Cheap AI Compute Won't Stay Cheap

Today’s AI compute pricing is not a stable baseline, it’s a promotional phase. Any product plan that assumes it holds forever is underwriting a risk it hasn’t actually priced.

Key takeaways

• Data centre and power buildout stories are a preview of a coming cost curve
• Current AI compute pricing reflects competition for market share, not a stable floor
• Products with thin margins on AI cost are exposed to a shock they haven’t modelled
• Founders should stress test unit economics at 2 to 3 times current compute cost
• This is a finance exercise, not a technical one, and it belongs in the board pack

Between building Obby, exiting Baluu, and now running FirstMotion, I’ve watched cost structures shift underneath businesses more than once, usually at the worst possible moment. Through VAQA’s finance and fundraising advisory work, I now spend a lot of time with founders whose entire product margin sits on an AI API bill they haven’t stress tested.

This piece sets out why I think compute pricing is heading up, not staying flat, and exactly how to model that risk before it shows up in your board pack.

Why I don’t believe today’s pricing is the real price

The AI infrastructure build out happening right now is enormous. Reports on data centre construction, power procurement and chip supply from hyperscalers and labs alike point in one direction: providers are spending aggressively to secure capacity ahead of demand, and a lot of current pricing looks like it’s being subsidised by competition for market share rather than reflecting the true cost of delivery.

That’s a familiar pattern to anyone who’s watched cloud computing, ride sharing or food delivery mature. Prices start low to win adoption, then rise once the market consolidates and providers need the unit economics to actually work. I don’t think AI compute will be the exception.

OpenAI, Anthropic and the major cloud providers are all racing to build out capacity, and that race costs real money. Power procurement alone has become a genuine constraint in several regions, with data centre operators competing directly with utilities and even other industries for grid access. None of that cost disappears; it eventually gets reflected somewhere in pricing.

The businesses most exposed right now

Not every company needs to worry about this equally. The exposure is concentrated in a specific pattern.

• Products where AI inference cost is a meaningful share of cost of goods sold
• Pricing models built around unlimited or high volume AI usage per seat
• Margins that already assume today’s per token or per call pricing as a constant
• Roadmaps that lean further into AI heavy features without re-pricing the plan

If any of those describe your business, a 2 to 3 times increase in compute cost isn’t a hypothetical stress test, it’s a plausible near term scenario that could meaningfully change whether the unit economics still work.

How I’d actually model this

This isn’t a complicated exercise, and that’s exactly why I think it’s inexcusable for founders to skip it. Here’s the version I walk through with clients.

1. Pull your current AI compute spend as a percentage of revenue for your top three product tiers
2. Recalculate gross margin at 2 times that compute cost, holding pricing flat
3. Recalculate again at 3 times, holding pricing flat
4. Identify the point at which any tier becomes structurally unprofitable
5. Decide now, not later, whether that triggers a pricing change, a usage cap or a product redesign

Doing this once isn’t enough. I’d suggest rerunning it every time you refresh your financial model, because the compute cost curve is genuinely moving and a stale assumption compounds quietly.

What this means for a fundraise

Investors who’ve been through a few cycles are starting to ask about this directly, and founders who haven’t modelled it tend to answer badly. A founder who can say “here’s our margin at current pricing, and here’s our margin at 3 times pricing, and here’s what we’d do about it” reads as a far stronger operator than one who hasn’t thought about it at all.

This connects directly to how I think about valuation discipline more broadly: a business whose margin structure quietly assumes a subsidy it doesn’t control is not as strong as its current numbers suggest. Investors doing real diligence will find that gap eventually, so it’s better to have already closed it.

The counterargument, and why I don’t buy it

Some founders will point out that AI compute costs have actually fallen over time on a per unit basis, as models get more efficient and competition intensifies between providers. That’s true, and it’s a fair point.

But efficiency gains and pricing power are two different forces pulling in opposite directions. Even if the cost to deliver a unit of intelligence keeps falling, providers still need to recoup enormous infrastructure investment, and pricing doesn’t have to track cost cleanly in a market this concentrated. I’d rather plan for the possibility that pricing power wins that tug of war for a while, even if efficiency eventually wins longer term.

Building a margin that survives the correction

I’m not predicting a specific price increase on a specific date. I’m saying that any plan built on the assumption that today’s AI pricing is permanent is making a bet it hasn’t acknowledged making.

The fix isn’t complicated. Run the 2 to 3 times stress test, know your breakeven point, and decide in advance how you’ll respond if it happens. That’s a much stronger position than discovering the answer during a live pricing shock with customers already on the platform.

This kind of margin stress testing sits at the intersection of Finance and Fundraising and Operational Efficiency, which is exactly where I spend most of my advisory time at VAQA. If your product economics haven’t been pressure tested against a higher compute cost scenario, get in touch and I’ll walk through the model with you directly.

Frequently Asked Questions

How likely is a 2 to 3 times increase in AI compute costs?

I can’t give you a precise probability, and neither can anyone else with confidence. What I can say is that the current pricing environment looks subsidised by competition for share rather than reflecting full delivery cost, which makes a meaningful increase a reasonable planning scenario rather than an edge case.

Should I renegotiate my AI vendor contracts now?

Locking in longer term pricing where a vendor offers it is worth exploring, but don’t treat a locked rate as a permanent solution. Model what happens at renewal too, since most enterprise AI contracts run on shorter cycles than the infrastructure investment behind them.

Does this apply to companies not using AI heavily yet?

If your roadmap includes more AI features later, it’s worth modelling before you build the dependency, not after. It’s far easier to design pricing around a realistic cost curve from the start than to retrofit it once customers expect a feature at a fixed price.

How does VAQA help with this kind of financial modelling?

Through the Finance and Fundraising advisory pillar, I work directly with founders to build stress tested unit economics models, including AI cost scenarios, ahead of board meetings or fundraising conversations. It’s practical modelling work, not theoretical advice.

Is this the same kind of work FirstMotion does?

No, FirstMotion is focused on AI Search and GEO strategy for B2B SaaS companies, helping them get found and cited by AI systems. The compute cost and margin modelling work is VAQA advisory work, so if that’s what you need, VAQA is the right place to start.

Tom Batting is a Forbes 30 Under 30 entrepreneur, founder of Obby and Baluu, and founder of FirstMotion. He advises founders and leadership teams through VAQA on board advisory, growth, go-to-market, AI implementation, operational efficiency, and finance and fundraising.