Only 802 of 3,969 announced US data centers are actually under construction. That gap is the most important number in nonresidential construction right now.
There are two conversations happening about AI data centers, and they almost never touch.
The first is happening in finance and technology media. It involves credit analysts, Federal Reserve presidents, short sellers, and writers like Ed Zitron, and it is fundamentally an argument about whether the capital being deployed into AI infrastructure represents the largest misallocation in modern economic history.
The second is happening in architecture, engineering, and construction. It involves MEP engineers, general contractors, electrical subs, and switchgear manufacturers, and it is fundamentally about the best three years many of these firms have ever had.
Conversation one is about whether the money is real. Conversation two is about whether the concrete gets poured.
These are the same conversation. Very few people are treating them that way, and the cost of that gap will be paid by firms who scaled their cost base against a pipeline they never underwrote.
The number that should reframe your backlog
CNN published a piece on August 6 with a counterintuitive premise: Americans are mobilizing against data centers, and yet surprisingly few are actually being built.
The supporting data is worth sitting with.
The United States ended last year with 5,427 data centers, according to Stanford’s AI Index Report. AI companies have announced plans for 3,969 more — a stated intention to roughly double the national footprint. Of those announced projects, 802 are under construction, per research firm Aterio.
That is not a rounding error. That is a 20% conversion rate from announcement to shovel.
The pattern repeats across every credible dataset:
- JPMorgan reports that roughly 60% of data center capacity planned for 2027 completion has not begun construction. Another 7% of started projects have already slipped.
- Goldman Sachs expects only about half of the AI computing capacity scheduled to activate between now and 2028 to hit its target date. The historical on-time rate for data center capacity is closer to 72%.
- Data Center Watch counted at least 75 US projects worth approximately $130 billion blocked or delayed in Q1 2026 alone.
- There are now 438 distinct data center developers with active US projects, according to Cleanview.
That last figure deserves emphasis. Four hundred and thirty-eight developers is not the signature of a disciplined, capital-constrained market. It is the signature of a land rush — and land rushes have a consistent historical ending.
Meanwhile the capital commitments keep climbing. JPMorgan puts AI infrastructure investment at roughly $750 billion this year. ABC chief economist Anirban Basu has cited hyperscaler AI-related construction spending near $450 billion in 2025, with 2026 projections in the $700–725 billion range. Columbia’s Stijn Van Nieuwerburgh, quoted in the CNN piece, prices a single state-of-the-art AI campus at around $8 billion — and offered the warning that ought to be the epigraph for this entire cycle: getting the timing right on buildouts this large is very hard, and the recurring pattern is over-excitement, excessive debt, and investments that go bust.
Two opposite failure modes that damage you identically
Here is what makes this difficult to reason about clearly.
The optimistic reading of that data is that supply is severely constrained, demand vastly exceeds deliverable capacity, and anyone with capability in this market holds pricing power for a decade.
The pessimistic reading is that two-thirds of the announced pipeline is vapor, that this is a press-release pipeline rather than a construction pipeline, and that when financing conditions change the vapor evaporates.
Both readings are correct. They describe different segments of the same market. And critically, both produce the same category of damage to firms in the delivery chain — just on different timelines.
Failure mode one is the delivery crunch, and it is happening now. You win the work and can’t staff it. You can’t source the switchgear. The utility interconnect slips fourteen months. Generation step-up transformer wait times have tripled, per JPMorgan. Since 2020, transformers and power regulators have registered the second-highest inflation of all 47 categories the Bureau of Labor Statistics tracks in its Producer Price Index. Your fixed-price exposure erodes margin while liquidated damages exposure climbs, because delayed commissioning on a 60 MW facility can cost an owner an estimated $14 million per month in unrealized revenue.
So you hire aggressively at premium wages to protect schedule. Your cost basis permanently ratchets upward.
This is uncomfortable, but it is a good problem. The work exists and eventually gets paid.
Failure mode two is the demand air pocket, and it is the one to actually worry about. Financing conditions shift. A significant counterparty renegotiates or defaults. The pipeline that was never real gets formally cancelled rather than quietly deferred. The 60% of 2027 capacity that never broke ground never breaks ground. Design backlog fills with projects that never reach GMP. And you are carrying two hundred people and forty million dollars of equipment against a pipeline that no longer exists.
The cruelty is in the sequence. Failure mode one compels you into exactly the position that makes failure mode two lethal. You must scale to serve the crunch. Scaling is what kills you when the crunch ends.
This is not a novel pattern. It is the standard shape of every construction-adjacent boom and bust on record.
The fragility is in the capital stack, not the buildings
If you want to locate the actual risk, stop looking at the concrete and look at how it is financed.
A large and growing share of AI data center construction is not funded off hyperscaler balance sheets. It runs through special purpose vehicles — separate legal entities that own the asset, carry the debt, and lease the facility back to the operator.
