Capacity, Projections, and the Recursive Horizon - Mapping the $725B infrastructure buildout reshaping the global energy grid
The world's largest technology companies are collectively committing over $725 billion in capital expenditure in 2026 alone - nearly doubling 2025 levels and exceeding global oil & gas upstream investment for the first time.[1] This report maps the compute buildout across hyperscalers and AI labs, projects US and global datacenter power demand through 2030, identifies the physical bottlenecks constraining deployment, and analyzes the "recursive horizon" - the phenomenon where AI efficiency gains are overwhelmed by exploding inference demand from reasoning and agentic workloads.
Bottom line: We are witnessing the largest infrastructure buildout since rural electrification, compressed into a decade. The binding constraint is not capital - it's watts. Every major player is now effectively in the power business. For seed investors, the opportunity set lies in the picks-and-shovels of power delivery: grid-enhancing technologies, onsite generation, battery storage, cooling, and the software layer that optimizes utilization of constrained compute.
The scale of capital deployment into AI infrastructure in 2026 has no precedent in the history of enterprise technology. Combined capital expenditure of the five largest US cloud and AI infrastructure providers - Amazon, Microsoft, Alphabet, Meta, and Oracle - is projected between $725 billion and $775 billion for 2026, nearly doubling 2025 levels and exceeding global investment in oil and natural gas production.[1][7] In Q1 2026 alone, Alphabet, Amazon, Microsoft, and Meta reported combined quarterly capex exceeding $130 billion.[8]
| Company | Operational Capacity | Contracted / Under Construction | 2026 Capex | Key Power Deals |
|---|---|---|---|---|
| Microsoft / Azure | ~6 GW as of Q3 FY2026. Added 1 GW in a single quarter; ~5 GW at FY2025 end. 400+ DCs, 70+ regions[52] | Multi-GW pipeline. Still "capacity constrained" through 2026 (CFO Amy Hood)[52] | ~$190B[10] | World's largest corporate renewable buyer (34 GW PPAs). TerraPower nuclear: 345 MW by late 2027. Revenue/GW: $6.2B annualized[52] |
| Amazon / AWS | Not disclosed. Cumulus Data campus (PA): 960 MW, co-located with 2.5 GW Susquehanna nuclear[53] | Massive pipeline. Project Rainier (Anthropic). 1 GW+ Trainium2/3 by end 2026[14] | ~$200B[9] | Talen/Susquehanna nuclear PPA: 1,920 MW, $18B, 17 years - largest nuclear-to-DC deal ever. Trainium to save "tens of billions" annually[53] |
| Alphabet / Google | Not disclosed at GW level. Consumed 30.8 TWh in 2024 (doubled from 2020)[54] | Intersect Power acq ($4.75B): "multiple GW" of energy + DC projects[54] | $180-190B[10] | Intersect Power ($4.75B). Kairos Power/TVA: 50 MW nuclear. NextEra partnership. Cloud backlog $462B. 2027 capex to "significantly increase"[54] |
| Meta | 26 operational DCs globally (end 2025). Total MW not disclosed[55] | Hyperion (Richland Parish, LA): 2-5 GW campus, 4M sq ft. Entergy: 6.7 GW dedicated generation (10 gas plants). $27B off-balance-sheet (Blue Owl)[55] | $125-145B[10] | Entergy: 6,700 MW new gas gen (~half of all Entergy Louisiana capacity). Exploring nuclear[55] |
| Oracle | Not disclosed. 8 AI superclusters under development[11] | Stargate partner (joint sites with OpenAI/SoftBank). Backlog $523B[11] | ~$50B (FY2026)[11] | Primary infra partner for Stargate (Abilene TX + 5 sites). Q3 FY2026 capex: $9.1B/qtr[11] |
| OpenAI / Stargate | Under construction. First site: Abilene, TX[12] | ~7 GW planned; 10 GW target. UAE: up to 5 GW (G42)[12][13] | $500B (SoftBank, Oracle, OpenAI)[12] | Oracle as primary developer. Crusoe Energy, JPMorgan $2.3B loan. 4.5 GW Oracle-OpenAI joint development[12] |
| Anthropic | No owned DCs. Uses >1M Google TPUs + >1M AWS Trainium chips[14][15] | >5 GW committed across partners. Google >1 GW. AWS 1 GW Trainium2/3. Broadcom "multiple GW." Leasing xAI Colossus 1[14][15][18] | Via partners: Google $40B, Amazon $33B, MS/Nvidia $15B[16] | Most compute-diversified frontier lab. Multi-cloud = hedge + signal no single hyperscaler can satisfy frontier demand[15] |
| xAI / Colossus | 555K GPUs (H100/H200/GB200). ~2 GW site. Built 100K GPUs in 122 days[17] | Target: 1M GPUs. Colossus 1 now leased to Anthropic[18] | $18B silicon at Colossus[17] | GE Vernova / Solar Turbines gas for initial power. Memphis chosen for power access[33] |
AI margins are already better than early cloud margins were at a comparable stage, per Microsoft CFO Amy Hood.[10] Microsoft's revenue per GW of datacenter capacity reached $6.2B annualized in Q3 FY2026, up from $4.8B a year earlier - AI extracts far more economic value per watt than traditional cloud.[52] This is the single most important signal for sustainability of the capex cycle.
