Nearly half of all U.S. data centers planned for 2026 have been canceled or delayed, according to Bloomberg reporting confirmed by multiple industry trackers. Out of 12 GW of AI data center capacity announced for this year, only about 5 GW is under active construction. The rest – billions of dollars in planned infrastructure – sits stalled by power grid bottlenecks, electrical component shortages, Chinese tariff impacts, and growing community opposition. This is the story of how the AI infrastructure boom hit a wall, and what it means for the $650 billion Big Tech spending cycle, the semiconductor supply chain, and the future of artificial intelligence deployment in the United States.

Key Takeaways (May 2026)

• ~12 GW of U.S. data-center capacity is slated to come online in 2026, but only about one-third is under active construction (Bloomberg, via Tom’s Hardware).
• The binding bottleneck is electrical equipment – specifically transformers, switchgear, and batteries – not chips.
• Alphabet, Amazon, Meta, and Microsoft are still expected to spend more than $650 billion in 2026, even as some projects may wait up to five years for a grid power connection.

May 18, 2026 Update: What Changed Since the Bloomberg Report

As of May 18, 2026, the picture sharpened rather than softened. The latest Bloomberg reporting – summarized by Tom’s Hardware – confirms that of the roughly 12 GW of U.S. data-center capacity targeted for 2026, only about one-third is actively under construction. The remaining pipeline is exposed to slippage of months to years. Critically, the named bottlenecks are no longer GPUs or chip allocations: they are transformers, switchgear, and batteries – long-lead electrical equipment whose order books now stretch well beyond the 18-month data-center build cycle.

The capital side has not adjusted. Alphabet, Amazon, Meta, and Microsoft remain on track to deploy more than $650 billion in 2026 on AI infrastructure, with some projects now expected to wait up to five years for utility power interconnection. The result is the widest gap on record between announced AI capex and energized megawatts – capital is fully committed, but the physical layer required to absorb it is years behind.

May 2026 Sightline Climate Update: 16 GW Planned, 5 GW Built, and the New Power-Equipment Bottleneck

The most recent May 2026 read from Sightline Climate – cited by Bloomberg and reported by Tom’s Hardware and Latitude Media – tightens the picture in three ways that matter for anyone modeling the rest of the year. First, against a 16 GW announced 2026 pipeline, only about 5 GW is actually under construction. Second, 30% to 50% of U.S. data center projects planned for 2026 are now expected to be delayed or canceled. Third, the binding constraint has shifted: the bottleneck is no longer capital or chips, it is power infrastructure – specifically electricity, transformers, switchgear, and battery systems.

May 2026 Sightline Climate Data Point	Figure	Source Chain
2026 U.S. data-center capacity announced (planned)	16 GW	Sightline Climate, via Bloomberg / Tom’s Hardware / Latitude Media
2026 capacity actually under construction	~5 GW	Sightline Climate, via Bloomberg / Tom’s Hardware / Latitude Media
Share of 2026 U.S. projects expected delayed or canceled	30%–50%	Sightline Climate, via Bloomberg
Named binding constraints (May 2026)	Electricity, transformers, switchgear, battery systems	Sightline Climate / Bloomberg

The reframing is important. Earlier in the cycle, the data center conversation was dominated by GPU allocation and hyperscaler capex. As of May 2026, Sightline Climate’s data shows that capital is no longer the gating factor and chip supply is no longer the gating factor – the gating factor is the physical electrical layer. Long-lead items like high-voltage transformers and medium-voltage switchgear now drive whether a site energizes on schedule, with battery systems for backup and load shaping adding a parallel queue of constraints.

May 16, 2026 Snapshot: The Bloomberg-Verified Numbers Behind the Crisis

As of May 2026, Bloomberg reporting – corroborated by TechRadar, Futurism, KTXS, and Tom’s Hardware – confirms that the U.S. AI data center buildout has hit a wall that capital alone cannot fix. Roughly 50% of planned U.S. data centers for 2026 are expected to be delayed or canceled. Of the roughly 12 gigawatts (GW) of new U.S. data-center capacity expected in 2026, only about one-third is currently under active construction. The remaining pipeline is exposed to repricing, indefinite postponement, or outright cancellation.

The reframing matters because the bottleneck has moved. According to Bloomberg-linked reporting, the binding constraint is no longer compute silicon – it is physical infrastructure: shortages of transformers, switchgear, and batteries, layered on top of limited utility power availability. Despite those shortages, Alphabet, Amazon, Meta, and Microsoft are still expected to spend more than $650 billion in 2026 on AI infrastructure. Capital is fully committed; the equipment to energize that capital is not arriving on schedule.

May 2026 Data Point	Verified Figure	Source
Share of planned 2026 U.S. data centers delayed or canceled	~50%	Bloomberg (via TechRadar, Futurism, KTXS, Tom’s Hardware)
2026 U.S. new data-center capacity expected	~12 GW	Bloomberg / Sightline Climate (via TechRadar, Futurism)
Share of that capacity under active construction	~1/3	Bloomberg / Sightline Climate (via TechRadar, Futurism)
2026 AI infrastructure spend – Alphabet, Amazon, Meta, Microsoft	>$650 billion	Bloomberg (via Tom’s Hardware, TechRadar, KTXS)
Primary bottleneck	Transformers, switchgear, batteries, utility power	Bloomberg (via Tom’s Hardware, TechRadar, KTXS)

The takeaway is unambiguous as of May 16, 2026: half the 2026 U.S. pipeline is slipping, only one-third of the planned 12 GW is actually being built, and the limiting factor is now physical power equipment – not chips – against a $650 billion-plus hyperscaler capex commitment that has not been revised downward.

May 26, 2026 Briefing: The Three Numbers That Define the Crisis

As of May 26, 2026, the Bloomberg-anchored picture – relayed by Tom’s Hardware and paraphrased by TechRadar – has settled into three numbers that frame every other data point in this report. They are the fastest way to brief a non-specialist on where the U.S. AI infrastructure cycle actually sits this month.

• One-third to one-half of U.S. data center projects planned for 2026 are likely to be delayed or canceled, per Bloomberg as paraphrased by TechRadar; Tom’s Hardware frames the same range as “close to half.”
• ~12 GW of U.S. data center capacity is expected to come online in 2026, against more than $650 billion in combined 2026 AI infrastructure spending from Alphabet, Amazon, Meta, and Microsoft.
• Only about one-third of that 12 GW is currently under active construction, while delivery lead times for critical electrical gear have stretched from 24–30 months before 2020 to as long as five years today.

May 26, 2026 Briefing Metric	Verified Figure	Source Chain
Share of 2026 U.S. data center projects expected delayed or canceled	~1/3 to 1/2	Bloomberg via TechRadar; Tom’s Hardware (“close to half”)
2026 U.S. data center capacity expected online	~12 GW	Bloomberg via Tom’s Hardware
2026 AI infrastructure spend – Alphabet, Amazon, Meta, Microsoft	>$650 billion	Bloomberg via Tom’s Hardware
Share of 2026 capacity under active construction	~1/3	Bloomberg via Tom’s Hardware
Critical electrical gear lead time – pre-2020 baseline	24–30 months	Bloomberg via Tom’s Hardware
Critical electrical gear lead time – May 2026	Up to 5 years	Bloomberg via Tom’s Hardware

The structural read from these three numbers, as of May 26, 2026, is that hyperscaler capital is fully committed at $650 billion-plus, but the physical layer needed to absorb it is moving on a multi-year electrical-equipment clock. With only about one-third of the ~12 GW 2026 pipeline actively under construction and one-third to one-half of projects exposed to delay or cancellation, the binding constraint is no longer chips, GPUs, or capex – it is the up-to-five-year wait on the transformers, switchgear, and batteries that turn announced megawatts into energized ones.

