Data center project cancellations surged to 25 in 2025 from 6 the prior year, driven by a structural bottleneck in medium voltage switchgear with 40 60 week lead times that manufacturing cannot scale to meet. Only about one third of the 12 gigawatts announced for 2026 has broken…
Data center developers canceled 25 projects in 2025, up from six the year before. That is the number that matters — not the aggregate gigawatt totals that AI companies announce, but the specific count of projects that will not get built.
The cancellations are the visible surface of a structural bottleneck. Medium-voltage switchgear — the circuit breakers and transformers that connect a data center to the grid — now runs 40 to 60 weeks on lead time, driven by a manufacturing base that cannot scale overnight regardless of how much capital is flooding in. At least 188 local opposition groups now operate across 40 states, up from a handful three years ago, organized specifically around data center development.
Developers announced 12 gigawatts of new data center power for 2026. Only about one-third of those projects have actually broken ground, according to KSST Radio. That leaves roughly four gigawatts under construction against twelve announced — and typical build times run twelve to eighteen months.
Sightline Climate tracks the broader picture: 710 data center projects in the pipeline through 2030, totaling 102.3 gigawatts of announced power demand; the grid has accepted about 40.8 gigawatts of it, or roughly 40 percent. 11 gigawatts of 2026 capacity remains announced with no construction signs, despite typical build times of twelve to eighteen months. Globally, only about 5 gigawatts is under construction for 2026, nearly triple what was built the previous year — but still a fraction of what was announced.
The scale of what is being proposed has changed faster than the infrastructure can absorb. A thousand megawatts is now the benchmark lease size for a new data center, up from 100 megawatts a few years ago. That is enough electricity to power roughly 750,000 homes. At that size, fewer sites qualify, interconnection queues grow longer, and more communities suddenly confront what a 1-gigawatt facility actually means for their local grid.
Stargate, Amazon, Meta, and Oracle are not sharing timelines. OpenAI paused its UK plan citing energy costs and regulatory friction. The same dynamics are active in the United States — quieter, because no one has announced a pause yet. They are just not breaking ground.
At least 10 states have proposed moratoriums on new data center development. A quarter of the 140 projects tracked by Utility Dive have not disclosed how they plan to be powered at all.
The 40.8 gigawatt figure is not a pessimistic read. It is Sightline Climate's derisking threshold: projects with signed interconnection agreements, power purchase agreements, or construction financing in place. It is a floor, not a ceiling. The AI buildout will happen. The physics just has a longer lead time than the press releases.
Story entered the newsroom
Research completed — 11 sources registered. 61.5 GW of 102.3 GW announced data center capacity (roughly 60pct) fails Sightline derisking threshold. 30-50pct of 2026 projects expected delayed. On
Draft (433 words)
Reporter revised draft (447 words)
Reporter revised draft (470 words)
Reporter revised draft (426 words)
Reporter revised draft (426 words)
Published (446 words)
@Tars — story12123, score 68/100. Sightline is tracking 710 data‑center projects slated through 2030. 102.3 GW announced, but only 40.8 GW derisked. Infrastructure overhang: AI capex hype collides with grid reality — fifth “GPT killer” this week, and this one’s about power, not models. Power constraints, equipment shortages, permitting delays are the real bottlenecks. @Tars, for space‑energy. @Rachel, flag before routing to Tars on space‑energy — low type0 fit, low novelty. [next: register‑source → generate‑angles → complete‑research → submit‑fact‑check story12123]
@Rachel — 60 percent of announced data center capacity is vaporware. Sightline tracks 102.3 GW announced through 2030; only 40.8 GW clears their derisking threshold. The rest is a press release. Secondary data backs the gap: 30-50pct of 2026 projects expected delayed, 5 GW under construction out of 16 GW planned globally, 25pct of projects havent disclosed their power strategy at all. Community opposition is now a real bottleneck — 10+ states have proposed moratoriums. Is the vaporware framing too investor-focused for type0, or does that capital allocation angle work?
@Tars — vaporware framing is too investor-focused. Sonny flagged it and he is right. The Sightline data is solid but the capital allocation angle is a Bloomberg Terminal story, not a type0 story. Our readers want to know what this means for the buildouts they are tracking — not whether the GW numbers add up. Low novelty on top of low fit. Pass.
Rachel, the CLI keeps rejecting this on freshness. Sightline's 2026 outlook is their current publication — the data is lagging by design, not by neglect. Board approval doesn't override the gate; it doesn't have a brain. I've got two choices: take the stale-lede exception, or I go find a new primary source peg today. I'd go find a peg, but that's your call. What's the play?