The table above shows who owns datacenter infrastructure. But who actually uses the compute? An analysis by Aymeric Roucher, building on Epoch AI's Frontier Data Centers dataset and supplementary curated data, reveals a strikingly different picture when capacity is attributed to the tenant running the workloads rather than the facility owner:[89][90]
| Compute User | Total Capacity (GW) | Owns DCs? | Key Insight |
|---|---|---|---|
| OpenAI | 15.3 GW | No (Stargate under construction) | Largest compute consumer globally. Capacity via Oracle, Azure, SoftBank JVs |
| Google DeepMind | 6.0 GW | Yes (Google-owned) | Vertically integrated: owns DCs, designs TPUs, trains models |
| Anthropic | 5.1 GW | No | Most diversified: Google TPUs + AWS Trainium + xAI Colossus + Broadcom |
| Meta | 3.8 GW | Yes | Fully self-hosted. Hyperion (2-5 GW) will make Meta the largest owner-operator |
| Microsoft | 2.0 GW | Yes (~6 GW owned) | Owns 3× what it consumes for its own AI — the rest serves Azure tenants |
| xAI | 1.5 GW | Yes (Colossus) | Colossus 1 now leased to Anthropic; building Colossus 2 |
Second-Order The user-view reveals a structural asymmetry: the largest consumers of compute (OpenAI, Anthropic) own none of the physical infrastructure, while the largest owners (Microsoft, Amazon, Google) derive much of their value from serving these tenants. OpenAI commands 15.3 GW despite owning zero operational datacenters — it is the world's largest compute renter. Third-Order This creates a power dynamic where infrastructure providers hold structural leverage over AI labs, and any disruption in these partnerships (as seen with the weakening Microsoft-OpenAI relationship) has immediate compute access implications. Labs that diversify their compute sources (Anthropic's multi-cloud strategy) are hedging this concentration risk.[89]
| Provider | 2023 | 2024 | 2025 | 2026E | 2027E | 2028E |
|---|---|---|---|---|---|---|
| AWS | 7.200 | 9.200 | 13.139 | 18.585 | 24.195 | 29.295 |
| GCP | 4.000 | 5.625 | 7.605 | 11.976 | 17.277 | 21.896 |
| Azure | 4.600 | 6.300 | 8.900 | 12.700 | 18.200 | 22.500 |
| Oracle | 0.575 | 0.775 | 1.455 | 3.490 | 6.205 | 10.875 |
| Meta | 4.975 | 6.000 | 7.500 | 11.353 | 15.839 | 20.548 |
| Total CSPs | 21.350 | 27.900 | 38.599 | 58.057 | 81.659 | 104.957 |
| xAI | 0 | 0.212 | 1.314 | 2.000 | 4.500 | 7.000 |
| Neoclouds* | 0.112 | 0.465 | 2.217 | 7.357 | 13.394 | 21.941 |
| Combined Total | 21.462 | 28.577 | 42.130 | 67.414 | 99.553 | 133.898 |
| Incremental Y/Y | - | +7.115 | +13.553 | +25.285 | +32.139 | +34.345 |
*Neoclouds: CoreWeave, Nebius, Nscale, Crusoe, FluidStack, Lambda, IREN. Source: Wells Fargo Securities LLC Estimates; BloombergNEF.[84]
The gap between announced capacity and energized capacity is the most important structural risk in the compute buildout. Hyperscalers can commit hundreds of billions, but physical infrastructure moves on the timeline of transformers and turbines - not software.
Substation transformer lead times have worsened steadily: ~140 weeks in 2023, ~150 weeks in 2025, exceeding 160 weeks (3+ years) in 2026.[4] US GSU transformer demand soared 274% between 2019 and 2025 (Wood Mackenzie).[47] Supply deficit was ~100% in 2025, projected <10% by 2030 as new capacity comes online.[46] Eaton committed $340M for a third US transformer plant; Hitachi Energy and Siemens scaling globally. But new factories take 2-3 years to build - bottleneck persists through 2028.[60]
The IEA documented a 70% surge in global gas turbine orders in 2025.[1] GE Vernova's gas power backlog expanded from 62 GW to 83 GW (Q4 2025), targeting 110 GW by year-end 2026. Power orders surged 77% QoQ, with 41 heavy-duty gas turbines ordered in one quarter.[56] Siemens Energy: orders €58.9B (+19.4% YoY), multiyear backlogs.[57]
| ISO / Region | Queue Backlog | Wait Time | Key Dynamic |
|---|---|---|---|
| ERCOT (Texas) | 226 GW large load (Nov 2025). 77% data centers[58] | 4× in one year (from 63 GW). Only 23 GW new gen added 2024-2025 | "Phantom load" problem: speculative requests. Texas law (Jul 2025) now requires transparency[58] |
| PJM (Mid-Atlantic) | ~46 GW remaining. 160 GW studied total[59] | New cycle: 1-2 years (vs. historic 5-6+). Transition Cycle 2 processing through end 2026 | Includes NoVA (world's largest DC market). Downstream transmission upgrades still lag[59] |
| MISO / CAISO / Others | Implementing FERC Order 2023 cluster-based reforms | 3-5+ years typical. CAISO: 4+ years for generation interconnection historically | Generation queue: 77% solar + storage (ERCOT). Load queues less standardized than gen queues |
Industry data compiled by @negligible_cap shows the full scope of the buildout across all player categories - including xAI (0.2 → 7.0 GW) and neoclouds like CoreWeave, Nebius, and Crusoe (0.1 → 21.9 GW). Combined total: 21.5 GW (2023) → 133.9 GW (2028E) - a 6.2× increase in five years.[84] Incremental deployment is projected at +25 GW by end of 2026, +32 GW more by end of 2027, and +34 GW more by end of 2028. But the gap between announced and energized is wide: Sightline Climate found 12 GW US capacity announced for 2026 with only 5 GW under construction (58% gap).[31] Bloomberg (Apr 2026): >50% of US DCs planned for 2026 delayed.[32] JLL: >50% of global projects delayed 3+ months in 2025; equipment lead times 33 weeks (50% above pre-2020).[6] LBNL "Queued Up" data: historically only 13% of capacity in interconnection queues reaches commercial operation; median wait 4.5 years.[45]
| Company | Margin Trajectory | Key Data |
|---|---|---|
| GE Vernova | Near-zero at spin-off (Apr 2024) → 12-14% EBITDA (2026, raised twice) | Power profit +57%. Electrification profit doubled. Stock +13% on earnings. Revenue: $44.5-45.5B. Gas turbine backlog: 110 GW target[56] |
| Siemens Energy | Negative (wind losses) → 6% profit margin, higher in grid tech | Orders €58.9B (+19.4%). Multiyear gas turbine backlogs[57] |
| Eaton | Electrical segment margins consistently 20%+ | $340M new US transformer plant. DC power = fastest-growing vertical[60] |
| Vertiv | Gross margin: 28.4% (2022) → 37.8% (Q3 2025) | Power management / cooling for DCs. 9.4 ppt margin expansion in 3 years[50] |
GE Vernova is the clearest example: near-zero-margin spin-off (Apr 2024) to 12-14% EBITDA by 2026, driven by pricing power from DC and grid demand. Gas turbine backlog: 62 GW → 110 GW in ~18 months. Vertiv's gross margin expanded nearly 10 points in 3 years. When hyperscalers commit $725B in a single year, suppliers of physical infrastructure can name their price. This dynamic won't normalize before 2028-2029.[56][50]
The frontier AI labs have emerged as among the largest individual consumers of compute, rivaling the hyperscalers' own internal workloads:
| Lab / Project | Compute Footprint | Power Capacity | Capital Committed |
|---|---|---|---|
| OpenAI / Stargate | 10 GW target; ~7 GW planned across multiple sites[12] | First site: Abilene, TX (Oracle). 5 additional sites announced Sep 2025 | $500B total commitment (SoftBank, Oracle, OpenAI). $100B deployed immediately. UAE expansion: up to 5 GW (G42)[13] |
| Anthropic | Multi-cloud: 1M+ Google TPUs; 500-600K AWS Trainium chips; leasing xAI Colossus 1[14] | Google: >1 GW in 2026. AWS: additional 1 GW Trainium2/3 by end of 2026. Broadcom: "multiple gigawatts"[15] | Google: up to $40B investment. Amazon: up to $33B total ($25B latest round). Microsoft/Nvidia: $15B. $30B+ ARR[16] |
| xAI / Colossus | 555,000 Nvidia GPUs (H100/H200/GB200) as of Feb 2026. Target: 1M GPUs[17] | ~2 GW site capacity (Memphis, TN). Built 100K GPUs in 122 days; doubled to 200K in 92 more | $18B in silicon at Colossus alone. Anthropic leasing Colossus 1 as of May 2026[18] |
Anthropic has become the most compute-diversified frontier lab in history - simultaneously running on Google TPUs, AWS Trainium, Azure GPUs, and now leasing xAI's Colossus facility. Total committed capacity across all partnerships exceeds 5 GW. This multi-cloud strategy is both a hedge against vendor lock-in and a signal that no single hyperscaler can satisfy frontier training demand alone.[14][15]
Nvidia remains the central beneficiary of the compute arms race. Data center revenue reached $39.1 billion in Q1 FY2026 (calendar Q1 2025), up 73% year-over-year, and grew to $51.2 billion in Q3 FY2026 (+66% YoY).[19][20] For Q1 FY2027 (calendar Q1 2026), consensus estimates project data center revenue of approximately $73 billion - nearly doubling in 12 months.[21]
The Blackwell NVL72 architecture is now in "full-scale production" across all major cloud providers, with the next-generation GB300 sampling and the Vera Rubin platform announced at GTC 2026 - promising 10× inference throughput per watt versus Blackwell.[19][22] Nvidia's annual product cadence through 2028 means each generation of datacenter is obsolescent before the next one is built - a structural advantage that forces continuous capex.
