There’s a story doing the rounds in startup land right now. AI inference costs are high today, but they’re falling fast – so build through the difficult early period, lock in your customers, and margin expansion will follow almost automatically as the cost curve drops.

It’s a reasonable bet. It’s not a guaranteed one. And most founders aren’t modelling what happens if it doesn’t play out the way they’re expecting.

In this article, we look at three risks sitting inside that assumption – vendor pricing strategy, switching costs, and energy – and why they belong on your risk register rather than your list of background assumptions.

The subsidy problem

The companies setting AI prices right now – OpenAI, Anthropic, Google – are not pricing to make money. They’re pricing to capture market share, and most are burning extraordinary capital to do it.

That’s a subsidy, and subsidies are worth taking. But unlike cloud infrastructure, which commoditised because anyone with capital could build a data centre, frontier AI has genuine structural moats: the talent to train models, proprietary data, safety infrastructure. This is not a market that structurally prices toward zero.

At some point, investors ask hard questions about profitability, and the pricing calculus changes. If your gross margin only works at today’s subsidised prices, that’s a risk worth naming.

The switching cost trap

There’s a second layer most founders miss, and it’s more immediate than the long-run cost curve question.

Once you’ve built on a specific model’s API – the prompt engineering, the fine-tuning, the workflows baked into your product – switching is genuinely painful. The AI companies know this. The playbook is as old as enterprise software: price generously to drive adoption, reprice once dependency is established.

If you’re building on top of a single provider’s API without a clear view of what migration would cost you in time and product rework, you’re exposed. It’s worth thinking about now, before the dependency is locked in.

The risk nobody’s modelling: energy

This is the one that gets the least attention, and it sits entirely outside the AI companies’ control.

The conventional view is that compute efficiency keeps improving, model sizes get leaner, and inference costs follow them down. That may well happen. But it assumes energy costs stay broadly stable – and that assumption is doing a lot of work that very few founders are examining.

Data centres running frontier AI models consume extraordinary amounts of power. That demand is already running into real grid constraints in major markets. New data centre capacity is being delayed not by a shortage of capital or hardware, but by a shortage of electricity. That’s a physical infrastructure problem, and it doesn’t resolve quickly.

Layer geopolitics on top of that. Energy prices aren’t set by markets alone – they’re exposed to supply decisions, sanctions, conflict and policy choices made by governments with priorities that have nothing to do with your token costs. A cost model that assumes stable energy prices is making a geopolitical assumption dressed up as a financial one.

And then there’s the regulatory pressure building in the background. Scrutiny of AI’s energy footprint is growing. Carbon costs, reporting obligations and operational constraints on data centres are already moving through policy pipelines in multiple jurisdictions. Any of them feeds back into the cost of running the models your product depends on.

What this means for your financial model

The point isn’t that costs will definitely rise. The point is that there are now several independent variables – commercial strategy, switching costs, energy, geopolitics, regulation – any one of which could move against the founding assumption. Most founders are modelling none of them.

That’s the bit that matters from a financial planning perspective. Model your gross margin at today’s AI costs. Model it with costs flat for three years. Model it with costs that rise modestly. If only one of those scenarios produces a viable business, that belongs on your risk register – and it’s something any serious investor is going to ask about.

None of this means don’t build. It means build with clear eyes about the bets you’re making, and make sure your numbers tell an honest story across a range of scenarios.

If AI prices don’t fall the way you’re expecting – is your business still a business?

Get your financial model stress-tested

Disclaimer: This article is for general informational purposes only and does not constitute financial, legal or tax advice. Please speak with a qualified adviser (hey, that’s us!) before making decisions based on your specific circumstances.