The Capex Reckoning: Why Wall Street Is Right to Push Back on AI Infrastructure Spend
A trillion dollars in projected data center capex. A Microsoft shareholder lawsuit. And a growing gap between AI revenue and AI infrastructure outlay. We think the bulls have this backwards.
Here is the position that will get you laughed out of a lot of rooms right now: the hyperscalers are spending too much, too fast, on AI infrastructure, and the market is only starting to grasp what that means.
We know what that sounds like. It sounds like the people who said the internet was overbuilt in 1999, right before it was, catastrophically. It sounds like the guy at the party who keeps pointing at the punch bowl. Nobody wants to be that guy. The mood in this industry, for the last two years, has been something close to euphoric, and euphoria has a way of making skeptics look foolish right up until the moment they don't.
But here is what we're tracking, and why we think the consensus has gotten ahead of reality. A Microsoft shareholder has now filed a lawsuit challenging the company's AI and Azure spending decisions. Total projected AI capital expenditure across the ecosystem is expected to reach $5.5 trillion through 2030, with roughly $4.1 trillion of that financed through debt markets JPMorgan, figures that would have seemed science fiction three years ago. And the gap between what AI infrastructure costs and what AI revenue actually delivers remains wide, poorly understood, and almost entirely unacknowledged in the mainstream conversation about this buildout.
We are not saying the AI infrastructure boom is over. We are saying the math deserves a harder look than it is currently getting.
The Consensus View, Steelmanned Honestly
Before we dismantle the bull case, we should state it fairly, because it is genuinely compelling.
The argument for sustained hyperscaler capex at current or accelerating levels goes roughly like this: AI is a platform shift comparable to the internet or mobile. Platform shifts require infrastructure investment that precedes revenue by years. The companies that built cloud capacity ahead of demand in the 2010s were rewarded enormously. Microsoft, Google, Amazon, and Meta are all seeing AI driven product improvements that are already generating meaningful incremental revenue. The demand signals from enterprise customers are real. The race to maintain competitive positioning creates a situation where underinvesting is more dangerous than overinvesting. And the depreciation schedules on GPU and data center assets, while significant, are manageable for companies with the balance sheets these players carry.
This is a serious argument. It is not crazy. And the people making it are not stupid.
The problem is not that this argument is wrong in principle. The problem is that it is being used to wave away questions that deserve direct answers, and that the numbers underlying the bet are increasingly difficult to square.
The Numbers That Should Make You Uncomfortable
Let us start with what we actually know about the spending trajectory.
Microsoft alone has committed to spending roughly 80 billion dollars on AI enabled data centers in fiscal year 2025. That is not a projection from an analyst. That is a figure Microsoft itself published Microsoft, framing it as evidence of commitment to American AI leadership. More than half of that spend was designated for US facilities.
Google has similarly committed to 75 billion dollars in capital expenditure for 2025, a figure that surprised even optimistic analysts when it was disclosed on Alphabet's Q4 2024 earnings call Alphabet. Meta signaled full year 2025 capex in the range of 60 to 65 billion dollars, up from 37 billion in 2024. Amazon has not broken out its AI infrastructure number cleanly, but AWS capital expenditure has been running at a pace that puts the group total well north of 300 billion dollars for 2025 across just these four companies.
The top five hyperscalers collectively spent $256 billion on capex in 2024, a figure that rose to $443 billion in 2025 and is projected to hit $602 billion in 2026, a 36% year-over-year increase. CreditSights Then you add in the broader ecosystem: the data center developers, the co-location providers, the power infrastructure build, the cooling systems, the land acquisition. JPMorgan projects hyperscaler capex alone could peak above $1.1 trillion in 2027 JPMorgan, with more than $2.1 trillion of data center financing flowing through high-grade corporate bond markets through 2030. When you aggregate the full capital stack, the one trillion dollar figure that keeps appearing in analyst notes starts to look not like hyperbole but like a reasonable near-term milestone.
Top 5 Hyperscaler Capex: 2024–2026
Combined capital expenditure across the five largest hyperscalers
Source: CreditSights (2026). 2026 figure is a projection.
Now here is the part of the conversation that keeps getting skipped.
The Revenue Gap Nobody Wants to Talk About
Microsoft's AI revenue, across Copilot products, Azure AI services, and related offerings, is growing. That is true and worth acknowledging. But the growth rate, while impressive in percentage terms, is operating off a small base at a moment when the capital base is not small at all.
