Stop Saying Half of 2026 US Datacenter Capacity Is Canceled
Summary
The “half of 2026 US data-center capacity is canceled” story rests on a denominator SemiAnalysis says fails basic checks. The underlying report counted 12 GW scheduled for 2026 but only 5 GW under construction; Bloomberg’s “half of US data-center capacity is delayed” framing then propagated as cancellation evidence. Amazon alone publicly built 4 GW in 2025 and could add 5 GW-plus in 2026, while CoreWeave has another 1 GW underway. Jeremie’s verdict: “There’s a massive issue in the denominator.”
Cancellations are substantial, but they largely prune speculative early-stage options rather than erase committed capacity. Requested US data-center load exceeds 1 TW against roughly 750 GW of current system peak load, because hyperscalers may explore ten sites for one final investment decision. These “cloud-coded projects” can advertise 10 GW and 500 MW next year while showing no permits beyond a “Contact us” page; SemiAnalysis’s forecast moved less than 5% across the displayed vintages.
Oracle’s Project Jupiter illustrates the real risk: an advanced, financed project can still miss schedule because power infrastructure is unprecedented at gigawatt scale. Its desired New Mexico gas-pipeline route remains unapproved, with only a secondary route approved, and the applicable regulatory process has no precedent for finishing inside two years. Ellie said trucked CNG’s only proven delivery is around 200 MW and may even be optimistic. This is not an early-stage “cloud-coded” project, Jeremie argued, but a roughly $100 billion Oracle/OpenAI commitment confronting execution reality.
SemiAnalysis forecasts more than 40 GW of behind-the-meter data-center additions by 2028 because grid delivery promises can be less dependable than self-generation. Jeremie said 3 GW of deals had been signed in the prior two weeks and offered roughly 18 months from signing to first capacity as a rule of thumb. A utility might turn a handshake for 500 MW in 2027 into 100 MW in 2028 and the full amount in 2032; behind the meter lets buyers “control your destiny,” subject to permitting and supplier execution.
The capacity race remains financed principally by hyperscalers and the AI labs behind them. AWS buildout is associated with Anthropic, Microsoft’s with OpenAI and some Anthropic demand, and Meta’s typically with MSL. Less-capitalized neoclouds struggle to fund both buildings and GPUs unless backed by firms such as Blackstone or KKR. Once a gigawatt lease is signed, financing can follow rapidly—Jeremie cited roughly $25 billion raised for Vantage about a month later—while first capacity generally comes around 18 months after signing and full ramp is expected to be much faster than five years.
Power-equipment scarcity is attracting enough suppliers and substitute technologies that execution speed matters alongside turbine nameplate capacity. Major OEMs are expanding, Bloom Energy is described as unusually “AI-pilled,” and automotive factories producing 100 GW-plus annually at 40%-50% utilization could redirect engines into generation. With labor rates “going to the moon” and revenue per megawatt rising, fast installation and predictable schedules can matter more than paying a premium for equipment or electricity.
Gas is the time-to-power answer for the 2020s, while solar, batteries and nuclear remain longer-duration diversification. Solar’s grid capacity value may be only 10%-20% of nameplate, depending on the region, and a gigawatt-scale behind-the-meter design could require 20,000 acres; nuclear and SMR agreements remain slow and often non-binding. Reyk described a potentially bearish-looking “peak turbine” dynamic in 2026, when over-purchased units may become underused or reach the secondary market, while stressing that this would not invalidate the broader behind-the-meter thesis.
Deep dive
1. The cancellation headline fails a basic capacity check
Jeremie traced the viral claim to a report saying 12 GW was scheduled to enter service in 2026 but only 5 GW was under construction; Bloomberg’s “half…delayed” framing then propagated as cancellation evidence. His reaction was that the number was “just not possible.”
The immediate sanity check was Amazon: it publicly built 4 GW in 2025 and, in Jeremie’s estimate, will probably add 5 GW-plus in 2026. Add CoreWeave’s roughly 1 GW under construction and the source would effectively make that company “a fifth of the market.” “Come on. There’s a massive issue in the denominator.”