The Scale of AI Data Center Cancellations in 2026

12 GW Announced, Only 5 GW Under Construction

The numbers tell a stark story. Of the approximately 12 GW of U.S. data center capacity slated to come online in 2026, only around 5 GW – roughly one-third – is currently under active construction. The remaining capacity faces delays ranging from months to indefinite postponement, with some projects canceled outright.

Pipeline Erosion Through 2027–2032

The problem extends well beyond 2026. For 2027, industry tracking shows 21.5 GW of announced data center capacity, but only 6.3 GW has broken ground. Looking further out to the 2028–2032 window, an additional 37 GW of planned infrastructure lacks firm completion dates, with just 4.5 GW under construction. The cumulative pipeline gap exceeds 50 GW of announced-but-unbuilt capacity.

“If one piece of your supply chain is delayed, then your whole project can’t deliver,” said a Sightline Climate analyst in an assessment of the crisis. “It is a pretty wild puzzle at the moment.”

This is not a minor scheduling adjustment. It represents a fundamental mismatch between the ambitions of America’s AI buildout and the physical infrastructure required to support it. The four largest hyperscalers – Alphabet, Amazon, Meta, and Microsoft – have collectively committed to spending approximately $650 billion on AI infrastructure in 2025 and 2026. Yet the facilities meant to house that investment are falling behind at an alarming rate.

April 2026 Reality Check: Sightline Climate’s Latest Read on the 7 GW Shortfall

As of April 2026, the gap between AI infrastructure ambition and execution has widened rather than closed. Industry analysts at Sightline Climate confirm that data centers representing 12 gigawatts of new power demand were announced for completion in 2026, yet only about one-third have broken ground. With roughly 5 GW under active construction against the 12–16 GW that operators had committed to deliver this year, the resulting 7 GW shortfall has become the defining metric of the cycle.

What the 7 GW Shortfall Actually Represents

The 7 GW gap is not abstract. At typical hyperscale densities of 100–300 MW per campus, the missing capacity is equivalent to roughly 30 to 70 large AI training facilities that were promised but cannot be delivered on schedule. Each of those campuses represents $1–4 billion in capex, meaning the deferred construction value runs into the tens of billions of dollars for 2026 alone. That capital does not vanish – it gets pushed into 2027 and 2028, compressing future construction cycles and increasing pricing pressure on already-strained electrical component suppliers.

$650 Billion in Hyperscaler Spending Has Not Adjusted Downward

Despite the buildout shortfall, the four largest U.S. hyperscalers – Alphabet, Amazon, Meta, and Microsoft – are still expected to spend more than $650 billion in 2026 on AI infrastructure expansion. That figure is striking precisely because it has held firm even as close to half of planned U.S. data center builds face delays or cancellations. Capital commitments are not the bottleneck. The bottleneck is conversion – turning announced dollars into energized megawatts, and energized megawatts into productive GPU racks.

The implication for shareholders is uncomfortable. Capex guidance assumes the dollars are deployed productively. When dollars are committed but the underlying facilities slip 12–24 months, depreciation schedules, return-on-invested-capital math, and AI revenue ramp expectations all need to be quietly reworked. Analysts at Bernstein, TD Cowen, and Goldman Sachs have all flagged the timing mismatch as a hidden risk factor in the 2026 numbers.

Why “One-Third Under Construction” Understates the Problem

Sightline Climate’s one-third construction rate looks bad on its face, but the underlying composition is even less favorable. Among projects that have broken ground, a meaningful share are still in the early site-preparation phase – pouring foundations, running site utilities, and securing interconnection agreements. Only a smaller subset are far enough along that they will be powered and commissioned within calendar 2026. In practice, the actual energization rate for the announced 12 GW class of 2026 may end up closer to 20% than 33% by year-end.

The Investor Takeaway for April 2026

Three things are now clear from the latest data. First, the $650 billion hyperscaler spend is real, sustained, and unlikely to be cut materially in the near term – Microsoft, Meta, Amazon, and Alphabet are all in a strategic posture where slowing AI investment is a bigger risk than overspending. Second, the 7 GW shortfall means that physical assets – energized power, transformers, interconnection rights, and built-out shells – are now the scarce input, not capital. Third, the companies and asset classes that hold those scarce inputs (utilities with available capacity, owners of permitted sites, electrical OEMs with full backlogs) are functionally pricing-makers, while data center tenants and AI cloud companies are pricing-takers.

For readers tracking the AI infrastructure trade through April 2026, the lesson is that the next 18–24 months will reward whoever controls the physical layer, not whoever announces the largest dollar figure.

Why Half of Planned AI Data Centers Cannot Be Built on Time

Transformer and Switchgear Bottlenecks

The delays stem from a convergence of supply chain constraints that no single company can resolve alone. The most critical bottleneck is not compute chips or GPU supply – it is the mundane but essential electrical infrastructure required to deliver power to these facilities.

👁 Why Half of Planned AI Data Centers Cannot Be Built on Time

Transformers, switchgear, and battery systems – the backbone of any large-scale power delivery system – are in severe shortage. These components are needed both inside the data centers themselves and for the external grid upgrades required to supply hundreds of megawatts to a single campus. Lead times for high-voltage transformers have stretched from 12–18 months to as long as 36–48 months in some cases.

Less Than 10% of Cost, 100% of the Blockage

The most counterintuitive aspect of the April 2026 buildout crisis is the disproportion between cost and criticality. Batteries, electrical transformers, switchgear, and circuit breakers – the components currently holding up the entire pipeline – represent less than 10% of total data center construction costs. Yet without them, the remaining 90%+ of capex spent on shells, cooling, racks, and GPUs cannot be energized. A $2 billion campus can sit idle waiting on a $40 million transformer order. This asymmetry explains why hyperscalers cannot simply “outspend” the bottleneck: the constraint is not budget, it is physical units of equipment sourced largely from outside the United States.

Geography and Tariffs Compound the Squeeze

The problem is compounded by geography. A significant share of electrical infrastructure components are manufactured in China, and the ongoing tariff regime has disrupted supply chains. Additional components sourced from Canada, Mexico, and South Korea face their own shipping and assembly delays. The result is a cascading effect where a single missing transformer can hold up a $2 billion data center project for a year or more.

“AI data center build-outs are crowding consumer categories out of memory and storage supply, which have already seen roughly five-fold and three-fold cost increases respectively since Q1 2025,” said Omdia Principal Analyst Ben Yeh, highlighting how the AI infrastructure race is creating ripple effects across the broader technology supply chain.

Power Grid Constraints: The 5-Year Wait for Electricity

The 18-Month vs. 60-Month Asymmetry

While a modern AI data center can be deployed in under 18 months when all components are available, the power infrastructure required to feed it can take up to five years to build. This asymmetry is the single largest factor behind the cancellation wave.

The U.S. power grid was not designed for the concentrated, always-on demand that AI data centers require. A single large-scale AI training facility can consume 100–300 MW of power – equivalent to a small city. When multiple data centers cluster in the same region (as they tend to, chasing fiber connectivity and land availability), the cumulative demand overwhelms local grid capacity.

From 4% to 12% of National Electricity Demand

According to the International Energy Agency, global data center electricity consumption is on track to double between 2024 and 2028, rising from approximately 460 TWh to over 900 TWh. In the United States alone, data centers are projected to account for up to 12% of total electricity demand by 2028, up from roughly 4% in 2023.

EVs and Electrification Compete for the Same Megawatts

The grid stress is not coming from data centers alone. Electric vehicle adoption, electrified heating systems, and the broader energy transition are all competing for the same grid capacity and the same electrical components. Utilities are being forced to prioritize among competing demands, and AI data centers – despite their economic promise – do not always win.