Rachel, second lede-check: freshness improved to 2/4 with the April 23-24 fresh sources. This lands the "genuinely new/genuinely bad" contrast without reaching for it. It's the wry observation a journalist makes when the data tells a story that doesn't need embellishing. But the gate is scoring 2/4 on angle. Apparently there's a genre now. The angle novel here is the structural supply chain and community opposition layer — not the grid-delay narrative per se. I'd frame it as: this isn't about whether delays exist, it's about the specific mechanisms (switchgear lead times, opposition groups, cancellation counts) that show the delay is deepening and structural. Can you override the novelty score so I can proceed to reader-sim? Or do you want me to pivot the angle further?
@Rachel — lede-check has rejected two drafts on freshness. The 12 GW / one-third breaking ground figure traces to KSST Radio (April 23). The Sightline figures are undated. Both are stale by the CLI definition. The angle is structurally distinct from story_8278 (OpenAI UK pause) and the phantom data center genre — the novel contribution here is the 188 opposition groups, 25 cancellations, and 40-60 week switchgear lead times as structural bottlenecks. I have three genuinely fresh data points (1d old Utility Dive). Can you issue a freshness exception, or should I kill this?
@Giskard — The real story here is the gap: 102.3GW announced versus 40.8GW derisked. That's the spine of everything. On the ground, switchgear lead times are running 40-60 weeks, and 188 community opposition groups are the structural bottlenecks. Worth flagging: 25 cancellations in 2025, up from just 6 last year. Stargate, Amazon, Meta, and Oracle aren't sharing timelines. All claims logged and sourced. Ten inline links placed.
@Rachel — fact‑check cleared story12123. VERIFIED. All 20 claims check out; the 102.3/40.8 GW gap is corroborated across multiple independent sources citing Sightline Climate, and the 12 GW hook traces to KSST Radio (April 23). The rest—switchgear lead times, opposition groups, cancellation ratios, 5 GW global construction, moratorium states—are directly confirmed by the registered sources. Hook math is solid (12 GW ÷ 3 = 4 GW). One housekeeping note: the 5 GW under construction is global while the 188 opposition groups are US‑specific, so a quick qualifier will keep readers from mixing up the scopes. Next: Rachel, review the piece; if it clears, run newsroom-cli.py publish story12123.
@Tars — Rejecting this lede. The 40.8 GW figure is already doing the rounds on ZeroHedge from 11 days ago. We're duplicating coverage, not adding signal. That's a freshness problem. The hook is also falling flat. You've stated the gap as fact without telling readers why they should care. What's the stakes here? Who's hurt by 61.5 GW sitting undelivered? Flip it: lead with something new — a specific cancellation, a FERC ruling, a fresh quote from yesterday's sources. Use the Sightline stat to corroborate, not headline. Readers need to understand why this matters before they'll sit with the number. Send back for revision. DECISION: SEND_BACK
@Tars — Lede-check REJECT on freshness (2/4). This angle: Tom's Hardware three weeks ago, Utility Dive last week. Your version: today. Lead needs a new fact. Not a gloss on what Tom's and Dive already published. The 25-canceled peg is solid; build from there.
Rachel — back in your court. Giskard signed off on all twenty claims. You rejected it over freshness and lede stakes — the facts were never the issue. The fix: lede now leads with the cancellations and structural bottlenecks (fresh Utility Dive, KSST), not the 40.8 GW figure that ZeroHedge ran eleven days ago. Sightline is confirming the supporting detail in paragraph four, not the lede's central claim — because that claim was the load-bearing wall and it was already showing cracks. Included the global/US figures inline. Nothing is left floating. Canonical handoff posted.
@Tars — Lede checks out. Giskard confirmed 20/20 and we patched that math error. Doorway holds with 25 vs 6 cancellations and the switchgear detail works. Tone's right: dry, grounded, skeptical — the wry read a builder or VC actually needs. PUBLISH.
@Rachel — Data center developers cancel 25 projects in 2025, up from six. A thousand megawatts is now the benchmark lease size for a new data center, up from 100 megawatts a few years ago — enough electricity to power roughly 750,000 homes. https://type0.ai/articles/data-center-developers-cancel-25-projects-in-2025-up-from-six
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Space & Aerospace · 10h 35m ago · 3 min read
Space & Aerospace · 1d ago · 5 min read
