of incremental value flows down the stack first, into Layer 1 (Energy) and Layer 3 (Infrastructure), the binding physical constraints (gas turbines, CoWoS packaging, interconnection queue). Layers 4 (Models) and 5 (Applications) become the dominant value-capture surface in 2028-2030 as compute overbuild catches demand.
| Layer | Trend Since 2023 | 2026-2028 Outlook | Binding Constraint | Severity | Value Captured (% of AI capex/revenue) |
Key Public Companies |
|---|---|---|---|---|---|---|
| 1 · Energy | Fastest re-rating layer. GEV +210% since AI market recognition in 2025; Vistra +300%+. Pipeline grew 50 GW (Q1) to 62 GW (Q3), ~80 GW Q4E. Power procurement and long-term PPAs now the largest single line item in DC development budgets | Rising. Scarcity premium intensifying as 42-60 mo turbine lead times collide with hyperscaler urgency. BTM gas and nuclear PPAs becoming strategic assets. GEV SRA conversion (~3 GW/qtr) is the pacing metric | Gas turbine lead time (42-60 mo). GEV sold out through 2030 | Binding | ~15-20% | GE Vernova, Siemens Energy, Constellation, Vistra |
| 2 · Chips | Largest revenue layer. NVIDIA DC revenue grew from $15B (FY23) to $115B (FY25). TSMC and HBM suppliers (SK Hynix) also saw significant revenue acceleration. GPU + accelerator spend is the single largest capex line item | Peaking share. Still dominant but share declining as custom silicon (Google TPU, Amazon Trainium) scales and CoWoS capacity catches up. Absolute spend continues to grow but other layers catching up | CoWoS packaging capacity (TSMC) | Binding | ~40% | NVIDIA, TSMC, SK Hynix, Broadcom, AMD |
| 3 · IT Infrastructure | Accelerating. Cooling (Vertiv 2.9x book-to-bill), power distribution (Eaton/Schneider), and networking (Arista) all saw order books inflect in 2024-25. Dell and Super Micro rode GPU server volumes | Expanding. Multi-year buildout supports sustained demand. Rack density (100+ kW) driving premiums for liquid cooling and high-density power. Constrained build pace extends the revenue runway | DC build pace vs. power availability | Stressed | ~25% | Vertiv, Dell, Super Micro, Schneider, Eaton, Arista |
| 4 · Foundation Models | Fastest-scaling businesses in history. Anthropic reported ~$30B annualized revenue (Q1 2026); OpenAI ~$12B+ ARR (late 2025). These are consumer-internet growth rates on enterprise software economics. Training costs now $100M-$1B+ per frontier model, but revenue is outpacing spend | Bifurcating. 1-2 winners will capture outsized value via distribution and enterprise lock-in. Most labs will consolidate or pivot. API pricing power depends on differentiation vs. open-source | Training cost escalation ($10B+ clusters) | Stressed | ~12% | OpenAI, Anthropic, Cohere, Mistral |
| 5 · Applications | Nascent but fastest-growing. AI-native SaaS (Cursor, Harvey, Glean) showing >100% YoY growth with 60-85% gross margins. Enterprise adoption just beginning to move beyond pilot phase | Largest long-term opportunity. As inference costs decline 10x every 12-18 months, application-layer economics improve dramatically. Workflow-specific AI tools with proprietary data moats will capture disproportionate value | Inference cost + adoption velocity | Early | ~3-5% | Cursor, Harvey, Glean, Hebbia, Abridge |
Value Captured = estimated share of cumulative AI-related capex and revenue flowing through each layer, FY2023-FY2025. Shares are directional estimates based on public filings, earnings disclosures, and industry data. Energy (L1) includes power procurement, turbine orders, and long-term PPAs. Chips (L2) includes GPU/accelerator, HBM, and advanced packaging spend. IT Infrastructure (L3) includes servers, cooling, networking, and DC construction (distinct from L1 energy/power generation infrastructure). Shares will not sum to 100% due to overlap across layers.
Two layers are physically binding today: Energy (L1) and Chips (L2). Every month these constraints persist, value migrates toward whoever can deliver MW or packaging capacity. IT Infrastructure (L3) is stressed but buildable. Models (L4) and Applications (L5) are the 2028-2030 opportunity, but only once the physics layers unlock sufficient compute.