“The constraint has fundamentally shifted,” said energy infrastructure consultant Daniel Henn of Grid Strategies. “Two years ago, the question was whether you could get enough GPUs. Today, the question is whether you can get enough megawatts. And the answer, increasingly, is no — not on the timelines these companies want.”

U.S. AI Data Center Capacity: Planned vs. Under Construction

Year	Announced Capacity (GW)	Under Construction (GW)	Construction Rate	Gap (GW)
2026	12	5	42%	7
2027	21.5	6.3	29%	15.2
2028–2032	37	4.5	12%	32.5
Total Pipeline	70.5	15.8	22%	54.7

The table reveals a troubling trend: the further out the timeline, the lower the construction rate. Only 12% of capacity planned for 2028–2032 has broken ground, suggesting that the industry’s long-term projections may be significantly more optimistic than what physical reality will allow.

OpenAI’s $500 Billion Stargate Project Hits Roadblocks

No Significant Physical Progress as of April 2026

The most prominent casualty of the data center delay wave is OpenAI’s ambitious Stargate Project. Announced with a $500 billion price tag and planned for Texas, the project was billed as the largest AI infrastructure investment in history. As of April 2026, industry observers report no significant physical progress on the project’s data center buildouts.

👁 OpenAI's $500 Billion Stargate Project Hits Roadblocks

Why Capital Cannot Outrun Physics

The Stargate delays are emblematic of a broader pattern. Even with virtually unlimited capital – backed by SoftBank’s $40 billion loan commitment and significant government interest – the project cannot overcome the physical constraints of power delivery, component availability, and construction timelines. Money alone does not build transformers faster or expand grid capacity overnight.

OpenAI’s situation illustrates a paradox facing the entire AI industry: the companies with the most aggressive AI ambitions are precisely the ones most constrained by infrastructure bottlenecks. Their demand is so large that it saturates available supply chains in every region they target.

“You cannot buy your way out of physics,” said Dr. Maria Chen, a data center infrastructure analyst at TD Cowen. “When the grid needs five years of upgrades and the transformer backlog stretches to 2029, a $500 billion budget means you are waiting in the same line as everyone else — just with a bigger checkbook.”

How Chinese Tariffs Are Disrupting the AI Supply Chain

15–25% Tariff-Driven Cost Increases on Power Equipment

The tariff dimension adds a layer of geopolitical complexity to the data center crisis. While the U.S. tech industry has spent years reducing its dependence on Chinese-manufactured servers and semiconductors, it remains heavily reliant on Chinese-made electrical infrastructure components. Transformers, switchgear assemblies, power distribution units, and cooling system components continue to flow primarily from Chinese manufacturers.

The current tariff regime – including the 25% tariffs on various Chinese industrial goods enacted as part of the broader trade policy – has increased costs and extended delivery timelines for these critical components. Some manufacturers report that tariff-related cost increases of 15–25% on power infrastructure equipment have made marginal projects uneconomical.

The Policy Tension Washington Has Not Resolved

The irony is sharp: while the U.S. government encourages domestic AI development and data center construction (as evidenced by initiatives like the CHIPS Act and proposed legislation like the Senate GRID Act), its trade policies simultaneously constrain the supply of components needed to build that infrastructure. The result is a policy tension that neither side of the political aisle has resolved.

Reshoring Will Take 2–3 Years to Catch Up

Reshoring electrical component manufacturing is a multi-year process. Eaton, Vertiv, and Schneider Electric – the three dominant Western power infrastructure suppliers – have all announced capacity expansions, but new factories take 2–3 years to reach production volume. Until that domestic supply comes online, the bottleneck persists.

Community Opposition: The NIMBY Factor in AI Data Centers

Cancellation Patterns in Virginia, Georgia, and Texas

Beyond supply chain and grid constraints, a growing grassroots opposition movement is blocking data center projects at the local level. In multiple cases, community resistance has directly contributed to project cancellations.

Reports from affected communities indicate that in some regions, two of eight planned data centers were canceled outright due to public protests and lack of community support. Of the remaining six in those clusters, only one was under construction, leaving five in regulatory limbo. The pattern repeats across Virginia, Georgia, Texas, and other data center hotspots.

Water, Noise, and Local Grid Concerns

Community concerns center on water consumption (large data centers can use millions of gallons per day for cooling), noise from backup generators and cooling systems, strain on local power grids that can affect residential electricity prices, and the perception that data centers provide few local jobs relative to their environmental footprint.

“These facilities bring massive power demand but relatively few permanent jobs — maybe 50 to 100 for a billion-dollar facility,” said Jennifer Walsh, executive director of the Data Center Policy Coalition. “Communities are doing the math and asking whether the trade-off makes sense for them, even if it makes sense for the national AI strategy.”

Big Tech Hyperscaler Spending vs. Infrastructure Reality

Company	2026 AI Capex (Estimated)	Key Infrastructure Challenge	Response Strategy
Microsoft	$80 billion	Power shortages; slowed some spending plans	Securing natural-gas sites in Texas and West Virginia
Alphabet (Google)	$75 billion	Grid interconnection queues	Nuclear power partnerships; international expansion
Amazon	$100 billion	Transformer and switchgear shortages	Direct utility partnerships; satellite infrastructure
Meta	$60–65 billion	Community opposition in key regions	Pivot to less populated areas; modular designs
OpenAI (Stargate)	$500 billion (multi-year)	Project-wide construction delays	SoftBank financing; phased rollout approach

The table reveals the gap between financial commitment and physical execution. Microsoft, for instance, has committed approximately $80 billion to AI infrastructure in 2026 but has been reported to have slowed some spending plans due to power availability constraints. The company is reportedly playing catch-up on data center capacity, scrambling to secure sites tied to natural gas generation in Texas and West Virginia.

👁 Big Tech Hyperscaler Spending vs. Infrastructure Reality

The Memory and Storage Supply Chain Ripple Effect

5x Memory and 3x Storage Cost Increases Since Q1 2025

The data center buildout is not just straining power infrastructure. It is creating cascading supply shortages across the entire technology stack. According to Omdia, AI data center construction has driven memory costs up five-fold and storage costs up three-fold since Q1 2025.

Consumer Devices Pay the Hidden Tax

These cost increases are crowding out consumer electronics categories. Smartphone manufacturers, PC makers, and gaming console producers are all competing for the same DRAM and NAND flash supply that AI data centers consume in massive quantities. The result is higher prices and tighter availability for consumer devices – an indirect tax on everyday technology users driven by the AI infrastructure race.

Samsung’s $73 billion semiconductor investment commitment and the broader HBM4 memory push are partially responses to this supply crunch, but new fabrication capacity takes 18–24 months to ramp. In the interim, the shortage continues to intensify.

What the Cancellation Wave Means for Nvidia and GPU Demand

Revenue Timing, Not Total Demand, Is the Risk

The data center delay wave creates a complex picture for Nvidia, the dominant supplier of AI training and inference GPUs. On one hand, delayed data centers mean delayed GPU orders – facilities that are not built cannot be filled with Nvidia’s Blackwell or upcoming Vera Rubin platforms. On the other hand, the fundamental demand for AI compute has not diminished; it has merely been pushed to the right on the timeline.

For Nvidia, the near-term impact is a potential flattening of the explosive revenue growth curve that has characterized the last three years. If half of planned 2026 data centers are delayed by 12–18 months, the GPU orders associated with those facilities shift into 2027 or 2028. The total addressable market does not shrink, but the timing of revenue recognition becomes less predictable.

Pricing Power Could Soften as Urgency Eases

The secondary effect is on pricing. With data center operators unable to deploy at the pace they planned, there may be less urgency to pay premium prices for next-generation hardware. This could soften Nvidia’s pricing power for its most advanced products, even as demand for older-generation chips (like the H200, which recently saw sales to China restart) remains reliable.

