Microsoft CEO Satya Nadella’s argument that businesses need to be able to easily switch between artificial intelligence models is correct but elides the fact, if implemented, his advice would direct investment toward Azure, Microsoft’s cloud business, which benefits regardless of which model wins. To avoid similar market abuse by cloud providers, portability at that level must exist, too, writes Javaid Sofi.
On June 14, Microsoft CEO Satya Nadella posted a long argument on X about how companies should think about maximizing value from artificial intelligence. The danger, he wrote, is that a few large models, like those from OpenAI and Google, capture most of the value going into AI and starve the firms and users that rely on them, the way an earlier wave of outsourcing hollowed out manufacturing towns. A few of these companies have made headlines for asking workers to reduce their use of AI because of burgeoning costs.
Nadella’s advice is that every company should build a “learning loop”: the system around a model that turns a company’s use of AI into an asset of its own. That includes its data, the record of which answers worked and which did not, and the tests and workflows its staff develop. As employees use the tools and correct them, the accumulated data and feedback make the system better at the company’s particular work. A company can do some of this itself by keeping those assets outside any one model and building a software layer that can connect to several models. However, that does not make models fully interchangeable. Models behave differently and use different formats, so switching can still require testing and rebuilding. Even so, Nadella’s point is that a company that owns the learning loop around the model is less likely to be held hostage by any one model vendor in the case that they raise prices or change the terms and conditions.
Much of this is sensible, but ask who benefits when the advice becomes the default way companies think about commercial AI. Making foundational models swappable pushes the value down the stack, to the compute where models run and the tooling where a company’s data is stored and put to work. This is the part that Microsoft sells through its cloud service, Azure. A warning about concentration and market power at the model layer quietly redirects the value toward the infrastructure layer beneath it: the layer where Microsoft is strongest, whichever model wins. The way Nadella defines the danger also narrows the range of solutions companies are likely to consider.
The cheapest part of the stack
Think about what “swap one model for another” does to the price of a model. If a buyer can replace your product with a rival’s at low cost, it becomes harder to charge a premium. As models grow more interchangeable, their providers lose pricing power. Nadella is telling every company to treat foundation models this way.
There is an old name for this: commoditize your complement. A complement is something customers buy alongside your product: the way cars need fuel, or game consoles need games. When the complement gets cheaper, your product becomes more attractive, and you keep more of what the customer is willing to spend on the combined products. Microsoft has run this move before. By licensing its operating system MS-DOS non-exclusively to every PC maker that wanted it, it helped turn the machines themselves into cheap, interchangeable boxes whose makers drove each other down on price. The hardware became a commodity, the operating system did not, and that is where the money went.
In Nadella’s vision for AI, the models will be the complement. They sit on top of compute, alongside the tools and data each company builds around them. Make the model cheap and swappable, and its providers lose their pricing power. The value flows to these other layers instead.
Where the money goes
Start with compute. Training and running large models take vast, specialized data centers to provide that cloud processing power, and only a few firms can build them at scale. In January 2025, Microsoft said it was on track to invest roughly $80 billion during that fiscal year to build AI-enabled data centers, and its spending continued to accelerate: capital expenditures, including finance leases, reached $37.5 billion in a single quarter by early 2026. That scale of capital is a barrier to entry for potential cloud-computing upstarts, and it helps explain why only a handful of firms compete seriously in cloud infrastructure. Three companies hold most of the market: Amazon Web Services at about 28 percent, Microsoft at 21 percent, and Google at 14 percent. Oracle is next at 4 percent. The same three firms continue to dominate even as demand surges. Microsoft told investors that Azure demand exceeded its available capacity. Capacity constraints favor the few firms able to finance infrastructure at this scale. Whatever happens in the model race, much of the compute beneath it will still run through a small group of cloud providers, and Microsoft is one of them.
Then the tools. A learning loop is the plumbing around a model: a place to keep your data, and the machinery to train models on it and put them to work. Microsoft sells that plumbing through Azure, billed by use. As a company builds that loop out, it leans harder on the surrounding services, regardless of which model sits inside.
The apparent paradox of Nadella’s proposal is that Microsoft remains heavily invested in particular models. Following OpenAI’s October 2025 recapitalization, Microsoft’s stake was valued at approximately $135 billion and represented about 27 percent of the company. Microsoft also offers its own models. On June 2, less than two weeks before Nadella’s X post, it unveiled seven in-house models at its annual developer conference.
