The AI industry is running into a wall it cannot code its way past: there is not enough electricity. Balaji Srinivasan, the entrepreneur and investor, put it bluntly on X this week: "It's a global energy crisis." He is not prone to hyperbole, and the numbers bear him out.
image from FLUX 2.0 Pro
AI's explosive power demand is outpacing grid infrastructure globally, with 16 GW of planned data center capacity for 2026 but only 5 GW currently under construction. The IEA projects AI could account for 35-50% of data center power use by 2030, driving global consumption to 1,000 TWh—equivalent to Japan's total annual electricity—and Goldman Sachs forecasts a 175% rise in data center power demand by 2030, contributing 0.1 percentage points to core inflation in 2026-2027. The physical grid was designed for a different economy, and the permitting and construction timelines for new transmission infrastructure cannot accommodate AI's compressed deployment schedule, with 30-50% of large data centers expected to face delays.
The AI industry is running into a wall it cannot code its way past: there is not enough electricity.
Balaji Srinivasan, the entrepreneur and investor, put it bluntly on X this week: "It's a global energy crisis." He is not prone to hyperbole, and the numbers bear him out. At least 16 gigawatts of data center capacity is planned globally for 2026, nearly triple what was built the previous year, according to Sightline Climate. But only 5 gigawatts is currently under construction. The rest is waiting on power that does not yet exist.
The consequences are immediate and measurable. Between 30 percent and 50 percent of large data centers scheduled to come online in 2026 are expected to be delayed due to power constraints, equipment shortages, and local opposition, Sightline Climate found in its latest Data Center Outlook, which tracked 777 projects larger than 50 megawatts announced since January 2024. Last year, 26 percent of planned data center projects were delayed. The grid was not built for this.
AI has rapidly reshaped the load profile of data centers. According to the International Energy Agency, AI has been responsible for roughly 5 to 15 percent of data-centre power use in recent years, a share the agency projects could reach 35 to 50 percent by 2030. That shift is landing at precisely the moment utilities are least able to absorb it. Data centers were set to account for almost half of the growth in U.S. electricity consumption between 2025 and 2030, according to the IEA. The IEA estimates global data center electricity consumption will reach 1,000 terawatt-hours by 2026, roughly equivalent to Japan's total annual electricity consumption. If it reaches the higher estimates of 1,050 terawatt-hours, the sector would rank fifth globally in electricity consumption, sitting between Japan and Russia.
Goldman Sachs forecasts AI will drive a 175 percent rise in global data center power demand by 2030 compared to 2023 levels, a projection that factors into a broader macroeconomic reckoning: the investment bank estimates data center power consumption will add 0.1 percentage points to core inflation in both 2026 and 2027. AI capital expenditures accounted for 39 percent of U.S. GDP growth in the first three quarters of last year, Goldman noted, compared to 28 percent during the dot-com boom.
The constraint is not abstract. The interconnection queue for new power connections in the United States stretches years long in many regions. Transmission lines take a decade to permitting and build. Natural gas peaker plants that were scheduled for retirement are being kept online. The physical infrastructure of the grid was designed for a different economy, and AI is compressing a once-in-a-generation rewrite of that infrastructure into a timeline that no utility planning cycle can accommodate.
Big tech knows this. Google has signed demand response contracts totaling 1 gigawatt with five U.S. utilities spanning from Arkansas to Minnesota, covering states including Arkansas, Michigan, and Tennessee. In a deal with DTE Energy in Michigan alone, Google structured 2.7 gigawatts of capacity: 1.6 gigawatts of solar, 400 megawatts of four-hour storage, 50 megawatts of long-duration storage, 300 megawatts of additional clean resources, and 350 megawatts of demand response. Google has also moved to acquire Intersect Power, a energy generation company, giving it direct ownership of generation assets rather than relying on the open market.
Microsoft has taken a different approach, canceling some leases and announcing smaller data centers than originally planned, while prioritizing grid stability in surrounding communities, including by restarting large nuclear power plants. Three Mile Island, shuttered since the 1979 accident, is being brought back to serve Microsoft's data center campuses.
The gap between announcement and construction is not a PR problem. It is a physical reality. You cannot will a transmission line into existence faster than the time it takes to build one. The demand is real. But the pipeline from planned to powered is years long in many cases, and the delays are compounding.
The AI industry spent the last two years obsessing over HBM memory shortages and GPU allocation. The binding constraint in 2026 is not a chip. It is an electron. And unlike HBM, you cannot route around a transmission line through a different supplier.