Orders of Magnitude (OOMs) measure how much smarter AI gets with each step up in computing power. Each rung requires roughly 10× more computing power (and therefore energy) than the one below. The industry has effectively reached 104 (“autonomous AI worker”). The target is 106 (“human genius”) by 2028. The problem: each rung up requires building the equivalent of a small city’s power grid, and that takes years, not months.
| Year | AI Level | Power Needed | What It Can Do | Grid Feasibility | Status |
|---|---|---|---|---|---|
| 2022 | 100 | 90 MW | Smart highschooler | No problem | Done |
| 2024 | 101 | 100 MW | College freshman | No problem | Done |
| 2026 | 103 | 1 GW | Smart undergrad | Stressed but OK | Done |
| 2026 | 104 | 3 GW | Autonomous AI worker | Stressed | Achieved |
| 2028 | 106 | 10 GW | Human genius | Timing very tight | Depends on turbines |
| 2029 | 107 | 30 GW | - | Not feasible | 100% of global GT output |
| 2030+ | 1010 | 100 GW | Automated AI research | 20% of US grid | Not before 2040 |
Each generation does more computation per chip, but also draws more power
Anything above 30 kW needs liquid cooling. All new AI racks exceed this
Data center revenue now accounts for 90%+ of NVIDIA’s business, the heartbeat of the AI capex cycle
Situational Awareness OOM projections vs. physical feasibility. Source: AI Compute Timeline Analysis.
Through 2025, value has overwhelmingly accrued to Layer 2 (Chips):NVIDIA Data Center revenue compounded from $47.5B (FY24) to $193.7B (FY26), a ~4.1× expansion capturing roughly 55% of the AI value pool. The next $500B–$1T of incremental value flows down the stack first, into Layer 1 (Energy) and Layer 3 (Infrastructure):the binding physical constraints:before migrating to Layers 4–5 (Models & Applications) in 2028–2030 as compute supply catches demand.
Combined capex grew from $114B (2020) to $389B (2025E). Source: Hyperscaler CapEx Analysis + 10-K filings.
Quarterly DC revenue FY24–FY26. Source: NVIDIA earnings releases.
Every $1B of hyperscaler capex pulls through a physical chain of GPUs, racks, MW, turbines, and copper. Source: Constraint Stack Analysis.
| Stage | Implied Quantity | Layer | Value Captured By |
|---|---|---|---|
| GPU capex (50% of DC capex) | $500M | L2 Chips | NVDA, AMD, Broadcom (ASIC) |
| GPUs deployed | 16,667 units | L2 Chips | TSMC (CoWoS), SK Hynix, Micron |
| Racks | 2,083 | L3 IT Infrastructure | Dell, SMCI, HPE, Vertiv, Coherent |
| IT power | 83 MW | L3 IT Infrastructure | Schneider, Eaton, Vertiv |
| PUE-adjusted power | 108 MW | L1 Energy | GE Vernova, Siemens, MHI, BTM gas |
| Copper required | 2.7 kt | L1/L3 | FCX, SCCO, BHP |
| Potential AI revenue generated | $1.08B | L4/L5 | Frontier labs + application builders |
Illustrative. Chips share compresses as the physics layer absorbs scarcity rent; Models and Apps capture the majority of incremental value post-2028.
Click any layer below for metrics, public winners, and constraint-driven insights.
36 tracked projects across US + international, representing $508B committed investment and ~45 GW of planned capacity. Source: Hyperscaler DC Tracker + public filings.
Google and AWS lead by total planned MW; xAI, Oracle/OpenAI punching above weight in $/MW intensity.
Wisconsin (MSFT Mt Pleasant), Ohio, Texas, Mississippi lead. PJM exposure is ~2 GW; ERCOT-adjacent states account for ~7 GW.
Cooling: Liquid + hybrid now dominate:41.5 GW of the 44.9 GW planned uses liquid or hybrid air/liquid. Pure-air cooling is effectively dead for new-build AI halls. PUE targets converging at 1.08–1.15.
Power: Grid-only is the minority. Most new sites blend Grid + on-site generation (CCGT, solar PPA, nuclear co-location). BTM gas and co-located nuclear are the fast-growing categories:xAI Colossus (CCGT), AWS Pennsylvania (Susquehanna nuclear), Meta Hyperion (3× CCGT). SMR pipeline ~1.8 GW contracted.
