Guide · Compute costs
How to budget AI compute without lying to yourself
Compute budgets fail when they treat a changing technical system as a fixed overhead line. The better approach is to map spend to workload, usage, and commitments, then forecast the product changes that are supposed to change unit cost.
By Serge Mochtchenkov, CFA · Fractional CFO for AI startups
Separate production from experimentation
Production inference, training, evaluation, development sandboxes, and internal tools behave differently. If they are all rolled into one cloud line, management cannot tell whether growth, experimentation, or inefficiency caused the variance.
Choose a driver for each material workload
API tokens, requests, GPU hours, jobs, documents, minutes, storage, or active customers may be the right driver depending on the product. Forecast volume and unit cost separately so management can see which side moved.
Model commitments explicitly
Reserved capacity, committed-spend discounts, GPU leases, and data contracts can reduce unit cost but create fixed cash obligations. The cash forecast should reflect payment terms and minimum commitments, not only accounting expense.
Add the technical roadmap
If engineering expects caching, routing, quantization, a smaller model, or provider migration to reduce unit cost, show the expected date and magnitude. Otherwise the finance model contradicts the engineering plan by assuming today's architecture forever.
Create variance ownership
Each material compute category should have an owner, budget, driver, and explanation when actual spend moves. The point is not to make engineers ask finance for permission to experiment; it is to make large cost movements visible quickly.
Stress-test fundraising
Run a downside case where usage grows faster than monetization, cost reductions arrive late, or provider pricing is worse than expected. If that scenario shortens runway materially, the financing plan should know before the market does.
Work with me
Need the model behind this decision? I build the financial system for AI-native startups preparing to raise, improve runway, or defend their unit economics. Book a free 30-minute intro call.
Related reading