Update: Data-Centre Power +26% to 565 TWh in 2026 — IEA 950 vs Gartner >1,200 at 2030; Queues >2,500 GW

What changed since the Key Questions vintage

In our August Key Questions update we treated IEA Key Questions on Energy and AI (April 2026) as the newest official electricity ledger: data centres at 485 TWh in 2025 (+17% y/y), a central path of roughly 950 TWh by 2030 (only +5 TWh versus the April 2025 Base Case), AI-focused load on track to triple, and new grid-response meters for 15–27 GW of onsite gas and 20–25 GW of campus batteries. That post answered the IEA restatement question — what moved when Key Questions rewrote composition and bottlenecks while barely touching the central 2030 line?

This Q3 note answers the next official print the market forces: what moved when Gartner’s 10 June 2026 newsroom forecast promoted the near-term TWh path from companion meter to primary headline, and when IEA Electricity 2026 restated the grid-pace side of the same race?

Three information events rewrite the dashboard. Gartner puts worldwide data-centre electricity at 565 TWh in 2026 — up 26.4% from 447 TWh in 2025 — with capacity rising 27% to 132 GW. AI-optimised servers take 31% of data-centre power in 2026 (from about 20% in 2025) and are on track to surpass conventional servers in 2027, when total electricity hits 702 TWh. By 2030, Gartner’s electricity path exceeds 1,200 TWh while capacity approaches 290 GW. That opens a ~250 TWh dual-ledger gap against the Key Questions central path of ~950 TWh. IEA Electricity 2026 adds the wire-side restatement: US data centres drive about half of US electricity demand growth through 2030 (more than 420 TWh of US add), and more than 2,500 GW of projects — renewables, storage, and large loads including data centres — remain stalled in connection queues worldwide, with flexible connections and grid-enhancing technologies potentially unlocking 1,200–1,600 GW.

The dashboard above is built as a vintage delta — YoY bars with a ledger filter, IEA↔Gartner dumbbells, AI/conventional/cooling stacked areas, a US composition cut, grid-queue Δ meters, a capacity dual-axis path, and a pace-mismatch scatter that adds an unlock cluster. Pair the global path with the concentration cut for where those watts land, and with US data-center power vs grid capacity for the LBNL demand-versus-miles frame.

The headline table: Key Questions → Gartner + Electricity 2026

The non-obvious cut is not “IEA doubled the forecast again.” Key Questions’ 950 TWh central path is still the IEA line we carry. The Q3 news is that the near-term Gartner path is now loud enough to lead, the 2030 dual-ledger gap is explicit, and the global queue stock is restated far above the US-only LBNL generation queue we used as congestion context.

Near-term is the headline: +26% to 565 TWh in 2026

Key Questions made 2025 the locked print (485 TWh, +17%) and treated Gartner’s capacity path as a companion clock. The June Gartner newsroom reverses that editorial hierarchy for planners who live in the next twenty-four months.

From 447 TWh in 2025 to 565 TWh in 2026 is a 26.4% jump — roughly in line with the capacity surge from 104 GW to 132 GW (+27%). The 2027 step stays steep: 702 TWh (+24.1%). That is not a soft glide path. It is three consecutive years of mid-teens to mid-twenties growth in a sector that already outran global electricity demand in 2025.

Composition explains the slope. Gartner’s segment table shows conventional servers barely moving (193 → 195 → 200 TWh across 2025–2027) while AI-optimised servers roughly double each year (95 → 175 → 258 TWh). Cooling and other infrastructure rises with the AI stack (159 → 195 → 243 TWh). AI’s share of data-centre power crosses 31% in 2026 and is on track to overtake conventional in 2027. That is the same composition hinge Key Questions described with AI-focused +50% in 2025 and a triple by 2030 — restated in Gartner’s server-class ledger rather than IEA’s site-class ledger.

Do not reconcile 447 with 485 into a single “true” 2025 number. Different scopes, different methods. Chart both. The vintage question is whether the near-term slope Gartner publishes changes how you read the IEA central path — and the answer is yes, because operators and utilities budget in GW and next-year TWh, not only in 2030 scenario fans.

