Microsoft starts replacing its partners' AI models with its own

Information Technology 4 min 80 sources

Microsoft is routing its products onto homegrown models that cut serving costs up to 89 percent, treating OpenAI and Anthropic as swappable parts. As the frontier model turns into a commodity, the money is moving to chips, distribution, and unglamorous suppliers.

Key takeaways

Microsoft is swapping its partners' AI models for its own homegrown ones, cutting the cost of running some features by up to 89 percent and treating OpenAI and Anthropic as parts it can route around.
As frontier AI models get cheap, open, and copyable, they stop being where the money is; the profit moves to the next scarce layer, chips, distribution, and specialized suppliers.
The quiet winners of the AI boom include Japanese firms that make toilets and food seasoning, because they also make hard-to-copy parts the chip supply chain depends on.

Microsoft starts swapping out its partners’ AI

Microsoft released two new in-house models on Wednesday and, more tellingly, published the production numbers behind them [18]. The models are a high-end image generator and a fast, cheap voice model built for high-volume enterprise work [18]. But the news is not the models. It is where they now run: Bing, PowerPoint, OneDrive, Excel, Dynamics 365, GitHub Copilot, and Azure [18].

Microsoft says the switch cuts the cost of running these features sharply. In PowerPoint, its own image model reduces graphics-processing-unit costs up to 84 percent versus OpenAI’s image model [18]. In its Dynamics 365 call-center product, used by customers including T-Mobile and EasyJet, it claims voice costs down up to 89 percent [18]. GPUs are the specialized chips that run AI, and renting them is the single biggest cost of serving a model. One in-house coding model even runs on Nvidia’s two-generation-old chips instead of the newest ones [18].

Chief executive Satya Nadella framed this in a post titled “Frontier Diffusion and Control” [18]. His argument: the smartest capabilities of a year ago are now ordinary, so Microsoft can copy them cheaply for the repetitive tasks that fill real product use [18]. Frontier models from OpenAI and Anthropic stay in the mix, he said, but as “interchangeable components” the company routes around whenever its own models match them [18]. Reuters reported in April that Microsoft’s exclusive license to OpenAI’s technology had already been loosened to a non-exclusive one [18].

The frontier model is turning into a commodity

Microsoft is not alone in reading the same shift. On Friday, Anthropic released Claude Opus 5, priced to deliver most of the intelligence of its costliest model at half the cost [58]. The company’s own framing was that the race is moving “from raw capability to the economics of daily use” [58]. Across Silicon Valley and Washington, the concept everyone suddenly cares about is distillation: training a small, cheap model to mimic a big, expensive one [25].

The open-source side pushes the same way. Chinese lab Moonshot has been releasing capable models for free download, closing the gap with US rivals despite chip export controls [23]. And a bloc of American giants, Nvidia, Microsoft, and Meta among them, signed an open letter warning against “premature restrictions” on open-weight models [19][21]. Those are the ones anyone can download and run. At a summit in China, the US and other nations backed open-source AI too [4]. Open weights mean the software layer is close to free.

Here is the thread tying these together. When the hardest, most expensive thing to build, the frontier model, becomes cheap, standard, and copyable, its price does not stay high. The value has to go somewhere else.

Where the money went instead

It went, first, to the chips. On Saturday, Samsung said it struck a deal with Broadcom to widen chip cooperation to more than 200 billion dollars through 2030 [35]. Intel, left for dead a year ago, put out a forecast that beat estimates on AI server-chip demand, and its shares jumped [40]. It also committed to mass-producing its most advanced manufacturing process in 2028, with sales up 25 percent from a year earlier [22]. A chip startup called Etched, which builds hardware for one narrow AI task, hit a 10.3 billion dollar valuation [1]. AMD unveiled a full server-rack system to challenge Nvidia head-on [70].

It went, too, to memory. Reuters reports that China’s CXMT and YMTC, which make the memory chips AI systems hoard, are flexing new pricing power as demand outruns supply [41].

The pattern is the same one Microsoft is acting on. The layer that was scarce, the smartest model, is filling up with cheap substitutes. So the scarcity, and the profit, moves to the next bottleneck: who can make the chips, who owns the reach to a billion users, who runs the plumbing.

Brussels fines Google, Washington threatens back

Separately, the EU fined Google 890 million euros, about a billion dollars, for breaking competition rules across Search and its app store, and ordered changes to both [67][20]. President Trump said the US would open a trade investigation in response, calling the fines illegal and a targeting of American companies [44][76]. It is a live test of whether Brussels can still discipline US tech giants when Washington is willing to hit back on trade.