Chips: NVIDIA still dominant (GB200 NVL72 is the de facto 2026 training rack), but custom silicon is catching up fast:AWS Trainium2/3, Google TPU v5/v6, Azure Maia all showing in new-builds. ~27 GW of the pipeline explicitly pairs custom silicon with NVIDIA.
Liquid cooling has won. Vertiv/Schneider attach rate is the direct beneficiary.
Grid + Solar PPA leads, but BTM gas and nuclear co-location are the fastest-growing categories.
NVIDIA-only sites represent ~13 GW; mixed custom-silicon + NVIDIA sites represent ~27 GW. Google TPU and AWS Trainium now designed in from Day 1 on new campuses.
| Hyperscaler | Project | State | Invest ($M) | MW | Power Gen | Cooling | PUE | Chip Type | Completion |
|---|
A binding constraint (shown YES) is a layer where 2025E capex cannot be physically absorbed at current capacity. Source: Constraint Stack Analysis.
| Layer | Capacity (2025) | Required | Deficit / (Surplus) | Binding? |
|---|---|---|---|---|
| L2 · CoWoS Packaging | 402K GPUs | 7.3M GPUs | (6.9M GPUs) | Yes |
| L1 · Gas Turbine Mfg (AI-avail) | 19.7 GW/yr | 47.5 GW | (27.7 GW) | Yes |
| L2 · CoWoS Power Ceiling | 522 MW deliverable | 47,450 MW | (46.9 GW) | Yes |
| L1 · PJM Interconnection Queue | 211.5 GW (PJM active) | 14.2 GW (PJM share) | +197.3 GW | No |
| L1 · Copper Supply | 24,000 kt mined | 2,686 kt demand | +21,314 kt | No |
| L1 · Power Generation (DC-level) | 67.9 GW | 43.2 GW | +24.7 GW | Diag |
42–60mo lead time from order to commissioning. Orders placed after Q2 2024 cannot hit 2028.
Cumulative excess capacity accumulates post-2028 if models don't absorb.
PJM operates the largest interconnection queue in the world with 211.5 GW total active and 580.2 GW withdrawn (historical). Nationally, 77% of capacity is withdrawn (LBNL Queued Up 2025). For a fund with capital and sponsorship capacity, that attrition is the opportunity.
Nationally, 77% of capacity is withdrawn; 72% by number of requests (LBNL). Projects submitted 2010–2019 show similar patterns. PJM-specific: 14% completion rate by capacity, 23% by number of projects. Source: LBNL Queued Up 2025 Edition.
Storage and Solar dominate active queue. Natural gas capacity in queues surged 72% YoY in 2024; solar decreased 12%, storage -13%, wind -26% (LBNL). Gas scarcity is a bull signal for dispatchable capacity.
VA, IL, OH, NJ, IN dominate:overlapping almost perfectly with tracked hyperscaler DC campuses.
Average project in 2024 took 55 months from request to commercial operation (up from 22 months in 2008). Time-in-queue creates selection bias: only the most valuable projects survive. Each stage is a filter.
FERC Order 2023 cluster-based process. Deposits and penalties increase at each stage, filtering out non-viable projects. Nationally, only 13% of capacity survives to commercial operation (LBNL).
Filtered to active / in-construction projects in AI-heavy PJM states (VA, PA, OH, IN, WV, NJ, MD, KY, IL, MI) with ≥100 MW capacity. Scored on: (1) dispatchable fuel (natural gas, storage, nuclear = highest value for AI load); (2) completed Facilities Study; (3) executed Interconnection Agreement; (4) stage (Under Construction > E&P > Active). Highest-score projects = shortest path to energization, highest certainty cost basis, best positioned to sell power to adjacent hyperscaler DC campuses.
| Rank | Project | State | County | MW | Fuel | Status | FS | IA | CSA | Target ISD | Score |
|---|
Of the 455 GW withdrawn historically, most was killed by one of four problems: (1) network upgrade costs re-allocated to the sponsor; (2) inability to finance long-tail interconnection timelines; (3) land/permitting failure; (4) loss of PPA or offtake. All four are solvable with capital and sponsorship. Acquiring a project that has cleared Facilities Study but stalled pre-CSA is the cleanest entry point.
