Issue 014: Capital, Cost and Control

For two years the enterprise question about artificial intelligence (AI) was what it could do. In 2026 that question has been settled often enough to stop being interesting. This week the news triangulated a different one. Anthropic raised $65bn (about £51bn) at a valuation near $965bn and filed confidentially for a public listing; Alphabet set out plans to raise $80bn (about £63bn) to build infrastructure. The capital going into AI is industrial in scale. And yet surveys of technology leaders this week put the share of AI initiatives that have met their business goals at under one in five, while reporting on enterprise spending shows firms rationing usage and shopping for cheaper models as their bills climb. Capital is pouring in at the top. Returns are not keeping pace at the bottom.

That gap is the story. The binding constraint on enterprise AI is no longer capability. It is cost, and whether you can govern it. A pilot that proves value but cannot control its spend is not a win. It is a liability waiting for a budget review.

Three forces meet in the same place. The model providers are spending to win, so prices and products churn constantly. Buyers are discovering that usage, not licences, drives the bill, and that agentic workloads consume tokens faster than anyone budgeted. And the people signing the cheques have started asking the obvious question: what did this return? The 19% figure is not a verdict on whether AI works. It is a verdict on whether most organisations are measuring it.

The fix is not a better model. It is three controls, none of them exotic, all of them usually skipped in the rush to ship. Meter the spend: instrument cost per team and per use case, so you can see where the money goes before finance does. Tier the model choice: route routine work to the cheapest model that clears the quality bar, and keep the frontier models for the work that needs them. Gate the funding: fund AI in stages tied to outcomes, and stop the pilots that do not pay.

The counter-argument is fair. Token prices keep falling, the argument runs, so the squeeze is temporary and this discipline is premature. Unit prices do fall. But usage and agentic workloads grow faster than prices drop, which is why the bills keep rising even as each call gets cheaper. Falling prices reward the disciplined buyer most, because the savings compound on top of a spend you already control. The habit is what survives the price war, whoever wins it.

So treat AI compute the way you treat any other metered resource that scales with use. Give it an owner, a budget, and a number it is supposed to move. The recommendation for this week: pick your single largest AI expense and write one sentence saying what business outcome it is meant to change. If you cannot, you have found your first audit. The Prompt Pocket below turns that into a repeatable triage you can run across every AI task on your plate.

Poe puts more than a dozen leading models, including GPT, Claude, and Gemini, behind one subscription and one simple chat interface. For the cost question this issue is about, that matters: instead of paying for several separate AI subscriptions, you choose the right model for each job in one place, and you can drop to a cheaper, faster model for routine work rather than reaching for a frontier model out of habit. It needs no application programming interface (API) or developer setup, which keeps it within reach of a non-technical reader.

It makes right-sizing easy. One place to compare models, switch to a cheaper one for routine work, and avoid paying frontier prices for jobs that do not need them. The interface is plain and quick to learn, and the free tier lets you test the habit before you pay for it.

The points system that meters your usage is not always easy to read, so it takes a little while to learn what a given model costs you per message, and heavy use of the top models burns an allowance quickly. Poe is built for individuals, not for company-wide governance, so it is a personal cost-control tool rather than an enterprise one. For metering a whole team, you will still want the controls in your provider billing console.

Try Poe →

The organisations that win with AI in 2026 will not be the ones that spent the most, or the ones that waited. They will be the ones that treated AI like any other resource that scales with use: owned, budgeted, and measured against a number it is supposed to move.

Three predictions for the next two quarters. First, at least one more frontier lab will formalise public-market steps before the fourth quarter, because Anthropic's confidential filing sets the precedent and investors will start pricing the option into the whole sector. Second, "AI adoption" will quietly disappear from serious board packs and be replaced by "AI cost per outcome," because the one-in-five figure has made raw adoption look like a vanity metric. Third, at least one more European government will block or screen a United States acquisition of critical digital infrastructure within six months, because the Netherlands decision plus the coming European Commission sovereignty package make screening the default rather than the exception.

The question for your week: find the single largest line in your AI spend and ask what number it is supposed to move. If you cannot answer in one sentence, that is your first audit. Hit reply and tell me what you found. I read every response. If the honest answer is "we are not sure what it returns," you are in the nineteen-in-twenty, not the one. The fix is not less ambition. It is a meter, a tier, and a gate.