The toilet-maker and the seasoning company

The clearest sign of where AI money travels is not in Silicon Valley at all. Three of this year’s quieter winners are Japanese firms that make, respectively, toilets, glass fiber, and food seasoning [77].

Toto, the toilet company, makes the ceramic parts that hold silicon wafers steady inside chip-making machines [77]. Nittobo makes the fine glass fiber woven into circuit boards [77]. And Ajinomoto, best known for the flavor enhancer MSG, makes a special insulating film, called ABF, that goes into the packaging around high-end chips [77]. Shares in two of them are up more than 60 percent this year [77]. None of them designs a chip or trains a model. They just happen to sit on a narrow, hard-to-copy layer that the whole boom now depends on. That is where the profit went.

02 · Lesson · why it matters

Why the money leaves the thing everyone is chasing

When the hard part gets easy, its value does not disappear. It slides quietly to whatever is still hard.

The prize everyone chased is turning cheap

A year ago, the most valuable thing in technology was the frontier AI model, the smartest system money could train. Companies poured tens of billions of dollars into building one and guarding it.

Now look at today. Microsoft copies those capabilities in-house for a fraction of the cost. Open models give away work that was a trade secret last spring. The industry’s new obsession, distillation, is a method for shrinking a big expensive model into a small cheap one. The thing everyone fought to own is becoming ordinary.

A price is paid for scarcity, not effort

It is tempting to think a thing is worth a lot because it was hard to make. That is not quite right. A thing commands a high price because it is scarce and needed at the same time.

The two can come apart. The frontier model took billions of dollars and rare talent to build. But once cheap substitutes exist, the scarcity is gone, and the premium goes with it, no matter how much genius went in. Effort earns your admiration. Only scarcity earns the money.

The money does not vanish. It moves.

Here is the part that is easy to miss. Demand for the whole system, AI features running inside real products, is as strong as ever. So when one layer goes cheap, the profit does not disappear. It moves to the next layer that is still a bottleneck.

If the model is no longer scarce, then scarcity, and money, shift to whatever is. The chips that run it. The reach to a billion users. The plumbing that ties it together. The one supplier no one can replace. The bottleneck moves, and the money follows it like water finding low ground.

Whoever owns the bottleneck owns the margin

Read Microsoft’s move plainly. It is not trying to win the layer that is commoditizing. It is doing the opposite. It treats the frontier models as parts it can swap out, and owns the layer that is still scarce: the products a billion people open every morning.

The same logic explains a stranger fact. A Japanese company that makes toilets is a winner of the AI boom, because it also makes a ceramic part that chip machines cannot run without. It does not design a chip or train a model. It simply sits on a narrow layer no one else can supply. That is enough.

The map was drawn by choices, and it can be redrawn

Which layer counts as “the scarce one” is not fixed by nature. For two years the model was king, and the chipmakers, the cloud providers, the obscure suppliers were the supporting cast.

That order is being rewritten right now, on purpose. Someone chose to build cheap substitutes for the crown jewel. The moment they did, the map of who holds the advantage started to redraw itself. What looks like the natural shape of an industry is really a temporary settlement, and settlements get renegotiated.

You are standing on a layer too

This is not only a story about companies. Every skill, every job, every small business sits on some layer of some larger system. It is valuable while it is scarce.

When tools, training, or a cheap machine make that skill common, the premium leaves, however honestly you earned it. The developer whose careful way of prompting a model was rare last year. The worker whose task a cheap model now handles. They did nothing wrong. The thing they stood on simply stopped being scarce.

So the comforting line, “I am worth it because I do the hard thing,” turns out to be only half true. The other half is whether the hard thing is still hard to get, and that is decided somewhere you do not sit. The map of where value lives is redrawn constantly, by choices made in rooms most of us are not in. No one, not Microsoft, not the seasoning company, not the reader, can see the whole board or keep their square forever.

03 · Lab · your turn

Follow the Money

Rehearse choosing which layer of a stack to own as scarcity, and the profit, migrates from one layer to the next.

04 · Hope · carry this

What a handful of billion-dollar labs guarded a year ago is now cheap enough for a small team, a distant lab, or an unlikely supplier to build on. Power that spreads this fast rarely stays locked in a few hands for long.