Real Alpha
Watch modeSubscribed

Possible · 4 min · Jun 5, 2026

Match model cost and capability to the problem

Satya Nadella says token efficiency depends on defining the outcome and measuring it well. For repeatable workflows, a smaller model trained on work traces may beat a frontier model used with prompting.

Satya Nadella

Satya Nadella is chairman and CEO of Microsoft.

Led Microsoft's cloud infrastructure and services transformation, and has been CEO since February 2014.

  • Reid Hoffman, Host · Co-founder, Manas AI

    Co-founded LinkedIn; one of OpenAI's first investors; Greylock partner.

About this segment

Satya Nadella says token efficiency depends on defining the outcome and measuring it well. For repeatable workflows, a smaller model trained on work traces may beat a frontier model used with prompting.

From “Satya Nadella: AI Is the Future of the Firm”, Jun 5, 2026. Watch the full episode

“one of the dictums I have is don't use frontier models for non-frontier problems”
Satya Nadellaat 34:04 · Possible34:04

Takeaways

  1. Define the outcome and rubric before spending tokens.
  2. Use frontier models for frontier problems, not routine workflows.
  3. Train smaller models on traces for repeatable tasks.

In this episode

Full episode
  1. 6:55Enterprises need to keep their tacit knowledge inside3:01

  2. 31:29Match model cost and capability to the problem3:32

  3. 51:27AI needs tangible benefits to earn public trust3:54Up next
  4. 5 more moments in the full episode

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

  1. 31:34

    Well, and I think that point intersected with your earlier one about, you know, kind of token capital and AI capital is that part of the skillset for the modern, you know, kind of human knowledge worker is to say, okay, how do I strategize on the use of AI, canvas orchestration, but also within kind of capital allocation balance, right? Because part of what we've seen is people can go crazy on spending lots of tokens in ineffective ways.

  2. 32:02

    It's like, you know, buying, you know, go back to bad Pentagon days, toilet seats for a million dollars. And so putting your thinking together with the cognition together with token management. And what do you think is part of the question around this skill of blending the cognitive coverage with also token intelligence in terms of how do you amplify the outcomes? It's a great one. In fact, I've been thinking about like one of the things.

  3. 32:30

    that I think one has to study deeply is what are all the things that become more valuable in the age of, let's call it token abundance. One is what you reference, which is how to use tokens becomes very valuable. Whoever figures out that, oh, I can use tokens more efficiently for an outcome that matters in the world is going to get ahead, right? By definition. So then how does one build the intuition for it, right?

  4. 32:56

    Which is, oh, this is where, what is that outcome? How do I measure it? What's a rubric? In fact, this is where I think the evals, if it's fascinating, how much time needs to be spent, especially for an RL regime, I think it's been the clearest, where if you really want, you want to set up the rubric and the eval dimensions or the rubrics scoring dimensions such that they're really capturing the high taste that you only can define.

  5. 33:28

    because if you did that, then you could be token efficient, right? In fact, the other side of the token efficiency is, in some of the examples we even showed at Build was, let's say you have, I don't know, you're a retailer or you're a packaged goods company and one of the things you do is handle sort of transactions around trade promotions. Right now, you get all these claims from all the retailers and you have to process them and the humans are doing it.

  6. 34:00

    You can easily automate that using an agentic workflow. And you could say, oh, I use a frontier model for it. But then one of the dictums I have is don't use frontier models for non-frontier problems, right? So this is not a frontier problem. You're not trying to discover some new material science. This is a repeatable, deterministic workflow, but that can benefit from all the mistakes humans can make and claims and so on.

  7. 34:28

    So therefore you do need, you know, intelligence in it. But you can take, you know, a model like an MAI 5B and use the traces to hill climb in your RLE to perform even, outperform even a frontier prompted, right? So that to me is token efficiency. So that type of human understanding of both the limits of the system, the characteristics of the system, I think, become high premium at this stage.