Inference is the P&L

The fastest way to misprice an AI company in 2026 is to underwrite it on training capability. Capability is converging. When two labs ship near-identical quality at a two-month lag, the model layer becomes a price-taker, and price-takers do not defend gross margin — they defend volume, and then they defend nothing.

What does not converge is the cost of serving. Batch efficiency, cache hit rate, context discipline, and the ratio of retries to first-pass answers are operational facts rather than research results. They differ by an order of magnitude between two products that look identical in a demo, and they are almost never in the deck.

Buyers pay for outcomes. A product that burns four agent turns to do one job has a four-times problem hiding inside a competitive per-token price, and the problem is invisible in any metric that ends in 'per thousand tokens'.

The teardown we run is simple and unforgiving. Take a real completed task — a merged pull request, a reconciled invoice, a drafted contract clause — and count every token spent reaching it, including the failed attempts, the retrieval calls, the reranking, and the human correction loop. Then divide by the price charged. In our sample, the spread between the best and worst product in the same category was roughly eleven to one.

Companies that measure themselves this way tend to do two unglamorous things: they cut turns, and they cache aggressively. Both are engineering work with no announcement attached, which is exactly why they are underpriced by the market.

Falling token prices are good news only if volume compounds faster. This is an arithmetic question and it has a clean answer per company, but it is usually presented as a narrative one.

Where usage does outrun price, revenue grows while unit cost collapses and the business looks briefly miraculous. Where it does not — and in most seat-based products it does not — the revenue line flattens while the story stays bullish for another two quarters, which is roughly how long the gap between the two takes to become visible in a board pack.

Fixed-price seats over variable inference is a margin trade the seller wins only if usage is predictable. Most agent products discover it is not. Our case file on this is blunt: the heaviest decile of users consumed about nine times the median, that decile grew fastest because heavy users evangelise, and blended margin fell for four straight quarters while headline growth held.

The fix is not clever pricing. It is a seat price plus metered overage, and a serious campaign to reduce turns so the overage is small enough that nobody resents it.

A wrapper over a commodity endpoint has a switching cost measured in weeks. We have watched customers pay that cost without much internal debate once model prices fell far enough for a small platform team to justify rebuilding.

The products that hold are the ones where corrections land in a system of record. Correction data compounds while model prices fall, and it is the only asset in this stack that gets more valuable as the underlying model gets cheaper.

A serving stack whose cost curve falls faster than list price, held by a company that owns the workflow, would break the price-taker framing outright. So would a durable capability gap that buyers actually pay for — not a benchmark gap, a renewal gap.

We watch for both in month-twelve renewal cohorts, not in launch benchmarks. Twelve months is roughly how long it takes for an enterprise to discover whether the product removed work or relocated it.