Hype Index: 'Half of All Planned U.S. Data Centers Have Been Canceled'

TL;DR: No. The primary source is one hedged sentence from Sightline Climate's February 2026 Data Center Outlook: "it wouldn't be surprising if 30-50% of the capacity slated for 2026 ends up delayed," measured across a global dataset of 777 projects. By the time it reached the internet it had become "half," "U.S.," and "canceled." I fetched 23 articles carrying the claim on August 15, 2026 and classified the 19 that loaded. Only 5 kept the source's own verb. Sightline's own database, as of that report, held nine canceled projects. Verdict: 75% noise / 25% signal.

What is the claim?

Stated the way it circulates:

Half of the U.S. data centers planned for 2026 have been canceled.

You have seen the softer version too: "nearly half canceled or delayed." The hard version is real and load-bearing. One Medium post from April 25 is headlined "Half of All Planned U.S. Data Centers Have Been Canceled," past tense. A viral tweet, quoted in Tech Fund's write-up of the All-In podcast, put it as "a 5-year backlog on grid transformers just killed half of America's 2026 AI data centers."

What does the primary source actually say?

I read it. The origin is Sightline Climate's first Data Center Outlook, published February 24, 2026, now hosted at Currence AI after the firm rebranded. The operative sentence, verbatim:

"In 2025, 26% of expected capacity slipped, and another 10% of projects pushed back their commercial operation dates without much notice. Given that track record, it wouldn't be surprising if 30-50% of the capacity slated for 2026 ends up delayed."

Three things in that sentence get destroyed downstream.

It says delayed, not canceled. It gives a range, 30 to 50, not a point estimate. And the dataset behind it is global: 190GW across 777 large projects above 50MW announced since 2024, of which "at least 16GW of capacity is slated to come online in 2026 across roughly 140 projects" with "only about 5GW currently under construction." Axios, which covered the report the day it dropped, titled its chart "Global data center capacity, by operation date" and led with "as many as half of the world's data center projects". The claim as it now circulates has swapped the verb, taken the top of the range, and changed the continent.

One caveat I owe you: the full report sits behind a request form. The 16GW, 5GW, 11GW and 777-project figures come from Sightline's own public summary page, which I fetched, but the project-level dataset underneath is reported by Sightline and not independently auditable by me or by any of the 20 outlets repeating it.

The number that kills the "canceled" framing came from Sightline directly. Axios: "The firm is tracking nine canceled projects in its database, so for now, most are facing delays, not outright cancellations."

What did the 19 articles do to it?

I fetched 23 URLs carrying the claim on August 15, 2026. Nineteen resolved. Japan Times and The Malaysian Reserve, both running Bloomberg's syndicated copy, returned Cloudflare challenges. Evadaily.com reset the connection. TechRadar served me a headline with an unreadable body, so I scored its headline and excluded it from the citation-chain tally. Bloomberg's own April 1 feature, the piece every aggregator points at, returns 403 to fetchers, so I never read it and I am not going to pretend I did. Its wording is reported by the outlets quoting it.

What the 19 said	Count
Kept the source's verb ("delayed" or "slip")	5
Said "canceled or delayed"	12
Said "canceled" or "killed" outright	2
Reported the full 30-50% range	11
Reported only "half" / "nearly half" / "about half"	7
Escalated past the source ("more than half")	1

That last row is Sandstone Group, which reported "more than half," a number that exists nowhere in the source and sits above the top of its range.

On hops, counting the 18 whose sourcing I could read: it splits exactly down the middle, 9 and 9, so the median article sits one and a half hops from the primary. Nine named Sightline and reached its published figure in a single step. Latitude Media and Axios both interviewed analyst Olivia Wang. TechSpot named the firm and printed the range. Five sat at two hops, citing only Bloomberg. Four never reached a primary number at all, and those are where the damage is: ZeroHedge routed the figure through a Canaccord Genuity analyst note, and tech-insider.org sourced it as "Bloomberg, citing Sightline Climate via The Standard" while simultaneously claiming "Tech Insider's own tracking put the figure at between 30% and 50%." It has no tracking. It has a copy of the sentence.

The distortion is not one bad actor. It is a verb quietly swapping at hop two and never swapping back, the same mechanism that turned a survey finding into "95% of AI pilots fail" and that I hit again tracing 15 personalization stats to their sources.

Is the rebuttal trustworthy?

Partly, and you should discount it deliberately. SemiAnalysis published "Stop Saying Half of 2026 US Datacenter Capacity Is Canceled" on June 18, 2026, arguing the under-construction figure is "off by multiples" based on satellite imagery, and offering this as its receipt: "Over the last 6 months, our YE2026 NA Hyperscaler Self-build forecast only moved by ~1%, and NA colocation <5%."

That is a real, checkable-in-principle claim and it is also a firm grading its own homework. SemiAnalysis sells a Datacenter Industry Model and an Energy Model to the industry it is scoring, plus retained advisory work. The ~1% is its own revision of its own unaudited forecast. A shop with commercial exposure to the buildout thesis reporting that its buildout forecast barely moved is evidence, but it is the weakest kind: self-reported, self-defined, and directionally convenient. Take it as a counterweight to the panic, not as a measurement.

What does this not prove?

It does not prove the pipeline is fine. The gap Sightline measured is real and unpleasant: roughly 11GW of 2026 capacity sitting in the announced stage with no visible construction, against build timelines of 12 to 18 months. Transformer and switchgear lead times are genuinely the binding constraint, and the moratorium wave Wang flagged ("we expect this trend to continue and meaningfully increase the risk of projects being delayed, withdrawn, and ultimately canceled") is a real forward risk that a February report cannot have scored yet. A delayed gigawatt and a canceled gigawatt look identical on a 2026 revenue forecast, and both are already showing up on somebody's power bill: PJM states' median residential price rose 10.26% in twelve months against 3.96% everywhere else.

It also does not prove Sightline is right. Nobody in this chain, including me, has audited the 777-project dataset. The honest position is that one firm published a hedged global delay projection, one rival firm with a product to sell disputed the construction count, and seventeen other publications repeated whichever version fit their thesis. That is what most AI-bubble evidence looks like when you open it.

Verdict

The underlying finding is signal: announced capacity is running roughly three times actual construction, and a lot of 2026 will land in 2027. The claim as it circulates is noise, manufactured by three specific edits to one sentence, and the "canceled" version is contradicted by the source's own count of nine cancellations.

Verdict: 75% noise / 25% signal. Delay is not cancellation, global is not American, and a range is not its top number. If a headline made all three swaps, it did not read the report.