Jordan suggested the spirit might remain directionally right because projects do slip, noting SemiAnalysis’s forecast changed less than 5% across the displayed vintages. Jeremie’s pushback—worth keeping—was that cancellations are indeed “pretty substantial”; the error is treating abandoned early-stage options as lost committed capacity.
2. A terawatt of requests contains many projects that were never real
Current US data-center load requests exceed 1 TW, while the entire system’s peak load is about 750 GW; Jeremie said the underlying chart was already six months old and requests had since more than doubled. That queue cannot all materialize within a few years, so extensive early-stage attrition is unavoidable.
A hyperscaler may examine ten locations for one final investment decision, sounding out counties, utilities, labor availability and supply chains at each. Power scarcity makes every buyer pursue multiple options aggressively, but counting all ten as independent committed builds manufactures cancellations later.
Reyk’s phrase for the weakest entries was “cloud-coded projects”: one developer announces a 10 GW campus and 500 MW next year, yet its website offers only “Contact us,” with no supporting permits. Human judgment quickly identifies a land claim without a real project; automated aggregation can mistake it for construction-ready supply.
SemiAnalysis instead models individual buildings, tenants, end users and timelines bottom-up. Reyk said the team has inspected 10,000-20,000 satellite images—down to interpreting gray pixels as concrete pads or vertical construction—and built agents to scan permit portals globally. The workflow consumed roughly $170,000 of Claude Code in one week, but permits, imagery and sector-wide triangulation remain necessary to connect data centers with Nvidia supply and AI-lab demand.
3. Project Jupiter shows where real execution risk lives
Ellie’s specimen was Oracle’s Project Jupiter in Doña Ana province, New Mexico, where the gas pipeline needed for behind-the-meter generation does not exist and the desired route lacks approval; only a secondary route is approved. The proceeding has fallen into a regulatory process with no precedent for resolution in less than two years, while local opposition makes acceleration unlikely.
Trucked CNG or LNG is not a gigawatt-scale escape hatch: Ellie put the largest proven CNG delivery around 200 MW and cautioned that even this may be optimistic. The area lacks enough producers and trucks, while every FERC filing she said she had seen showed little or no progress. Earlier turbine difficulties had already prompted a switch to Bloom fuel cells.
Jeremie rejected treating Jupiter as an early-stage “cloud-coded” project. It has financing and a roughly $100 billion Oracle commitment contracted for OpenAI, but “we’re still in the early innings of bringing the first gigawatt-scale data centers to market”; contractors and suppliers routinely promise optimistic schedules, and novel designs, municipalities and infrastructure multiply execution errors.
4. Behind-the-meter power replaces grid uncertainty with build risk
SemiAnalysis’s forecast of more than 40 GW of behind-the-meter additions by 2028 begins with a grid-generation shortfall against data-center demand rising by tens of gigawatts annually. Gas-pipeline access became a central site-selection criterion during 2024 and especially 2025. Traditional operators often prioritized tier-one, grid-connected markets such as Northern Virginia for reliability, while newer AI-native operators such as Crusoe moved earlier toward sites with fuel access.
Crusoe’s publicly announced 672 MW Abilene deal with Microsoft was the concrete proof point. Colossus I and II also demonstrate that self-generation works at scale, although Jeremie called them imperfect examples because of permitting issues. Behind the meter is not delay-free; it shifts risk from utility studies toward power-plant construction, equipment and permits.
The grid alternative can be worse because utilities have no binding obligation to honor indicative schedules. Jeremie’s recurring developer anecdote: a handshake for 500 MW in 2027 becomes 100 MW in 2028 and 500 MW in 2032 after network upgrades, competing requests and missing generation are recognized. Self-generation can put buyers “in control of your destiny,” at least from a power standpoint and subject to execution.
Contract form changes the clock: turnkey leases make developers such as QTS responsible for everything, powered-shell arrangements leave more CapEx to tenants, and the Oracle–VoltaGrid deal at the Shackelford County, Texas, site can be a pure PPA without a data center. Jeremie’s rule of thumb was roughly 18 months from signing to first capacity, followed by an increasingly rapid phased ramp—not five years.