But Microsoft’s larger disclosed business sits below the models. It does not report revenue from Copilot or its own models separately, but one estimate puts its revenue from Copilot at $2-4 billion per year. Azure, by contrast, surpassed $75 billion in revenue in fiscal year 2025, up 34 percent. That scale helps explain Foundry. Microsoft placed its new in-house models in the same catalog where customers can discover, compare, and deploy more than 11,000 models from providers, including OpenAI, Anthropic, Meta, Mistral, DeepSeek, and xAI. Foundry offers customers a wide choice of models, but the surrounding tools, billing, and deployment remain tied to Azure. Microsoft still has reason to compete in models, especially when using its own models reduces what it pays outside providers. But whichever model a Foundry customer chooses, Microsoft still supplies the platform around it.
Microsoft’s relationship with OpenAI has also loosened. Under an amended agreement announced in April 2026, Azure remains OpenAI’s primary cloud and OpenAI products will appear there first. But Microsoft’s license to OpenAI technology is now non-exclusive, and OpenAI can serve its products through other clouds. Microsoft remains a major shareholder. The point is not that Microsoft no longer cares which model wins. It is that Azure and Foundry can support models from many suppliers, including Microsoft itself.
What owning the loop does not fix
“Own your loop” sounds like independence. But when a company builds its datasets, applications, and agents in a public cloud, they become wired to that provider’s storage and tools. Each provider-specific service can make a later move harder. Shifting a mature system to a rival cloud can mean rebuilding pipelines and retraining the people who run them. The barrier is well documented: after a two-year investigation, the United Kingdom’s competition regulator concluded that technical barriers, egress fees, and software licensing lock cloud customers into their existing providers. So, the freedom to swap models comes with deeper dependence on the platform beneath them, something Nadella’s advice never mentions.
Go back to his outsourcing comparison, because it does more than Nadella means it to. The lesson he draws from it is that firms should own their loops so the value does not drain away.
But look at who kept the value during the earlier wave of globalization. A World Intellectual Property Organization study found that Chinese labor captured about one percent of the price of an iPhone 7, while Apple captured 42 percent. The factory was essential, but Apple owned the design, the software, the brand, and the relationship with the customer. Read that way, the comparison cuts against Nadella’s own framing rather than for it. If models become interchangeable, they start to look like the factory, and the cloud starts to look like the port that everyone else’s work has to pass through. By naming the danger at the model layer, Nadella makes a fix there look sufficient. It also directs companies toward the cloud, where Microsoft is strongest whichever model wins.
His advice is still worth taking. A company that can move between models is less exposed to any one AI lab’s price increases or sudden changes in terms. It can keep its data and its evaluations in portable formats, avoid tools that lock it to a single model, and test alternatives often enough to keep the option real. The same discipline should reach one layer down, to the cloud beneath the models: limit the services that run on only one provider and keep a second one in use where that is practical.
But there is only so much a customer can do alone. No single company can make its cloud provider easy to leave, and the costs that hold it in place are not ones it set. That is where regulation comes in. The European Union’s Data Act requires cloud providers to help customers move, remove technical and contractual obstacles to switching, and, from January 2027, stop charging fees for taking data out. Britain’s competition regulator identified many of the same problems. Its inquiry recommended that the regulator consider strategic-market-status investigations of Microsoft and Amazon. The CMA has relied for now on steps the companies agreed to take on egress fees and interoperability, It has opened a separate investigation into Microsoft’s business-software ecosystem. That investigation could address licensing practices that make Microsoft software more expensive or less attractive to run on AWS or Google than on Azure.
Clouds do not need to become identical. Customers simply need to be able to leave. That would force providers to compete for their next workload on price and service instead of relying on the cost of moving the last one. Businesses would keep more bargaining power, and more of the gains could reach their customers.
Nadella is right that a company should own what it learns from AI. But ownership counts for little if the data and the systems built on it cannot leave the cloud where they were made. The real test is not whether a company can swap one model for another. It is whether it can lift the whole system around the model and move it. Until it can, the value Nadella says he wants to spread will keep pooling in the small group of clouds beneath the AI economy, including the one Microsoft has spent years and tens of billions building.
Author’s Disclosure: The authors report no conflicts of interest. You can read our disclosure policy here.
Articles represent the opinions of their writers, not necessarily those of the University of Chicago, the Booth School of Business, or its faculty.
Subscribe here for ProMarket’s weekly newsletter, Special Interest, to stay up to date on ProMarket’s coverage of the political economy and other content from the Stigler Center.