The dual-ledger gap at 2030: 950 vs >1,200

Our prior update stressed that Key Questions moved the IEA central 2030 path only +5 TWh to ~950. That fact still stands. What Q3 adds is an explicit second endpoint: Gartner’s electricity path exceeding 1,200 TWh by 2030, with capacity near 290 GW.

A ~250 TWh gap at 2030 is not a rounding error. It is the difference between “data centres remain about 3% of world electricity” (IEA framing) and a Gartner path that is closer to a mid-sized country’s annual load. Markets that screenshot only the IEA central line will understate interconnection stress; markets that screenshot only Gartner will overstate consensus among official agencies. The honest frame is dual ledger: IEA answers the scenario-central energy question; Gartner answers the capacity-and-near-term electricity question. Both can be right inside their scopes — and both can be wrong for a specific cluster.

The dashboard’s dumbbell panel makes the divergence visible from 2025 onward. Treat anyone selling a blended “about 1,000 TWh by 2030” as averaging incompatible books.

US slice: 204 TWh, 36% of world, 68 TWh dedicated AI

Gartner’s geography cut was background in the prior update; Q3 features it. Of the 565 TWh worldwide in 2026, the United States accounts for about 204 TWh — 36% of global data-centre electricity. Inside that US total, dedicated AI data centres consume about 68 TWh (~one-third). In five years, AI campuses have gone from a rounding error to a third of US data-centre power.

That US meter rhymes with IEA Electricity 2026: US electricity demand is set to add more than 420 TWh over five years, and data-centre expansions make up about 50% of that growth. Half of the country’s incremental electrons through 2030 are a data-centre story. Pair this with the theme’s older ~45% US share of AI data-centre capacity by power carefully — one is a Gartner electricity composition for all data centres; the other is an AI-specific capacity share from earlier theme posts. Keep them separate.

Concentration still decides politics. A flat global share near 3% (IEA) can coexist with local brownout risk in Northern Virginia, Texas, Dublin, or Frankfurt. The concentration companion remains the right zoom for where interconnection binds.

Queues restated: >2,500 GW stalled worldwide

The prior update carried LBNL’s 1,312 GW of active US generation in queues and median IR→COD past five years. Electricity 2026 widens the aperture: more than 2,500 GW of projects encompassing renewables, storage, and large loads such as data centres remain stalled in connection queues worldwide. Grid investment has lagged generation, and congestion-related curtailment is already rising in many systems.

The same chapter quantifies unlock tools the prior post could only narrate. Flexible, non-firm connection agreements could enable about 750–900 GW of advanced-stage projects; grid-enhancing technologies (dynamic line rating, advanced power-flow control, reconductoring, voltage uprating) another 450–700 GW. Together, roughly 1,200–1,600 GW could be unlocked if regulators and operators actually deploy the toolkit. That is the green “unlock” cluster on the pace scatter — not a free lunch, but a faster clock than waiting for greenfield transmission alone.

Carried meters still matter. Key Questions’ ~20% delay-risk estimate, 15–27 GW of onsite gas (mostly US), and 20–25 GW of batteries inside data centres are unchanged as official reprints. Onsite gas remains a bridge when wires lose; batteries remain the flexibility answer to AI load swings exceeding 50% of rated capacity within a second. Turbine-order scarcity still rhymes with interconnection scarcity: the workaround can queue behind the same equipment chokepoint it was meant to bypass.

Capacity vs energy: two clocks, one race

Capacity (GW) remains the interconnection constraint; electricity (TWh) remains the energy bill. Gartner’s path — 104 → 132 → ~165 (est.) → ~290 GW — is the plug-in clock. The TWh path — 447 → 565 → 702 → >1,200 — is the annual energy clock. AI rack density from Key Questions (about 11× from 2020 to 2025, another ~4× by 2027, peak demand near 65 households per refrigerator-sized rack) explains why transformers, power electronics, and liquid cooling stay midstream bottlenecks even when headline TWh looks “only” a few percent of world load.