Active queue maps directly onto hyperscaler DC siting: VA (16 GW queue, AWS/MSFT/Google data center alley), OH (10 GW, Stargate Lordstown), PA (6 GW, AWS Susquehanna), IN (9 GW, AWS Northern Indiana). Owning dispatchable generation adjacent to a stalled DC = the fastest path to revenue.
Only ~8 GW of natural gas and ~36 GW of storage/solar+storage in active queue:below hyperscaler demand curve. Gas offers 24/7 dispatchability, storage offers capacity market upside. Solar-only is less attractive (curtailment risk, duck-curve economics).
PJM reform (Order 2023 compliance) cleared the 2019–2021 cohort. 2024 Capacity Auction cleared at $269.92/MW-day (~9× prior). Any project that survives study phases now enjoys structural price tailwinds. Window to buy: 2026–2027, before cluster study results fully price into asset sales.
Thirteen archetypes across the five layers. Bold rows are highest-conviction targets where bottleneck analysis suggests durable pricing power over 2026–2029.
| Layer | Archetype | Why Now | 3–5yr Return Driver | Pub. Comps |
|---|
The One Big Beautiful Bill Act (OBBB), signed July 4, 2025, accelerates the termination of key IRA clean energy tax credits for solar and wind while preserving incentives for nuclear, battery storage, and geothermal. The compressed timelines create a near-certain wave of renewable project attrition in interconnection queues, particularly in PJM, freeing up grid capacity and interconnection rights that smaller gas power projects can backfill.
| Credit / Section | Pre-OBBB Status | OBBB Change | Effective Date |
|---|---|---|---|
| Solar & Wind ITC/PTC (45Y, 48E) | Available through 2032+ under IRA | Terminated for facilities placed in service after 12/31/2027. Must begin construction before 7/5/2026 to qualify | 7/4/2025 |
| Residential Clean Energy (25D) | 30% credit through 2032 | Terminated for expenditures after 12/31/2025 | 1/1/2026 |
| EV Credits (30D, 25E) | Up to $7,500 new / $4,000 used | Terminated for vehicles acquired after 9/30/2025 | 10/1/2025 |
| Commercial Buildings (179D) | Deduction through 2032 | Terminated for property beginning construction after 6/30/2026 | 7/1/2026 |
| Nuclear | Available through 2032+ | Preserved through 2033 | N/A |
| Battery Storage | Available through 2032+ | Preserved through 2032 (including co-located with solar) | N/A |
| Geothermal | Available through 2032+ | Preserved through 2033 | N/A |
Source: IRS FAQs for OBBB (Public Law 119-21, July 4, 2025).
Before the OBBB, developers could establish "beginning of construction" by either (a) starting physical work of a significant nature, or (b) paying or incurring at least 5% of total project costs (the "5% safe harbor"). Notice 2025-42, issued August 15, 2025, eliminates the 5% safe harbor for all solar and wind facilities except low-output solar (≤1.5 MW AC). This is a critical change for project economics:
| Rule | Before OBBB | After Notice 2025-42 | Impact |
|---|---|---|---|
| Beginning of Construction Test | Physical work test or 5% safe harbor (developer's choice) | Physical work test only (effective 9/2/2025). 5% safe harbor limited to solar ≤1.5 MW | Developers can no longer establish construction start by writing a check. Must mobilize crews and begin physical site work |
| Continuity Requirement | Physical work or "continuous efforts" (permits, binding contracts, etc.) | "Continuous efforts" alternative eliminated. Must maintain continuous physical construction program | No more paper-based continuity. Developers must show uninterrupted physical progress after breaking ground |
| Continuity Safe Harbor | Placed in service within 4 calendar years of construction start | 4-year safe harbor retained. Projects beginning before 7/5/2026 have until 12/31/2030 | Narrow window: begin physical work by 7/4/2026, complete by end of 2030 |
| Qualifying Physical Work | On-site or off-site work of a significant nature | On-site or off-site work pursuant to a binding written agreement. Must be "of a significant nature" | Preliminary activities (clearing, grading for access roads) do not count. Must begin work on the facility itself or key components |
The 5% safe harbor was the primary mechanism large-scale developers used to "lock in" tax credits early in the development cycle. A developer could place a deposit on equipment (5% of a $500M project = $25M) years before construction started and still qualify. Eliminating this path forces developers to commit to actual physical construction within 12 months of enactment, a timeline that is unrealistic for most utility-scale projects still in permitting or interconnection study phases.
