For two years, we have all been using AI at a price that was never real.
The frontier model companies needed adoption. They got it — by subsidising usage so heavily that what cost $200 in token terms was billed at $20 a month. It was deliberate, it worked, and it is now ending. The metered era has arrived. API costs are rising. Flat-rate plans are disappearing for businesses. And the gap between what AI actually costs to run at scale and what organisations budgeted for is about to become a very uncomfortable conversation in a lot of boardrooms.
That is the surface problem. The deeper one is that most organisations were not really using AI properly anyway.
A recent AWS survey found that 64% of UK organisations claim some level of AI adoption. Sounds encouraging. Except only 24% say AI is genuinely embedded in their core processes and decision-making — and that number has moved by exactly one percentage point in the last twelve months.
What the other 40% have, mostly, is Copilot licences nobody uses and a pilot that impressed someone at a demo eighteen months ago.
"The understanding of the power and the capability behind what they have — beyond taking meeting notes and writing email summaries — is very, very small."
That is the real adoption gap. Not a shortage of capability — the technology is extraordinary — but a shortage of organisations that know what to do with it. AWS estimates this is costing the UK £35bn in unrealised efficiency by 2030. And the UK government's response is to announce a £500m Sovereign AI Fund to back new AI startups.
There is something slightly backwards about that. You do not solve a deployment problem by funding more supply.
Here is what changes when you move from dabbling to deploying.
When AI is a chat interface you open occasionally, the token cost is trivial. When AI is labour — running unattended, executing workflows, writing and testing code, managing pipelines while you sleep — the cost profile looks completely different. The architecture of usage has changed. AI that was consuming tokens for twenty minutes a day is now consuming them for eight hours. And the more rigorous you need the output to be, the more tokens it burns to get there.
Production-grade code is the clearest example. It is not enough for AI to write something that looks right. It needs to write the test first, then write the code to pass that test, then run end-to-end browser testing, then produce the documentation and architecture records. Each stage costs. What started as "isn't this cool, I built something in ten minutes" becomes a rigorous engineering process — as it should — and a rigorous engineering process has a real cost.
"I need to get something I can actually ship. And to have trust in the thing I can actually ship, I've got to 10X my token usage."
This is not a problem with AI. It is AI being used properly. But it does mean that every organisation's AI budget, if it was set during the exploration phase, is probably wrong by an order of magnitude.
Rishi Sunak has been told privately by CEOs that flat headcount with growing revenue is the new benchmark for success. Coinbase said it more directly: they can now do in days what used to take weeks. They explicitly do not need the same number of people. Unless you are a senior "player-coach" — someone who can both manage and build — they do not want you in their organisation.
This is the AI dividend. It is real. It is arriving. And it creates a tension that nobody has cleanly resolved.
The UK government is investing in AI to grow the economy. The organisations adopting that AI are reducing their workforces to do it. The jobs that disappear are visible now. The jobs that emerge to replace them are theoretical. The period between those two things is a gap that policy is not really addressing — and optimistic statements from people like Sam Altman, who has obvious reasons to be optimistic, do not close it.
"Flat is the new up makes it sound okay. It makes it sound like a steady state. But that's not what's happening. For them to stay flat, they are reducing headcount."
Down to stay flat. That is the honest version of the story.
For a while, the narrative was that AI would hollow out professional services. Why pay consultants to think when you can prompt your way to an answer?
That narrative has quietly reversed. Anthropic has launched an enterprise services arm backed by Goldman Sachs and BlackRock. OpenAI already has OpenAI Frontier doing similar work. Both are in the business of closing the gap between what AI can do and what organisations are actually getting out of it.
This should not be surprising. The gap between an impressive demo and a deployed, production-grade system has always required human intelligence to bridge. AI has not changed that. If anything, it has made it more visible — because the demos are now so impressive that the gap between demo and production feels more jarring than it used to.
"Behind all of the magic and the hype, there is still professional services underneath it that need to help you get from where you are to where you want to be. And that's just never going to change."
What has changed is what that professional services engagement looks like. It is less about methodology and more about opinionated, experienced deployment. Organisations do not need someone to explain what AI is. They need someone who has already built what they are trying to build.
One more thing worth paying attention to: the model that everyone is referring to as Mythos.
OpenAI's unreleased model has been appearing on benchmarks. Forty companies reportedly have access to it ahead of any public launch. The suggestion — credible enough to take seriously — is that it cannot be released publicly yet because the token cost would be prohibitive for general consumption. Not a commercial choice. An infrastructure constraint.
If true, that represents a structural shift in how frontier capability reaches the market. The era of everyone getting the new model on launch day may be ending. The most capable AI is going to land with a small number of organisations first, and the competitive gap between those organisations and everyone else will widen before it narrows.
"Maybe the releases start, as they start getting more intelligent, with much longer gaps before they hit us — and maybe they do get released to a select few organisations first."
Distribution, as ever, is most of the game.
The free ride is over. Token costs are real, headcount implications are real, and the gap between organisations that have deployed AI seriously and those that have run a pilot is becoming a gap that compounds.
The question is not whether your organisation should be using AI. That conversation is finished. The question is whether you are using it at the level where it actually changes your economics — or whether you are still in the phase where it impresses people in meetings.
Most organisations are still in the second category. That is the opportunity, and it is closing faster than most people think.
The Next Thing Now is the Dootrix podcast — weekly conversations about AI, software, and what comes next.