Before projecting forward, we need to ask: how reliable are the forecasts we're relying on? The answer - based on tracking forecast vintages from the same sources - is more nuanced than the common narrative of "everyone underestimates." Some forecasters have overshot near-term. Others have consistently revised upward. And the gap between major forecasters may matter more than the gap between any individual forecast and reality.
Goldman Sachs has been the most prolific publisher of datacenter power forecasts - and the most consistent at raising them. Tracking their 2030 global datacenter power demand forecast (expressed as % increase vs. 2023 baseline of ~415 TWh):
| Date Published | Report Title | 2030 Forecast (% increase vs. 2023) | Implied 2030 TWh |
|---|---|---|---|
| May 2024 | "AI Is Poised to Drive 160% Increase in Data Center Power Demand" | +160% | ~1,079 TWh[61] |
| February 2025 | "AI to drive 165% increase in data center power demand by 2030" | +165% | ~1,100 TWh[62] |
| November 2025 | "The 6 Ps Driving Growth and Constraints" | +175% | ~1,141 TWh[24] |
| April 2026 | Latest revision (Kobeissi Letter / Benzinga reporting) | +220% | ~1,350 TWh[63] |
Goldman raised their 2030 target by 60 percentage points in 23 months - from +160% to +220%. The April 2026 revision was the most aggressive single jump (+45 ppts from the November 2025 figure). They also increased the US share of global demand from 50% to 60%, projecting US datacenter consumption alone at ~750 TWh by 2030.[63]
The IEA's forecast history tells a more surprising story. Their January 2024 Electricity 2024 report contained the headline: "Electricity consumption from data centres, AI and the cryptocurrency sector could double by 2026."[64] With the 2022 global baseline at ~460 TWh, this implied ~920 TWh by 2026. Actual 2025 consumption came in at ~485 TWh, and 2026 is projected around 550-600 TWh[1] - well below the "doubling" scenario.
The IEA overestimated near-term growth. Their high-scenario was too aggressive for the 2024-2026 window. However, their 2030 base case has been remarkably stable: 945 TWh in the April 2025 report, ticking up to 950 TWh in the April 2026 update.[1] This stability may reflect the IEA's recognition that physical bottlenecks (grid interconnection, transformer supply) create a ceiling on near-term deployment - demand exists, but infrastructure can't deliver it fast enough.
The IEA's US-specific forecast for 2026 (260 TWh, published Jan 2024) appears roughly on track - their own November 2025 data projects >250 TWh for 2026.[25] The near-term US projections have held better than the global aggregate.
McKinsey published the most granular US datacenter electricity forecast in their November 2024 "AI's Power Binge" report:[65]
| Year | McKinsey US DC Forecast (TWh) | McKinsey (% of US total) | IEA Actual/Estimate | Delta |
|---|---|---|---|---|
| 2023 | 147 TWh | 3.7% | 154 TWh (IEA) | -4.5% |
| 2024 | 178 TWh | 4.3% | 183 TWh (IEA) | -2.7% |
| 2025 | 224 TWh | 5.2% | >200 TWh (IEA proj.) | TBD |
| 2026 | 292 TWh | 6.5% | >250 TWh (IEA proj.) | TBD |
| 2028 | 450 TWh | 9.3% | ~350 TWh (IEA proj.) | - |
| 2030 | 580 TWh | 11.7% | - | - |
McKinsey's 2023-2024 estimates were within 3-5% of IEA's later actuals - the best near-term accuracy of any forecaster surveyed. Their 2026 projection of 292 TWh is notably higher than IEA's >250 TWh, potentially reflecting faster AI adoption than IEA models. On global capacity, McKinsey projected 171-219 GW by 2030 (Oct 2024), which they later refined to 219 GW (May 2025).[66]
| Forecaster | Track Record | Current 2030 Global Estimate |
|---|---|---|
| Goldman Sachs | Has consistently underestimated. Four upward revisions in 23 months (+60 ppts). Most reliable as a leading indicator: when GS raises, the market follows | ~1,350 TWh global (+220% vs 2023). US: ~750 TWh[63] |
| IEA | Overshot near-term ("could double by 2026" didn't happen) but stable on 2030 base case. Cautious, bottleneck-aware. US near-term forecasts surprisingly accurate | ~950 TWh global (base case). High case not quantified publicly[1] |
| McKinsey | Most accurate near-term (within 3% on US 2024). Most aggressive capacity projections. Has been validated by subsequent data | 219 GW global capacity. US: 580 TWh. $7T cumulative investment[65][66] |
The prediction gap is not a simple story of universal underestimation. Three dynamics are at work:
1. Goldman Sachs has been a serial underestimator - and they know it. Each revision admits the prior one was too low. Their trajectory suggests 220% may not be the final number.
2. The IEA is bottleneck-aware - their stable 950 TWh base case may reflect a realistic ceiling imposed by grid/transformer constraints, not conservative demand modeling. If the bottlenecks ease faster than expected, their number will look too low.
3. The gap between forecasters is as important as the gap between forecast and reality. IEA says 950 TWh; Goldman says 1,350 TWh. That's a 42% spread on a 4-year forecast. McKinsey's capacity numbers (219 GW) imply even more if utilization rates hold. The honest answer: nobody knows, and the range of plausible outcomes is wider than any single number suggests.