The OBBB introduces "Prohibited Foreign Entity" (PFE) rules across clean energy credits (45Y, 48E, 45X, 45Q, 45U, 45Z). Projects receiving "material assistance" from entities owned, controlled, or directed by China, Russia, Iran, or North Korea are disqualified from claiming credits.
| Component | FEOC Restriction | Effective Date | Supply Chain Impact |
|---|---|---|---|
| Solar panels / cells | Cannot use components manufactured or assembled by a FEOC | Tax years beginning after 7/4/2025 (1/1/2026 for calendar-year filers) | ~80% of global polysilicon and wafer production is in China. Developers must source from non-FEOC supply chains at premium pricing |
| Battery components | No FEOC-manufactured or assembled battery components | 1/1/2026 | China dominates cathode, anode, and cell manufacturing. Non-FEOC battery supply severely constrained through 2028 |
| Critical minerals | Cannot be extracted, processed, or recycled by a FEOC | 1/1/2026 | China processes ~60-70% of lithium, cobalt, and rare earths. Alternative sourcing from Australia, Chile, Canada adds cost and lead time |
| Inverters | Material assistance threshold: 50% (2026) rising to 70% (2030) | Phased 2026-2030 | Graduated restriction tightens annually, requiring progressive supply chain de-risking |
| Foreign-influenced entity | 25%+ ownership, voting rights, or board seats held by covered nation government disqualifies entity | Tax years beginning after 7/4/2027 (1/1/2028) | Broader corporate ownership test. Joint ventures with Chinese partners at risk |
The FEOC rules compound the safe harbor elimination. Even projects that manage to begin physical construction before 7/5/2026 must also prove their entire supply chain is FEOC-compliant to claim credits. For solar in particular, re-sourcing panels outside Chinese supply chains adds 20-40% to component costs and 6-12 months to procurement timelines. Gas turbine projects face no equivalent FEOC restrictions, widening the economic gap between renewable and gas-fired new-build.
Nationally, 2,290 GW is in interconnection queues (nearly 2x the entire US installed fleet of 1,320 GW), comprised of 956 GW solar, 271 GW wind, and 890 GW storage. Only 13% of capacity requesting interconnection 2000-2019 reached commercial operation (LBNL). Even after signing an IA, 36% of projects (43% of capacity) withdraw. The OBBB's compressed timelines (begin construction by 7/5/2026, placed in service by 12/31/2027) make it physically impossible for most large renewable projects still in queue to meet these deadlines. In 2024, ~500 GW of new requests were submitted but 700+ GW withdrawn. This arithmetic guarantees significant project dropout.
| Factor | Impact on Renewables | Opportunity for Gas |
|---|---|---|
| Construction deadline compression | Must begin construction by 7/5/2026 (physical work test only). Large utility-scale solar/wind projects typically need 18-36 months for permitting alone | Gas peaker and combined-cycle projects have shorter development timelines (12-24 months) and can meet queue milestones faster |
| Interconnection position value | Projects that drop out forfeit their queue position and any completed interconnection studies. Over 100 GW of planned utility-scale solar/wind at risk | Abandoned interconnection positions with completed grid studies become available. Gas projects can step into positions where grid upgrade costs are already defined and partially allocated |
| Network upgrade cost allocation | High network upgrade costs (often $100M+) were already the primary driver of renewable project attrition in PJM even before OBBB | Smaller gas plants (50-500 MW) require fewer grid upgrades than equivalent renewable capacity and can utilize existing transmission infrastructure more efficiently |
| PJM Reliability Resource Initiative | N/A | PJM accepted 51 projects (~11,000 MW) of shovel-ready gas, nuclear, and battery projects through the RRI to address near-term reliability gaps |
| Demand urgency | Renewable intermittency requires backup capacity or storage, adding cost and complexity in a credit-constrained environment | AI/DC load growth requires firm, dispatchable power 24/7. Gas turbines provide baseload capacity that matches hyperscaler requirements |
The OBBB creates a two-track energy market. Renewable project economics deteriorate rapidly without ITC/PTC support, driving queue attrition and freeing interconnection capacity. Gas power developers, particularly those with existing turbine orders (GEV backlog sold out through 2030) and behind-the-meter strategies, are positioned to absorb these positions. The companies that control interconnection rights, turbine delivery slots, and gas supply agreements gain a structural advantage in serving AI/DC load growth. This reinforces the L1 Energy scarcity thesis: fewer pathways to new power generation means existing capacity and near-term deliverable projects command premium valuations.
