Your AI Budget Is Already Wrong. Here's How to Fix It.

If your team set an AI tools budget in Q4, the number is already wrong — not because prices went up, but because the prices on the invoice are not what it costs to run those tools.

The subsidy that will end

In May 2026, Simon Willison — who documents his AI tool use publicly — disclosed that his $200/month Claude subscription generates $2,180 in API compute. A 10x subsidy, paid by Anthropic to build market share. In the same month, Anthropic reported a $47B annual revenue run-rate, five times what it was at the end of 2025, with one enterprise client spending $500M a month. Uber exhausted its entire annual AI budget in months — not because anyone overused a toy, but because teams integrated AI deeply into real workflows and the volume caught finance off guard.

These numbers describe a market in an unsustainable pricing phase. Consumer AI subscriptions are priced far below cost. That pricing exists to accelerate adoption. When the subsidy narrows — and the revenue trajectories show it will — the tools your team uses for $20–100/month will cost more, switch to usage-based billing, or both.

The planning question that matters

There's one calculation worth running before your next planning cycle: what do your current AI tools cost at invoice price, and what would they cost if providers moved to cost-recovery pricing?

You are not hedging against a risk. You are measuring how much of your current workflow depends on a price that does not represent real cost. That gap — between what you pay now and what the work actually requires to run — is the number worth knowing before you make any AI-dependent decisions about headcount, workflow design, or procurement.

Priya: the marketing budget problem

Priya is head of marketing at a 55-person professional services firm. Five team members use Claude Pro at $20/month; two use Microsoft Copilot at $30/month. Total: $160/month. She budgeted $2,000 for the year and assumed that was conservative.

Using Willison's ratio as a benchmark — $200/month subscription generating $2,180 in compute — the same seven people doing the same work could cost $2,000–4,000/month at cost-recovery pricing, or $24,000–48,000 annually.

The firm does not need to allocate that now. But it changes what Priya should be building into her team's AI decisions: which use cases deliver enough value to survive professional pricing, and which are worth doing only because the cost is currently close to zero.

Omar: the invisible spend problem

Omar is an operations lead at a 300-person software company. Customer success, product, and ops each use AI tools independently, with no central visibility into aggregate spend. His rough count: about $1,200/month across subscriptions, plus API integrations the ops team built six months ago that bill on usage rather than flat rate.

He does not know what each team uses AI for. He does not know which uses would survive a 5x cost increase. And he does not know whether any of the API integrations have usage exposure that doesn't show up in monthly subscription invoices.

The Uber pattern is not recklessness. It is building AI deeply into real operations without modelling what "deeply" means financially when volume scales.

The number worth running this week

The window where AI tools are priced below cost is a useful one. You can build workflows, test integrations, and develop habits while the economics are favourable. The risk is building dependencies on tools that become unaffordable — not because prices suddenly spike, but because you never ran the real number and the next planning cycle catches you with commitments you can't maintain.

Willison's 10x ratio is not a prediction about your specific tools. It is a reference point. Run it against your current spend. Know what the gap is. You do not have to change anything today — but you should own that number before someone else surfaces it for you.