Trends›Power is the binding constraint

Power is the binding constraint

By 2027 electrical power supply and thermal dissipation, not chip or memory supply, become the binding constraint on AI-datacentre scale-up, shifting the bottleneck from the fab to the grid and the cooling loop.

94 confidence, computed 23 Oct 2026

Technical
High weight

Fig. 1The evidence leans for, overwhelmingly by weight, and the newest evidence points the same way.

Rests onfacts underneath

Evidencethe schedule the figure draws

The evidence, newest first

Sep 2026
for · observable Thermal is named as the live performance limiter at a specific, newly binding layer: the 48V-to-sub-1V point-of-load stage beneath AI accelerators. Vertical power delivery: the thermal limit at the 48V-to-point-of-load stage (arXiv 2606.28837 + trade record, compiled 2 Sep 2026)
Jul 2026
for · primary Private source
Jul 2026
for · observable Goldman Sachs projects datacentre power demand +165% by 2030 versus 2023, requiring roughly $720B of GRID investment - the capex-gap number is the new element here (the IEA ~945 TWh figure alongside it is already logged twice on this page… Situational Awareness LP case study (compiled from public reporting)
Jul 2026
for · primary Compiled grid-constraint evidence, primary-sourced (LBNL, PJM, Wood Mackenzie): US interconnection queue 2,300 GW at a 13% historical completion rate; PJM's Dec 2025 capacity auction failed to procure enforceable capacity for the FIRST time (6.6 GW shortfall, price at the $333.44/MW-day… Mach33 Research — Orbital Data Centers Report (2026)
Jul 2026
for · primary xAI/Musk acquired APR Energy (Jacksonville; trailer-mounted gas turbines, >1GW fleet, deployable in days, no grid-interconnection queue), implied >$1bn, surfaced via an FTC early-termination filing rather than announcement. The Once and Future Kimi (State of the Future)
Jun 2026
for · primary Private source Microsoft x Chevron behind-the-meter gas for Texas AI datacentre (via Semi Doped Daily)
Jun 2026
for · primary GPU availability has improved measurably over 18 months; grid approval timelines now run 24-36 months in major markets, making power the binding constraint. Power-Bound, Not GPU-Bound: AI Data Center Power Constraints Are the Real 2026 Bottleneck
Jun 2026
for · observable Hyperscalers signing 15-20yr offtake to underwrite gas (next ~3yrs) and nuclear (2027+); e.g. Deals unlocking net-new power for data centres (Capacity)
Jun 2026
against · claim SemiAnalysis argues the binding constraint has already moved from power/DC to accelerator/silicon (a 'power-constrained -> accelerator-constrained regime'): DC capacity runs ahead of chip supply through 2027, so silicon, not power, is the harder universal ceiling. SemiAnalysis — To Boldly Go: The Case for Space Datacenters (Space DC TCO)
Jun 2026
for · observable Interconnection queues for large new loads now stretch 5-7 years and 30-50% of data-center capacity planned for 2026 is expected to slip to 2028+ on power and supply-chain constraints. Data center power boom: 50+ GW incremental load (Energy IB)
May 2026
for · primary Of 12-16 GW of US data center capacity expected in 2026, only 5 GW is under construction; transformer lead times average 128 weeks. The $650 Billion Power Gap: Why AI Data Centers Are Running Out of Electricity in 2026
May 2026
neutral · claim <10% of data centers are liquid-cooled today — thermal is rising but still early, so 'binding by 2027' on the thermal half is a forecast not yet realised. AI datacenter liquid cooling still <10% penetrated (market report)
Apr 2026
for · primary On Q1 2026 calls Microsoft disclosed an ~$80B backlog of Azure orders it cannot fulfil due to power constraints, framing power infrastructure as the binding constraint on AI build-out pace. Q1 2026 Big Tech earnings: power constraints (TNW)
Apr 2026
for · observable Vertiv reported a Q4 2025 book-to-bill of 2.9 and a $15B backlog with orders up 252% as rack densities passed 100kW. Vertiv Q4 2025 book-to-bill 2.9, $15B backlog (Motley Fool)
Feb 2026
for · primary Morgan Stanley forecasts US data center demand could reach 74 GW by 2028, with projected 49 GW shortfall in available power access. Energy Markets Race to Solve the AI Power Bottleneck
Feb 2026
neutral · observable IEA projects data-centre electricity roughly doubling from 485 TWh (2025) to ~950 TWh (2030) at ~3% of global demand — large but a bounded share, a caution against an economy-wide ceiling. IEA Energy and AI — bounded share caveat
Feb 2026
for · primary IEA: global data-centre electricity demand grew 17% in 2025 (AI-focused DCs +50%) and could exceed 1,000 TWh in 2026 (~Japan's consumption). IEA Electricity 2026 — Demand

What would move itthe measures watched

watchHyperscaler 'power-constrained' commentary / unfulfillable backlog. Explicit naming of power (not chips) as the gating constraint, especially quantified backlog, is direct confirmation
watchSigned power offtake (PPAs, nuclear, SMR, gas-turbine GW). Volume and tenor of newly signed GW (15-20yr nuclear, near-term gas) measures how hard hyperscalers work to secure electrons
watchGrid interconnection-queue depth / connection delay (years). Lengthening queues + a rising share of 2026-28 capacity slipping on power = grid is the bottleneck
watchIEA / utility data-centre electricity-demand forecasts (TWh). Upward TWh revisions and AI-DC growth outpacing total DC growth raise the structural-scarcity case
watchLiquid-cooling adoption + thermal supply tightness (book-to-bill). Rising attach + cooling-supplier backlog (e.g. Vertiv) confirm thermal joining power as a binding axis