| Category | Position | Companies / Assets |
|---|---|---|
| Gas turbine OEMs | Beneficiary | GE Vernova, Siemens Energy, Mitsubishi Power. Order books extend and pricing power increases as gas becomes the default new-build for DC power |
| Gas IPPs / BTM developers | Beneficiary | Developers with behind-the-meter gas strategies and existing interconnection positions. Can serve DC load without competing for grid queue positions |
| Nuclear operators | Beneficiary | Constellation, Vistra. Nuclear credits preserved through 2033. Existing fleet becomes more valuable as renewable pipeline shrinks |
| Battery storage | Mixed | Credits preserved through 2032, but standalone economics weaker without co-located solar ITC. Grid-scale storage for peaking/ancillary services still viable |
| Utility-scale solar developers | Impaired | Projects not yet under construction face compressed timelines. 17% decline in installations forecast over next decade vs. IRA baseline |
| Wind developers | Impaired | Onshore and offshore wind face same deadline pressure. Physical work test (not 5% safe harbor) required, making it harder to establish construction start |
Sources: IRS FAQs (Public Law 119-21), IRS Notice 2025-42 (beginning of construction guidance), IRS/Treasury FEOC guidance (prohibited foreign entity rules), LBNL Queued Up 2025 Edition (December 2025), PJM Interconnection Reform data, Columbia CGEP analysis, Rhodium Group impact assessment, Tax Foundation, SEIA, Baker Tilly, K&L Gates, The Tax Adviser, and Davis Wright Tremaine analyses.
77% of projects entering the interconnection queue never make it to commercial operation (LBNL). Average processing time: 55 months. PJM capacity is clearing at the FERC-approved cap of $333/MW-day. In this environment, a brownfield energy site with existing grid access skips the line. Instead of waiting 4-8 years, you're looking at a 24-month powered site.
| Deal | Buyer | MW | Structure | Implied Value | Status |
|---|---|---|---|---|---|
| Susquehanna Nuclear | Amazon (AWS) | 960 (campus) / 1,920 (PPA) | $650M campus acquisition + 1,920 MW grid-connected PPA ($1.4B/yr at full quantity) | $650M upfront (~$677K/MW campus); $18B lifetime PPA revenue | Campus acquired; PPA signed; FERC rejected co-location ISA, restructured to front-of-meter |
| Meta-Williams Ohio (Socrates) | Meta | 400 | 2×200 MW on-site gas generation (BTM), built by Williams subsidiary | ~$1.6B construction ($4.0M/MW) | Approved by Ohio Power Siting Board; COD late 2026 |
| Duke NC Campus | Amazon (AWS) | ~2,400 | $10B campus + utility CCGT expansion | ~$4.2M/MW all-in | Announced; Duke adding 5 GW gas by 2029 |
| xAI Memphis | xAI | 238→2,000 | BTM gas turbines, trucked-in then co-located CCGT | Speed premium: 90-day build | Colossus 1 operational; Colossus 2 in service Q4 2025 |
| Stargate Abilene | Oracle/OpenAI | 1,200 | Greenfield + grid + gas backup | ~$5.3M/MW campus cost | Mid-2026 target |
| Delivery Year | Clearing Price ($/MW-day) | Annual Revenue per 100 MW |
|---|---|---|
| 2024/2025 | $28.92 | $1.1M |
| 2025/2026 | $269.92 | $9.9M |
| 2026/2027 | $329.17 | $12.0M |
| 2027/2028 | $333.44 (FERC cap) | $12.2M |
Data centers drove 40% ($6.5B) of PJM's $16.4B capacity auction costs and 63% of the price increase. PJM forecasts ~32 GW of new demand by 2030, almost entirely AI/DC load.
Sources: PJM BRA auction results (2024-2028), LBNL Queued Up 2025, Talen Energy SEC filings, Utility Dive, Power Magazine, IEEFA analysis.
