There is a question that almost never comes up in corporate AI strategy conversations, and it should. What happens to your AI-native operating model when the people selling you the intelligence start charging what it actually costs?
The economics of the AI industry are stranger than most boards realise. The frontier labs are losing extraordinary amounts of money. OpenAI posted roughly a $7bn loss in the first quarter of this year. xAI lost $6.4bn on $3.2bn of revenue in 2025, numbers we only know because they surfaced in the SpaceX S-1. These businesses are funded on the assumption that getting corporates dependent on their tools now will pay off later.
And the repricing has already started. GitHub Copilot moved from flat-rate to token-based billing this month, and heavy users have seen monthly bills go from $39 to over $800. Uber's COO said publicly in May that AI investment is getting harder to justify, with the company reportedly burning through its annual AI budget in four months. The era of all-you-can-eat AI is closing.
So should corporates slow down? No. But they should understand what is actually going on, because the picture is more interesting than "prices will go up".
Two opposing forces
Here is the tension. The cost of a given level of AI capability is collapsing. Analysis from Epoch AI shows the price of matching a fixed capability level falls somewhere between 9x and several hundred times per year. The intelligence that cost $60 per million tokens in 2023 costs pennies today.
At the same time, the price of the frontier is rising. The newest flagship models from OpenAI and Google launched at two to three times the price of their predecessors, the first generational price increases we have seen. And the subscription products wrapped around these models are being repriced from flat rates to usage-based billing, which is where the real bill shock lives.
Both forces are real. Which one you are exposed to depends entirely on how you have built. If your operations are wired directly into one vendor's frontier model and its proprietary tooling, you are exposed to the rising line. If you have built so that the model is a swappable component, you ride the falling one. That is the whole game.
The model is not the moat. Stop buying it like one.
The uncomfortable truth for the frontier labs is that the gap between their models and the best open-weight alternatives is narrow and closes fast. The leading open models now sit a few months behind the frontier on most capability benchmarks. When a frontier lab ships a breakthrough, fast followers reach most of that capability within a quarter or two, partly through legitimate research and partly through distillation practices the frontier labs publicly object to (although I have my suspicions that the frontier labs objecting to these practices have also started to do the very same thing themselves - dbrogle has done some great work uncovering this). Whatever you think of the ethics, the strategic fact stands: frontier capability does not stay scarce.
What actually makes today's best AI tools so useful is less the raw model and more everything built around it. The best agentic tools run a loop: plan, act, check the result, correct, repeat until the job is done. They give the model an environment where it can execute code, read files, and verify its own output against reality rather than guessing. And because an agent that can write code can integrate with anything that has an API, it no longer needs a pre-built connector for every system it touches. That combination is what turned chatbots into agents that do real work, and none of it lives in the model weights.
And that layer is commoditising in front of us. The connectivity standard for plugging AI into tools and data, the Model Context Protocol, is now governed by the Linux Foundation and adopted by every major lab. Open-source agent frameworks have hundreds of thousands of users and support dozens of model providers interchangeably. The harness is becoming infrastructure. The model is becoming a component.
Which means the durable asset for a corporate is neither the model nor the harness. It is the proprietary data, context, and process knowledge you wire into them. Nobody can commoditise what only you have.
Most core systems should not run on LLMs anyway
There is a second confusion worth clearing up. When people picture an "AI-native" retailer, they often imagine large language models running the core machinery: demand forecasting, replenishment, allocation, pricing. That is not how it works, and it should not be.
Look under the bonnet of any serious supply chain AI vendor and you will find classical machine learning doing the forecasting. Gradient boosting and specialised statistical models, which are accurate, reproducible, auditable, and cost fractions of a penny per prediction. The canonical retail forecasting benchmark, run on Walmart data, was won top to bottom by exactly this kind of model. The vendors themselves are explicit that generative AI sits on top as the interface layer, helping planners interrogate the forecast and act on it, not generating the forecast itself.
LLMs change the economics of building, understanding, and using these systems. They do not replace them. A team with good AI tooling can now build in weeks what used to take a systems integrator a year. That is the real revolution for mid-sized corporates, and it is largely immune to token repricing, because the expensive intelligence is used at build time, not on every transaction.
What this means in practice
For a corporate setting out an AI strategy today, the hedge against the end of subsidised AI is not caution. It is composability.
Put your weight into the data layer.
Clean, well-governed, well-structured data is the asset that appreciates regardless of how model economics play out. It is also the prerequisite for switching models, because a model swap is only viable if your context and knowledge are not trapped inside one vendor's product.
Treat the model as a swappable component.
Route AI workloads through an abstraction layer rather than hard-coding any single provider. This is established practice now, not exotic architecture. The day the cost-benefit flips on any given workload, switching should be a configuration change, not a re-platforming programme.
Match the tool to the task.
Use frontier models where the capability genuinely pays for itself, today that is complex agentic work and software development. Use cheaper or open models for the high-volume, simpler tasks that make up most enterprise AI usage. Keep deterministic core systems on traditional machine learning, with LLMs as the layer that builds and explains them.
Build the measurement muscle.
You cannot swap what you cannot evaluate. A simple evaluation suite for your key AI workloads is what turns "the cheaper model is probably fine" into a decision you can defend.
The corporates that get hurt in the next phase of this market will be the ones that confused a subsidised price with a permanent one, and built their dependency accordingly. The ones that win will have used the subsidy for what it is: a temporary discount on learning how to run an AI-native operating model, captured in an architecture they control.
The technology is real. The current prices are not. Build for the first fact, not the second.
