Alp Notes

The CapEx Trap: Capital Efficiency in the AI Era (Part 2)

In Part 1, I argued that AI breaks the traditional SaaS finance model by flipping a core assumption: infrastructure now comes before revenue, not after.

Once you see this shift, the main question is:

If capital is committed upfront, where does the cash actually go — and who ends up carrying the risk?

To answer that, you have to follow the cash.

Today, in the AI era, money leaves the system faster than value is created. This causes a mismatch: hyperscalers take on risk, suppliers get paid early, and standard profit metrics don’t show the real picture.

Cash Flow Tells the Story Earlier Than Earnings

AI companies might seem profitable in reports even when they are not truly healthy yet.

Why? Because the largest costs hit cash first, not earnings.

In SaaS, costs scaled with customers. In AI, capital is committed first to create future demand.

You can see this clearly in Microsoft’s recent results. In Q1 FY26, Microsoft reported ~24% year-over-year growth in operating income, driven by cloud and AI demand. At the same time, it disclosed record quarterly CapEx of ~$34.9B, largely tied to AI infrastructure.

The takeaway is simple:

In the AI era, profitability often shows up on the income statement long before it shows up in cash.

The "Depreciation Shield" Misleads Investors

This cash reality is blurred by accounting.

Recently, several major tech companies extended the useful life of their servers (e.g., from 4 to 6 years).

This creates a risky gap. Accounting says assets last longer, but technology changes faster. The takeaway: when asset lives are stretched, and innovation is fast, free cash flow is the most reliable measure.

Risk and Value Are Misaligned

Looking at cash flow shows that risk and reward are found in different parts of the business:

  1. Suppliers (NVIDIA) monetize early: They capture cash upfront with high margins and zero downstream utilization risk.
  2. Hyperscalers (Microsoft/Google) absorb risk: They commit capital and own the depreciating assets, depending on future pricing power to earn a return.
  3. Applications lag: They must prove willingness-to-pay to justify the infrastructure already built upstream.

As a result, value is taken early in the process, before the risks are fully known later on.

Defensive vs. Offensive CapEx

Finally, not all cash outlays are equal. We must distinguish between:

The hard truth is that most current spending is defensive. It keeps companies in place but lowers their return on invested capital.

The New Finance Playbook

When companies invest in infrastructure before earning revenue, their financial strategies need to change:

  1. From Margin to Capital Efficiency: Blended margins matter less; ROIC and Payback Period matter more.
  2. From Static to Unit Economics: Track "Cost per Token" and "Revenue per Workflow" rather than just ARR.
  3. From Confidence to Discipline: In SaaS, scale fixed bad economics. In AI, scale magnifies them.

Closing Thought

The current AI expansion is the biggest capital cycle in tech history. But winning won’t depend on who bought the most GPUs first.

Success will come to those who spent their capital wisely, not just those who tried to grow the fastest.

In a world where intelligence is everywhere, the real limit isn’t computing power. It’s how efficiently companies use their capital.

— Yiğitalp Y.

References

  1. Sequoia Capital — AI’s $600B Question
  2. Goldman Sachs — AI: In a Bubble?
  3. The Cube Research — Resetting GPU Depreciation
  4. Neuberger Berman — AI Capex: It’s Not All or Nothing
  5. Microsoft Q1 FY26 Earnings Release
  6. Meta Platforms FY2024 Earnings Release