Every Company Can Operate at the Frontier — If It Owns the Evals

After a three-hour keynote sweep across the Microsoft estate, Satya Nadella walked into a crossover taping of No Priors and Latent Space at Microsoft Build. Sarah Guo, Elad Gil, and host swyx were waiting; Nadella said he already listens to both shows. The congratulations were brief. Guo asked for the most important reflection from a morning of announcements, and he did not name a product.

The biggest takeaway, he said, is to stop conceptualizing this moment as a single-model or even a single-platform race. It is an ecosystem play. AI-native startups and traditional enterprises alike should be able to participate as first-class citizens — not only consuming someone else's intelligence, but pointing to AI they themselves create. Microsoft's job, as he framed it, is the recipe: the stack, the tooling, and what is actually valuable if you want to do that.

Nadella has spent more than three decades at Microsoft and lived through four major platform shifts. His definition of a platform has stayed the same: it has to create more value above itself than it captures. On Windows, Adobe and Autodesk grew into giants on Microsoft's foundation. In the cloud, other people's businesses ran in Microsoft's rooms. He even jokes with Jensen Huang that Microsoft built DirectX and NVIDIA built CUDA on top of it — "I got the short end of that," he tells Huang, wishing the company had recognized the pattern sooner.

That history is why he keeps first-party products and enablement in the same breath. Successful platform builders have always shipped their own apps — Windows, then SaaS and cloud — but those apps should not become a ceiling on everyone else's success. The network effects around intelligence, he argues, are shallower than they look: models learn from a few novel samples, not from locking up oceans of data. The durable assets sit with the customer: private evals, context, and tools.

He offered a one-line mission for the whole developer conference: can everybody operate at the frontier with their own frontier intelligence? Without that path, companies cannot see terminal value in compounding on a platform that keeps getting better. "Without it, why have a developer conference?" he asked. You could just invite the room to worship at the altar of one model — and that, he said, is not a developer conference.

Guo asked about the training strategy behind the MAI models Microsoft had just announced. Nadella started with lineage. Mustafa Suleyman's team, he said, is obsessed with clean pre-training: high data quality, full ablations, and stripping out the junk that now saturates the web. That is also his critique of many open-weight models — they look strong on one or two benchmarks and then falter in practice. It is why Microsoft's forward-deployed engineers were excited about a small 5-billion-parameter MAI model: how does something that size keep hill-climbing?

His answer is a cognitive core wrapped in a hill-climbing scaffold. Companies should not use the model only as a generalist. They should attach their own tools, collect agent traces, stand up reinforcement-learning environments, and, most important, build a private eval. Public evals are interesting, he said, but they are no longer critical, because they can all be maxed. The scarce thing is the eval only you have. He later called that private eval perhaps "the biggest IP" a company will own.

He also redrew what "frontier" means. In a Land O'Lakes demo at Build, the team used GPT-5.5, collected traces, then took a 5-billion-parameter reasoning model and beat the larger system. Add a little temporality, he argued, and operating at the frontier is not the same as always calling the biggest model. Distill the frontier into traces, climb on a smaller specialist, and you are still at the frontier. The acid test he wants every company to run: you have a private eval and you are using model A — can you switch to model B and keep climbing? If you can, you are in control. If you cannot, you are a tenant.

That is why the harness decision matters as much as the model. For Nadella, a harness is the loop across models, data, and tools. GitHub Copilot, Security Copilot, the science-discovery work, and the GitHub harness Microsoft is reusing across products and exposing in Foundry are all multi-model harnesses with progressive tool disclosure, so they stay token-efficient. Two years of painful lessons sit in one sentence: the magic is preparing the context layer so an agent's plan can execute efficiently. You can bring an open harness, your own tools, and your own context. Microsoft's existence proof, he said, is the security agent shown at Build, which found vulnerabilities that Mythos had not.

Looking back three years — the NVIDIA partnership, then two years of MAI — Nadella said two things are now obvious. The scaling-laws paper and the OpenAI bet delivered: throw a lot of compute at transformers, and the crude slogan that intelligence is the log of compute mostly holds. What the industry underestimated is the real-world complexity of deploying these systems so they create value people will actually pay for. Benchmarks are interesting. The true eval is when someone out in the world can do a unique thing only they can measure.

That is why he has little patience for the complaint about "token max." It is an artifact, he said, of an industry that has not yet proven that every token is buying value along the way. Coding is the clearest proof of both the upside and the mess. It works so well that Microsoft now has to rebuild the IDE. A person running a hundred agent sessions cannot dump that cognitive load into a chat window. You need a canvas. In a fully agentic world, he noted, UI matters more, not less.

The next place he expects similar amplification is glue work — the judgment-heavy coordination that still eats human capital: aligning information, chasing follow-ups, stitching departments together. Long-running, durable agents with delegated authority can scale that work the way coding agents scaled software. His near-term picture is specific. Six months from now, he said, people will sit down in the morning, open a new agent development environment, and ask what last night's autopilots did on their behalf — and whether they did it right.

Backstage, the hosts had been arguing that a company's IP used to be the length of human tenure. Now it is evals and the know-how of applying agents. Nadella did not throw out human capital. People will still be the ones who find the gaps. The compounding happens when token capital and human capital leave traces together — in Teams, in workflows — and those traces train not a generalist model but a company-veteran agent. Human capital never made it onto the balance sheet because tacit knowledge could not be captured. Agent traces, he thinks, can. Guo joked that the SEC will need accounting standards for token expertise.

