For most of my career, the scarce input in building a company was capital. At Uber we raised $3.5bn because growth was a land grab and money bought territory. The constraint on what you could build was how much you could raise and how fast you could deploy it.
In AI, the scarce input has changed. It's compute — the chips, the data centres, and the power to run them. Capital still matters, but increasingly it's just the thing you convert into compute. And unlike capital, compute can't always be bought at any price on the timeline you need it.
Why the demand keeps compounding
The industry assumed demand for compute would plateau as models matured. The opposite is happening. Training frontier models takes more each generation. But the bigger shift is inference: every agent that runs a workflow, every model that reasons before it answers, every product that moves from a chat box into a company's daily operations consumes compute continuously, not once. Usage that compounds is exactly what founders want — and it means compute demand compounds with it.
What it changes for founders
Compute is cost of goods, not overhead. In software, the marginal cost of a new customer was close to zero. In AI-first companies it isn't — every user carries an inference bill. The founders I back know their cost per task, per customer, and per dollar of revenue, and they know which way each is trending. Gross margin is a design decision, made in architecture, model choice, and pricing — not something to fix after Series B.
Efficiency is a moat. Two companies can ship the same product with very different economics. The one that routes simple work to smaller models, caches aggressively, and reserves expensive compute for the tasks that need it can price lower and still grow margin. That advantage compounds, and it's hard to copy quickly — which is what a moat is.
Fundraising has a new question. Investors used to ask how you'd spend the money. Now they ask how much of it becomes compute, and what that compute buys. The strongest pitches I've seen this year answer that directly: how capital converts to capacity, capacity to usage, and usage to revenue with improving unit economics. A clear compute plan has become part of a fundable story.
Compute is also geography
International expansion used to be about regulators, partners, and local teams. For AI companies it's increasingly about where compute sits. Data residency, sovereign AI programmes, energy prices, and grid capacity now shape which markets you can serve well and on what terms. The question "where are we launching next?" now comes with a second one: "where will that market's workloads run?" Founders who treat infrastructure location as a strategic decision, not a procurement detail, will expand faster and with fewer surprises.
What I look for
I'm not looking for founders who've solved compute — almost nobody has. I'm looking for founders who see it clearly: who treat it as a core input to the business, design for efficiency from the first architecture decision, and can explain how their margins improve as they scale. The platform shift is real. But like every platform shift I've seen, the winners won't be the ones who spend the most. They'll be the ones who turn a scarce input into the most value.