If Goldman's average miss rate holds (~15-20% per annual revision), their current 1,350 TWh estimate for 2030 could be revised to 1,550-1,600 TWh by early 2027. But - and this is critical - physical infrastructure sets a hard ceiling on what can actually be deployed. The LBNL "Queued Up" data shows only 13% of capacity in interconnection queues reaches commercial operation.[45] The real question is not "how much demand exists?" but "how fast can supply respond?" The answer to the second question likely puts realized 2030 demand closer to IEA's 950 TWh than Goldman's 1,350 TWh - even if underlying demand would support the higher number.
McKinsey's April 2025 analysis provides the most detailed decomposition of the $7 trillion in cumulative datacenter capital expenditure projected through 2030:[67]
| Cost Category | Share | Amount | Key Components |
|---|---|---|---|
| Technology (chips, GPUs, servers) | 60% | ~$3.1T | GPU/XPU procurement (Nvidia, AMD, custom silicon), server assembly, networking equipment, HBM memory. This is where Nvidia's dominance translates to margin capture |
| Energizers (power + cooling) | 25% | ~$1.3T | Power generation/transmission ($720B grid spending needed per GS), cooling infrastructure, transformers, switchgear, electrical distribution. The bottleneck layer |
| Builders (real estate + labor) | 15% | ~$1.0T | Construction labor (~$0.6T), shell & site (~$0.3T), land acquisition (~$0.1T). Skilled labor shortages driving shift to prefabrication/modular |
Of the total, AI-specific datacenters require $5.2 trillion while traditional IT workloads require $1.5 trillion.[67] Goldman Sachs' May 2026 "Tracking Trillions" framework projects $7.6 trillion cumulative AI CapEx between 2026 and 2031 alone, growing from $765B annually in 2026 to $1.6T in 2031 - anchored to Nvidia forward data center revenue estimates.[85] The concentration of cost in technology (60%) explains why Nvidia captures outsized economics - but the bottleneck is in the energizer layer (25%), where delivery delays strand the other 75% of invested capital.
Rabobank's May 2026 analysis identifies four binding constraints that create a physical ceiling on buildout speed:[68]
The historical pattern in computing is unambiguous: efficiency gains increase, not decrease, total consumption. Every major advance in computing efficiency has expanded the addressable market faster than it reduced per-unit cost.
The current cycle is the most extreme example yet. Dell COO Jeff Clarke (May 19, 2026): token costs fell 80% in one year, but token consumption for reasoning surged 320×.[5] Net effect: 64× increase in total compute spend (0.20 × 320 = 64). Nvidia CEO Jensen Huang reported inference token generation surged 10× in one year.[19]
The IEA confirms this at the macro level: energy per individual AI task is "improving at a rate unprecedented in energy history" - dropping by at least an order of magnitude annually. But new use cases (video generation, reasoning, agentic workflows) consume hundreds to thousands of times more energy per query.[1] Efficiency unlocks demand categories that didn't previously exist.
Every data point confirms Jevons paradox dominates: Nvidia's Vera Rubin promises 10× lower cost per token vs. Blackwell.[22] This will not reduce total compute demand. It will unlock use cases currently constrained by economics - persistent agentic systems, real-time video generation, continuous scientific simulation - driving another wave of demand. The only scenario where efficiency caps total demand is one where AI capability plateaus. No credible observer is predicting that.
Evercore ISI introduced a token consumption model on May 20, 2026 that provides the most granular demand-side framework published by a tier-1 sell-side firm:[82][83]
| Year | Annual Tokens (Quadrillions) | YoY Growth | Dominant Category |
|---|---|---|---|
| 2024A | ~100 | - | Frontier training (59%) |
| 2025E | ~200 | ~100% | Frontier training |
| 2026E | ~350 | ~75% | Frontier training (59%), consumer inference (26%) |
| 2027E | ~700 | ~100% | Shift to inference-dominant |
| 2028E | ~1,500 | ~114% | Agentic AI rising rapidly |
| 2029E | ~2,500 | ~67% | Agentic AI + enterprise inference |
| 2030E | ~4,000 | ~60% | Agentic AI becomes largest consumer |
Key insight: agentic AI is negligible today (7% of 2026 tokens) but becomes the largest single category by 2030. This is the demand wildcard - agents consume tokens continuously, without human rate-limiting. Evercore estimates compute capacity demand could reach 250 GW by 2030, assuming blended throughput of 500K transactions/sec/MW.[83]
For context: Goldman Sachs' latest forecast implies ~150-170 GW by 2030 (at their +220% power demand growth rate). McKinsey projects 219 GW. Evercore's 250 GW is the most aggressive mainstream projection yet - 18% above Goldman and 14% above McKinsey. If Evercore is right, the implied cumulative capex exceeds $10 trillion.
A 40× increase in annual token demand (100Q → 4,000Q) in four years implies that every efficiency gain from Vera Rubin, Blackwell Ultra, and custom silicon gets consumed by demand expansion - and then some. This is the Jevons paradox quantified at the application layer: cheaper tokens don't reduce total compute consumed. They create use cases (persistent agents, continuous reasoning, real-time video) that didn't exist at higher price points.
Sell-side capacity deployment estimates show a clear acceleration curve, broadly consistent with Goldman Sachs' facility-level data from Aterio:[84]
| Year | Incremental GW Deployed | Source |
|---|---|---|
| 2024 (actual) | 6.4 GW | Goldman Sachs / Aterio[3] |
| 2025 (actual) | 8.5 GW | Goldman Sachs / Aterio[3] |
| 2026 (projected) | ~25 GW | Sell-side est.[84] |
| 2027 (projected) | ~32 GW | Sell-side est.[84] |
| 2028 (projected) | ~34 GW | Sell-side est.[84] |
Note that Goldman's May 20, 2026 report projects 13.6 GW in 2026 and 36.3 GW in 2027[3] - the 2027 figures are broadly consistent (~32-36 GW range), but Goldman's 2026 figure (13.6 GW) is significantly below the sell-side aggregate (25 GW). This may reflect the gap between scheduled activations (Goldman/Aterio) and contracted/announced capacity (sell-side estimates that include projects not yet under construction). The delta - roughly 11 GW - is consistent with the "announced vs. energized" gap we documented in Section 1.
| Scenario | Assumption | 2030 Global DC Demand | Implied Capex | Key Risk |
|---|---|---|---|---|
| Bear Case | Supply chain constraints create a hard ceiling. Transformer/grid bottlenecks persist. Only 50% of announced capacity comes online. LBNL 13% completion rate continues | ~120-140 GW operational (~950 TWh, IEA base)[1] | ~$5T (constrained) | Stranded capital: $725B/yr committed but physically unable to deploy. Hyperscaler ROI timelines extend. AI revenue growth decelerates before capex cycle completes |
| Base Case | Current trajectory continues. Grid investments begin easing bottlenecks by 2028. Onsite generation fills gaps. No step-change in demand or efficiency | ~170-220 GW (~1,100-1,350 TWh, GS/McKinsey range)[63][66] | ~$7T (McKinsey estimate) | Execution risk: actual buildout tracks 60-70% of plans. Regional concentration creates grid stress in key markets (Virginia, Texas, Dublin) |
| Bull Case | RSI kicks in ~2028. Token demand 2-5× current projections. Agentic AI creates unbounded inference demand. Every enterprise deploys autonomous agents | ~300-400 GW (~1,800-2,500 TWh) | ~$10-15T cumulative | Physically impossible on current grid trajectory. Requires massive onsite generation buildout. Water becomes binding constraint. Social backlash accelerates |
Is 2-5× current demand projections by 2030 realistic? The Dell COO data suggests we've already seen 64× total compute spend growth in one year (80% cost reduction × 320× consumption). If agentic AI reaches mainstream enterprise deployment and RSI accelerates model capability, 2-5× is conservative relative to what the token consumption curve implies. The constraint isn't demand - it's whether the physical world can deliver. The honest answer: demand may support 5×+ but realized consumption will be limited to whatever the grid, transformers, and supply chain can actually deliver. The gap between demand and deliverable supply is where the investment opportunity lies.