5. Financing concentrates construction around hyperscalers and AI labs
Reyk framed the buildout as a race among labs expressed through their capital providers: AWS capacity is largely driven by Anthropic, Microsoft by OpenAI plus some Anthropic demand, and Meta typically for MSL. Hyperscalers possess the investment-grade financing needed to turn billion-dollar power, buildings and GPU orders into construction.
Neoclouds face a much harder capital stack unless unusually well funded or backed by firms such as Blackstone and KKR. The market therefore divides between hyperscalers and a narrow set of heavily capitalized challengers, even when many more developers can announce sites.
Once a credible gigawatt lease exists, funding can arrive quickly. Jeremie cited DigitalBridge raising roughly $25 billion for Vantage about a month after a lease: the whole amount was already secured, leaving the developer to build as fast as possible. First capacity may arrive roughly 18 months after signing, with the full ramp varying by deal but expected to be much quicker than five years.
6. Equipment substitution makes gas the 2020s bridge
Jordan’s chart showed the top eight suppliers making up just over half, roughly, with 25 additional names listed. Ellie described an equipment market spanning major OEMs—including Siemens, GE Vernova and Lenovo—and smaller suppliers. Castings and blades remain bottlenecks, but continued factory expansions suggest suppliers are responding to AI demand rather than allowing the scarcity narrative to stop the market.
Established players are competing with aeroderivative solutions, recycled aircraft technology, boilers and other new or revived technologies. Permitting can redirect technology as readily as supply. Ellie said the Nebius New Jersey facility switched to Bloom fuel cells and thought it had previously used Bergen turbines; Bloom’s lower NOx and SOx emissions, possibly alongside its manufacturing footprint, may help with speed to power. Sites may also relocate toward states such as Texas, mix generation with net metering, remain fully islanded, or later connect turbines as grid peakers—a relationship she called more “symbiotic” than either/or.
Jeremie’s Bloom framing captured the changing economics: it is “not very good at backup.” He claimed that, when run extremely hot, it reaches 15,000 degrees Celsius and said, as far as he knew, it takes two days to go from zero to full output. But no power is worse. As cloud-, lab- and model-layer revenue per megawatt rises, electricity cost becomes less decisive; Jordan added that even paying double for turbines barely moves total project CapEx beside the GPUs.
Labor rates “going to the moon” strengthen the case for modular, fast-deployment systems with predictable installation bills. Automotive factories offer another reservoir: Jeremie estimated more than 100 GW of annual production running at only 40%-50% utilization, while Ellie noted Tesla and Ford moving battery-storage capacity toward data centers. “That unlocks gigantic capacity.”
7. Renewables diversify later, while surplus turbines create false bearish signals
Grid-connected solar may add 50-60 GW of nameplate annually, but Jeremie put its effective capacity contribution around 10%-20%, sometimes below or above depending on the region. Behind-the-meter solar plus batteries can work, yet a gigawatt campus might need 20,000 acres, complicating land assembly and sacrificing the time-to-power advantage.
His best land example: once owners learn the buyer is planning a $100 billion project, they may demand ten times the expected price. Solar’s logistics improve with multi-year planning, but interconnection queues and transmission constraints remain; nuclear and SMRs are slower still, with many agreements non-binding or contingent on regulatory milestones.
His timing view was: “In the 2030s we’re gonna see a gigantic diversification of energy sources… For the 2020s, we’re gonna be very much in the gas world.” Ellie added that carbon capture, utilization and storage may increasingly be co-located with behind-the-meter generation.
Reyk described what the team calls “peak turbine” in 2026: some buyers over-purchased before solving permits or data-center construction, leaving equipment underused or potentially later offered on the secondary market. Seeing a favorite project’s turbines for sale will look bearish and prompt “is it over?” His answer remained no—the failures reveal project-selection risk, not the end of behind-the-meter growth.