The AI Infrastructure Trade: How it Fails

A technology paradigm rarely announces its ceiling

THE BREAKLINE | ISSUE 005

What Breaks the AI Trade?

The AI trade has absorbed a remarkable amount of bad news. Capital spending has climbed faster than many investors expected. Data centers have been delayed. Power has become harder to secure. Communities are fighting new projects. Software stocks have sold off over fears that AI will eat the very businesses buying cloud services. Free cash flow has come under pressure at some of the largest technology companies in the world.

And yet demand keeps coming. Google Cloud grew 82% last quarter and ended June with $514 billion of contracted backlog. Microsoft says its AI business has passed a $37 billion annual revenue run rate; up 123%. Meta continues to grow advertising revenue while spending heavily on AI. Oracle finished fiscal 2026 with $638 billion of remaining performance obligations, much of the recent increase tied to large AI contracts.

So perhaps the interesting question is no longer whether the AI buildout is under stress. It clearly is.

Four issues into this series, we have followed the trade from capital commitments into chips, memory, networking, power equipment, data centers and eventually revenue. The evidence has become increasingly physical for why the trade is going right. Now we run the chain backward. Now we ask: what has to go wrong for this trade to fail?

The AI trade won’t fail because the system is under stress; it already is. It fails when that stress interrupts the conversion of demand into durable returns.

That distinction matters because delays, higher costs, and ugly quarters are not enough on their own. We’ve witnessed these already. The trade breaks only when one part of the system stops feeding the next.

1. Demand That Cannot Become Capacity

Let’s start with the easiest part of the thesis to defend: demand is real.

Google's Cloud backlog is difficult to dismiss as enthusiasm, and management continues to say demand exceeds available capacity. Microsoft has been saying much the same thing while its AI revenue grows rapidly. Oracle echoes the other two all while CoreWeave accumulates $100 billion in backlog and continues adding large new customer commitments.

That is the bull case in its simplest form: customers want more compute than the industry can currently provide. Scarcity supports pricing power, reinforces backlog and gives suppliers time to expand. Scarcity, however, cuts both ways.

Backlog is demand waiting for something to happen and capacity constraints are showing up across the infrastructure market. Power has to arrive. Transformers have to arrive. HBM has to arrive. Networking has to connect the machines. Utilities have to approve interconnections. Permits have to survive local politics. Buildings have to get built. Until then, backlog is a only claim on future capacity and not materialized revenues.

We already know that conversion can stall. Data Center Watch counted at least 75 U.S. data-center projects worth roughly $130 billion that were blocked or delayed during the first quarter of 2026 alone. That sounds bearish, but delay by itself does not invalidate the trade. Projects slip during every large infrastructure cycle. Customers move workloads. Developers change sites. Capital gets reallocated. A delayed data center does not evaporate AI demand.

The more useful question is what customers do when waiting becomes too expensive. They can move workloads to another cloud or another geography, accept older or less capable hardware, optimize models to use less compute, defer projects, build internally, or conclude that the workload was never valuable enough to justify the cost.

Scarcity supports the trade while customers remain willing to compete for scarce capacity. The real danger appears when customers stop waiting for more compute and begin finding ways not to need it.

That gives us a cleaner set of signals to watch. Cancellations matter more than delays. Shrinking commitments matter more than slipping construction dates. Efficiency gains matter when they stop being optimizations and start becoming substitutions.

The bend is a customer waiting another six months for resources. The break is a customer discovering that they did not need the capacity in the first place.

2. Capacity That Cannot Earn Its Cost

The industry could solve its capacity problem and still destroy capital. That makes monetization the harder test and probably the most important one.

Alphabet, Microsoft, Meta and Amazon are spending at levels that would have looked absurd only a few years ago. Yet the evidence that AI is producing revenue is also becoming harder to dismiss.

Meta is a useful example because it shows why capex alone tells us very little. The company has continued increasing infrastructure spending while advertising revenue, impressions and pricing have all strengthened. In fact, revenue grew 28% from the previous year. However, free cash flow dropped 91% from $8.6 billion to $784 million. If AI is improving targeting, engagement and ad economics, then large capex is not evidence of failure. It is expensive evidence of monetization.

Markets can tolerate enormous investment when revenue follows. They can tolerate falling free cash flow when management can show that today's capital produces tomorrow's earnings. What becomes harder to tolerate is a widening gap between the cost of incremental compute and the cash flows that compute produces.

