Big Tech AI Capex Surge|the accounting gap nobody prices in

The Paradox at the Center

Three of the largest companies on earth are simultaneously accelerating capital spending at a pace that has no historical precedent in the technology industry. Amazon, Alphabet, and Microsoft are each committing hundreds of billions of dollars to AI infrastructure — and the market, broadly speaking, has rewarded that ambition. The consensus reads this as a race for dominance, where the winner captures the next platform shift. That framing is not wrong. But it is dangerously incomplete.

Here is the tension that deserves more attention. AI capex industry-wide has surged from $250 billion in 2024 to an estimated $650 billion this year — a figure that now equals roughly two percent of U.S. GDP. That is not a rounding error. That is a structural reallocation of capital at a scale comparable to the buildout of the interstate highway system. And yet the financial models most investors are using to evaluate these companies treat that spending as traditional capital investment. It is not.

The conventional assumption is that when a company builds a data center or purchases servers, it is acquiring an asset that generates returns over a decade or more. Steel mills and rail lines depreciated over forty to forty-five years. The industrial logic was simple: build once, operate for generations. The AI infrastructure buildout is operating on entirely different physics.

The Three-Year Clock

According to Research Affiliates CEO Chris Brightman, whose firm oversees approximately $200 billion in investment strategies, AI hardware loses its economic value in roughly three years — not the five or six years that hyperscalers are reporting on their income statements. That gap between accounting life and economic life is not a footnote. It is the core of the margin problem.

Brightman's analysis of Nvidia's H100 GPU illustrates the trajectory precisely. In its second year of operation, a single H100 generated $36,000 in annual profit — a 137 percent return on investment. By year four, the same chip was losing more than $4,400 annually, a negative return of 34 percent. The physical hardware had not failed. The chip could still run. But newer generations from Nvidia and AMD were delivering dramatically more computing power per watt, and since hyperscalers face hard energy constraints, they were forced to swap out functional equipment for more efficient alternatives. The old hardware became economically stranded long before it wore out.

This is the structural feature that most market commentary glosses over. These companies are not building railroads. They are stocking shelves. Brightman's analogy is precise: the hyperscalers now operate more like supermarkets than traditional industrial enterprises, except their inventory is GPU clusters rather than produce. The turnover cycle is relentless, and unlike a supermarket, they are reportedly selling the resulting AI products — large language models, vector databases, cloud inference — at a loss.

The accounting treatment amplifies the distortion. Microsoft, Amazon, and Alphabet each book depreciation on this hardware over five to six years. But if economic obsolescence arrives in three, then reported earnings are being overstated relative to true economic earnings during the high-spend phase. Margins look better on paper than they are in practice. This is not fraud — it is a structural mismatch between accounting convention and technological reality, and it matters enormously for how these stocks should be valued.

The Reversal — Why They Cannot Stop

Here is where the analysis needs to turn. The natural response to the capex trap argument is that rational executives would simply slow spending once returns deteriorated. That is the wrong model for what is actually happening.

Each of the three hyperscalers is locked into what amounts to a coordination failure. If Microsoft slows AI infrastructure investment while Google accelerates, Microsoft risks losing enterprise cloud contracts to Google Cloud, because customers increasingly select cloud providers based on AI capability. The same logic runs in reverse. None of the three can defect from the spending race without ceding competitive position to the others. This is not irrational exuberance — it is a Nash equilibrium where the dominant strategy for each player, taken individually, produces an outcome that is collectively expensive for all of them.

The Anthropic angle illustrates this dynamic from a different direction. Claude Opus 4.7 launched this week, distributed immediately across Microsoft Azure, Google Cloud, and Amazon Web Services at no additional cost to end users. All three hyperscalers benefit from offering cutting-edge model access, which drives AI credit consumption on their platforms. None of them owns the model outright. They are paying for the distribution infrastructure that delivers a competitor's product, because the alternative — not offering it — would make their cloud platforms less attractive. The infrastructure spending and the model ecosystem are entangled in ways that make unilateral restraint strategically dangerous.

Notably, Anthropic's approach with Opus 4.7 also signals where competitive differentiation may actually emerge. The model deliberately limits its own cybersecurity capabilities — what Anthropic calls differential reduction. This is not a technical limitation. It is a calculated positioning move targeting enterprise clients who face regulatory scrutiny and liability risk. The implication for the hyperscalers hosting it is subtle but important: the enterprise AI market may ultimately be won not by the most powerful models, but by the most governable ones. That shifts the competitive axis in a direction that favors platforms with strong compliance infrastructure.

Scenario Branching — Two Paths from Here

The evidence points toward continued margin compression for all three companies in the near term, but only if AI monetization fails to scale faster than hardware replacement costs. That condition is genuinely uncertain, and both directions deserve honest treatment.

The downside path is mechanically straightforward. If AI product revenues — cloud inference, enterprise software subscriptions, advertising uplift — do not grow fast enough to outrun the three-year hardware obsolescence cycle, then reported earnings will increasingly overstate true economic returns. Analyst price target reductions are already appearing. Robert W. Baird recently cut its Microsoft target from $540 to $500, maintaining an outperform rating but acknowledging the pressure. The stock's 200-day moving average sits at $455, well above the current trading range near $393 — a structural gap that reflects exactly this uncertainty about whether current spending levels are justified.

The recovery path, however, is grounded in a different structural argument. The hyperscalers are not passive hardware buyers. They are increasingly designing their own silicon. Amazon's Trainium chips, Google's TPUs, and Microsoft's Maia accelerators are all attempts to escape the Nvidia replacement cycle by controlling the hardware roadmap internally. If custom silicon achieves the compute-per-watt efficiency gains that currently force GPU swaps, the economic life of AI infrastructure extends significantly. The three-year clock slows. The accounting gap narrows. And the capex that currently looks like a margin trap becomes a genuine long-duration asset.

The timing of that transition is the variable that matters most. Evidence leans toward the custom silicon path gaining traction over a two-to-three year horizon — but only if the energy constraint remains the binding limitation rather than software or model architecture breakthroughs that require entirely new hardware paradigms. If the constraint shifts, the calculus resets again.

Microsoft's earnings trajectory offers a near-term read on where the balance currently sits. Revenue grew 16.7 percent year over year last quarter, with earnings per share beating consensus by a meaningful margin. That is not the profile of a company being crushed by its capex. But the 200-day moving average gap and the analyst target compression suggest the market is beginning to discount a future where the beat-and-raise pattern becomes harder to sustain. The next two earnings cycles will be more diagnostic than the last four.

The single most important number to watch is not revenue growth. It is the ratio of AI revenue to AI capex — a figure none of the three companies currently discloses in a form that makes direct comparison possible. Until that transparency arrives, the gap between the accounting story and the economic reality will remain the central unresolved question for all three stocks.