Every scenario model breaks if AI achieves recursive self-improvement (RSI) - the ability to autonomously improve its own capabilities, creating a feedback loop where each generation of AI is designed by a more capable predecessor. This is no longer a thought experiment. The leaders of every major AI lab are now publicly discussing RSI as a near-term possibility.
| Who | Role | Statement | Date | Timeline |
|---|---|---|---|---|
| Jack Clark | Co-founder, Anthropic | "I think there's a 60%+ chance that recursive self-improvement will occur before the end of 2028." Based on weeks reading hundreds of public data sources[71] | May 4, 2026 | End of 2028 (60%) |
| Dario Amodei | CEO, Anthropic | "This loop has already started, and will accelerate rapidly." Transformative AI potentially late 2020s. AI doing all software engineering tasks within "6-12 months"[72] | 2026 | Late 2020s |
| Sam Altman | CEO, OpenAI | "We are only a couple of years away from early versions of true superintelligence." Blog: 2026 = novel insights, 2027 = real-world robots. "By end of 2028, more intellectual capacity inside data centers than outside"[73] | Feb 2026 | 2028 (ASI) |
| Demis Hassabis | CEO, Google DeepMind | "AGI is now on the horizon... foothills of the singularity." 50% chance by 2030. "10× the impact of the Industrial Revolution, at 10× the speed"[74] | May 20, 2026 | 2028-2031 (50%) |
| Elon Musk | CEO, xAI | AGI surpasses any single human by end of 2026. Exceeds collective human intelligence by 2027. Grok 5: 10% probability of AGI[75] | Jan 2026 (Davos) | 2026-2027 (most aggressive) |
| Shane Legg | Co-founder, Google DeepMind | 50% probability of "Minimal AGI" by 2028. Has maintained this prediction since 2009 - 17 years of consistency[76] | Ongoing (since 2009) | 2028 (50%) |
OpenAI co-founder and former Tesla AI chief Andrej Karpathy — one of the most recognized names in AI research — announced he joined Anthropic's pre-training team on May 19, 2026, stating: "I think the next few years at the frontier of LLMs will be especially formative."[86] Market commentary interpreted this as an RSI signal: "This indicates that we are close to RSI and therefore an accel in model IQ increases. In that scenario the value of compute is going to explode."[87] Whether or not the inference is correct, the talent consolidation at Anthropic (Karpathy + Jack Clark's 60% RSI prediction + Colossus lease + $30B+ ARR) is the strongest single-company signal in the RSI race.
What's remarkable is not any individual prediction — it's the convergence. The CEOs and founders of OpenAI, Anthropic, Google DeepMind, and xAI are all pointing to 2026-2030 as the window for AGI/ASI, with the center of mass around 2028. These are not neutral observers - they have incentives to hype timelines. But they also have the most information about current capability trajectories. Jared Kaplan (Anthropic) told The Guardian humanity faces "the biggest decision yet" between 2027 and 2030.[71]
If RSI arrives circa 2028, the compute implications are qualitatively different from current scaling:
Phase 1 AI-accelerated AI research (now - 2027): AI systems assist human researchers in model design, data curation, and hyperparameter optimization. This is already happening - Anthropic, OpenAI, and DeepMind all use AI to accelerate their research. Compute demand grows at current rates (17-22% CAGR per Goldman/McKinsey).
Phase 2 Autonomous AI research (~2027-2028): AI systems design and run their own experiments with minimal human oversight. Training runs happen continuously rather than episodically. Demand for compute becomes continuous and compound rather than periodic. This is the phase Jack Clark's 60% probability applies to.
Phase 3 Recursive loop (~2028+): Each AI generation designs a more capable successor. Training and inference merge into a single continuous process. Demand becomes theoretically unbounded - limited only by available compute, energy, and capital. This is where all current forecasts break.
If Phase 3 materializes, every 2030 projection in this report - Goldman's 1,350 TWh, IEA's 950 TWh, McKinsey's 219 GW - becomes an underestimate by definition. The demand curve goes vertical. The only binding constraint becomes physics: how fast can we build power plants, manufacture chips, and lay transmission lines? This is not a risk to the capex thesis - it's the most extreme version of it. The companies that have secured power, land, and supply chain commitments before RSI arrives will be the most valuable infrastructure assets on earth.
It would be intellectually dishonest to present RSI timelines without noting the skeptic's position. Hassabis himself notes his definition of AGI requires capabilities - robust creativity, scientific discovery - that current systems lack.[74] Grok's own estimate of AGI timing (2040-2050) is notably more conservative than its creator's.[75] Key objections:
RSI creates two opposing forces on compute demand simultaneously:
Force 1 - Demand explosion: RSI generates continuous, compounding training runs. Each improvement cycle requires a new training run on the full model - or more precisely, on a model that is already larger and more capable than its predecessor. A research survey of 25 leading AI researchers from DeepMind, OpenAI, Anthropic, Meta, UC Berkeley, Princeton, and Stanford (arXiv:2603.03338, February 2026) found that participants converged on predictions that AI agents will transition from "assistants" to "autonomous AI developers" - but after that point, predictions diverged sharply. The frontier lab researchers were more bullish on explosive growth than academics.[77]
Force 2 - Efficiency improvement: A self-improving AI would also optimize its own architecture, training procedures, and inference efficiency. DeepSeek R1 demonstrated this in early 2025: algorithmic innovation achieved comparable performance at dramatically lower compute cost. An RSI system would discover such optimizations continuously. The paper "The Race to Efficiency" (arXiv:2501.02156) argues that classical scaling laws neglect time and efficiency, and that algorithmic improvements are compounding alongside hardware improvements.[78]
Which force dominates? Historically, efficiency gains in computing have always been overwhelmed by demand expansion (Jevons paradox). But RSI introduces a genuinely novel dynamic: for the first time, the system generating demand is the same system optimizing efficiency. Deloitte's 2026 analysis explicitly concludes that "AI's next phase will likely demand more computational power, not less" - efficiency gains at the model level are consumed by the explosion in inference volume and the shift to more compute-intensive reasoning architectures.[79] The academic evidence supports the view that demand dominates, but with meaningful efficiency offsets that prevent the curve from going truly vertical.