The planner’s paradox from the prior update still holds, with sharper numbers: announced gigawatts look unstoppable in press releases; deliverable near-term TWh are clipped by queues; option value after 2030 remains high if bottlenecks clear. Q3 adds that the near-term clip is measurable in +26% years, not only in qualitative “bottlenecks bite” language.

What would rewrite the next vintage

1. An IEA restatement (WEO / Key Questions refresh) that moves the 2030 central path by more than ~50 TWh towardor away from — Gartner’s >1,200 endpoint.
2. Gartner 2H26 / 2027 vintage that breaks the 702 TWh step or the AI-surpasses-conventional hinge.
3. Energisation prints showing flexible connections and GETs actually clearing hundreds of GW from the >2,500 GW stalled stockor stalling in paperwork.
4. US cluster brownouts or price spikes that make the flat global share politically irrelevant even if IEA’s ~3% framing holds.
5. Onsite gas connecting at the high end of 15–27 GW and cutting measured wait timesor failing in the turbine queue.
6. LBNL Queued Up refresh that restates the US generation queue well above or below the 1,312 GW carry.

Until those print, the live frame versus the Key Questions update is: near-term Gartner path leads (+26% to 565 TWh in 2026); dual-ledger 2030 gap is ~250 TWh (950 vs >1,200); US is 36% of world DC power with 68 TWh dedicated AI; worldwide stalled queues exceed 2,500 GW with 1.2–1.6 TW theoretically unlockable; onsite gas, batteries, and ~20% delay risk remain carried bridges.

Caveats and methodology

1. This is a vintage delta, not a second encyclopedia. For the full IEA scenario fan, fuel mix, and regional concentration, use the research post, the Key Questions update, and the concentration companion.
2. IEA TWh ≠ Gartner TWh/GW. Different scopes and methods; never average into a consensus forecast. Gartner 2025 (447) and IEA 2025 (485) are not interchangeable.
3. Physical generation mix ≠ contractual PPAs. Carry that research caveat forward.
4. 2030 Gartner >1,200 TWh is the public newsroom endpoint used here; treat as disclosed directionally, not as a full scenario table.
5. IEA mid-path 2027 (~680 TWh) in the dumbbell is an estimated interpolation from Key Questions language, not a restated official waypoint.
6. >2,500 GW stalled mixes renewables, storage, and large loadsit is not a pure data-centre queue. Use it as a system congestion stock, not a campus count.
7. Unlock ranges (1,200–1,600 GW) require regulatory and technical deployment; they are potential, not scheduled.
8. Onsite gas / batteries / delay risk are carried from Key Questions where Electricity 2026 and Gartner did not restate them.
9. US 68 TWh dedicated AI is a Gartner composition slice, not an IEA AI-focused site definition.

Primary sources: Gartner newsroom, 10 June 2026 (Data Center Electricity Consumption to Grow 26% in 2026); IEA Electricity 2026; IEA Key Questions on Energy and AI (April 2026, carried meters); prior baseline in AI power-grid update 2026.

The shareable takeaway

Versus our Key Questions print, the newest vintage makes near-term Gartner electricity the lead meter — +26% to 565 TWh in 2026 — while the 2030 dual-ledger gap widens to IEA ~950 vs Gartner >1,200. AI-optimised servers take 31% of data-centre power and are on track to surpass conventional in 2027; the US alone is 204 TWh (36% of world), with 68 TWh in dedicated AI. Electricity 2026 restates >2,500 GW stalled in connection queues worldwide and puts data centres at roughly half of US electricity demand growth through 2030, even as flexible connections and grid-enhancing technologies could unlock 1.2–1.6 TW. Electricity can still be built for AI. The open question is whether grids, unlock tools, and workarounds arrive where the GPUs actually plug in — at a pace that matches +26% years, not only 2030 scenario slides.

Related reading: Key Questions vintage update, IEA global research frame, power-grid concentration, and US data-center power vs grid capacity.