The 18-Month AI Cliff

Somebody has to pay for all this compute

There is a story hiding behind Salesforce's admission about its Claude bill that is considerably bigger than Salesforce, and bigger than Claude.

What Salesforce has discovered is that if you hand a very capable model to a large R&D organisation, they use it. Enthusiastically. Their deputy CFO, Mike Spencer, put it about as plainly as a finance executive ever does: "Roughly about six months ago we unleashed Claude in our R&D cycle. It's part of the reason we didn't raise margin guidance on the year because we're covering some of the token spend."¹

Worth being precise about that, because the headline version going around is that AI costs are eating Salesforce's margins. They are not, or at least not yet. Q2 operating margin came in at 20.5% and full-year guidance sits at 20.1%. Nothing collapsed. What happened is that a margin increase they could otherwise have announced went into token spend instead. Marc Benioff had already flagged an expected 300 million dollars with Anthropic across 2026.¹ That is not a company in trouble. It is a company quietly absorbing a new cost line big enough to show up in guidance, which is a different and more interesting thing.

Their response is exactly what you would expect from anyone who has run an infrastructure budget. Spencer described moving into "refinement mode" and applying "prescription model choice for task at hand", with the observation that "you don't need to use the latest and greatest model for every single task". They are running OpenAI, Cursor and Claude alongside each other, and trialling Grok, choosing on cost-effectiveness rather than raw capability.¹

That is sensible engineering discipline and I have no argument with it. It is also an early warning signal for the entire industry, because it is the first visible instance of a large sophisticated customer looking at the invoice and asking whether the expensive model was actually worth it.

The build-out was the easy part

For the last two years the operating assumption has been: secure the compute, and the revenue will follow. That assumption has driven an extraordinary infrastructure programme. GPUs, power, land, data centres and long-dated capacity contracts, all committed now for facilities that in many cases are still concrete and cranes.

The Dallas Fed puts the financing requirement for that build-out at between 3 and 5 trillion dollars over the next three to five years.² The number is arresting, but the number is not the point. The timing is.

Roughly 500 to 600 billion dollars of the investment since 2023 has been funded out of hyperscaler retained earnings, and those firms have only more recently turned to public and private debt markets.² Read that sequence carefully. The comparatively painless phase, where you pay for capacity out of profits you already made, is largely behind us. The financed phase is just starting.

And a commitment behaves very differently from an operating asset. Then the building switches on. The lease starts. The power bill arrives monthly. Financing costs arrive whether or not the racks are busy. Somebody has to consume the capacity, and somebody has to pay for having consumed it.

We have spent two years talking about AI as deflationary because it raises productivity. That may well turn out to be true. But you have to finance the productivity machine before you get the productivity.

This is why it shows up in the price of money

AI is not doing this in isolation, and that is the part I think gets underweighted. The US government is borrowing heavily. Corporates are borrowing heavily. Data-centre developers are borrowing heavily. AI companies are committing enormous sums to infrastructure. All of it competes for the same pool of long-duration capital.

The Dallas Fed work is specifically about this mechanism rather than the headline number. It describes AI-related debt supplying duration through three channels: direct bond issuance, synthetic duration via pay-fixed swaps, and the potential crowding-out of financial issuers. Its conclusion is measured and worth quoting exactly, because measured is more useful here than dramatic: this supply "should on the margin bias yield levels higher and the yield curve steeper". The analysis projects as much as 360 billion dollars of new duration supply in 10-year equivalents across 2026, from roughly 300 billion dollars of estimated AI-related investment-grade issuance.²

Meanwhile the US ten-year Treasury touched 4.818% in early September 2026, its highest level since November 2023, settling around 4.79%.³ I want to be careful here, because this is exactly the point where commentary usually overreaches: the Dallas Fed does not put a basis-point number on how much of that is AI, and neither will I. Yields move on jobs data, inflation and Fed expectations. The honest claim is narrower and still significant. The AI build-out has become a real part of the long-duration supply picture, and it is directionally pushing the same way.

