The Moving Finish Line of AI CapEx

Programming note: ARPU will return next Friday to look at enterprise AI.

Show Me the Money

Earnings season occasionally produces a perfect A/B test in market psychology. This past week, two tech giants delivered similarly stellar results—and Wall Street reacted as though one had discovered a new profit engine and the other had set a large pile of cash on fire.

First, Alphabet reported its Q2 numbers. Google crushed expectations across the board, with Cloud revenue jumping 82% year-over-year. But then management dropped the detail Wall Street was actually waiting for: Alphabet was raising its 2026 capital expenditure forecast to as much as $205 billion. Spending that much real, physical cash pushed Alphabet's quarterly free cash flow into negative territory, and investors promptly wiped 6% off the stock.

A few days later, Microsoft reported its numbers. Azure revenue surged 43%, pushing the cloud business past $100 billion in annual revenue, and the stock jumped 15%. Microsoft avoided Alphabet's fate by keeping its free cash flow firmly in the green. Its underlying investment plans remained largely unchanged, although an accounting change lowered its reported CapEx forecast to a more polite $175 billion.

The contrast revealed Wall Street's new rulebook. Spending enormous sums on AI is still fine. The only requirement is that the money quickly reappear somewhere else as revenue and cash flow.

This tension now sits at the heart of every debate about the AI build-out. To anchor those debates, Goldman Sachs recently mapped out the infrastructure implied by current deployment trends. Its baseline model estimates that the required compute, data centers and power infrastructure through 2031 will cost $7.6 trillion.

Naturally, the next question is whether AI can generate enough revenue to cover that bill. But the comparison is not so simple, because the bill itself is still moving—and depends on a few variables that often get buried in headline news.

The Replacement Cadence

Start with the most important variable in Goldman's estimate: the economic useful life of AI silicon.

When it comes to AI data centers, buildings last 20 years and power grids last 25, but AI chips turn over on much faster cycles. Because silicon accounts for the lion's share of AI capital expenditure, the replacement cadence of processors is the most influential lever in the entire build-out equation.

The core issue is a mismatch between accounting schedules and operational reality. Most tech giants depreciate chips over five to six years. But AI accelerators face rapid economic obsolescence. Nvidia's relentless annual release cadence means each new generation delivers massive leaps in performance per watt. Inside a power-constrained data center, running legacy hardware carries a heavy opportunity cost: you are occupying valuable rack space and electricity with chips that yield far less compute than newer alternatives.

To see how this changes the economics, consider a single $50,000 accelerator:

• Accounting Assumption (5-Year Life): The company records a $10,000 annual depreciation expense.
• Operational Reality (3-Year Life): If rapid obsolescence forces a replacement after three years, that annual cost recovery burden leaps to $16,667.
• Fleet Scale: Across a cluster of 100,000 processors, that $6,667-per-chip difference raises the annual capital recovery burden by roughly $667 million.

In practice, however, obsolescence does not necessarily mean retirement. Strong secondary rental prices for older Nvidia A100s suggest that trailing-edge chips can continue generating meaningful value long after their initial deployment. For now, at least, the economic life of AI hardware may be proving longer than the release cycle alone would suggest.

The Physics of the Grid

The second variable in the build-out equation is the physical reality of deployment.

Next-generation AI facilities do not just cost more—with construction costs rising from a historical average of roughly $10 million per megawatt to between $15 million and $20 million—they also move on a completely different clock than the chips inside them.

Accelerators are manufactured in months; power grids take years. U.S. energy projects currently face median grid interconnection queues exceeding 5 years, while high-voltage power transformers carry 3-year lead times.

Microsoft CEO Satya Nadella summed up this bottleneck plainly in an interview last year, noting that while Microsoft could obtain the GPUs it needed, it lacked "warm shells to plug into."

This mismatch creates a massive financial drag. When chips arrive before the building or the power connection is ready, capital sits idle. A $50,000 GPU sitting in a warehouse box earns zero revenue, yet the clock on its 5-year economic useful life ticks down anyway. By the time electricity finally reaches the facility, deferred revenues and accumulating interest have already degraded the original investment profile.

In other words, the physical bottleneck does not merely delay the build-out; it raises the amount of capital required to complete it.

The Other Half of the Equation

Which brings us back to the real test facing the AI build-out.

Investors are right to ask whether AI can generate enough revenue to justify trillions of dollars of infrastructure spending. But revenue is only half of the return equation. The other half is capital velocity: how quickly purchased equipment can be installed, energized, utilized and paid back before a more efficient generation arrives.

That makes the $7.6 trillion figure less a forecast than a moving range. If older chips can be cascaded into less demanding workloads, facilities arrive on schedule and new capacity is quickly filled, the industry may extract far more value from each dollar invested. If silicon turns over faster than expected, grid delays leave equipment idle and developers resort to costly power workarounds, the ultimate bill rises even before demand disappoints.

The AI investment cycle therefore does not fail only when customers refuse to pay. It can also fail through timing—when the next round of capital arrives before the last one has been converted into productive, revenue-generating capacity.

Signal Stack

The operating reality beneath the headlines.

• Buying the AI Infrastructure Dip (Morgan Stanley) – MS argues that cheaper, more efficient Chinese open-weight models are not a threat to compute demand but a reinforcement of it—efficiency gains lower the cost of AI, which in turn increases total usage under Jevon's Paradox, leading the firm to conclude that "demand for compute is likely to vastly exceed supply" regardless of which country's labs get more efficient first.
• What You Need to Know About AI Energy Use (Epoch AI) – The largest US power plant generates roughly 7 gigawatts, but training frontier models could require 10 gigawatts by the end of the decade—meaning no single power plant can meet a frontier training run's demand, forcing AI companies toward decentralized training across multiple fiber-connected data centers instead.

📺 On Our Channel

Who Is Actually Paying for AI?

Goldman Sachs projects AI infrastructure spending will hit $765 billion this year alone—more than the inflation-adjusted cost of the US Interstate Highway System built over four decades. Behind that number sits one unresolved question: whether consumers or enterprises will ever pay enough to justify it. With 95 percent of enterprise AI pilots failing to show measurable P&L impact, and hyperscalers pushing capex to 94 percent of operating cash flow, we broke down the buildout across 6 layers—from the physical bottlenecks to the funding escalation to who ultimately bears the risk.

Watch ARPU's deep dive on YouTube (36 Mins)

You received this message because you are subscribed to ARPU newsletter. If a friend forwarded you this message, sign up here to get it in your inbox.