Nadella: AI Buyers Will Weigh Token Cost Against Outcomes

Microsoft CEO Satya Nadella wrote in his annual letter to shareholders, published on 21 October 2026, that “customers will increasingly measure the cost of tokens against the outcomes they produce.” He names this the “cost-to-outcome” frontier and says Microsoft is “focused on optimizing it.”

The same letter is the source of our earlier piece on Microsoft’s harness and swappable models. This piece looks at the cost-to-outcome idea on its own: what it means, what Microsoft says it is doing about it, and five figures published on 20 and 21 October that bear on it.

Business Pill · TOKENS PER WATT

A short Business Engineer explainer on tokens per watt, the efficiency measure behind the cost of AI output. It teaches the general idea only and says nothing about Microsoft or any company in this piece.

What Nadella Wrote

The passage comes early in the letter. Nadella writes that “inference is making marginal cost an important factor in software economics as usage scales,” and that, as a result, customers will weigh the cost of tokens against the outcomes those tokens produce.

He then widens it into a market claim: “Any high-value activity or process that can be made faster, less expensive, or more effective through intelligence becomes part of that opportunity.”

On what Microsoft is doing, the letter points to three things, in separate passages. It says Microsoft is “bringing capacity into service faster, increasing utilization, and continuously improving performance and cost at every layer of the stack and across the stack as a whole.” It says Microsoft is “evolving our business models to reflect the value and usage of our services.” And in a later section, the letter says Microsoft’s multimodel architecture, in which the harness is separate from the model, “optimizes the cost-to-outcome frontier by allowing us to use the right model for the right job.”

Elsewhere in the letter, under the heading “Our why,” Nadella writes: “Our success depends on our customers’ success. We want every organization to own its own continuous learning system so it retains its core IP. And we want to empower every person, amplifying their agency and ambition.” As we read it, that passage pairs with the cost-to-outcome line, because it says the learning system, and the IP it holds, should belong to the customer. For background on how Microsoft states its purpose, see our page on Microsoft’s mission statement.

On the hardware side, the letter sets the goal as “the highest performance and utilization per dollar and per watt,” and says Microsoft’s Maia 300 chip “delivers 40 percent more tokens per watt compared to leading GPU-based systems.” That comparison is Microsoft’s own.

What Cost-to-Outcome Means

In plain terms, a token is a small chunk of text that an AI model reads or writes, and AI providers price their services by the token. A cost-per-token view asks how cheap each chunk is. A cost-to-outcome view asks how much it cost to get a finished piece of work: a resolved support ticket, a reviewed contract, a shipped bug fix.

The two can move apart. A cheaper model that needs many attempts can cost more per finished task than a pricier model that gets it right first time. As we read it, that gives a buyer two levers: lower the cost of each token, or get more finished work from the tokens already being paid for.

Five Figures From 20-21 October

Four companies published figures on 21 October, and researchers posted a paper on 20 October, that touch one side or the other of that equation. The table is our framing; each figure is the company’s or the researchers’ own, from the letter or as reported in our earlier pieces, and they measure different things, so they are not comparable with each other.

The Nvidia figure comes from its release at GTC Berlin. The Meta figure is a goal Mark Zuckerberg gave on Meta’s earnings call, as reported by TechCrunch. The research figure comes from a paper posted to arXiv on 20 October. Tesla’s chief financial officer Vaibhav Taneja said on the earnings call: “We’re revising our CapEx guidance to be approximately 20 billion for the current year,” according to a Benzinga transcript, and Tesla’s update says: “We plan to reach 400 MW of SI training compute by the end of the year.”

How We Read It

As we read it, three of the five figures, from Microsoft, Nvidia and Meta, are about lowering the cost of AI computing. The research paper is the one figure on the other lever, a better result from the same model. Tesla’s figures are about how much it spends to build capacity.

As we read it, the letter connects cost-to-outcome to both levers: its chips on the cost side, and its multimodel architecture on the outcome side.

The letter also says Microsoft is changing its business models “to reflect the value and usage” of its services, but it does not say how.

The Structural Read

As we read it, cost per token and cost per outcome can move apart: a cheaper model that needs more attempts can cost more per finished task.

As we read it, the letter links the idea to two things Microsoft sells: its own chips, which it says deliver “40 percent more tokens per watt,” and its multimodel architecture, which it says lets it “use the right model for the right job.”

As we read it, the research paper in the table is the one figure from 20 and 21 October that measures the outcome side directly, holding the model fixed.

Three Implications

FOR ENTERPRISE BUYERS As we read it, the letter’s frame invites a different question for vendors: not the price per million tokens, but the cost of a finished task, which the letter itself does not quantify.

FOR MODEL AND CHIP MAKERS Nvidia, Microsoft and Meta all published efficiency figures on 21 October, in tokens per dollar, per watt or per megawatt, or as a cost-efficiency goal. As we read it, those are inputs to cost-to-outcome, not the outcome itself.

FOR SOFTWARE VENDORS The letter says Microsoft is “evolving our business models to reflect the value and usage of our services,” without saying how. As we read it, that is the pricing question to watch in Microsoft’s next disclosures.

The Business Engineer Lens

This story maps onto the Business Engineer framework The CFO’s Guide to the Token Economy.

The framework’s core line: the new job of finance is to “manage the ratio between the two — work produced per token burned. That ratio is the gross-margin equation of the agentic era.”

As we read it, Nadella’s “cost-to-outcome” frontier names the same ratio from the vendor side. The framework was published in June 2026 and does not discuss the letter.

What Is Not Established

The letter gives no cost-to-outcome figure for Microsoft or its customers, and does not say how such a measure would be calculated. It does not describe specific pricing changes.

The figures in the table are each company’s own claims, on different measures. The Meta figure is a stated goal, not a result. The research paper is one study on three benchmarks, and it reports task scores, not token counts or cost.

Connecting the Coverage

On 21 October we covered the same letter for its claim that Microsoft’s harness makes AI models swappable; our harness piece also carries the Google research paper as an update. The same day, Jensen Huang put Vera Rubin at 30 times Blackwell, which our Vera Rubin piece covers, and Meta set its Muse cost goal, covered in our Muse piece. Tesla’s compute plan is in our Tesla compute piece, and its capex cut is in our Tesla results piece.

What’s new here: Microsoft’s CEO has put a name, the cost-to-outcome frontier, on a measure that weighs token cost against the work it produces, and says Microsoft is optimizing for it.

The Bottom Line

Nadella’s letter says customers will increasingly measure the cost of tokens against the outcomes they produce, and names this the cost-to-outcome frontier. As we read it, the five figures published on 20 and 21 October include more on the cost of computing than on outcomes, which the letter does not put a number on.