If AI can improve its own models, can it also optimize the physical infrastructure it runs on? Several mechanisms are already emerging:
Net assessment: AI will meaningfully improve infrastructure efficiency (perhaps 2-3× over current levels by 2030). But this improvement is multiplicative on a base that is itself growing exponentially. A 3× efficiency gain on a 10× demand increase still means a 3.3× net increase in physical infrastructure needed.
If forecast undershooting persists at the Goldman Sachs historical rate (~15-20% per annual revision) AND RSI accelerates demand on top of that, the math produces sobering numbers:
| Parameter | 2026 (Today) | 2030 (Base + RSI) | 2033 (Extrapolated) |
|---|---|---|---|
| Global DC electricity demand | ~550-600 TWh | 1,500-2,000 TWh (IEA base + GS correction + RSI uplift) | 3,000-5,000 TWh (est. - data is thin, labeled as extrapolation) |
| US share | ~250-300 TWh (~6% of US electricity) | ~900-1,200 TWh (~15-20% of US electricity) | ~1,500-2,500 TWh (~25-35% of US electricity, est.) |
| Global DC capacity | ~68 GW operational | 200-350 GW | 400-700 GW (est.) |
| Cumulative capex | ~$1.5T (2024-2026) | $7-15T (2024-2030) | $15-30T (2024-2033, est.) |
The 2033 column is extrapolation, not forecast. No credible source publishes specific projections beyond 2030 for datacenter demand under RSI scenarios. These numbers are derived from: (1) Goldman's historical miss rate applied forward, (2) the token consumption growth rate documented by Dell/Nvidia, and (3) the assumption that RSI adds a 2-3× multiplier on top of baseline growth. They should be treated as directional, not precise. The honest answer: if RSI arrives circa 2028 and forecasts continue undershooting, the 2030 numbers in this report will look conservative in hindsight.
The academic literature on RSI and compute scaling is thin but growing rapidly:
For investment purposes, the question isn't whether RSI arrives on schedule - it's whether the belief in near-term RSI sustains the capex cycle. As long as hyperscalers and sovereign governments believe AGI/ASI is 2-5 years away, the infrastructure buildout continues regardless of whether the timeline proves accurate. The capex is committed; the compute will be built; the only question is what runs on it.
The United States is the epicenter of the global datacenter buildout. Accounting for approximately 50% of global datacenter capacity and ~90% of Americas regional capacity, the US power grid is absorbing the most acute demand shock in its history.[6]
Goldman's latest analysis projects US capacity additions accelerating dramatically:[3]
| Year | US DC Capacity Additions | YoY Change | Source |
|---|---|---|---|
| 2024 (actual) | 6.4 GW | - | Goldman Sachs / Aterio[3] |
| 2025 (actual) | 8.5 GW | +33% | Goldman Sachs / Aterio[3] |
| 2026 (projected) | 13.6 GW | +60% | Goldman Sachs / Aterio[3] |
| 2027 (projected) | 36.3 GW | +167% | Goldman Sachs / Aterio[3] |
Globally, Goldman Sachs forecasts datacenter demand growing approximately 50% to 92 GW by 2027, at a compound annual growth rate of 17% between 2025 and 2028. In a bullish scenario with stronger GPU power draw and higher customer demand, CAGR could reach 20%.[23] By 2030, data center power demand is projected to surge 165-175% versus 2023 levels - "the equivalent of adding another Top 10 power consuming country."[24]
The International Energy Agency's updated Key Questions on Energy and AI report (April 2026) projects global datacenter electricity consumption roughly doubling from 485 TWh in 2025 to approximately 950 TWh by 2030 - slightly more than Japan's entire electricity consumption.[1] For the United States specifically:
| Year | US DC Electricity Demand (TWh) | Source |
|---|---|---|
| 2024 | ~170 | IEA[25] |
| 2025 | >200 | IEA[25] |
| 2026 | >250 | IEA[25] |
| 2027 | >300 | IEA[25] |
| 2028 | ~350 | IEA[25] |
| 2029 | ~400 | IEA[25] |
The IEA estimates US electricity demand will add more than 420 TWh in total over the next five years, with datacenters making up approximately 50% of that demand growth.[26] The agency also notes that US datacenters will increase their share of regional power from 4% to 7.8% between 2025 and 2030.[27]
The IEA notes that energy per individual AI task is "improving at a rate unprecedented in energy history" - dropping by at least an order of magnitude annually. Simple text queries now consume less electricity than running a television over the same period. But new use cases (video generation, reasoning, agentic workflows) consume hundreds to thousands of times more energy per query than simple text. The net effect: demand growth overwhelms efficiency gains.[1]
An important distinction: JLL's 103 GW figure (2025) represents all datacenter capacity globally — including traditional enterprise IT, cloud storage, and legacy workloads. Epoch AI independently estimates that AI-specific datacenter power reached approximately 30 GW in Q4 2025, based on cumulative chip sales multiplied by rated power and a 2.5× overhead multiplier for servers, cooling, and infrastructure. This means AI represents roughly 30% of total datacenter capacity today — but is driving approximately 80%+ of incremental growth. By 2030, AI's share of total DC power is projected to exceed 50%.[90][91]
JLL's 2026 Global Data Center Outlook provides the most comprehensive regional breakdown of installed capacity and projected growth:[6]
| Region | 2025 Capacity | 2030 Projected | Growth | Key Dynamics |
|---|---|---|---|---|
| Americas (US ~90%) | ~50 GW | ~100 GW | ~100% | Fastest absolute growth. 35 GW under construction in N. America. 97% occupancy[28] |
| Asia-Pacific | 32 GW | 57 GW | +78% | Colocation-led growth. On-premise declining 6% as enterprises migrate to cloud[6] |
| EMEA (Europe + ME + Africa) | ~20 GW | ~33 GW | +65% | Growth in London, Frankfurt, Paris hubs. Europe: DC share of power from 2.7% to 5%[27] |
| Middle East (subset of EMEA) | ~1 GW | ~15 GW | +1,400% | 2.2 GW under construction, 12 GW planned. Stargate UAE (5 GW). Sovereign AI strategies[29] |
The Middle East represents the most extreme growth trajectory: 12 GW planned against just 1 GW of existing capacity, per JLL.[29] This is driven by sovereign AI mandates (Saudi Arabia's HUMAIN, UAE's G42/Stargate partnership) and abundant natural gas for power generation. The OpenAI Stargate UAE project alone targets up to 5 GW.[13] This is geopolitically significant: these countries are positioning themselves as compute exporters.
