When AI Ate the Budget

404 Media reports that SAP — one of the largest enterprise software companies on the planet, a company whose entire current sales motion is convincing other companies that AI will make them leaner — has frozen most travel and most hiring in order to cover the cost of its own AI infrastructure.

Sit with the org chart on that one. The productivity gains have not shown up. The invoice has. And the invoice is being paid out of the two line items that are made of people: the ones you would have hired, and the trips they would have taken to talk to each other.

The public story about AI and jobs runs in a specific order: companies adopt AI, the AI does the work, the humans become redundant, the headcount comes down — cause, then effect, with a productivity gain in the middle to justify it. What's happening here is simpler and considerably worse. The spend came first, the savings are a projection in a slide deck, and the headcount is collateral. Nobody at SAP was replaced by a model. They were replaced by a bill.

The generous readings deserve a hearing, because there are at least two. One: this is what conviction looks like on a cash-flow statement — every company entering a heavy capex cycle freezes discretionary spend, and a freeze is evidence of a bet, not of a bad one. Two: SAP has been restructuring since early 2024, when it announced a program touching some 8,000 roles, so it isn't obvious those reqs would have opened in a world without any of this. Both are fair. Neither changes the shape of the problem, because the problem isn't SAP's judgment. It's what happens to a software business when compute becomes a line item that scales.

And here it's worth separating two bills that look alike, because conflating them is how this argument usually goes wrong. SAP's internal spend — tooling, R&D, the infrastructure it runs to build with — is an operating cost like any other, and a company that size can absorb an operating cost. The one that matters structurally is the other one: inference as cost-of-goods, sitting inside a shipped product.

Because enterprise software has spent forty years being an extraordinarily good business for one specific reason: the marginal cost of the next customer is approximately zero. You write it once, you sell it a million times, the gross margin sits at eighty-plus percent and everyone gets a nice quarter. That property is the entire foundation of the SaaS valuation model.

Inference doesn't work like that. Every AI feature you bolt onto the product drags a variable cost back into a business that was architected around not having one. Compute per query, per user, per session, forever — and it scales with success. The better your AI feature performs, the more of it people use, the more it costs you. That is the opposite of every incentive the industry has internalized since roughly 1985.

The honest counter is that cost per token has fallen steeply and keeps falling. True — which is why the claim here isn't "SaaS margins are dead." It's that margins now ride a race between usage growth and unit-cost decline, and nobody credibly knows which one wins. That's its own kind of news. A business whose margin was a property has become a business whose margin is a result, and those two things get valued very differently once someone notices.

SAP isn't badly run. SAP is running into arithmetic.

The obvious next move is to ask what this looks like at SAP's customers, and the obvious answer is wrong. A customer buying SAP's AI features pays a price, not a compute bill. The vendor eats the inference — that's the entire mechanism just described. Downstream exposure is real, but it's different in kind: price increases, feature metering, seats that quietly become consumption tiers. Customers don't inherit SAP's problem. They get billed for SAP's solution to it.

The industry's answer to all of this is that it's capex, not waste — buy the capacity now, amortize it over years, harvest the returns later. Maybe. But that accounting leans on GPUs having a five- or six-year useful life, a claim that people who actually operate clusters tend to find funny, given that a three-year-old accelerator is mostly a space heater with a networking problem. Stretch the depreciation schedule and this quarter looks fine. The hardware is indifferent to what schedule you filed it under.

None of this means the technology is fake. Some of it is genuinely useful, which is exactly what makes the cost structure a real problem rather than a bubble that pops cleanly and lets everyone go home. Fake things stop costing money when you stop believing in them. Useful expensive things just keep billing.

So look at where the money actually went, because it didn't evaporate into productivity — it moved. Out of payroll and into GPUs and the companies that rent them by the hour. Out of the application layer, where software profits have lived for forty years, and down into the infrastructure layer underneath it. When variable cost re-enters software, the application stops being a rent business and becomes a metered one, and metered businesses belong to whoever owns the meter. SAP has the lock-in, the entrenched contracts, the two-year migration cost that makes it un-rippable — it is about as strong as an application-layer company gets, and it is being quietly reduced to a reseller of somebody else's compute. If SAP can't hold the margin there, nothing above it can.

So the honest 2026 version of the AI efficiency story reads: we spent the salary budget on compute, and we'll get back to you. A freeze on hiring, a freeze on travel, and a roadmap slide where the savings are supposed to appear. Somewhere in Walldorf there's a req that quietly closed so a cluster could stay warm.

Nobody got automated. They got outbid.