A Quinn Emanuel client alert on AI data center financing documents the scale: roughly $13 billion from Blue Owl and JPMorgan into an SPV owning the Oracle/OpenAI facility in Abilene, Texas; a $38 billion debt package covering two data centers in Texas and Wisconsin; an $18 billion loan for a New Mexico site; and Meta’s Hyperion facility in Louisiana, a $30 billion private credit transaction structured through an SPV called Beignet Investor.
Private credit lending to AI-related companies has gone from near zero to over $200 billion in a few years. Morgan Stanley has projected private credit could supply another $800 billion in data center financing.
This changes who you are actually doing business with.
When Microsoft is your owner, you are contracting with a AAA-rated entity holding tens of billions in cash. When the counterparty is a project-specific SPV, you are contracting with a bankruptcy-remote shell whose only asset is the partially completed building you are standing in, and whose only revenue is a lease from a tenant whose only revenue is a compute contract from a company that has never turned a profit.
Your mechanic’s lien attaches to that building. In a downside scenario, you are positioned behind a project finance lender holding a first-position mortgage and an intercreditor agreement you were never shown.
The counterparty concentration compounds this. Publicly reported OpenAI compute commitments alone include roughly $300 billion to Oracle, $38 billion to Amazon, and $22 billion to CoreWeave. Those take-or-pay contracts are the credit support making the debt underwritable — which means the entire structure inherits the credit quality of the compute buyer.
Oracle illustrates the mechanics. Its five-year credit default swap spread reached roughly 1.25 percentage points, a three-year high. It has carried over $100 billion in debt with free cash flow gone negative — notably the only major hyperscaler funding this buildout primarily with debt. S&P has flagged the OpenAI concentration directly: if OpenAI cannot meet its obligations, Oracle holds long-dated leases with no easy exit. There is also a duration mismatch, with reported lease and capacity commitments running fifteen to nineteen years against customer contracts closer to five. Oracle bondholders sued in January over losses tied to the buildout.
In March, CNBC reported OpenAI declining to expand its flagship Stargate site with Oracle, preferring next-generation chips at new locations. That is the duration mismatch converting from theory into a headline.
Ed Zitron’s case, taken seriously
Ed Zitron writes the newsletter Where’s Your Ed At and hosts Better Offline. Politico described him as this boom’s most acerbic gadfly. He is routinely dismissed by people in the infrastructure business who have not actually read him.
They should. Not because he is necessarily right — I will give the counterarguments — but because his thesis describes the specific mechanism that would eliminate your backlog. Understanding it is risk management, not ideology.
His core claims:
The unit economics of inference don’t work. Zitron argues LLM costs run contrary to essentially every model of selling software. Traditional software carries near-zero marginal cost per user; generative AI carries substantial, persistent marginal cost per query. He reads the early-2026 industry shift toward token-based billing as the tell, arguing that once enterprises were made to pay something closer to true cost, many pulled back within months because they could not demonstrate ROI. He has cited SemiAnalysis data suggesting subscribers on $200 monthly plans can consume thousands of dollars in tokens.
The revenue required to justify the capex does not exist and cannot plausibly appear in time. By his estimate, the buildout and compute commitments across OpenAI, Anthropic, Nvidia, and Oracle imply the need for something on the order of $2–3 trillion in annual AI revenue by 2030. He notes that OpenAI and Anthropic together account for roughly 89% of AI startup revenue, and that their combined projected 2026 revenue would need to grow several hundred percent within a few years to close that gap.
The demand is substantially circular. This is his load-bearing claim: that most capacity is being absorbed by the model labs themselves rather than by broad enterprise adoption. If accurate, the demand justifying the buildout is largely labs consuming capacity funded by the same hyperscalers constructing it — which is not demand in any economically meaningful sense.
The debt is opaque and the contagion path runs through pensions. Private credit funds are financing these facilities, and those funds are substantially backed by pension capital. That is a systemic transmission channel most people have not priced.
There is no clean bailout. This is his sharpest and least-appreciated point. Project financing means the capital is effectively spent. The only routes to making lenders whole are buying out the debt or manufacturing revenue that does not exist. Unlike 2008, there is no obvious backstop mechanism.
And the failure modes do not require everything to go wrong. A facility that never breaks ground because financing collapses. One that stalls mid-construction. One that opens and finds no customers. His argument is that one or two prominent failures reprice risk across the entire sector.
The counterarguments are also serious
The demand crunch is empirically visible right now. Amazon has stated it expects capacity to trail customer demand despite increasing spend. Second-quarter hyperscaler earnings calls pivoted from discussing aggregate capex to discussing time-to-energy — how quickly capacity can be energized and monetized. That is rationing language, not glut language.
Zitron has also been directionally early before. He has called this a bubble for years while the buildout accelerated. For anyone making decisions on a twelve-month horizon, being early is operationally indistinguishable from being wrong.
His unit economics argument assumes relatively static efficiency, and inference cost per token has fallen dramatically and repeatedly through distillation, quantization, better serving infrastructure, and specialized silicon.
And the physical constraints genuinely cut against a glut. It is difficult to overbuild a market where transformer lead times have tripled, turbine slots are sold out for years, and there is a documented national shortage of people qualified to pull fiber-optic cable.