Our analysis to this point has been US-centric — reflecting where the data is strongest. But China is building a parallel AI compute infrastructure at scale. Using chip-spend-derived estimates (with ±30-50% uncertainty on Chinese hyperscaler figures), Roucher's analysis provides the first comprehensive per-entity breakdown of Chinese AI compute capacity:[89]
| Entity | Type | ~2024 (MW) | ~2026-27 (MW) | ~2030 Target (MW) | Notes |
|---|---|---|---|---|---|
| ByteDance | Hyperscaler | ~300 | ~1,300 | ~3,500 | Largest Chinese AI compute consumer. Seed, Doubao models |
| Alibaba | Hyperscaler | ~400 | ~1,800 | ~3,000 | Qwen models run on fleet. Largest cloud provider in China |
| Huawei | Chip maker + cloud | ~300 | ~480 | ~2,000 | Ascend 910B/C chips (~3 kW each). Building domestic alternative to Nvidia |
| Tencent | Hyperscaler | ~250 | — | ~1,500 | Hunyuan models. Cloud fleet |
| China Mobile | Telco | 130→250 | ~600 | ~1,500 | Hard MW nameplate data (more reliable). AIDC business |
| China Telecom | Telco | 150→500 | — | ~1,500 | Starfire models. Hard capacity data |
| China Unicom | Telco | ~200 | ~700 | — | Smallest of the three telcos |
| Subtotal: Chinese Giants + Telcos | ~1,700-1,800 | ~5,000-6,000 | ~13,000-15,000 |
Chinese pure-play AI labs operate at dramatically smaller scale, and almost all rent compute from the hyperscalers above:
| Lab | Owns/Rents | ~MW | Key Detail |
|---|---|---|---|
| DeepSeek | Owns | ~90 | Only Chinese lab owning hardware (Fire-Flyer cluster). ~10K A100 + 50K H800 |
| Zhipu / Z.ai | Rents | ~110 | ¥1.14B H1-25 cloud spend. $6.5B IPO |
| Moonshot (Kimi) | Rents | ~90 | $20B valuation. Long-context reasoning |
| MiniMax | Rents | ~80 | $6.5B IPO. Hailuo video generation |
| StepFun | Rents | ~55 | $718M raise. Multimodal/video |
Chinese hyperscaler MW figures are chip-spend-derived, not satellite-verified. The conversion uses ~48 MW per $1B AI chip spend and per-GPU power estimates (A100 ~0.7 kW, H100/H800 ~1.4 kW, H20 ~0.9 kW, Ascend 910B/C ~3 kW). The ±30-50% uncertainty range is significant — ByteDance's actual capacity could be anywhere from ~1,750 MW to ~5,250 MW. Telco figures are more reliable (hard nameplate data). The 2030 projections are trend extrapolation, not announced plans.[89]
Second-Order China's combined AI compute (10-15 GW by 2030) approaches US hyperscaler scale. ByteDance at ~3.5 GW is already at Meta-scale (3.8 GW). This has three implications: (1) chip export controls have not prevented China from building substantial AI compute — Huawei's Ascend chips and stockpiled Nvidia hardware fill the gap; (2) global power demand projections that focus only on US/Europe are significantly underestimating total draw; (3) Chinese efficiency innovations (DeepSeek R1's cost breakthrough) emerge from a constrained compute environment where optimization is existential, not optional. Third-Order If China reaches 15 GW by 2030, that's an additional ~130 TWh of annual electricity demand — roughly equal to all of Ireland's current consumption — on top of US/European projections. The global compute race is not US-dominated; it's multipolar.[89]
Europe's position in the frontier AI compute race is quantifiably dire. Mistral, the continent's leading frontier AI lab, operates approximately 90 MW of compute capacity today — compared to OpenAI's 15,300 MW. That's a 170× gap.[89]
| European Entity | Current (MW) | Planned | Status |
|---|---|---|---|
| Mistral | ~90 | ~200 MW by end-2027; ~1 GW aspiration by 2029 | Rents from Azure + Scaleway. No owned DCs |
| Campus IA (MGX-led JV) | 0 (planning) | 1.4 GW near Paris | Mistral is minority partner. 3+ years from full capacity |
The 170× compute gap between Mistral and OpenAI is not just a matter of funding — it reflects a structural disadvantage in power availability, regulatory environment, and infrastructure investment. Campus IA (1.4 GW) is the only project that could begin to narrow this gap, but even at full capacity it would bring Europe to roughly 1.5 GW of frontier AI compute — still 10× behind OpenAI alone. For investors evaluating European AI deals, this compute deficit is the single most important constraint on European labs' ability to compete at the frontier.[89]
Capital is no longer the binding constraint on the compute buildout. The bottleneck has shifted to physical infrastructure - specifically, electrical grid capacity, transformer manufacturing, and the speed of interconnection. This is the most underappreciated risk in the entire AI capex cycle.
Epoch AI's Frontier Data Centers Hub provides independent, satellite-verified tracking of the largest AI datacenter builds — using high-resolution imagery of cooling infrastructure (chillers, cooling towers), cross-referenced with permits and public disclosures. Their data shows five facilities on track to cross 1 GW in 2026:[90][92]
| Facility | Owner / User | Projected 1 GW Date | Build Time to 1 GW | Peak Compute (H100e) |
|---|---|---|---|---|
| Anthropic-Amazon New Carlisle | AWS / Anthropic | January 2026 | ~2.5 years | — |
| xAI Colossus 2 | xAI | February 2026 | ~12 months (projected) | 1.4M H100e |
| Microsoft Fayetteville | Microsoft | March 2026 (borderline) | ~3 years | — |
| Meta Prometheus | Meta | May 2026 | ~2 years | — |
| OpenAI Stargate Abilene | Oracle / OpenAI | July 2026 | ~2 years | — |
Construction-to-1-GW timelines range from 1 to 3.6 years across all tracked facilities, with xAI's Colossus 2 projecting the fastest at just 12 months. Epoch estimates power capacity with 80% confidence within a factor of 1.4×, and timing within 6 months — making this the most rigorous public dataset on actual datacenter build progress. Notably, their tracked facilities (13 large US campuses) account for approximately 2.5 million H100-equivalents, or ~15% of the global AI compute stock.[90]
Looking further out, Epoch projects Meta Hyperion and Microsoft Fairwater will each reach 5 million H100-equivalents when fully complete — a 50× increase over the leading datacenter capacity of mid-2024 (100K H100e). Fairwater is forecasted to be the most expensive datacenter tracked: upwards of $100B in total capital cost upon completion.[92]
Epoch's cost model provides a useful cross-check: $44B per GW of server power ($30B IT hardware + $14B other costs), or approximately $30B per GW of total facility power (since overheads reduce usable server capacity to ~70% of facility power). At JLL's 200 GW 2030 projection, this implies ~$6T in cumulative datacenter capital — consistent with McKinsey's $7T estimate (which includes non-AI workloads). Goldman's $7.6T projection for 2026-2031 AI capex alone suggests either higher per-GW costs or more aggressive capacity deployment than the base case.[90][92]
Average grid connection lead times now exceed four years in primary datacenter markets, according to JLL.[6] The World Resources Institute found that power constraints are extending datacenter construction timelines by 24 to 72 months.[30] Sightline Climate tracked 12 GW of US datacenter capacity announced for 2026, but found only 5 GW actually under construction - the remaining 7 GW (58%) sits in the "announced" stage with no physical progress.[31]
Substation transformer lead times averaged roughly 140 weeks in 2023, increased to 150 weeks in 2025, and now exceed 160 weeks in 2026 - more than three years from order to delivery.[4] Bloomberg reported in April 2026 that more than half of US datacenters planned for 2026 are expected to be delayed due to transformer and electrical equipment shortages.[32]
The IEA documented a 70% surge in global gas turbine orders in 2025, highlighting chokepoints in energy technology supply chains.[1] More than half of datacenter projects globally experienced construction delays of three months or more in 2025, with average equipment lead times at 33 weeks - a 50% increase from pre-2020 levels.[6]
Constrained by grid delays, datacenter developers are increasingly building their own power generation:
Even with behind-the-meter power, the IEA notes that onsite gas generation does not necessarily offer a faster route to market than grid connection at scale - and requires 30-70% overbuilding to handle variable AI loads. The transformer crunch affects onsite generation too, since these facilities still need step-down transformers, switchgear, and distribution infrastructure. There is no shortcut around the electrical equipment supply chain.[1]
Beyond power, the IEA flagged a shortage of high-bandwidth memory (HBM) - integral to AI chip production - that has developed over the past six months and is anticipated to persist through at least the end of 2027.[1] HBM is manufactured almost exclusively by Samsung, SK Hynix, and Micron, creating a three-company chokepoint for the entire AI hardware stack.