That matters because when long-term money gets more expensive, everything that depends on long-term money gets more expensive. Including data centres.

Which is why Oracle interests me more than Salesforce

Salesforce has a cost-control problem, and cost-control problems are solvable. You route cheaper models, you cap usage, you renegotiate. Oracle has placed itself somewhere structurally more exposed: right in the middle of the build-out.

The figures are genuinely remarkable. Oracle closed FY2026 with 638 billion dollars in remaining performance obligations, up 363% year on year. Capital expenditure ran at 55.7 billion dollars in FY2026, with roughly 70 billion in net capex guided for the following year and 90 to 95 billion gross. It carries about 130 billion dollars of debt against 32 billion in cash, at net leverage of about 2.67 times, having raised 43 billion in debt and 5 billion in equity during FY2026.⁴

If AI demand continues anywhere near the currently assumed trajectory, that is a career-defining piece of positioning. The tension is structural rather than clever: a data centre has an economic life measured in decades, while the technology consuming it changes in months. The industry has been making long-duration infrastructure bets against short-duration assumptions about demand. Note that only about 12% of that 638 billion backlog is expected to convert to revenue in the next twelve months, around 76.6 billion.⁴ The rest is a promise about years that have not happened yet.

The counterweight, because it is a real one

I would be doing the lazy version of this argument if I stopped there, because there is a detail that cuts against me and it deserves airtime.

Most of Oracle's RPO growth in the last two quarters came from large AI contracts where the customer either prepaid for the GPUs or bought and supplied the hardware directly. That prepaid and customer-supplied portion now totals about 75 billion dollars, which materially reduces the capital Oracle itself has to raise to build the capacity.⁴

That is a genuinely better risk position than the headline capex figure suggests, and anyone making the "Oracle is dangerously levered to AI" case needs to deal with it honestly. It does not dissolve the exposure. Roughly 70 billion of net capex still has to be funded, the debt is real, and prepayment moves risk onto the customer rather than deleting it. But it does mean the structure is smarter than the scary numbers alone imply.

What the next eighteen months actually asks

The AI industry has been operating on a first-to-market-wins philosophy, and there was a defensible logic to that while compute was the binding constraint. You could build a great deal of infrastructure on the assumption that demand keeps compounding. Eventually that assumption has to turn into invoices somebody pays without complaint.

The question is changing under everyone's feet. For two years it was "can we get enough compute". The buildings are now coming online, and the question becomes "can we sell enough AI to pay for them". Salesforce has just shown us, in public, in a quarterly call, what the first half of the answer looks like: a large sophisticated customer deciding the frontier model is not worth it for every task.

If demand does keep compounding, this will stand as one of the largest productive infrastructure investments in history and the people who committed early will look prescient. If it does not, the consequences will not stay politely inside the AI industry. They will surface as margins, debt service, utilisation rates, capital costs, and eventually as something the rest of the economy notices.

None of which is a prediction of collapse, and I am not making one. It is an observation that the industry spent two years asking one question and is about to be handed a much harder one.

For two years the question was how much compute we could get. The next eighteen months will ask who is going to pay for it.

What I would actually do about it

If you are buying rather than building, the practical read is straightforward and it is the same conclusion Salesforce reached the expensive way. Know what your inference actually costs, per workflow, not as a monthly total that lands on someone else's cost centre. Route by task rather than by habit, because the frontier model is the right answer far less often than the demo suggested. Design the boundary so the model is not reasoning over things your runtime could have decided for free.

That last one is where architecture and economics stop being separate conversations, which is more or less the argument I have been making for the last several posts. The cheapest architecture is the one that does not ask a model to think when it does not need to. That was a performance argument eighteen months ago. It is turning into a balance-sheet argument.

• MCP and A2A Are Not Integration