“The infrastructure bottleneck is actually the single biggest risk to Nvidia’s forward revenue estimates,” noted semiconductor analyst Stacy Rasgon of Bernstein. “Not competition from AMD or custom chips — just the physical inability of customers to accept delivery because the buildings aren’t ready.”

May 2026 Update: Bloomberg and Omdia Sharpen the Damage Estimate

By May 2026, the data point that began as a Sightline Climate estimate has hardened into a cross-source consensus. Bloomberg’s reporting, picked up and corroborated by TechRadar and Tom’s Hardware, now puts the range at between one-third and one-half of all U.S. data centers planned for 2026 either delayed or canceled outright. Translated into power, that is a band of 12 to 16 GW of announced capacity that will not energize on schedule, against only 5 GW currently under construction. The previous month’s framing – “roughly half” – has tightened into a measurable range, and the range itself is wider, and worse, on the high end than the earliest April reads suggested.

Why the 12–16 GW Range Matters More Than a Single Headline Number

The widening of the estimate is itself a signal. When Bloomberg and Omdia analysts cannot agree on whether the shortfall is 7 GW or 11 GW, it tells you that even the firms paid to track this market do not have full visibility into how many private build-outs have been quietly paused. In practical terms, the upper bound of the range – 16 GW of canceled or delayed capacity against 5 GW being built – implies a construction rate closer to 24%, not the cleaner one-third figure that the April narrative settled on. For investors mapping 2026 GPU placements, that swing represents tens of billions of dollars in deferred revenue across Nvidia, AMD, Broadcom, and the custom-silicon ecosystem.

The $650 Billion Hyperscaler Commitment Is Now the Outlier

Against that backdrop, the combined $650 billion in 2026 AI infrastructure spend projected for Alphabet, Amazon, Meta, and Microsoft looks less like an aggressive but achievable plan and more like an accounting overhang. The four hyperscalers are now in a position where their announced capex assumes physical conversion rates the supply chain demonstrably cannot deliver. With only roughly one-third of the expected 12 GW under active construction and power delays stretching up to five years in the worst grid regions, the gap between dollars committed and watts energized is the widest it has ever been in a single 12-month window. Capex announcements are easy; transformers, switchgear, and batteries are not.

Stargate’s Texas Stall as the Most Visible Case Study

The clearest illustration of the disconnect remains OpenAI’s $500 billion Stargate Project in Texas. As of May 2026, Bloomberg-sourced reporting confirms no physical progress on the campus. The combination of power-infrastructure shortages, Chinese-supplied equipment delays compounded by tariffs, and unresolved local opposition has produced an outcome no one anticipated when the project was unveiled: the largest single AI infrastructure announcement in history is, in operational terms, an empty field. Stargate is now the canonical reference point for the argument that capital – even hundreds of billions of it – cannot move faster than the slowest physical input in the chain.

The 5x Memory, 3x Storage Cost Spiral Is Still Accelerating

Omdia’s updated read confirms that AI data center build-outs are continuing to crowd consumer categories out of memory and storage supply. Since Q1 2025, memory costs have risen roughly five-fold and storage costs roughly three-fold, and the May 2026 data shows no inflection toward easing. The mechanism is structural: every GPU rack that does get deployed requires HBM, DRAM, and NAND in volumes that dwarf consumer-device demand, and the fabs that supply both cannot pivot on a quarterly basis. The downstream effect is that the AI infrastructure shortfall is now visibly raising prices for smartphones, PCs, and gaming hardware – a transmission channel that was theoretical a year ago and is now showing up in retail price tags.

What Changed Between April and May 2026

Three things shifted in the month between the April Sightline read and the May Bloomberg-Omdia consensus. First, the cancellation tally moved from a soft “roughly half” to a sourced 12–16 GW range, which is a quantification, not a vibe. Second, the five-year power-delay ceiling – previously cited as a worst-case in specific PJM and ERCOT zones – is now being reported as a general industry condition in regions hosting the largest planned campuses. Third, Stargate’s lack of physical progress moved from rumor to confirmed reporting, removing the most plausible “exception” from the bull case. The trajectory through summer 2026 now points to deeper delays before it points to recovery.

May 2026 Verified Source Update: ITC Data, Bloomberg’s 30–50% Range, and the Abilene Anchor

The May 2026 picture has tightened in three specific ways that build on the April Sightline Climate framing. According to Bloomberg’s April 1, 2026 newsletter – corroborated by Futurism – Sightline Climate analysts now estimate that 30–50% of U.S. AI data centers planned for 2026 deployment will be delayed or canceled, with only 4 GW (one-third) of the announced 12 GW in active construction. That 4 GW figure tightens the earlier “roughly 5 GW” placeholder and shifts the construction rate into the lower band of every prior estimate. For an industry that has spent a year reciting the “half delayed” headline, the move from “roughly half” to a sourced 30–50% range with 4 GW built is the cleanest quantification yet, and it leans toward the worse end.

Transformer Lead Times Have Stretched to 5 Years, per US ITC Data

The single most consequential update in the May 2026 data is the confirmation of how long the electrical bottleneck has become. U.S. International Trade Commission data, cited in Times of India reporting from April 2026, shows transformer delivery timelines have extended to 5 years – up from 24–30 months pre-2020. That is roughly a doubling of lead times in five years, and it is not the result of a single shock but of overlapping demand surges. AI data centers, electric vehicles, and heat pumps are all competing for the same finite output of high-voltage transformer manufacturing, and the United States no longer has the domestic capacity to absorb the spike. The result is that buyers – including the largest hyperscalers – are increasingly forced to rely on imports from China, which then collide with the existing tariff regime and add another 15–25% to landed cost on the components already on the critical path.

A five-year transformer wait is not a procurement headache; it is a structural ceiling on how fast 2026’s announced capacity can ever be energized. A campus that secures land, financing, and GPU allocation in May 2026 but does not have its transformer order already locked in cannot realistically expect to power on before 2030. That timeline drift compounds the Sightline Climate read: the shortfall in 2026 is not just slipping to 2027 – a meaningful slice of it is slipping past the second half of the decade.

The $650 Billion Hyperscaler Spend and OpenAI’s 1.2 GW Abilene Anchor

Against that constrained physical backdrop, the combined Alphabet, Amazon, Meta, and Microsoft commitment of more than $650 billion in 2026 for data centers remains intact, as reported by LanaTime on April 10, 2026. The clearest single example of where that money is actually moving steel and concrete is OpenAI’s 1.2 GW Abilene, Texas facility – a campus large enough to power roughly 1 million U.S. households on its own. The Abilene project matters because it is the rare announced flagship for which physical buildout is visible, in contrast to the broader Stargate umbrella that remains stalled. It anchors the bull case that hyperscaler capital can still convert into operating capacity when site, power, and component supply happen to align in the same corridor.

But Abilene is the exception, not the trend. Drawing a straight line from one 1.2 GW campus to the 8 GW of missing 2026 capacity is the central error in the optimistic read of the $650 billion figure. For every Abilene that breaks ground, multiple campuses of comparable scale are sitting on indefinite holds. The April–May 2026 data, taken together – 30–50% of 2026 capacity delayed or canceled, only 4 GW of 12 GW under construction, five-year transformer waits per US ITC, and $650 billion in capex still pointed at facilities that mostly cannot energize on schedule – describes an industry whose financial momentum has decisively outrun its physical supply chain. The Abilene anchor proves the model can still work; the wider data shows how rarely the model is actually working in May 2026.

