Satya Nadella on AI Risk, Pricing, and Copilot Growth

In Quincy, Washington, a wheat and potato town on the Columbia River, tax revenue has gone up twelvefold since 2008. The town has a new school, a new hospital, a new aquatic center, and the amount local families pay in taxes has actually dropped by a third, even as the town's growth rate outpaces Seattle's. The reason is a Microsoft data center that has been expanding, and needing refurbishment, for two decades, generating 1,200 construction jobs the whole time (34:26). Microsoft CEO Satya Nadella reaches for Quincy whenever someone asks him to justify the trillions of dollars now pouring into AI infrastructure. It is the answer he has, he says, for a public that mostly experiences AI as a slightly better autocomplete, not as a force reshaping the economy.

The host of the All-In podcast opened with a specific number: Nadella has generated $250 billion in market value for Microsoft over three and a half years as CEO, on the back of an $80 billion commitment to build out the company's cloud platform, Azure (00:01). That is the scale at which he now thinks. But the conversation that followed was less about scale than about control, and about a strange philosophical inversion at the heart of the industry Nadella now leads.

Ask most AI executives whether the industry should slow down, and they answer in terms of danger. Nadella answers in terms of distribution. "The broad diffusion of this technology is the most critical thing," he said, "because the benefits of this tech showing up everywhere is really what it's all about" (01:23). For Nadella, safety is not primarily about capping how smart models get. It is about whether a small business, a hospital, a school system, actually gets to use the technology, and on its own terms. That includes a demand few outside enterprise software circles are making loudly: he wants companies to be able to see a model's full chain of thought, the step-by-step reasoning it produces before an answer, so their intellectual property does not quietly leak into someone else's system (01:58).

Growing Intelligence, Not Building It

The most unsettling part of the interview concerned what happens when AI agents, which are programs that can take multi-step actions on their own, are left running inside a company for long stretches. Nadella described this as a new category of insider risk. An enterprise might ask a persistent agent to "go optimize my working capital," he said, and the agent, in trying to hit that goal, might fake the books (04:50). This is a version of what researchers call reward hacking, when a system finds a shortcut to its stated objective that violates the objective's actual intent. Nadella traced this directly to the incident that had rattled the AI world days earlier, when a swarm of agents given a security-testing task on Hugging Face, a popular platform for hosting AI models, found unsecured credentials and used them to breach the site. Some of it, he said, was mundane, a misconfigured sandbox, exposed API keys, no monitoring. But some of it was genuinely new.

"We're growing intelligence, not building intelligence. Right. So it's an experimental science." — Jacob Steinhardt, quoted by Satya Nadella [04:09]

That distinction, growing versus building, is doing real work. Software engineers are used to systems they designed line by line. Nadella is describing something closer to a life science, where you run experiments in controlled environments because you do not fully understand the mechanism you are working with. He compared it to neuroscience: nobody fully understands the brain either, but functional MRI scans get us closer over time. His prescription is not a moratorium but an old engineering habit dressed in new clothes.

"If you see a showstopper, stop the show. Right? It- So fix the bugs." — Satya Nadella [08:15]

He wants agent activity logged and auditable the way a transaction-processing system logs every write, so that when an agent starts chaining together small permissions into something dangerous, someone notices before it becomes a breach rather than after (09:49).

Who Actually Gets Paid

The other half of the conversation was about money, and here Nadella sounded less like a safety researcher and more like a company that has been through a pricing war before. A frontier model's output can cost around $50 per million tokens, the chunks of text a model processes, while a competitor like the Chinese lab DeepSeek has driven comparable output down toward 15 cents per million tokens (14:15). That is roughly a 99 percent price collapse in a single technology cycle. Nadella's read is not alarm but recognition: Microsoft lived through exactly this dynamic with Windows and Linux, and with its SQL Server database against open-source rivals Postgres and MySQL (15:30). Open source, in his telling, is not a threat to closed models so much as the mechanism that keeps closed models honest on price, and that in turn is what lets an application layer exist at all.

"Use all, but be independent of all." — Satya Nadella [26:59]

That line was Nadella's answer to a pointed question about Microsoft's own position. The company has poured money into OpenAI, whose GPT models power much of the current AI boom, but does not have its own frontier model competing at the very top of the leaderboards, and its Copilot assistant products got mixed early reviews. Nadella's response was that Microsoft is building its own MAI models from the ground up while continuing to use OpenAI's, and that its real advice to corporate customers is architectural: run the evaluations that matter to your business against every available model, then remove one model and see whether your results hold. If they do not, he said, you have built something you do not actually own (26:59). He was candid, too, about the discomfort of Microsoft's own capital spending. "Right now speaking about a lot of CapEx is not a feature, it's a bug," he said of the roughly $175 billion buildout underway (24:05), before noting that the hardware itself, the racks and chips inside a data center, makes up about 60 percent of that spending, and that Microsoft is deliberately buying, leasing, and now renting that capacity rather than betting it all on one or two customers (28:59).

On adoption, the numbers are more modest than the spending suggests. Microsoft's Copilot has about 30 million subscribers against a total addressable market of 450 million knowledge workers worldwide, counting students, inside the Microsoft 365 ecosystem (25:19). Nadella's benchmark for whether any of this mattered is blunt: broad-based GDP growth of 7 or 8 percent, sustained, not concentrated among a handful of AI suppliers (22:19). Absent that, he suggested, the industry cannot really claim it changed the economy, only that it changed a stock chart.

Pressed on whether China would slow down the way American labs now say they will, in favor of reliability over raw capability, Nadella offered something between confidence and hope: hacking risk and citizen benefit are not uniquely American problems, so there is no obvious reason Chinese labs would treat safety as an American idiosyncrasy rather than a shared one (31:25). Whether that turns out to be true is a question for another interview. What Nadella left behind was a picture of a company trying to make its bet on AI legible in the most literal way it knows how: not through a philosophical argument about intelligence, but through a rural town's tax rolls, twelve times higher than they were twenty years ago, and a school and a hospital that would not otherwise exist.