Your AI ROI Runs On a Subsidized Rate

The ROI math under most enterprise AI projects is built on a price that isn't real yet. The number you're modeling against is a subsidy, set to win the market, not to cover the cost of serving you. And the gap between those two prices is the part nobody put in the business case.

Two sets of vendor financials this week point at the same thing from opposite ends: cost structure scaling right alongside revenue. OpenAI's leaked audited accounts show losses widening in absolute terms even as the top line scales. Databricks grew sales 80% to roughly $6.9 billion annualized and watched its margins shrink anyway, because the swarm of agents doing the data work costs real money to run.

One of those gets told as a startup growth story that profitability eventually catches. The audited numbers don't support that reading. R&D and operating expense aren't investment-phase line items you grow out of, they're the cost of running the product. When cost structure scales in proportion to revenue, growth isn't a path to profit. It's a faster way to find out the price is wrong.

The Databricks margin line is the more useful signal because it isn't vendor-specific. It's agentic AI unit economics showing up on someone's income statement. Every enterprise standing up agents at scale is generating the same compute draw. Right now a vendor is eating it. That's a choice, not a law.

Watch what the biggest player does about it. Microsoft weighing DeepSeek for Copilot Cowork gets framed as a capability decision. Read it as cost containment instead: the model layer is commoditizing faster than enterprise contracts assumed, and a cheaper model under the same product protects margin without raising the sticker. The shift to usage-based pricing for its enterprise agent is the same move, more direct. Usage-based pricing is how the vendor hands the compute risk back to you.

Usage-based pricing is how the vendor hands the compute risk back to you.

So the repricing isn't theoretical or a year out. It's already in motion across the people who set the prices.

This is where the difference between "we'll use AI to transform the customer experience" and "we'll take this from three days to fifteen minutes" starts being financial exposure directly. A concrete ROI case survives a price change, because you can re-run the math at the new rate and see if it still clears. A vague transformation case can't, because there was never a real number under it to stress.

I don't know the timing, whether the repricing lands in two quarters or six, or which line items move first. What I'd want before committing critical infrastructure to any of these tools is a single answer: re-run the case at double the unit price. If it still clears, build. If it only works at today's rate, you haven't bought an ROI. You've bought a subsidy with an expiration date you don't control.