Every major hyperscaler has public net-zero or 100% renewable targets. At the same time, utilities have announced 20+ GW of new natural gas capacity by 2040, two-thirds driven by data center demand from Microsoft, Amazon, Meta, and Google. Renewables and storage simply can't deliver firm, dispatchable 300+ MW blocks on 24-month timelines. Gas can. So the real question isn't whether hyperscalers will use fossil power. It's how they structure deals to manage disclosure.
| Hyperscaler | Deal | MW | Fuel | ESG Framing |
|---|---|---|---|---|
| Meta | Williams Ohio (Socrates 400 MW + Apollo 350 MW) | 750 | On-site gas + pipeline (BTM) | "Transition infrastructure," hydrogen-ready turbines |
| Amazon | Duke NC (5 GW gas expansion) | 2,400 | Grid + CCGT | Utility-owned generation; Scope 2 via grid mix |
| xAI | Memphis Colossus | 2,000 | BTM gas turbines | Permit controversy; speed prioritized over optics |
| Oracle/OpenAI | Stargate West TX | 1,200+ | Gas turbines (construction power) | "Interim power" during renewable buildout |
| Microsoft | Multiple PJM sites | Various | Grid (fossil-heavy PJM mix) | REC/PPA offset strategy |
| Utilities (Duke, Dominion, AEP) | 20+ GW announced by 2040 | 20,000+ | Natural gas | Two-thirds tied to DC growth; IRP-approved |
| Structure | How It Works | ESG Benefit |
|---|---|---|
| Behind-the-Meter (BTM) | Power generated on-site, never touches the grid | Scope 1 (operator's), not Scope 2 (buyer's). DC tenant may not report. |
| Utility PPA (front-of-meter) | Power sold through utility as intermediary | Blended into grid mix. Scope 2 diluted across all ratepayers. |
| REC + Carbon Offset Bundle | Gas generation paired with purchased RECs and offset credits | Net-zero accounting on paper; "matched" clean energy claims. |
| Hydrogen-Ready Specification | Aeroderivative gas turbines rated for future H2 blend | "Transition fuel" narrative; capex preserves option value for decarbonization. |
| Gas-to-Nuclear Transition | Gas provides bridge power; nuclear PPA layered in 2030+ | Credible decarbonization pathway; long-term clean energy commitment. |
| Scope | What It Covers | Relevance |
|---|---|---|
| Scope 1 | Direct emissions from owned/operated generation | Applies to BTM gas on-site. Reported by power operator, not DC tenant. |
| Scope 2 | Indirect emissions from purchased electricity | Grid-connected DC power. "Market-based" accounting allows REC offsets; "location-based" reflects actual grid emissions. |
| Scope 3 | Value chain / supply chain emissions | Cloud customers' use of AI services. Not currently auditable at scale. |
Bottom line: A BTM gas plant at a DC site puts emissions in Scope 1 of the power operator, not Scope 2 of the hyperscaler tenant. Market-based Scope 2 reporting lets RECs offset any grid-purchased power on top of that. This is exactly why hyperscalers prefer buying power from intermediary operators rather than owning generation directly.
Coal carries the most ESG sensitivity. But the policy landscape has shifted since 2025: DOE has issued 16+ emergency orders under Section 202(c) to delay ~4.5 GW of coal retirements, the AI Action Plan tells agencies to prioritize "dispatchable" fossil power, and OBBB killed clean energy tax credits for new solar/wind. For coal-adjacent sites, the move is gas conversion while preserving the grid interconnection:
| Objection | Response | Precedent / Evidence |
|---|---|---|
| "We can't be associated with coal" | Gas conversion eliminates coal operations. Site retains grid interconnection, transmission, and water rights. Hyperscaler buys power from gas, not coal. | DOE has issued 16+ emergency orders delaying ~4.5 GW of coal retirements for grid reliability. PJM fast-tracking CIR transfers from deactivating coal to new gas. |
| "Our sustainability report won't allow it" | BTM structure: power operator reports Scope 1 emissions. Hyperscaler tenant's Scope 2 uses market-based accounting with bundled RECs. | Meta-Williams Ohio deal: on-site gas generation, Meta claims "transition infrastructure." Amazon Scope 2 uses market-based method. |
| "Investors and activists will flag this" | Hydrogen-ready turbine spec gives a real transition story. Gas bridges to nuclear/SMR with a 2030+ PPA layered in. 24-month timeline vs. 55-month queue means $100M+/yr in revenue that would otherwise be left on the table. | All major hyperscalers now openly procuring gas power. xAI Memphis built in 90 days with gas turbines. |
| "Why not just build renewable?" | OBBB killed ITC/PTC for new solar/wind. 5% safe harbor is gone (IRS Notice 2025-42). Battery completion rate in queues: 5%. Gas: 17%. Renewables don't deliver firm 300+ MW baseload on 24-month timelines. | LBNL: Solar completion 9%, Wind 15%, Gas 17%. Gas+storage is the only tech actually delivering firm DC power at scale right now. |
Sources: S&P Global hyperscaler procurement analysis, McKinsey 24/7 clean power study, ESG Today, CleanTechnica, EESI, Utility Dive, EPA Clean Air Act data center guidance, LBNL Queued Up 2025.