A Microsoft shareholder derivative lawsuit filed in early 2025 alleged that the company's board failed to adequately oversee AI related capital expenditures Bloomberg, arguing that the spending lacked sufficient return on investment analysis and exposed the company to material financial risk. Shareholder lawsuits are common and many go nowhere. But this one is worth taking seriously not because it will necessarily succeed in court, but because it articulates clearly the question that analysts have been dancing around: where, precisely, is the return?
The depreciation math alone is sobering. High end AI accelerators, primarily Nvidia H100 and H200 GPUs and the Blackwell generation following them, depreciate over three to five years on typical hyperscaler accounting schedules. A data center building has a longer useful life, but the compute inside it does not. Goldman Sachs estimates AI company capex consensus for 2026 has already been revised upward to $527 billion from $465 billion at the start of the Q3 2024 earnings season Goldman Sachs, a $62 billion upward revision in a matter of weeks, with no corresponding upward revision to near-term revenue forecasts. When you are spending at that pace on equipment that will be obsolete or substantially devalued within half a decade, you need a revenue engine that justifies that burn rate, and you need it to materialize on a timeline that aligns with the asset life.
The bull response is usually one of two things: either "just wait, the revenue will come," or "the hyperscalers have such strong core businesses that they can afford to be patient." Both of these are partially true and both of them are partially a way of avoiding the question. Being able to afford a bet is not the same as the bet being smart. And "just wait" is a position that works right up until the moment patience runs out, whether in a board room, an earnings call, or a shareholder meeting.
The Microsoft Lawsuit Is a Signal, Not Just a Lawsuit
We want to spend more time on the shareholder lawsuit because we think the industry is underweighting its significance.
The legal theory is straightforward: derivative lawsuits like this one argue that a company's board of directors failed in its fiduciary duty by approving spending that was not in shareholders' best interests. What makes this particular suit notable is the target. Microsoft is not a struggling company making desperate bets. It is one of the most valuable companies in human history, run by people with long track records of capital allocation discipline. If a shareholder can credibly argue that even Microsoft's board is not asking hard enough questions about AI capex returns, that is a statement about the entire industry's relationship with this spending cycle.
Think about what it takes for a shareholder lawsuit on capital allocation to gain any traction at all. Courts are extremely reluctant to second guess business judgment. The bar for demonstrating that a board acted improperly in making strategic investment decisions is high. For a plaintiff's attorney to file this suit and believe it has merit, they have to believe that the gap between spend and return is large enough, and the board's oversight thin enough, to clear that high bar.
We are not predicting this lawsuit wins. We are saying that its existence is a data point about where institutional sentiment is heading. The unspoken agreement that hyperscaler AI capex is beyond question is starting to crack.
What a Pullback Actually Looks Like, and Who Gets Hurt First
Here is where the contrarian view gets genuinely complicated, because a pullback in hyperscaler spending would not be uniform, and it would not hit everyone at the same time.
The first place to feel it would be the co-location and wholesale data center market. Companies like Equinix, Digital Realty, Iron Mountain, and a long list of private operators have made capacity commitments, land purchases, and power agreements based on demand projections that assume the current trajectory holds. Morgan Stanley estimates $3 trillion in AI-related infrastructure investment through 2028, including $2.9 trillion in global data center construction costs alone. Morgan Stanley If Microsoft or Google decided tomorrow to slow their commitment growth by twenty percent, the ripple through pre-leased capacity, construction pipelines, and power purchase agreements would be immediate and significant.
The second wave would hit the supply chain. Nvidia's revenue is the most obvious casualty of any pullback, and the market knows this, which is why Nvidia's stock has become one of the most watched sentiment indicators in tech. But the exposure runs deeper: cooling manufacturers, power distribution equipment makers, fiber infrastructure providers, and the utilities themselves have all made investments premised on continued growth in data center load. ARK's Big Ideas 2026 report projects data center system spending reaching $600 billion in 2026 and $1.4 trillion by 2030, compounding at 29% annually since the launch of ChatGPT. ARK Invest Utility companies in Virginia, Texas, and other high concentration markets have filed interconnection requests and capacity plans that assume demand keeps climbing at that pace. If it does not, ratepayers and utility shareholders are holding the bag.
The third and most subtle effect would be on the real estate and land market surrounding major data center corridors. We have watched land values in Northern Virginia, Phoenix, and Dallas appreciate sharply as operators and developers competed for sites with adequate power access. A meaningful slowdown in hyperscaler commitment would reprice those assets in ways that would affect everyone from pension funds to regional banks.
The Strongest Objection to Our View
The most serious pushback to everything we have written above is this: the demand is real, it is accelerating, and anyone who slows down now will simply hand market share to whoever does not.
