The invoices are ending Silicon Valley’s favorite motto that built the AI boom.
Right before Facebook went public, Mark Zuckerberg shared the “move fast and break things” motto in his 2012 letter to investors. The idea was to build and launch quickly, fix problems later, and learn from mistakes. Startups across Silicon Valley embraced this approach. During the web and mobile boom, the focus was on gaining users first, with monetization, compliance, and governance left for later. When ChatGPT entered the market, the industry repeated this strategy on a larger scale. Companies scrambled for computing power and began spending heavily on licenses, pilots, and APIs, treating AI as a land grab.
At Facebook’s developer conference in April 2014, Zuckerberg joked about a new motto: “move fast with stable infra.” With 1.28 billion users, breaking things had become expensive.
Why the Era Is Over
The motto worked until mistakes became more expensive because of the token costs, and the results stopped matching expectations. In a PwC survey of 4,454 CEOs, 56% said they had seen no significant financial benefit from AI so far.
Costs have also been harder to predict than anyone planned. A Harness study found that 72% of organizations had had an unexpected AI cost spike, and more than half said no one clearly owns AI spending. Harness sells AI cost-management software, so it has a stake in how big the problem looks. Cloudera’s research points to another problem: segmented data systems and silos make it hard for companies to measure what AI really costs them to run.
Then the CFOs Reached for the Cap
In April, Uber’s CTO Praveen Neppalli Naga said the company had used up its entire 2026 AI budget after just four months of Claude Code. Teams had been ranked on an internal leaderboard by how much AI they used. By June, the company limited each employee to $1,500 a month per AI coding tool. In August, Naga wrote that Uber was coming to the end of the so-called “tokenmaxxing era,” when burning tokens meant productivity.
Gartner's advice to CFOs is to treat AI cost as an architecture issue rather than a budget line item. One unmonitored prompt library alone can "risk running up millions of dollars in additional expense".
That’s the new rule. Finance asks the question nobody asked loudly in 2023: what did we get?
How the Ledger Rewrites the Stack
Finance teams can't parse a transformer model, but they can easily understand a unit price. Unit price is now the driver of architecture decisions. So how do you deal with that?
Route tasks wisely. Don’t just send every job to the largest model. In his annual letter on October 21, 2026, Satya Nadella introduced the “cost-to-outcome” frontier: “After all, Super Intelligence is changing every organization’s inputs (how they work) as well as their outputs (the products and services they sell). At the same time, inference is making marginal cost an important factor in software economics as usage scales. As a result, customers will increasingly measure the cost of tokens against the outcomes they produce. We call this the “cost-to-outcome” frontier, and we are focused on optimizing it.”
For engineers, that means sending classification, extraction, and summarisation to a smaller, cheaper model, caching repeated prompts, and using the frontier model only for the work that needs it.
Count everything. FinOps for AI is the top priority in the State of FinOps 2026 Report, and AI expense management is a skill teams currently need to build. In practice, that means tagging every model call by team and feature and giving each agent its own budget and alert, so a runaway loop triggers an alarm before it reaches the invoice.
A new software market is emerging to meet this demand. Max Gontar, CEO of the AI-powered finance software company Kairos, explains:
The first wave of enterprise AI was all about wild spending on experiments in the hope of seeing what the models could write/generate. The current state is that CFOs are getting what they actually need: software that runs within their existing infrastructure to find working capital leaks, deliver tangible ROI and automate finance workflows as a result (that's where AI really is good at). The new challenge for the CFO is the management of AI infrastructure spend. They now have to work out how many tokens it makes sense to spend on each process, what to cut and what to rebuild. The ROI of every function in the business will be more transparent than it's ever been.
Watch who owns the numbers. In the FinOps Foundation’s 2026 survey of 1,192 practitioners, 78% of FinOps teams report to the CTO or CIO, and only 8% report to the CFO. So finance didn’t take control, but it did get veto power. The CFO asks, “What did this return?” and engineering manages the dashboard with the answer.
Balance Sheets vs. the Spark
There’s a problem, though. If every task has to meet a cost limit, who pays for work you can’t predict? Cost-to-serve metrics reward what’s measurable now, but exploration usually isn’t measurable in the short term.
There's no 2026 data at all to show that cost discipline is in any way reducing the amount of exploratory research so it's a valid concern rather than a fact. The moral guidelines, the risk frameworks, the ethical manifestos all sit in a PDF somewhere and are presumably being read. A runaway agent loop doesn't do that. It shows up as a number on a bill and numbers get attention. The ledger will only protect what it can count.
But Big Tech Is Spending More Than Ever
Here's the obvious objection: if the era is over why are the giants still throwing money at it? Forbes has the four biggest spenders (Amazon, Meta, Microsoft, Alphabet) at $700-$725bn of capital expenditure in 2026, up 75% on $410bn in 2025.
This only makes the argument clearer. The sellers are the hyperscalers. They build capacity and fund it with huge cash flows, betting on future demand. Their customers are the buyers, and buyers pay attention to invoices. If a buyer cuts back, chooses cheaper models, or limits usage, it affects the demand that sellers are counting on.
There's a bit of a tension there. Capacity is being built on the basis that the enterprises will be spending a lot more, whilst the same enterprises are in fact capping their usage, moving to cheaper models and asking what they got for the money. One side of that equation is either completely wrong or just way ahead of its time.
The Audit Is Still Coming
Facebook abandoned 'move fast and break things' when it started to get expensive. A decade on and enterprise AI is running into the same wall, just with a much bigger bill. It's not over yet. Finance has the veto, they just can't yet see what they're giving up in terms of accuracy and ambition. 'Move fast and break things' died because of the invoices. The next battle is over what gets counted.