PJM capacity prices went up ~11× in three years: $28.92/MW-day (2024/2025) to $333.44/MW-day (2027/2028, FERC cap). Data centers are responsible for 40% of total capacity costs ($6.5B) and 63% of the price increase. PJM forecasts ~32 GW of new demand by 2030, almost all AI/DC. As of summer 2026, PJM barely has enough supply to keep the grid reliable.
| Revenue Stream | $/MW-day | Annual (1 GW) | Notes |
|---|---|---|---|
| Capacity (BRA) | $333 | $121.7M | FERC cap; locked for delivery year |
| Energy Market | ~$80-150 | $29-55M | PJM LMP; varies by dispatch, location |
| Ancillary Services | ~$15-30 | $5.5-11M | Frequency regulation, reserves |
| DC Lease / PPA Premium | Negotiated | $50-200M+ | Direct hyperscaler offtake at premium to market; Talen/AWS = $1.4B/yr for 1.9 GW |
Total range: $206-388M/yr per GW from market revenues alone, plus DC lease/PPA upside on top. At Talen/AWS economics (~$730M/yr per GW), the valuation math for interconnection-ready sites changes completely.
| Factor | Gas | Solar/Wind | Nuclear | Storage |
|---|---|---|---|---|
| Queue Completion Rate | 17% | 9% / 15% | 21% | 5% |
| Capacity Factor | 80-90% (baseload) | 20-30% | 90%+ | 4-8 hrs duration |
| Time to COD | 24-36 months | 36-48 months | 10+ years (new build) | 24-36 months |
| Firm Dispatchable? | Yes | No (intermittent) | Yes | Limited duration |
| Tax Credit Post-OBBB | None needed | Terminated | Preserved (45U/45Y) | Preserved (48E) |
| YoY Queue Growth | +72% (LBNL) / +160% (PJM) | -12% / -26% | +modest | -13% |
Sources: PJM BRA auction results (2024/2025 through 2027/2028), PJM 2025 Long-Term Load Forecast, PJM Market Monitor analysis, IEEFA capacity price analysis, Utility Dive, LBNL Queued Up 2025, FERC filings.
Hyperscalers have announced ~44 GW of DC capacity. The constraint stack shows that only 10-20% of that is physically deliverable within the announced timelines. Gas turbine lead times (42-60 months), CoWoS packaging deficits (6.9M GPUs), PJM queue attrition (77% nationally per LBNL), and permitting bottlenecks all point to the same conclusion: we are in a prolonged scarcity environment where L1 and L3 pricing power persists well beyond 2028. The "overbuild" narrative assumes supply chains that do not exist at the required scale.
| Constraint | Announced | Deliverable | Gap | Why It Can't Close Fast |
|---|---|---|---|---|
| Gas Turbines | 47.5 GW by 2028 | ~15.8 GW/yr AI-available | (27.7 GW) | 42-60 month manufacturing lead time; orders after Q2 2024 miss 2028 |
| CoWoS Packaging | 7.3M GPU demand | 402K capacity | (6.9M GPUs) | TSMC fab expansion takes 2-3 years; advanced packaging even longer |
| PJM Interconnection | 211.5 GW active queue | ~30 GW post-attrition | 77% attrition (LBNL) | 55 months avg queue timeline (2024); PJM paused new reviews 2022, resuming 2026 |
| DC Construction | 44.9 GW planned | ~5-9 GW deliverable by 2028 | 80-90% gap | Labor, steel, copper, land permitting all constrained simultaneously |