Where I land
Zitron is probably wrong about the timing and probably right about the structure.
The buildout is not stopping next quarter. But the financing architecture — SPV project finance, private credit, take-or-pay contracts backstopped by unprofitable counterparties, circular vendor financing echoing Nortel and Lucent in the late 1990s, nineteen-year leases against five-year contracts — is fragile in a specific technical sense. It has no shock absorber.
It does not require AI to stop being useful. It requires a credit event.
And you do not need to resolve the philosophical question about AI’s trajectory to manage that exposure. You need to read your contracts.
The political layer is already moving
Thirty-eight states currently offer data center tax incentives, according to the National Conference of State Legislatures. Lawmakers in at least 28 of those states introduced bills in 2026 to curb them. At least nine considered outright repeal.
Illinois enacted a two-year pause on state data center tax incentives effective July 1. Arizona paused its incentives in June. Oklahoma’s Data Center Customer Ratepayer Protection Act now requires facilities with 75 MW or greater peak demand to pay their full utility service cost, and Tennessee enacted a comparable statute. New York moved to pause new facilities at or above 50 MW. Gallup found in May that 71% of Americans oppose AI data center construction in their local area, including 48% strongly opposed.
There is a second-order effect here that most AEC firms have not connected.
When a utility builds generation and transmission capacity to serve a promised data center load, that investment enters the rate base. If the load never materializes, that infrastructure becomes a stranded asset — and under conventional ratemaking, residential and small business customers absorb the cost. States are legislating to prevent exactly that outcome.
Which means ratepayer protection and project pipeline are in direct tension. Every protective statute raises the effective cost of a new project and pushes marginal projects below the hurdle rate. It is good policy and bad backlog. And the incentive package a developer priced into a pro forma eighteen months ago may simply not exist when they break ground.
What this actually means for firms
The industry data already shows a structure worth worrying about.
Associated Builders and Contractors reported overall contractor backlog at 8.8 months in April 2026, a ten-month high — a reading Basu characterized as driven by a narrow subset of the membership. Contractors with data center work carry approximately 12.2 months of backlog. Those without carry 8.3. Forty-two percent of contractors above $100 million in revenue are under contract for data center projects; only 7% of smaller contractors are. Firms above $100 million posted their highest backlogs since 2021. Firms under $30 million posted their lowest.
Moody’s has been blunt that data center development is the primary driver behind essentially any observed improvement in overall construction metrics.
Read that again. The headline numbers routinely cited as evidence of a healthy construction industry are, to a substantial degree, one sector. If data center construction contracts by a third, the entire nonresidential picture moves from modest growth to recession — and the firms currently outperforming fall furthest, because they are the ones who scaled.
For design firms, the exposure is not primarily bad debt. It is capacity commitment against a phantom pipeline. Master service agreements with no minimum volume are being carried as backlog when they are actually pricing agreements with an option the other party holds for free. Programmatic multi-site work feels secure and is actually single-point-of-failure concentration in disguise.
For contractors, working capital exposure is enormous and asymmetric. You are fronting significant cash for long-lead equipment. If an owner slows payment during a liquidity event — not defaults, merely slows — you can be technically profitable and functionally insolvent inside ninety days. And lien rights against project-financed assets are weaker than most subs assume when they sign a lender’s direct agreement as a formality.
For manufacturers, the risk is the bullwhip. Eaton, Schneider Electric, Hitachi Energy, and GE Vernova are all committing enormous capital to new capacity — capital underwritten against demand forecasts built on the announced pipeline rather than the under-construction pipeline. Worse, in a shortage buyers routinely order the same transformer from three suppliers to guarantee delivery. When lead times normalize, two of those three orders vanish. Every order book in this sector currently contains phantom demand that cannot be distinguished from real demand.
Four questions worth answering this quarter
Regardless of your view on AI, these are answerable now and expensive to answer later.
1. What percentage of your backlog is committed versus optioned? Track signed-and-funded, signed-but-optional, and MSA-implied work in three separate buckets that are never summed in a board report. A number of firms are currently reporting an MSA-inflated backlog figure to their lenders.
2. Who is the actual contracting entity on each agreement? Hyperscaler or SPV. Parent guarantee, letter of credit, or nothing. Equity fully funded at closing or drawn over time. These questions are routine in project finance and strangely rare in construction procurement.
3. What breaks at a 30%, 50%, and 70% pipeline reduction? Not to predict, but to locate the breaking point. Knowing which covenant fails at which threshold converts a future panic into a plan.
4. What is your DSO trend on data center work specifically, versus everything else? This is the single earliest reliable distress signal available to you. Owners under financial pressure slow payment long before they announce anything.
The best time to negotiate demobilization compensation, payment security, escalation clauses, and non-refundable deposits is while the shortage still gives you leverage to ask. That window closes the moment lead times normalize — and it closes quietly, before anyone calls it a downturn.
I’d be interested in what people on the delivery side are seeing. Are lead times still stretching in your market, or beginning to ease? Are payments still arriving on schedule? Aggregate data lags by months. The people pulling cable know first.