On the "end of software" debate, he unpacked the old SaaS recipe. For twenty years vendors built a data model, wrote business logic, put a UI on top, added a little configuration, and sold per-user subscriptions. That vertical stack can now be re-litigated. He would not reinvent a general ledger; the entity-relationship model is a robust thing you want to feed, not replace. Under Power BI's dashboards sits a rich semantic model someone already suffered to build — measures, logic, structure — and that is exactly what agents need. Unbundle the package, rebundle it, and find a new business model. The data model is the part you do not knock down.

Microsoft 365 is his live example. When Exchange, SharePoint, and Lync Server moved to the cloud, the company thought it was lifting servers; usage went 10x and 100x because customers had been buying a subscription, not counting machines. Work IQ now exposes what he calls the most important database in a company that was never used as a database — email, Teams, documents, meeting transcripts, trapped inside apps. He will open a GitHub repo and ask it to pull last week's related design reviews and tell him how to change the code. The agent reads the transcripts and comes back with a plan. Previously you would never have thought to use Microsoft 365 for that. Agents may soon generate more M365 usage than human employees, which means the architecture built to serve a mailbox cannot serve an agent. The system has to be rewritten.

On pricing he expects layers, not a winner. Per-user subscriptions survive because budget owners need certainty; they are really a bundle of usage entitlements. Consumption comes next. Outcome-based pricing will be loudly requested and quietly refused: "most people love outcomes until they have an outcome," because sharing the upside feels like giving away a royalty. GitHub Copilot was priced per seat for interactive complete-and-task use, before anyone imagined launching 10,000 agents that run all day, so the company added a consumption meter. Internally, he wants enterprises to live through one full budget cycle before declaring SaaS dead. Generate versus buy is a marginal-cost question once you count maintenance, the security holes AI will find, and the tokens required to fix them. There will be little patience for inflexible vendors — and a path back for anyone who shows up with value and a new commercial model.

Asked what he is building himself, Nadella was willing to be the punchline. Software has become so easy that even "the incompetence of a CEO" can ship. He still remembers joining Microsoft when everyone had to read Dave Cutler's code to learn C++. Now a CEO can inspect a repository and feel some agency over the artifact. Last week he wired Work IQ to a long-running Foundry agent, stored memory in a managed backend he called Raven, and published the result to Teams — a chief-of-staff autopilot, built end to end. The threshold, he said, is now low enough that an amateur at CEO altitude can complete a project like that.

The hosts floated a rumor of the future org chart: in four or five years, only four engineering roles remain — people who manage agents, forward-deployed engineers, security engineers, and a small crew on large-scale infrastructure. Nadella would rather experiment than predict, but he already sees the shape. LinkedIn collapsed design, product, and front-end into a "full-stack builder" role that keeps an edge in one craft while expanding scope. Infrastructure got harder, not easier: even the Excel team found that building a reinforcement-learning environment in which a reward can actually be learned is one of the gnarliest infrastructure problems they have, and an end-user app team now needs distributed-systems talent. Specialists will not vanish. The generalist, though, is where he expects the maximum return — the same person who used to produce a Word doc or a spreadsheet can now, in the same sentence, produce an app. Guo called it a golden age for idea people with agency.

Elad Gil, citing an essay by his partner Mike Vernal — who started at Microsoft — asked how a company already worth more than a trillion dollars gets more ambitious. Nadella quoted CTO Kevin Scott: making hard things easier is leverage; true ambition is making the impossible possible. What organizations lack is a new conceptual model of what can now be built. His favorite example is the Azure networking team, the same group that added more Azure capacity in the last 15 months than Microsoft built in its first 15 years. They looked at that curve and said it would not work unless they reconceptualized the job. They manage more than 500 fiber operators and a wide-area network where cables get cut and mailboxes fill with incidents. They built an agentic system, gave it a character named Miles, and started asking the company for tokens instead of headcount. They had taken their work and made it meta. If someone in the 1980s had said four billion people would wake up and type, the old model would have hired four billion typists. They were never typing. They were doing knowledge work.

Guo turned to the data-center buildout that is redefining what it means to be a hyperscaler — not only on the balance sheet, but in the towns that host the campuses. Nadella split the answer in two. The scale is extraordinary, and he feels lucky to be in it. The other half is non-negotiable: unless the industry is principled about benefits that communities can feel, there will be no permission to keep building. That cannot be a campaign. People have to see that energy prices are not going up — that a better grid and more supply bring them down over time — that water is replenished in closed loops, and that construction jobs, operating jobs, training, and tax base are real.

He welcomes skepticism and hard questions. Permission, he said, is earned or it is not; it is that simple. History is unkind to projects that consume a lot of energy and create little value, and kind to those that do both. This time the share of the economy is too large for "trust us, the future is going to be glorious." Tangible benefits have to show up — health outcomes, the ability to start a company, a local store run more efficiently — and he wants to see politicians winning elections because they advocated for those benefits. His window is 12 to 18 months. Wealth creation will spread across startups and incumbents; in healthcare he already points to companies such as OpenEvidence.

Education is the obvious good where he has not yet seen matching impact. He recently met the founders of Alpha School and found himself rethinking what education even is. A Stanford CS class's AI guidelines caught his eye too: students still had to learn how to apply softmax, not type "fix my training run." Concepts remain essential. What has to be rebuilt are incentives, credentials, how employers value those credentials, and the jobs they unlock. The way people get information, educate themselves, and stay current has changed too much for the old pipeline to hold. Maybe the next big startup, he said, is someone who builds a new university — or a new pedagogy — that takes a person through a curriculum and into highly valuable economic opportunity. That has felt impossible for a long time. It was the note they chose to end on: it might now be possible.