Alp Notes

The CapEx Trap: Why AI Breaks the Old Rules of SaaS Finance (Part 1)

Over the past year, I’ve noticed something odd: most AI discussions barely mention depreciation or CapEx planning, even though AI has sparked the biggest capital spending cycle in recent corporate history, with yearly spending expected to top $300 billion.

This gap caught my attention because CapEx discipline used to be very important.

Early in my career, while working as a management consultant in San Francisco around 2017, I had the opportunity to work on strategic finance and FP&A processes for both leading SaaS companies and large hardware providers, often companies that were customers of each other. In practice, that meant seeing the same CapEx decisions from both sides of the table: software companies consuming infrastructure, and platforms and hardware vendors supplying it.

Back then, SaaS was booming and cloud-native software felt as transformative as AI does now. The financial logic was clear: SaaS prioritized flexibility and variable costs, infrastructure providers invested in durable assets, and the system functioned smoothly. That perspective shapes how I view today’s very different AI cycle.

The SaaS CapEx Model (In Simple Terms)

For most of the last decade, CapEx planning in software was intentionally boring, and that was actually a good thing.

In the classic SaaS model, infrastructure followed demand.

You acquired customers. You monitored usage. You added capacity incrementally.

If demand slowed, you scaled back. If demand accelerated, you expanded.

Cloud flexibility made this possible and easy to manage.

From a finance point of view, the logic was straightforward:

Gross margins grew as fixed costs were covered. Depreciation schedules were mostly routine.

This model influenced how many of us learned to approach software finance.

AI changes this model entirely.

The Break: When Infrastructure Comes First

The main change in AI economics is simple but significant:

Now, infrastructure comes before revenue, and that’s intentional.

In SaaS, capacity was added to serve existing or clearly forecasted customers. In AI, capacity is set up before demand is clear, and sometimes even before the use case is fully worked out.

You can’t just spin up AI infrastructure the way you would with a cloud instance.

Training and running advanced models means committing to:

Once you’ve made these investments, it’s hard to reverse them.

This completely reverses the usual SaaS logic.

CapEx is no longer something you do after earning revenue. Now, it’s a strategic bet that revenue will come later.

Or more simply:

We are no longer buying capacity to serve customers.

We are buying capacity to search for customers.

Why This Matters

That one change: putting infrastructure first and revenue later is where the usual SaaS finance thinking starts to break down.

What was once an optimization problem is now a risk management problem. What used to be elastic becomes committed. And what used to quietly compound margins now tests balance sheets.

This doesn’t mean AI investment is wrong. It means the financial strategies that worked in the SaaS era no longer fit as well.

In the next part, I’ll dig deeper into why this shift matters, especially when you consider asset lifetimes, depreciation, and cash flow. I’ll also explain who takes on the risk and who benefits in the AI CapEx cycle.

— Yiğitalp Y.

References

  1. Sequoia Capital — AI’s $600B Question
  2. Goldman Sachs — AI: In a Bubble?
  3. Stanley Laman — Why GPU Useful Life Is the Most Misunderstood Variable in AI Economics