This is not a trivial objection. The cloud wars of the 2010s demonstrated clearly that the penalty for underbuilding capacity is severe and hard to recover from. AWS's early lead in cloud infrastructure was partly a function of having built more than anyone thought was necessary, and that excess capacity became the foundation of a dominant market position. The hyperscalers know this history. They lived it. The executives making these capex decisions are not naive.
But here is the distinction we think matters: building ahead of demand in a market you are actively creating is different from building ahead of demand in a market you are waiting for someone else to monetize. In the cloud era, AWS was selling compute and storage directly, and every server rack they built had a clear path to revenue through a product they controlled and customers who understood what they were buying.
The AI infrastructure bet is more complex. Intersect360 Research projects the AI infrastructure market growing 60% in 2025 before reaching $520 billion by 2030 Intersect360 Research, impressive headline numbers that nonetheless obscure a significant portion of the current GPU buildout oriented toward training and inference workloads that are not yet clearly priced, not yet clearly owned by a specific customer, and not yet generating the kind of recurring revenue that justifies the capital stack. The hyperscalers are simultaneously building the infrastructure, developing the models, trying to sell the compute to enterprise customers, and competing with the very AI companies they are funding and partnering with. That is a lot of bets riding on the same stack of chips.
“Building ahead of demand in a market you are creating is different from building ahead of demand in a market you are waiting for someone else to monetize. The cloud era analogy only holds if the revenue model is as clear now as compute and storage were then. It is not.”
What Would Change Our Mind
We said at the outset that we try to be intellectually honest about this. So here is what would have to be true for us to abandon the contrarian position.
First, we would need to see AI revenue at the major hyperscalers start compounding at rates that close the gap with infrastructure spend within two to three years. Not "AI is contributing to growth." Specific, auditable AI revenue lines growing fast enough to make the depreciation math work. Microsoft has started breaking out Copilot commercial revenue more granularly, and that is a step in the right direction. But the numbers need to move substantially.
Second, we would need to see enterprise adoption of AI tools reach a maturity level where workloads are stable, predictable, and generating consistent utilization rates on the deployed infrastructure. Right now, a lot of the deployed capacity is underutilized relative to peak theoretical demand. Wells Fargo Investment Institute notes that AI and machine-learning deals represented more than 89% of all VC deal value in Q1 2026, against a total VC deal pool of $267 billion Wells Fargo Investment Institute, a concentration that signals enormous investor appetite but also an industry that is still overwhelmingly in the experimentation phase rather than the scaling-recurring-revenue phase. If enterprise AI workloads mature into something that looks like the cloud workload curve of 2016 to 2020, the bull case gets a lot stronger.
Third, we would need to see the power constraint problem in the US resolve itself in a way that actually enables the planned buildout. The irony is that one of the biggest risks to the infrastructure bet is not financial. It is physical. There is not enough grid capacity in the markets that matter most to absorb the planned load growth on the timelines the hyperscalers are projecting. If power becomes the binding constraint rather than capital, the spending levels become self-limiting regardless of balance sheet strength.
None of those conditions are impossible. Some of them may materialize. We are watching closely.
The Part Nobody Wants to Say Out Loud
The data center industry has had a very good run telling a very simple story: AI needs compute, compute needs data centers, therefore build. That story is not false. It is incomplete.
What it leaves out is that the relationship between infrastructure investment and financial return in AI is genuinely unproven at scale. The technology is real. The use cases are real. The demand for experimentation is real. What is not yet real, in a way that the capital committed can be justified against, is the monetization engine that makes a five-and-a-half trillion dollar infrastructure bet, JPMorgan's number through 2030, with over $2.1 trillion flowing through investment-grade corporate bond markets alone JPMorgan, look like a reasonable wager rather than an extraordinary one.
The Microsoft shareholder lawsuit will probably not change any of that on its own. But it represents something important: the beginning of a more serious institutional conversation about what, exactly, the hyperscalers owe their investors by way of accounting for where all this money is going and what it is supposed to come back as. That conversation has been overdue. It is starting now.
We could be wrong. The AI revenue curve could inflect sharply upward and make all of this look obvious in retrospect. But we think the more likely outcome over the next eighteen months is a period of increased scrutiny, slower commitment growth from at least one or two of the major players, and a repricing of risk in parts of the data center supply chain that have been priced as if the current trajectory is certain.
It is not certain. Nothing about this is certain. And the sooner the industry is willing to say that out loud, the better positioned everyone will be when the reckoning actually arrives.