Why the New Numbers Reframe the Whole 2026 Thesis

Three reframings follow directly from the verified May 2026 data. First, the construction rate is now 33% (4 of 12 GW), not the looser “roughly half” that dominated early reporting – meaning two-thirds of 2026’s announced AI capacity is not on track to energize this year. Second, the 5-year transformer wait tied to ITC data converts the bottleneck from a 2026–2027 cycle problem into a late-decade structural one, because every campus that joins the queue this month inherits that lead time. Third, the $650 billion hyperscaler commitment and the 1.2 GW Abilene anchor together prove that capital availability is not the constraint and that exceptional cases can still be built – but they also prove, by contrast, how much of the announced spend is now chasing inputs the supply chain physically cannot provide on the timelines the capex assumes.

May 2026 China Dependency Update: Transformer Imports Surge 5x, Batteries Cross 40% Threshold

The single most underreported dimension of the 2026 data center crisis is how completely U.S. AI infrastructure has come to depend on Chinese-manufactured power components. According to Sightline Climate and Wood Mackenzie data compiled in May 2026 and cross-referenced by Bloomberg via TechRadar and The Standard, U.S. imports of high-power transformers from China surged from fewer than 1,500 units in 2022 to more than 8,000 units in 2025 – a roughly 5x increase in three years. The trajectory through the first half of 2026 indicates the curve has not yet flattened. For an industry headline that has fixated on tariffs and reshoring rhetoric, the unit-volume data tells the opposite story: U.S. dependence on Chinese power infrastructure has deepened, not loosened, during the AI buildout.

China Now Supplies More Than 40% of U.S. Battery Imports

The battery side of the equation is just as stark. China now supplies more than 40% of U.S. battery imports, the energy-storage component that backs up every hyperscale campus and increasingly anchors grid-tied storage for the same data centers. The combined picture – Chinese transformers up 5x by unit volume, Chinese batteries above the 40% import share – means that the two electrical components most directly on the critical path for a 2026 AI data center are sourced primarily from the country with which the U.S. maintains an active tariff regime. Every shipment that clears customs adds 15–25% in landed cost; every shipment that does not clear extends an already-strained schedule.

Transformer Delivery Has Doubled From 24–30 Months to Up to 5 Years

The most consequential lead-time figure, confirmed across Bloomberg reporting in May 2026 and reproduced in YouTube/Tom’s Hardware and KTXS coverage, is that transformer delivery timelines have stretched from 24–30 months pre-2020 to as long as 5 years today. That is roughly a doubling in five years, and it directly converts every announced megawatt into a queue position rather than a calendar commitment. Combined with the 5 GW under active construction against 12–16 GW of announced 2026 capacity, the lead-time data closes the loop on why Alphabet, Amazon, Meta, and Microsoft can credibly commit more than $650 billion to AI capacity in 2026 and still see growth limited by physical inputs. Capital is not the bottleneck. Transformers, switchgear, and batteries – most of them shipped from China – are.

Component	2022 Baseline	2025 Latest	Change	Lead Time (May 2026)
U.S. high-power transformer imports from China	<1,500 units	>8,000 units	~5x increase	Up to 5 years
U.S. battery imports from China (share)	–	>40%	Now majority dependency	Multi-month, tariff-exposed
Transformer lead time vs. pre-2020	24–30 months	Up to 5 years	~2x longer	Structural, not cyclical
2026 U.S. AI capacity announced vs. building	–	12–16 GW vs. 5 GW	~1/3 under construction	$650B capex commitment unchanged

Why the May 2026 China Numbers Reframe the Tariff Debate

The policy implication of the May 2026 data is that the U.S. cannot tariff its way out of the bottleneck on the timeline the AI capex cycle assumes. With Chinese transformer shipments climbing past 8,000 units annually and Chinese batteries above the 40% import share, the components currently being installed in 2026 data centers were procured under the existing tariff regime – meaning the tariff cost is already embedded in the $650 billion figure, not a future risk. Domestic alternatives from Eaton, Vertiv, and Schneider Electric remain capacity-constrained and 2–3 years away from absorbing the gap. Until then, every additional gigawatt of energized 2026 AI capacity will continue to be unlocked, in practice, by a Chinese supply chain that the policy environment is simultaneously trying to discourage.

Historical Context: How Data Center Booms Have Stalled Before

Lessons from the Dot-Com and Cloud Buildouts

The current crisis has historical precedent. During the dot-com boom of 1999–2000, speculative data center construction outpaced actual demand, leading to a wave of empty facilities and bankruptcies. The 2010–2012 cloud computing buildout saw similar growing pains, with power and cooling constraints forcing delays in early hyperscale facilities.

👁 Historical Context: How Data Center Booms Have Stalled Before

Why 2026 Is Bigger and Harder Than Any Prior Cycle

What makes the 2026 situation distinct is the scale. The AI data center buildout dwarfs previous cycles in both capital investment and power demand. A single modern AI training cluster consumes more electricity than an entire early-2000s data center campus. The grid infrastructure was never designed for this density of demand, and upgrading it requires timelines measured in years, not months.

The other distinguishing factor is geopolitical complexity. Previous data center booms occurred in an era of relatively free global trade. Today’s buildout must navigate tariffs, export controls, and supply chain bifurcation between the U.S. and China – constraints that add both cost and delay to every component in the chain.

5 Predictions for the AI Data Center Market Through 2028

Based on the current trajectory of delays, supply chain constraints, and demand growth, the following predictions emerge for the AI data center market:

1. The 2026 capacity gap will push $150–200 billion in infrastructure spending into 2027–2028. Canceled and delayed projects do not disappear – they get rescheduled. The demand backlog will create a compressed construction surge once power and component supply catches up, potentially creating a new wave of bottlenecks.

2. At least two major hyperscalers will announce nuclear power partnerships by end of 2026. Alphabet has already signaled interest in nuclear, and the reliability and density of nuclear power make it the only realistic long-term solution for data center campuses requiring 500 MW or more. Microsoft and Amazon are likely to follow.

3. Electrical component reshoring will become a national security priority. The dependence on Chinese-manufactured transformers and switchgear is too acute to ignore. Expect executive orders or legislative action requiring domestic sourcing for data center power infrastructure, similar to existing requirements for defense-related manufacturing.

4. Modular and pre-fabricated data center designs will gain significant market share. Companies like Crusoe Energy and others are pioneering smaller, distributed data center models that can be deployed closer to power sources with shorter construction timelines. This model sacrifices some efficiency for speed of deployment.

5. International data center investment will accelerate faster than domestic. Countries with available power capacity and favorable regulatory environments – including the Nordics, the Middle East, and parts of Southeast Asia – will attract a disproportionate share of new AI data center investment as U.S. bottlenecks persist.

The Investment Implications: Who Wins and Who Loses

Power Infrastructure Suppliers and Utilities Win

The data center delay wave creates clear winners and losers across the technology and energy sectors. Power infrastructure companies – Eaton, Vertiv, and Schneider Electric – stand to benefit from sustained high demand and pricing power for their products. Their order backlogs are multi-year, and every delay in data center construction only increases the urgency (and willingness to pay premium prices) when projects do move forward.

Utility companies with available capacity in data center corridors are in a strong position. The 51 major U.S. investor-owned utilities planning a combined $1.4 trillion in grid investment over the next five years are effectively building the infrastructure that data center operators desperately need. Those that can deliver power connections fastest will command the highest premiums.

REITs and AI Cloud Tenants Face Revenue Risk

On the losing side, data center REITs and construction companies face revenue uncertainty as projects slip. Companies like Equinix, Digital Realty, and QTS (now part of Blackstone’s portfolio) must manage investor expectations around delayed capacity additions. CoreWeave, which has built its business model around GPU cloud services housed in leased data center space, faces particular pressure as its $30 billion capex plans depend on facilities that may not be ready on schedule.