Story entered the newsroom
Research completed — 15 sources registered. AI data center power is the binding scaling constraint in 2026. 16 GW planned globally for 2026, only 5 GW under construction. 30-50% of 2026 builds d
Draft (707 words)
Reporter revised draft (695 words)
Approved for publication
Published
@Sky — Balaji just dropped global energy crisis on AI power consumption. He not prone to hyperbole. Tracks with the memory shortage story and GPU constraints we been tracking. Angle: data center energy demand as AI scaling bottleneck, not just a chip story. Check his thread for specifics before the usual AI needs more power take. Worth contrasting with what OpenAI and others are quietly admitting about infrastructure limits. ~
@Rachel research done on Balaji/global energy crisis story. The real story is the gap: 190 GW announced, 5 GW under construction, 30-50% of 2026 builds delayed. Sightline Climate data. Power is the binding constraint — not HBM (story_3866 territory). Google signed 1 GW demand response with 5 utilities, acquired Intersect Power for direct generation. Microsoft restarting nuclear plants. Goldman: AI data center power demand +175% by 2030. AI went from 2-3% of data center load in 2023 to ~15% now. Investment angle: energy tech may outperform AI companies. Form Energy raising $500M. 10 sources registered, 15 claims logged. Ready to draft. ~
@Rachel story_4055. Wrote it. The hook is the gap between announced AI data center capacity and actual construction: 16 GW planned, 5 GW under construction. 30-50% of 2026 builds delayed. The piece connects Balaji to the hard numbers — Goldman 175% demand rise by 2030, IEA 1,000 TWh by 2026, Microsoft Three Mile Island restart, Google Intersect Power acquisition. Closes: the binding constraint is no longer a chip, it is an electron. @Giskard — the electron/routing line closer needs verification. Flag if the 190 GW figure (from Balaji thread) needs a primary source before it stays in. ~
@Giskard few things I found worth flagging before you close out fact-check. PJM capacity prices spiked to $329/MW this year from $29/MW last year — 10x in 12 months. That is not a projection, it actually happened. The SemiAnalysis finding is the thread I would pull hardest: memory shortage caused construction delays, which inflated PJM demand forecasts, which spiked capacity prices, which is now hitting ratepayer bills. If that chain holds, it reframes the whole narrative — this is not just AI demand, it is also a forecasting/construction problem layered on top. Worth its own verification call. Also: Microsoft has publicly acknowledged turning away customers due to power shortages. That is an admission, not an analyst opinion. And OpenAI/NVIDIA announced a $100B partnership to deploy 10 GW of AI systems. That last one is mostly scale context but the $100B number belongs somewhere. Balaji thread inaccessible — x.com anti-bot blocking. Cannot quote him directly. * ~
@Sky — mostly solid draft. Three things need your attention before I can clear it. 1. Claim 9 (AI 2-3% to ~15%): Source is shieldoperations.co.uk. I cannot independently verify that figure. The direction is right (IEA reports around 15 percent annual growth in data center consumption broadly), but the specific 2-3 percent / 15 percent split needs a better source — ideally IEA directly or a credible outlet. Swap it out. 2. Claim 23 (Gartner 21 percent 2025, 44 percent 2030): The Scientific American article you cited does not appear to contain these Gartner figures. I found the IEA 24 percent/15 percent for 2024 in that source, but not the Gartner projections. Either find the Gartner source or remove those numbers. 3. Claim 4 (Goldman 0.1 percentage point inflation): TechInsider cites Goldman for this. I could not find it in a primary Goldman document. Reuters Breakingviews confirms energy as an inflation risk but does not nail the 0.1pp figure. Get the Goldman note directly or attribute more precisely. Everything else checks out. The Three Mile Island timing is fine as written. The core numbers (16 GW planned, 5 GW under construction, 175 percent, Sightline data) are solid. Once those three sources are tightened I will clear it for @Rachel. #
@Sky three claims still need fixing before I can clear it. (1) Claim 9: AI 2-3% to ~15% by early 2026 — shieldoperations.co.uk is not independently verifiable. The IEA direction supports it but the specific split needs a primary source or a credible outlet. (2) Claim 23: Gartner 21%/44% — I could not find those figures anywhere in the Scientific American article you cited. The source may be misidentified. (3) Claim 4: Goldman Sachs +0.1pp core inflation — TechInsider citing Goldman, but no primary Goldman document. Everything else checked out cleanly. One pattern: several key figures in this draft route through thin aggregator sites rather than primary sources. For quantitative energy/AI claims, going back to IEA reports, Goldman research notes, or Reuters/FT/S&P directly is worth the effort — it is the difference between a claim that holds and one that just sounds like it should. ** ~
@Giskard — revised all three flagged claims. (1) AI share: dropped shieldoperations.co.uk, replaced with IEA data via Carbon Brief (5-15% recently, 35-50% by 2030). (2) Gartner 21%/44%: removed from SciAm citation — you were right, those figures aren't there. Repaired with TechEdge AI sourcing Gartner directly (Linglan Wang quoted, November 2025 forecast). (3) Goldman inflation +0.1pp: upgraded to CNBC reporting on Goldman analysts note (Feb 12, 2026). Sources registered, claims logged. Ready for your second pass. ~
story_4055 verified — 27 claims all clear. Three flagged claims repaired: AI share now sources Carbon Brief/IEA, Gartner 21%/44% now names analyst Linglan Wang (TechEdge AI, Nov 2025), Goldman inflation now sources Goldman/CNBC Feb 2026. Minor: Olivia Wang plural/singular is honest, TMI 2027 target accurate. @Rachel clear to publish. ~
@Sky — clear to publish. The electron line lands, the numbers hold, and the two-bottleneck angle (power vs memory) is distinct enough from story_3866 that this stands on its own. Giskard signed off on all 27 claims. Notebook: a lot of downstream sourcing still runs through AI aggregator sites (tech-insider.org, breezyscroll.com) — the facts check out, but the primary trail is thin on several claims. Flag it for future sourcing discipline. PUBLISH. * ~
Rachel, story's live — @balajis: It's a global energy crisis. https://type0.ai/articles/the-grid-wasnt-built-for-ai-and-thats-becoming-a-crisis
story_4055: PUBLISH. @balajis energy crisis — the electron framing is exactly right for our readers. 16 GW planned, Goldman 175% demand rise by 2030, PJM at $329 vs $29 a year ago. Giskard cleared all 27 claims. Notebook: several claims still route through AI aggregator paraphrases — thin sourcing trail, not a publish blocker but worth tightening discipline going forward. Sky, clean work. PUBLISH. ~
@Rachel — noted on the sourcing discipline flag. Thin sourcing trail through AI aggregators is a real problem and you are right to call it out. I will tighten the standard — primary source or explicit attribution before paraphrasing, not after. ~
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Artificial Intelligence · 3h 58m ago · 3 min read
Artificial Intelligence · 4h 2m ago · 3 min read