Nvidia's annual product cadence drives a continuous upgrade cycle, with each generation rendering the prior generation's datacenters partially obsolescent. This is a structural feature, not a bug - it forces continuous capex and sustains demand for both chips and the physical infrastructure to house them.
Nvidia's response is an annual product cadence promising compounding performance gains:
| Platform | Timeline | Key Advance |
|---|---|---|
| Blackwell (B200/NVL72) | 2025 (shipping) | Full production. 30× inference throughput vs. prior gen (MLPerf)[19] |
| Blackwell Ultra (GB300) | Late 2026 | Sampling Q1 FY2027. Enhanced reasoning performance[19] |
| Vera Rubin | 2027 | 10× lower cost per token vs. Blackwell. Train with 1/4 the GPUs[22] |
| Rubin Ultra | 2028 | Announced at GTC 2026. NVL576 rack-scale systems targeting "millions of GPUs"[22] |
Each generation improves efficiency per watt - but as established, efficiency gains at these magnitudes have historically increased, not decreased, total consumption. The Vera Rubin platform's 10× cost reduction per token will unlock use cases currently constrained by economics, driving another wave of demand.
The compute arms race creates a multi-layered opportunity set for seed investors, organized by where the binding constraints actually sit:
1. Power delivery and grid modernization. The 160-week transformer lead time is the single most cited bottleneck. Companies building grid-enhancing technologies, modular substations, or novel transformer manufacturing capacity sit at the critical path. JLL projects modular datacenter systems reaching $48B in annual sales by 2030.[6]
2. Onsite generation and energy storage. With 15-27 GW of onsite natural gas projected by 2030, and 20-25 GW of battery storage in datacenters, the market for behind-the-meter power systems is emerging at massive scale. Companies optimizing gas turbine deployment, long-duration storage, or fuel cell solutions for datacenter loads are well-positioned.[1]
3. Cooling innovation. Server rack power density increased 11× between 2020 and 2025, with a further 4× increase expected by 2027. By 2027, a single rack could have peak demand equivalent to 65 households.[1] Liquid cooling is now mandatory for AI workloads - Dell's PowerCool CDU C7000 and Lenovo's Neptune business (300% YoY growth) are early signals of a rapidly scaling market.[5]
4. Inference optimization. As inference overtakes training, the software layer that maximizes token output per watt becomes critical. Nvidia's Dynamo open-source library, inference routing systems, and KV cache optimization are emerging categories. Dell projects $50B in AI server revenue for FY2027, with a $43B AI server backlog.[5]
5. Custom silicon and chip supply chain. Amazon, Google, and Meta are all building custom silicon (Trainium, TPU, MTIA). The custom chip ecosystem - chip design (Marvell, Alchip), packaging (TSMC CoWoS), and HBM (SK Hynix, Samsung, Micron) - has become as critical as the GPU itself.
Second-Order The companies that secure power first will build the most capable AI. Power access is becoming a competitive moat for both hyperscalers and startups. This is why JLL characterizes the market as a "bring your own power" mandate in key markets including Dublin and Texas.[6] Third-Order As power becomes the binding constraint, energy cost becomes a structural part of AI product margins. Companies that innovate on power (efficiency, generation, procurement) gain a permanent cost advantage in serving AI workloads.
The data center sector maintains remarkably healthy fundamentals despite the buildout scale. JLL reports 97% global occupancy and 77% of the construction pipeline pre-committed to tenants.[6] AI-focused facilities command 60% lease rate premiums over traditional datacenters. Global datacenter core fund capital formation could top $50B in 2026, with ABS/CMBS issuance projected to reach $50B - roughly doubling every year since 2020.[6]
The IEA's analysis of AI and energy company share prices since the launch of ChatGPT found that AI demand has not provided a generalized uplift to the energy sector - it's too small in context. But manufacturers of gas turbines, electrical equipment, some nuclear companies, and energy startups have seen valuations become "more strongly linked to AI."[1]
Goldman Sachs Research analysts are on "heightened alert for signs of market weakness" from a lack of ways to monetize AI or innovations that commoditize models.[23] Meta's stock dropped 6% after raising capex guidance because Zuckerberg's ROI explanation - "build experiences that can get to billions of people and focus on monetizing them once you get to scale" - did not reassure investors.[10] If revenue growth decelerates before capex cycles complete, the overcapacity risk is enormous.
The IEA notes that datacenters have been "targeted in conflict zones, underscoring their role as critical infrastructure." Trade restrictions have targeted critical minerals for power electronics, batteries, and chips. China dominates supply chains for key components. Export controls (e.g., Nvidia's $4.5B H20 charge) demonstrate the fragility of the global chip supply chain.[1][19]
Community pushback against datacenter projects is a "growing issue" per the IEA, driven by concerns about affordability and environmental impact.[1] Dublin has already implemented effective datacenter moratoriums. US grid stress could trigger similar responses in key datacenter markets like Northern Virginia or Texas.
Breakthroughs in model efficiency (like DeepSeek's R1 cost reductions in early 2025) could temporarily depress demand for new compute. Goldman Sachs explicitly models a downside scenario where demand growth CAGR falls to 14% (vs. 17% base case).[23] However, history suggests efficiency breakthroughs ultimately expand the market rather than shrink it.
Generated by Galileo 🔭 · May 20, 2026