“The infrastructure bottleneck creates a two-tier market,” said Samantha Lee, managing director at Goldman Sachs’ technology infrastructure practice. “Companies that already have built capacity and power connections become dramatically more valuable. Companies still trying to build face escalating costs and uncertain timelines.”

Regulatory Response: The Senate GRID Act and Beyond

Federal Permitting Frameworks Take Shape

Washington is beginning to respond to the data center infrastructure crisis, though the legislative response remains fragmented. The bipartisan Senate GRID Act, introduced in early 2026, aims to regulate Big Tech’s power consumption and establish permitting frameworks for data center construction. However, the legislation faces opposition from industry groups who argue it could slow deployment further.

👁 Regulatory Response: The Senate GRID Act and Beyond

The policy challenge is multifaceted. Regulators must balance the economic benefits of AI data center development (jobs, tax revenue, technological competitiveness) against the strain on local infrastructure and communities. They must also navigate the tension between encouraging domestic AI development and maintaining trade policies that restrict component imports.

State-Level Tightening in Virginia, Georgia, Texas, and Arizona

At the state level, Virginia – which hosts the largest concentration of data centers in the world – has begun imposing stricter environmental and infrastructure requirements on new facilities. Georgia, Texas, and Arizona have similarly tightened permitting processes in response to community concerns. These state-level actions, while individually modest, collectively contribute to longer development timelines.

The Global Competitive Dimension

China, Europe, and the Middle East Move Faster

The U.S. data center delay crisis does not exist in a vacuum. While American projects stall, competitors are moving forward. China continues to build AI data center capacity at a rapid pace, using domestic manufacturing of electrical components and a regulatory environment that prioritizes infrastructure speed over community input.

In Europe, TikTok’s €12 billion Project Clover is advancing its Lahti, Finland data center as part of a broader European data sovereignty push. The Nordic countries, with their abundant renewable energy and cool climates (which reduce cooling costs), are attracting increasing interest from hyperscalers looking for alternatives to constrained U.S. sites.

The Middle East is emerging as another alternative. Saudi Arabia and the UAE have both announced multi-billion-dollar data center development programs with dedicated power infrastructure. For companies whose AI workloads are not latency-sensitive to U.S. users, these international alternatives offer faster deployment timelines and, in some cases, lower operating costs.

The Strategic Risk to U.S. AI Leadership

The risk for the United States is that infrastructure delays could erode its competitive position in AI. If the most advanced AI systems are trained and deployed from data centers outside the U.S. – not by choice but by necessity – it undermines the strategic rationale behind domestic AI investment incentives and export control policies.

Related Coverage

• Big Tech’s $650 Billion AI Infrastructure Bet: Inside the Capex Surge Reshaping the Global Economy
• The $1.4 Trillion Grid Overhaul: How 51 Utilities and $300B in AI Capex Are Reshaping America’s Power System
• CoreWeave’s $30 Billion Capex Gamble: Inside the AI Cloud Company’s Debt-Fueled Rise
• Nvidia Vera Rubin Platform: Inside the 336B-Transistor Chip and 5x Blackwell Leap
• Senate GRID Act Targets Data Center Energy Crisis
• Samsung’s $73 Billion Semiconductor Bet: Inside the Record Investment
• AI Chips 2026: The Guide

May 12, 2026 Update: Bloomberg, TechRadar, and The Standard Tighten the Picture

The reporting cycle through May 12, 2026 has hardened the case that the U.S. AI infrastructure shortfall is structural, not transient. Three fresh data points – sourced to Bloomberg and republished by TechRadar and The Standard on the same day – sharpen the contours of the 2026 buildout gap and explain why the hyperscaler spend cannot translate into proportional capacity this year.

One-Third to One-Half of 2026 U.S. Data Centers Delayed or Canceled

Bloomberg reporting, as carried by TechRadar on May 12, 2026, now puts the official range at between one-third and one-half of all U.S. data centers planned for 2026 being delayed or totally canceled. The framing matters: prior April reads hovered around “roughly half,” but the new range explicitly acknowledges the lower bound at 33% as well – a sourced quantification rather than a directional estimate. For investors and operators, the practical implication is that even the optimistic scenario for 2026 leaves a third of announced U.S. capacity unbuilt, while the pessimistic scenario takes that number to 50%.

$650 Billion Hyperscaler Spend Confirmed Intact for 2026

The same Bloomberg cycle, cited by The Standard on May 12, 2026, confirms that Alphabet, Amazon, Meta, and Microsoft are set to spend more than $650 billion in 2026 on AI capacity. That commitment is reported as fully intact, even with the cancellation tally moving up. The signal is that capex guidance from the Big Four hyperscalers has not been revised downward in response to the buildout shortfall – meaning the gap between dollars committed and watts energized is, by construction, widening rather than closing as the year progresses.

Transformer Delivery Now Stretches Up to 5 Years vs. Sub-18-Month Deployment Cycles

The single most consequential update is the formal confirmation of how mismatched the timelines have become. Bloomberg, citing Sightline Climate via The Standard on May 12, 2026, reports that transformer delivery times have stretched from 24 to 30 months pre-2020 to as long as 5 years today. Against that backdrop, AI data center deployment cycles themselves run under 18 months from groundbreaking to energization when components are available. The asymmetry – a 60-month wait for a transformer to feed an 18-month-buildable facility – is the cleanest single explanation for why the 2026 capacity shortfall cannot be closed by adding capital. The bottleneck is calendar time on a physical component, and no announcement can compress it.

May 12, 2026 Data Point	Number	Source
Share of 2026 U.S. data centers delayed or canceled	One-third to one-half (33–50%)	Bloomberg via TechRadar, May 12, 2026
2026 AI capacity spend by Alphabet, Amazon, Meta, Microsoft	More than $650 billion	Bloomberg via The Standard, May 12, 2026
Transformer delivery time, pre-2020	24–30 months	Bloomberg citing Sightline Climate via The Standard, May 12, 2026
Transformer delivery time, May 2026	Up to 5 years	Bloomberg citing Sightline Climate via The Standard, May 12, 2026
AI data center deployment cycle	Under 18 months	Bloomberg citing Sightline Climate via The Standard, May 12, 2026

Taken together, the May 12, 2026 data tightens the narrative from “supply chain pressure” to a quantifiable structural mismatch: 33–50% of 2026 U.S. data centers delayed or canceled, more than $650 billion in hyperscaler spend still on the books, and transformer lead times running roughly three times longer than the data centers they will eventually power. The numbers reframe 2026 from a year of expected AI capacity expansion to a year of measured slippage with no near-term resolution.

May 15, 2026 Bottom Line: Three Verified Numbers That Define the Crisis

As of May 2026, three figures – each sourced to Sightline Climate and reported by Bloomberg, with corroboration in Tom’s Hardware and TechRadar – anchor the entire U.S. AI data center capacity story. These are the load-bearing numbers an operator, investor, or policymaker needs to internalize before reading anything else about the 2026 cycle.

1. Roughly 50% of Planned 2026 U.S. Data Center Builds Are Delayed or Canceled

About half of all U.S. data center builds planned to come online in 2026 are now projected to be delayed or canceled, against a baseline of approximately 12 gigawatts (GW) of capacity originally expected to energize this year. The estimate comes from Sightline Climate, as reported by Bloomberg. The implication is direct: even before factoring in lead-time risk on power equipment, the headline pipeline has already lost roughly half its volume for the calendar year.

2. Only About 4 GW of the 12 GW Is Under Active Construction

The construction-rate figure is starker than the cancellation tally. Of the 12 GW of 2026 U.S. capacity announced, only about 4 GW – roughly one-third – is under active construction, according to Sightline Climate data carried by Bloomberg. The remaining two-thirds of the pipeline is still pre-construction and therefore exposed to further cancellation, repricing, or indefinite postponement. The 4 GW number tightens the earlier “roughly 5 GW” placeholder and sets the construction rate at a sourced 33%.

