As OpenAI and Anthropic IPOs near, Microsoft’s Satya Nadella tells every company why frontier AI models aren’t the future
As OpenAI and Anthropic march toward what could be the largest IPOs in history, Microsoft chief executive Satya Nadella has stepped in with a message that cuts against the whole premise of the moment. In a long post on X over the weekend, Nadella argued that the companies that win the AI era won’t be the ones who pick the best model. They’ll be the ones who build something a model can never sell back to them—a learning system trained on their own work, judgment, and institutional memory. The frontier model, in his telling, is just the engine. The car is what you build around it.The timing makes it sting. Anthropic and OpenAI filed confidentially for their IPOs within a week of each other, carrying private valuations of roughly $965 billion and $852 billion. The entire pitch to investors rests on the idea that frontier models are the prize. Nadella is telling every company that the prize sits somewhere else.
Two kinds of capital every company will need
His case turns on a split between what he calls human capital and token capital. Human capital is the knowledge, judgment, relationships, and pattern recognition your people carry. Token capital is the AI capability a firm actually builds and owns. The mistake, he warns, is assuming the second quietly erodes the first. “Human capital does not become less valuable as token capital grows. It only becomes more valuable,” he wrote, adding that without human direction, “you have compute running in circles.“The thing binding them is what Nadella calls a learning loop—a system that captures every interaction, correction, and outcome, then feeds it back so the AI gets sharper at your specific business. Picture a sales team whose AI drafts proposals. Without the loop, reps edit eighty out of a hundred drafts every month because the model keeps missing the pricing logic, and next month it misses again. With the loop, the system learns why one rep always leads with a particular objection and how your bundles really work. By the five-hundredth proposal, it barely needs editing. That accumulated judgment becomes proprietary IP no competitor can download from anyone’s website. Nadella calls it a “hill climbing machine,” an asset that compounds rather than a subscription that renews.
A warning borrowed from the globalization playbook
The sharpest part of the argument is political. Nadella drew a direct line back to the first wave of globalization, when outsourcing flattered GDP figures on paper but hollowed out entire industrial economies, with consequences still being felt. He doesn’t want AI to run the same script. “The last thing any of us wants is a world where every company across every sector is ceding value to a few models that eat everything they see,” he wrote. “If all the value is accrued by only a few models, the political economy will simply not tolerate it.”It’s a tidy argument, and a self-serving one, which Nadella doesn’t bother hiding. Microsoft wants enterprises building these loops on Azure, fine-tuning there, parking proprietary data there, and making any future switch expensive. As SFGate noted, the framing pulls the conversation away from “who has the best model,” where Microsoft leans on partners like OpenAI, toward “who built the smartest system,” where Microsoft happens to sell the infrastructure underneath. Days earlier, the company had launched seven of its own MAI models and a “Frontier Tuning” pitch built on almost exactly this logic.
Not everyone is buying the loop
There are skeptics. OpenAI’s counter-bet is simpler—keep making the base model good enough that elaborate loops stop being worth the trouble, and just write a better prompt. And building a real learning loop is genuinely hard. It means solving three problems at once: the infrastructure to capture and retrain on live data, the governance to turn messy proprietary conversations into clean training material, and the discipline to keep checking whether the model actually improved or just memorized. Plenty of teams have rushed to fine-tune, rented pricey GPUs, then realized a sharper prompt would have done the job by lunchtime.Nadella reinforced the message with a jab at “tokenmaxxing,” the reflex to throw the most powerful model at every task, telling staff not to use frontier models for non-frontier problems. With three trillion-dollar listings now asking the public to bet that frontier models are the future, one of the industry’s most powerful figures is quietly making the opposite case—that the real moat is everything those models can’t see.