The software selloff this year complicates that equation. The so-called SaaSpocalypse gave us one of the first broad examples of investors treating AI as a destroyer of existing software economics rather than simply another product category. Better models and AI coding tools raised uncomfortable questions about how much traditional SaaS functionality customers will continue paying for.

That is simultaneously evidence for AI and a risk to the AI trade. If AI can replace expensive software workflows, its economic value is real. But software companies are also major consumers of cloud infrastructure. Destroy enough incumbent software revenue before AI-native revenue replaces it, and some of the demand funding the infrastructure cycle begins to weaken.

Oracle shows the financing side of the same tension. Its infrastructure business is growing rapidly, but the build has also required extraordinary capital spending and substantial external financing. That does not mean Oracle is broken; it tells us where the burden of proof sits.

Heavy capex is simply an input. The failure condition is persistent investment that cannot generate returns above the cost of the capital funding it. Spending itself is not the problem. Spending that does not convert into durable economic output is.

3. Infrastructure Society Will Not Absorb

The physical AI trade has another counterparty that does not appear neatly in an earnings model: everyone living near it. In March 2026, Gallup found that 71% of Americans opposed construction of an AI data center in their local area, with nearly half strongly opposed.

That moves community resistance beyond a collection of colorful town-hall stories. Politicians are responding, and projects across the country are already being delayed or blocked over electricity costs, water use, land use and local infrastructure burdens.

Again, opposition alone is not enough to break the thesis. America has spent two centuries fighting over infrastructure. Railroads, highways, pipelines, transmission lines, factories and power plants all imposed local costs while producing benefits elsewhere.

What makes AI unusual is the concentration and speed of the electricity demand. Hyperscale data centers need enormous amounts of power, often in places where grid capacity is already tight, and they need it far faster than utilities normally build generation and transmission.

The obvious answer is to bring your own power, but every version of that answer comes with a constraint. Solar works, but hyperscale demand makes rooftop panels nearly irrelevant. Utility-scale arrays require enormous amounts of land, and generation disappears every evening unless storage or another source fills the gap.

Wind solves the land-use problem differently, but intermittency remains. The wind has retained its unfortunate habit of ignoring GPU utilization schedules.

Natural gas is dispatchable and familiar, but it can turn a data-center fight over electricity prices into a local fight over air pollution, emissions and fuel infrastructure.

Nuclear may ultimately fit data-center demand better than almost anything else because it provides dense, continuous power with a small physical footprint, but time becomes the constraint. New reactors and first-of-a-kind small modular designs require capital, licensing, construction and patience. A reactor that works beautifully in 2032 does little for a cluster waiting for power in 2027. Let alone that there is no permanent solution for nuclear waste and Three Mile Island, Chernobyl, and Fukushima still haunt the industry.

Moving generation behind the meter therefore does not eliminate the bottleneck. It changes its form: land and storage for solar, intermittency for wind, emissions and fuel exposure for gas, time and capital for nuclear.

Power is increasingly the constraint underneath the other constraints. The war in Iran has added another source of energy-price volatility to a system already absorbing rapidly growing electricity demand. Large hyperscalers can tolerate higher power costs for a long time. Marginal projects cannot.

The risk becomes structural when power costs, interconnection constraints and political restrictions remove enough economically viable sites that compute supply can no longer expand with demand. At that point, local resistance stops being a zoning problem and becomes a capacity ceiling.

4. Capital That Stops Funding the Build

Issue 002 of this series argued that the AI buildout had acquired the structure of a prisoner's dilemma. Microsoft cannot slow down because Google might not. Google cannot slow because Amazon might not. Meta cannot assume tomorrow's dominant models and platforms will belong to somebody else.

The risk of overspending is measurable. The risk of strategic irrelevance is much harder to tolerate, which is why the logic of continued investment still holds.

But there is another interpretation worth testing. What if nobody blinking first eventually stops looking like conviction and starts looking like a coordination failure?

Each company may be making a rational decision in isolation while the industry collectively builds more capacity than the economics can justify. That possibility matters more as financing structures become more creative.

Oracle has tapped debt and equity to accelerate infrastructure development. Neoclouds are building businesses around enormous capital requirements and long-duration compute contracts. CoreWeave has paired extraordinary demand and backlog with extraordinary capital needs. The capital structure of the trade is becoming as important as the demand story itself.

Now Wall Street is trying to turn compute into an investable asset class. Nvidia's August agreements with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR are designed to mobilize enormous amounts of third-party capital for AI infrastructure. The premise is straightforward: if hyperscalers and AI labs want more compute than their own balance sheets should finance, institutional capital can own the assets and collect the cash flows.

