The first thing I check on an AI architecture is no longer the model.
For most of last year it was. Which model, what context window, what the token bill looks like at volume. I would spend the first hour of a review on that page, because it was the line with the biggest number next to it, and a big number feels like a decision.
On 30 July 2026 OpenAI repriced GPT-5.6 Luna by 80%. Input went from $1.00 per million tokens to $0.20. Output went from $6.00 to $1.20. Terra came down 20%, from $2.50 and $15.00 to $2.00 and $12.00. Sol, the prestige tier, did not move at all.
Every hour anyone spent last quarter optimising a model choice on price was deleted by one update to a pricing page.
The thesis
Most enterprise AI budget is going into the layer that will be a utility bill inside 18 months. The two things that will still be yours after that happens are your data contracts and your evaluation sets, and almost nobody is funding either.
That is not a call to spend less. It is a call to spend the same money one layer down.
The numbers behind the claim
Menlo Ventures surveyed 495 US enterprise AI decision makers between 7 and 25 November 2025. Total enterprise AI spend for the year came to $37 billion, up from $11.5 billion in 2024 and $1.7 billion in 2023.
Split that $37 billion. Applications took $19 billion. Infrastructure took $18 billion. Inside infrastructure, foundation model APIs took $12.5 billion, model training infrastructure took $4.0 billion, and storage, retrieval and orchestration took $1.5 billion.
Read that ratio again. Enterprises spent 8.3 times more renting model capability than building the layer that feeds it.
Now put that next to the price curve. Andreessen Horowitz measured the cost of a fixed level of model performance falling roughly 10x per year: GPT-3 level quality cost $60 per million tokens in November 2021 and $0.06 three years later, a 1,000x collapse. GPT-4 level quality fell about 62x from its March 2023 launch. The 80% cut on 30 July 2026 is not an event. It is one visible step on a curve that has not broken once in five years.
You cannot build an advantage on a line item that falls 10x a year. Everyone else gets the same cut on the same day.
The switching data says nobody is locked in either
Perplexity published enterprise usage across January to December 2025. The leading model’s share of queries fell from 47.5% to roughly 23% inside twelve months. Early in the year two models held 91.5% of queries. By the end, four models each held more than 10%. Across organisations, 43.6% used more than one model at some point in the year, and 53% of users who pick a model switched between models at least once inside a single working day. The top 50 enterprise accounts averaged 30 models each, against seven for a typical account.
Menlo’s share numbers land in the same place: Anthropic 40%, OpenAI 27%, Google 21%. OpenAI held 50% in 2023.
A market where the leader halves its share in a year, where half of power users change model before lunch, and where the price of the thing drops 80% overnight, is a market with no defensible position in it. That is the definition of a commodity input. It is a fine thing to buy. It is a terrible thing to build a strategy on.
What does not commoditise
Two things. Both boring. Both underfunded.
The first is the data contract: an explicit agreement between the team that produces a table and the team that consumes it, covering schema, ownership, freshness and how a breaking change gets announced. Without it, someone renames a column upstream, your retrieval layer quietly returns worse context, your answers degrade, and your AI team spends three weeks debugging a model that never changed.
The second is the evaluation set: your own labelled examples of what a correct answer looks like for your business, run automatically against every model, prompt and retrieval change.
Here is the part people miss. Evaluation is filed under governance, so it gets treated as a brake. It is the opposite. The eval set is the mechanism that converts a price cut into cash.
When Luna drops 80%, you can only take that money if you can prove the cheaper model still does your job. With a golden set of a few hundred labelled cases you run it in an afternoon and migrate. Without one, you have a meeting, then a pilot, then a quarter of drift, and you keep paying the old price on the old model because nobody can sign off that the new one is safe. The company with evals harvests every price cut in the industry automatically. The company without them watches the curve go past.
The evidence that this layer is genuinely missing
Gartner surveyed 1,203 data management leaders in July 2024 and concluded that through 2026 organisations will abandon 60% of AI projects that are not supported by AI ready data. A separate Gartner survey of 248 data management leaders found 63% either lack appropriate data management practices for AI or are not sure whether they have them.
On the evaluation side, Gartner reported on 30 March 2026 that 15% of generative AI deployments currently include LLM observability, and projected 50% by 2028. Eighty five percent of deployments running today cannot tell you why an answer came out the way it did.
Meanwhile the build and buy ratio flipped hard. In 2024 enterprises built 47% of their AI use cases in house. In 2025 they built 24% and purchased 76%. That is a rational response to falling model costs and improving vendor products, and it has a side effect nobody priced: when you buy the application and rent the model, the only part of the stack you still own outright is your data and your definition of correct. Those are exactly the two lines that got $1.5 billion out of $37 billion.
I run this pattern in my own products. D30 pulls three statement fact files out of ASX annual reports, and the model that reads the PDF is not the asset. The articulation gates are. Eight of them, checking that the balance sheet balances, that the cash flow ties to the movement in cash, that the tagged store reconciles to the statements. If the numbers do not articulate, the fact file fails, whatever model produced it. That gate lets me change the model underneath without changing my confidence in the output. Same logic at D23, where clients pay for Superset to be run correctly rather than for the software. Same at SearchFIT, which has to query answer engines it does not control and would be worthless without a stable definition of what counts as a citation.
Where this breaks
The frontier is not commoditised, and my thesis does not cover it. Sol held at $5.00 and $30.00 on the same day Luna fell 80%. That gap is the market telling you that hard, long horizon reasoning still has scarcity value. If your use case genuinely sits there, model selection decides whether the thing works at all, and no eval set rescues a model that cannot do the task.
The thesis also over-fires for small companies. If you have three AI use cases and no data platform, building formal data contracts is over-engineering. Buy the application, use the vendor’s evaluation tooling, revisit at scale.
And evaluation sets rot. A golden set built on last year’s task distribution will happily approve a model that fails on what your customers ask today. An eval set is a living asset with an owner, not a compliance artefact you build once. If you are not editing it monthly, it is already lying to you.
What I would do on Monday
Print the AI budget. Mark every line by whether it survives a model swap. If more than half the spend dies when the model changes, you are funding the commodity.
Build one golden set this week. One hundred to three hundred labelled examples for your single highest volume task. Not a platform, not a vendor selection. A spreadsheet and a script are enough to start.
Put a contract on the three tables that feed retrieval. Named owner, schema, freshness target, and a rule that breaking changes get announced before they ship.
Re-run last quarter’s most expensive workload on the cheapest current model against that golden set. If it passes, take the 80% now. That single exercise usually pays for the eval work in one billing cycle.
Change the metric in the monthly pack from tokens consumed to cost per accepted output. Tokens are a price you do not control. Acceptance is a quality you do.
Close
The first thing I check now is the eval set. If there is one, I know the team can move when the price moves, and the model on the architecture diagram is a detail. If there is not, I know the exact conversation we will have in nine months, because the price will have fallen again and they will still be paying the old rate.
The model is rented. The definition of correct is owned.
Brightlume does this work with enterprise teams. If the gap between the AI pilot and the P&L is the problem you have, talk to me.