3. The Real Bottleneck Is Power Equipment – Not Chips

The most important reframing in the May 2026 reporting is that the binding constraint has shifted off of compute silicon. The critical shortages are now in transformers, switchgear, and batteries – the electrical equipment required to deliver power to a data center campus. Bloomberg, citing Sightline Climate with industry reporting reproduced by Tom’s Hardware and TechRadar, makes the case that these shortages are delaying buildouts even as Alphabet, Amazon, Meta, and Microsoft are expected to spend more than $650 billion in 2026 on AI infrastructure. Capital is fully committed; the components to convert that capital into energized capacity are not.

May 2026 Verified Metric	Number	Source
Share of planned 2026 U.S. data center builds delayed or canceled	~50%	Sightline Climate via Bloomberg
2026 U.S. data center capacity originally expected online	~12 GW	Sightline Climate via Bloomberg
2026 U.S. capacity actually under active construction	~4 GW (~1/3)	Sightline Climate via Bloomberg
Primary bottleneck components	Transformers, switchgear, batteries	Bloomberg, Sightline Climate, Tom’s Hardware, TechRadar
2026 AI infrastructure spend, Alphabet + Amazon + Meta + Microsoft	>$650 billion	Bloomberg, Sightline Climate

Together, these three verified data points reframe the 2026 cycle in a single sentence: half the pipeline is gone, one-third is being built, and the limiting factor is power equipment – not silicon – against a $650 billion capex commitment that has not been revised downward.

May 17, 2026 Update: Sightline Revises Pipeline to 16 GW, China Parts Shortage Named

Reporting in mid-May 2026 from Sightline Climate, cited by Bloomberg and summarized by Tom’s Hardware and TechRadar, refines the 2026 picture. The full announced pipeline for U.S. data centers in 2026 is now pegged at 16 gigawatts, with only approximately 5 GW actually under construction. That implies roughly 30% to 50% of the 2026 pipeline is on track to be delayed or canceled – a wider deferral band than earlier in the year, and consistent with the slower-than-expected ramp through Q1 and Q2.

The Bottleneck Has Shifted From Chips and Capital to Electrical Equipment

The most important reframing in the May 2026 reporting is what is now binding. Bloomberg, citing Sightline Climate, identifies the constraint set as transformers, switchgear, batteries, and grid power availability. Funding is not the limit. Semiconductors are not the limit. The limit is the electrical equipment and the utility interconnection capacity needed to convert capex into energized megawatts. For an industry that spent most of 2024 and early 2025 worrying about GPU allocation, that is a clean inversion of the scarcity story.

$650 Billion in 2026 Capex Still on the Table, Even as Slippage Widens

Coverage referencing Tom’s Hardware and Bloomberg reiterates that Alphabet, Amazon, Meta, and Microsoft are expected to spend more than $650 billion in 2026, even as a large share of planned U.S. data center builds slip. The named causes for the slippage are shortages of power infrastructure and parts from China – meaning the supply chain for transformers, switchgear assemblies, and battery cells is exposed to both manufacturing capacity limits and trade-policy frictions on Chinese components.

May 2026 Refined Figure	Value	Source
Announced U.S. data center capacity for 2026	16 GW	Sightline Climate via Bloomberg (Tom’s Hardware, TechRadar)
Under active construction	~5 GW	Sightline Climate via Bloomberg
Share of 2026 pipeline expected delayed or canceled	~30%–50%	Sightline Climate via Bloomberg
Named binding constraints	Transformers, switchgear, batteries, grid power availability	Bloomberg citing Sightline Climate
Named supply-chain cause	Shortages of power infrastructure and parts from China	Tom’s Hardware / Bloomberg
Big Four 2026 AI capex	>$650 billion	Tom’s Hardware / Bloomberg

The May 2026 update does not contradict the earlier 12 GW figure cited in this article; it broadens the lens. Sightline’s 16 GW number captures the full announced 2026 pipeline including projects that were already softening, while the ~5 GW under construction is the harder, build-ready subset. The investable takeaway is the same one this article has flagged since the 7 GW shortfall first appeared: capital is not the scarce input – energized power equipment is.

May 2026 Sightline Climate Deep Dive: 9 Canceled Projects, 4–7 Year Grid Waits, and the Maine Moratorium

The most granular May 2026 reporting from Sightline Climate, surfaced by Bloomberg and echoed by Tom’s Hardware, TechRadar, and Latitude Media, sharpens the 2026 U.S. data center picture in three concrete ways that earlier coverage only sketched. First, against 16 GW of announced 2026 capacity, only 5 GW is currently under construction – the cleanest reconciliation yet of the prior 12 GW vs. 16 GW framings, with the 16 GW number now treated as the official pipeline figure. Second, Sightline has now directly tracked 9 canceled projects in its 2026 dataset, moving the cancellation tally from an analyst estimate to a specific count. Third, the binding bottleneck has been named with new precision: the shortage is in transformers, switchgear, busways, and battery backup systems, layered on top of utility interconnection queues that now stretch 4 to 7 years in the markets where most of the announced 2026 capacity is actually sited.

30%–50% of 2026 U.S. Capacity Now Expected to Be Delayed or Canceled

The headline range from Sightline Climate as carried by Bloomberg in May 2026 is that 30% to 50% of 2026 U.S. data center capacity is now expected to be delayed or canceled. The underlying arithmetic is the same one this article has tracked all year, but the figures are now stable across sources: 16 GW planned for 2026, only 5 GW under active construction. That construction rate – roughly 31% against the full announced pipeline – is the load-bearing number behind the “half delayed” headline, and it is the version Tom’s Hardware and TechRadar have reproduced in their May 2026 summaries of the Bloomberg cycle.

Grid Interconnection Waits of 4–7 Years in Northern Virginia, Phoenix, and Dallas

The single most operationally important new figure in the May 2026 Sightline read is the named lead time on utility interconnection in the three U.S. markets that host the largest share of announced 2026 capacity. Grid interconnection waits in Northern Virginia, Phoenix, and Dallas are now reported at 4 to 7 years, per Sightline Climate data cited by Bloomberg. That range is what converts the announced 16 GW pipeline into a five-to-late-decade energization curve rather than a 2026 calendar event. A campus that joins the Northern Virginia interconnection queue in May 2026 cannot realistically expect to draw utility power before 2030 to 2033, regardless of how quickly the building itself is constructed or how much capital the operator commits.

The Bottleneck Component Set Now Includes Busways and Battery Backup

Earlier coverage in this article tracked transformers, switchgear, and batteries as the three named bottleneck components. The May 2026 Sightline Climate update broadens that list to include busways and explicitly calls out battery backup systems as a distinct constraint from grid-tied storage. Busways – the heavy-current bus duct that distributes power from switchgear to rack rows inside a hyperscale facility – have joined the long-lead category for the first time, reflecting that the bottleneck is no longer at a single point in the electrical chain but spans the entire power-distribution path from substation to server. The implication is that even sites that successfully clear the utility interconnection queue can still slip on internal distribution if any one of transformers, switchgear, busways, or battery backup systems is out of sequence.

9 Tracked Cancellations and the Maine House’s 82–62 Moratorium Vote Through 2027

Community opposition has moved from a backdrop risk to a measurable cause of project attrition. Sightline Climate has now directly tracked 9 canceled projects in its 2026 dataset, as reported by Latitude Media in May 2026. The most visible policy signal in that bucket is the Maine House’s 82–62 vote to impose a data center moratorium through 2027 – a state-level pause that takes Maine off the map for new hyperscale siting for the rest of this cycle. Maine is not, by itself, a major data center corridor, but the 82–62 margin matters as a template: it shows a clear bipartisan majority in a state legislature is willing to vote a multi-year moratorium even before community opposition reaches the scale seen in Virginia, Georgia, or Texas.

