Microsoft spent billions on OpenAI — now it's building its own way around it

Information Technology 3 min 16 sources

At Build 2026, Microsoft unveiled seven in-house AI models, a Windows-level cage for AI agents, and a new assistant — a clear move to lean less on the partner it helped make famous. DeepSeek lined up a $7bn raise, and Europe's power grid emerged as the real limit on the AI boom.

Key takeaways

At Build 2026, Microsoft unveiled seven of its own AI models and a new assistant — a quiet move to lean less on OpenAI, the partner it helped make famous.
It also built a Windows-level cage to contain AI agents, while cheaper rival DeepSeek lined up a $7bn raise.
The real ceiling on the AI boom isn't chips — it's the power grid.

The biggest tech event of the day was a developer conference, and the subtext was a quiet divorce.

Microsoft builds its own brain

At Build 2026, its annual conference for software developers, Microsoft’s keynote was almost entirely about artificial intelligence [4]. The headline: the company unveiled seven of its own AI models, including a flagship reasoning model called MAI-Thinking-1 — its first model built to “think” through hard problems step by step rather than answer in one shot [1].

Why this matters goes beyond the models. Microsoft has poured billions into OpenAI, the maker of ChatGPT, and has leaned on OpenAI’s technology to power much of its AI. These new in-house models are explicitly designed “to lessen reliance on OpenAI and lower costs for developers” [3]. In plain terms: the company that helped make its supplier famous is now building its own version so it depends on that supplier less. When you both fund a partner and could be undercut by it, owning your own option starts to look less like a luxury and more like insurance.

A cage for the agents

The most consequential release for people who actually run software was MXC — Microsoft Execution Containers. It’s a control layer built into Windows itself that lets a company spell out exactly what an AI “agent” is allowed to do: which files it can touch, which programs it can run, what it’s forbidden from [2]. An agent here means an AI that doesn’t just answer questions but takes actions on its own — opening apps, moving files, completing multi-step tasks.

For two years the industry raced to make these agents more capable while largely dodging the question that worries security chiefs: what happens when one goes wrong [2]? MXC is Microsoft’s attempt at an answer — a sandbox, or sealed playground, where an agent can act but can’t reach beyond the fences you set. OpenAI and Nvidia have already signed on [2]. If you build or secure software, this is the release worth reading closely; agent safety is shifting from a slogan to a setting you configure.

Microsoft also launched Scout, an AI assistant pitched as a coworker “that never logs off,” and a tool that lets developers write plain-English descriptions to generate tests for how an AI should behave [7][16][14]. Not to be outdone, OpenAI released new versions of its Codex coding tools aimed at white-collar office work [8].

The cheaper challenger raises big

The competitive backdrop sharpened. DeepSeek — the Chinese AI lab that rattled the industry by building capable models far more cheaply — is lining up its first outside fundraising, slated to draw about $7 billion [5]. That number is a marker of how much money is chasing the idea that the winning model won’t necessarily be the most powerful, but the one that’s cheapest to run at scale. Microsoft’s own push to “lower costs for developers” is aimed at exactly that fight.

The investment frenzy spilled past AI, too. Rocket-engine startup Impulse Space raised $500 million at a $4.26 billion valuation, and fusion-energy startup Focused Energy pulled in a $240 million round for laser-driven fusion [10][9]. The money is hunting the next hard-tech bet.

The story nobody’s covering: the grid is the real ceiling

Underneath the model launches sits a limit no keynote can wish away. Big Tech’s AI ambitions are now “a major power test for Europe” — the data centres that run these models consume enormous amounts of electricity, and Europe’s grid may not be able to feed them all [6]. A data centre is a warehouse of computers; training and running large AI models can draw as much power as a small city. The race everyone covers is about models and money. The race almost no one covers is about megawatts — whether there’s enough electricity, cooling, and grid capacity to physically run what’s being built. If you want to know which AI projects will actually get built over the next few years, watch the power contracts, not the demos. That constraint may decide more than any benchmark.

02 · Lesson · why it matters

Why companies build the thing they could just buy

Depending on one supplier is cheap until the day it isn't — and the bill comes as lost leverage, not a line item.

A strange thing to spend billions on

Microsoft already had access to some of the best AI in the world. It had funded OpenAI, wired ChatGPT into its products, and could simply keep paying for it. Today it announced seven of its own models instead — built, in its own words, to “lessen reliance on OpenAI.”

On the surface that’s wasteful. Why spend a fortune rebuilding something you can rent? The answer is the whole lesson, and it has almost nothing to do with AI.

Renting hides a second price

When you buy a capability from one supplier, you pay two prices. The first is obvious: the money on the invoice. The second is invisible and grows quietly: dependence. Every month you rely on that supplier, you get a little more locked in. Your products are built around their tool. Your staff learn their system. Switching gets harder. And the supplier knows it.

That second price doesn’t show up until the supplier uses it. Then it arrives all at once — as a price hike you can’t refuse, a change in terms you didn’t want, a feature they give your competitor first, or worst of all, the supplier deciding to sell what you sell and become your rival. The cheaper option you chose years ago turns out to have a clause written in invisible ink: I can squeeze you later.

Leverage is the real currency

Economists call this the danger of a single point of dependence. The deeper idea is about leverage — who can hurt whom, and how much. When you have one supplier for something essential, they hold leverage over you. The fix isn’t always to stop buying from them. It’s to have a credible alternative, so they can’t squeeze you without losing your business.

That’s what Microsoft is really buying with its own models. Not just cheaper inference — a seat at the table where, if OpenAI’s price or priorities turn against it, it can walk. The in-house models may never fully replace what it buys. They don’t have to. Their existence changes the negotiation. The mere fact that you could leave is what stops a supplier from squeezing you.

When it’s worth the cost — and when it isn’t

This doesn’t mean build everything yourself. Most things you should happily rent: the supplier does it better and cheaper, and the dependence is harmless because alternatives are plentiful. You don’t manufacture your own electricity or write your own spreadsheet software.

The calculation flips when two things are both true. First, the thing is core — if it fails or gets cut off, your whole business stumbles. Second, the supplier is concentrated — there are few alternatives, and they could become a competitor. AI is exactly that for Microsoft: core to its future, and supplied by a partner who is also, increasingly, a rival. That’s the combination that justifies paying to build what you could buy. Core plus concentrated equals own it.

The pattern to carry

This reaches into ordinary life. A freelancer who gets ninety per cent of their income from one client is in Microsoft’s position before today — comfortable, efficient, and quietly at that client’s mercy. A country that buys all its gas from one neighbour has handed that neighbour a lever it can pull in a crisis. A team that depends entirely on one person who “owns” a system has a single point of failure wearing a friendly face.

In each case the cheap, easy path is to keep renting from the one source. And in each case the wise, expensive move is to build a second option before you need it — not because the supplier is evil, but because dependence itself is the risk. The question to ask about anything you rely on is simple: if this one source turned against me tomorrow, how badly would it hurt — and do I have anywhere else to go? If the answer is “a lot” and “nowhere,” you’ve found the thing worth building yourself.

03 · Lab · your turn

Rent or Build

Rehearse the dependence trap: renting is cheaper every round until the supplier's leverage turns, and only the costly habit of building an alternative leaves you a way out.