In theory, that solves a major constraint. In practice, the announcement is still closer to architecture than finished building.

The agreements are memorandums of understanding. The capital has not all been deployed. The structures depend on utilization, customer credit, financing terms and assumptions about what expensive Nvidia hardware will be worth several years from now. None of that makes the idea dubious, but it does make the market's reaction worth testing against eventual execution.

A financing mechanism is not the same thing as financed infrastructure. A residual-value guarantee is not the same thing as residual value. An MOU is not a commitment. It most certainly isn’t $500 billion of completed projects.

Debt, leasing, and asset-backed finance are not evidence of a bubble on their own. The concern rises when each new layer of capital requires more protection than the last: higher returns, stronger guarantees, more collateral, or more equity because operating cash flows can no longer justify the build.

That is when financing stops expanding the trade and begins rationing it. We could see capital concentrating towards companies that can continue to fund the build internally and away from those that need the leverage more. Rationing financing moves the trade towards credit quality and proven systems and away from exploration and balance-sheet dependent expansion

5. When More Compute Stops Buying Progress

There is one failure condition that would probably take longer than the others to recognize: more compute may eventually stop producing enough progress.

This is where the history of artificial intelligence becomes useful. AI has gone through winters before, periods when expectations outran technical progress, promised capabilities failed to arrive, funding contracted and enthusiasm disappeared faster than the underlying research.

The research did not stop. The money did, and that distinction matters for this trade.

An AI winter does not require scientists to run out of ideas. It requires progress to remain disappointing for long enough that customers and capital providers stop paying today for breakthroughs expected tomorrow.

That gives the scaling-wall argument an economic mechanism. Diminishing returns alone are not enough. Every mature technology faces them, and better hardware, improved algorithms, model efficiency and new architectures can continue lowering the cost of useful intelligence even if brute-force scaling becomes less productive.

The more serious risk is a sustained period in which neither more compute nor better methods produce enough additional capability to justify the next round of infrastructure spending. If that happens, the feedback loop reverses. Capability gains slow. Commercial expectations fall. Customers become less willing to pay for marginal compute. Financing becomes harder. New clusters get postponed. GPU orders slow. HBM demand weakens. Networking projects shrink. Power-equipment orders soften.

An AI winter would therefore move backward through the entire supply chain. Power shortages can delay the build. Financing can slow it. Community opposition can move it somewhere else. A genuine stagnation in economically useful AI progress attacks the reason for building in the first place.

There is not convincing evidence that we are there today, but the transformer revolution will not scale forever. No architecture does.

A technology paradigm rarely announces its ceiling. Instead, it gets progressively more expensive to extract smaller gains. The warning will probably arrive well before models stop improving: each generation will require materially more compute to produce materially less new capability. At first, better chips, algorithms and efficiency can hide that deterioration. Eventually they may not. If customers begin deciding that the next increment of intelligence is not worth the next increment of compute, infrastructure demand will follow.

That is how an AI winter would reach this trade. Research would not stop. Progress might not even stop. Capital would simply lose patience with the rate of progress. And all successful growth companies become value companies. Looking beyond the winter, history suggests the thaw would eventually come from the same place this spring did: a new idea that changes the economics again.

What Still Holds

A bear case is easy to write when the numbers get large. The harder job is distinguishing evidence of excess from evidence of scale.

Right now, the strongest parts of the original AI infrastructure thesis remain intact. Demand continues to outrun supply. Cloud backlogs remain enormous. AI revenue is growing rapidly. Meta is finding measurable value inside its existing advertising machine. Oracle and the neoclouds continue signing large customer commitments. Power, memory, networking and data-center capacity remain scarce enough to constrain delivery.

Most importantly, customers are still paying. That is the evidence a failure thesis has to overcome.

High capex, delayed projects, weaker free cash flow, community opposition and more complicated financing structures all deserve attention. None of them, in isolation, proves that the trade is failing. They become meaningful only when they start interfering with the conversion chain.

Demand has to become contracted business. Contracts have to become financed capacity. Capacity has to become compute. Compute has to become usage. Usage has to become revenue. Revenue eventually has to become durable returns.

Four issues into this series, nearly every link has been strained somewhere, and some have bent considerably. The trade breaks when one stops feeding the next.

Now we know where to look.

Jason Bartlett

Jason Bartlett is CEO and President of Veche, Inc, parent company to Avalanche Markets. He works extensively in U.S. energy market finance and economics and is a member of the board of Thinking About Thinking, Inc--a 501c3 research and convening organization dedicated to advancing ideas about intelligence.