Grid-Connected Still Dominates, but Onsite-Power-Only Sits at Just 3% of 2026 Projects

The May 2026 Sightline data also resolves a question that has been open since the natural-gas-on-site narrative gained traction in late 2025: how much of the 2026 pipeline is actually planned to bypass the utility grid? The answer is very little. More than half of expected 2026 U.S. data center projects are still planned to be grid-connected, while only about 3% are designed for onsite power only. The remainder rely on hybrid configurations that still depend on the utility for at least part of their load. The takeaway is that the workaround narratives – behind-the-meter gas plants, dedicated nuclear small modular reactors, captive renewable build-outs – are real but tiny in 2026’s pipeline composition. The dominant outcome remains tied to the same utility interconnection queues that are running 4 to 7 years long.

May 2026 Sightline Climate Data Point	Value	Source Chain
2026 U.S. capacity announced (planned)	16 GW	Sightline Climate via Bloomberg, Tom’s Hardware, TechRadar
2026 U.S. capacity under construction	5 GW	Sightline Climate via Bloomberg
Share of 2026 capacity delayed or canceled	30%–50%	Sightline Climate via Bloomberg
Tracked canceled projects (Sightline dataset)	9 projects	Sightline Climate via Latitude Media
Grid interconnection wait – Northern Virginia, Phoenix, Dallas	4 to 7 years	Sightline Climate via Bloomberg
Named bottleneck components	Transformers, switchgear, busways, battery backup systems	Sightline Climate via Bloomberg
Share of 2026 projects grid-connected	More than 50%	Sightline Climate via Latitude Media
Share of 2026 projects on onsite power only	~3%	Sightline Climate via Latitude Media
Maine House moratorium vote (through 2027)	82–62 in favor	Latitude Media, May 2026

Read together, the May 2026 Sightline Climate update reframes the 2026 U.S. data center crisis as a multi-front constraint problem rather than a single supply-chain story. The components bottleneck has widened from transformers, switchgear, and batteries to also include busways. The utility-side bottleneck is now quantified at 4 to 7 years in the three highest-density U.S. data center markets. The community-opposition bottleneck has produced a specific count of 9 canceled projects and a state-level moratorium vote in Maine of 82–62 through 2027. And the workaround scenario – bypassing the grid entirely – applies to only about 3% of the 2026 pipeline. Every input that could meaningfully accelerate the 2026 buildout has now been measured and quantified, and none of them are moving fast enough to close the gap against the $650 billion in hyperscaler capex that remains pointed at this year’s announced capacity.

FAQ: AI Data Center Cancellations and Delays in 2026

How many U.S. AI data centers have been canceled or delayed in 2026?

Nearly half of all U.S. data centers planned for 2026 – representing approximately 7 GW out of 12 GW of announced capacity – have been canceled or delayed. Only about 5 GW is currently under active construction, according to Bloomberg reporting and Sightline Climate tracking data.

What is causing the AI data center delays?

The primary causes are shortages of electrical infrastructure components (transformers, switchgear, and batteries), power grid constraints that can take up to five years to resolve, tariff impacts on Chinese-manufactured components, and growing community opposition to data center construction in key regions.

Is OpenAI’s Stargate Project affected by the delays?

Yes. OpenAI’s $500 billion Stargate Project in Texas has shown no significant physical progress as of April 2026, despite massive financial backing from SoftBank. The project faces the same power infrastructure and component shortage constraints affecting the broader industry.

How do the delays affect Nvidia’s GPU business?

Data center delays push GPU orders to the right on the timeline. Facilities that are not built cannot be filled with GPUs. While total demand has not decreased, the timing of Nvidia’s revenue recognition becomes less predictable, and pricing power may soften as deployment urgency decreases.

Which companies benefit from the data center bottleneck?

Power infrastructure companies like Eaton, Vertiv, and Schneider Electric benefit from sustained demand and pricing power. Utilities with available capacity in data center corridors are also well-positioned. Companies with existing built-out data center capacity become dramatically more valuable.

When will the data center capacity crisis be resolved?

No firm recovery timeline exists. Power infrastructure upgrades can take up to five years, and electrical component reshoring is a multi-year process. Industry analysts expect the bottleneck to persist through at least 2028, with a compressed construction surge likely in 2027–2028 as supply catches up with demand.

How long does it now take to get a transformer for an AI data center?

According to Bloomberg, citing Sightline Climate via The Standard on May 12, 2026, transformer delivery has stretched from 24–30 months pre-2020 to as long as 5 years today. Because AI data center deployment cycles themselves run under 18 months, the transformer queue – not the building – is the binding constraint on energizing new capacity.

How much are Alphabet, Amazon, Meta, and Microsoft spending on AI capacity in 2026?

Bloomberg, cited by The Standard on May 12, 2026, reports that Alphabet, Amazon, Meta, and Microsoft are set to spend more than $650 billion in 2026 on AI capacity. That figure remains intact despite the May 2026 confirmation that between one-third and one-half of planned U.S. data centers for 2026 will be delayed or canceled.

Why has Sightline Climate revised the 2026 pipeline up to 16 GW?

The 16 GW figure published in May 2026 by Sightline Climate (via Bloomberg, Tom’s Hardware, and TechRadar) reflects the full announced pipeline for U.S. data centers in 2026, including projects that had not been fully captured in earlier 12 GW counts. Against that 16 GW, only about 5 GW is actually under construction, implying roughly 30% to 50% of the year’s pipeline is on track to be delayed or canceled.

What role do Chinese components play in the AI data center delays?

According to May 2026 reporting referencing Tom’s Hardware and Bloomberg, a named cause of the slippage is shortages of power infrastructure and parts from China. Transformers, switchgear assemblies, and battery cells all rely on Chinese-manufactured inputs, so manufacturing constraints and trade-policy frictions on Chinese components directly throttle how quickly U.S. data center shells can be energized – even when capital and chips are available.

How long are grid interconnection waits in the biggest U.S. data center markets right now?

Per Sightline Climate data cited by Bloomberg in May 2026, grid interconnection waits in Northern Virginia, Phoenix, and Dallas are now running 4 to 7 years. That timeline applies even to campuses with full capital, GPU allocation, and broken-ground construction – the queue is set by the utility, not the operator, which is why grid power is the binding constraint across the 2026 pipeline.

How many 2026 U.S. data center projects has Sightline Climate confirmed as canceled?

Sightline Climate has directly tracked 9 canceled projects in its 2026 dataset as of May 2026, reported by Latitude Media. That is the floor for cancellations rather than the ceiling – the broader 30%–50% delayed-or-canceled estimate still applies to the wider 16 GW pipeline, with additional projects expected to slip from delayed into the canceled category through the second half of the year.

Are most 2026 U.S. data centers planned to bypass the grid with onsite power?

No. Sightline Climate data carried by Latitude Media in May 2026 shows that more than 50% of expected 2026 U.S. data center projects are still planned to be grid-connected, while only about 3% are designed for onsite power only. The remainder are hybrid configurations that still depend on the utility for at least part of their load, which is why utility interconnection waits drive the schedule for the bulk of the 2026 pipeline.

What did the Maine House moratorium vote do, and why does it matter?

The Maine House voted 82–62 in May 2026 to impose a data center moratorium through 2027, per Latitude Media. Maine is not a top-tier U.S. data center market, but the margin of the vote matters as a template: it shows that a clear bipartisan majority in a state legislature is willing to pause new hyperscale siting for a multi-year window, signaling that community-opposition risk has moved from informal protest into binding state policy.