Something unusual is happening to Microsoft from two directions at once. A few days ago, Microsoft presented Copilot as 'a new OS for work', acknowledging the inversion of gravity that increasingly puts traditional workplace apps as 'satellites' of agentic workbenches. Meanwhile, a new generation of small, local AI models is quietly pulling a meaningful share of computing back onto the device—away from the Azure cloud business Microsoft spent fifteen years building. Neither move has cost the company anything yet. But the parallel shift on both ends could, eventually, cost it a great deal—at least reshape the ground Microsoft, and a large part of the corporate enterprise workplace, has been built on.
The sky has followed computing for decades. Windows 95 filled its four panes with blue sky and clouds; Windows XP left a green hill under an empty sky nobody had to explain. Later, the sky swallowed the machine itself: computing moved into "the cloud," somewhere distant and apparently weightless, the machinery hidden behind something almost celestial—a Rubens ceiling for the digital age. One company borrowed the blue for its name; another simply drew the cloud as its logo.
The New OS
On September 25, Microsoft unveiled its biggest overhaul of Copilot yet—Home, Code and Autopilot folded into one app—and Satya Nadella described it in a sentence that outweighs the product: "We're building Copilot as a new OS for work that spans every model, every form factor, and every task." Every form factor is the phrase to notice. If the new operating system spans every form factor, then Windows is considered one of the form factors from now on, and no longer the thing standing above them. The word 'Windows' was not mentioned at all.
Burying a forty-year-old franchise by simply not naming it is its own kind of announcement. An operating system, at bottom, decides what a person sees first when the machine turns on. Windows earned that role by choosing which application got their attention. Handing it to Copilot instead is either the most confident thing Microsoft has done in years, or the moment it walked away from the one asset no competitor can copy. Windows doesn't disappear. It goes back to being the frame around the picture, not the picture itself.
Leaving the Fortress
Naming the shift is not the same as owning it. Copilot's own recent history is one of renaming: Chat and Cowork just merged into something called Home—nine days after Anthropic merged its own Claude Chat and Cowork into one app, itself arriving two months after OpenAI folded Codex and Atlas into a single ChatGPT app alongside Chat and Work—and an agent called Scout was rebranded Autopilot within months of shipping. Work grounded through what Microsoft calls Microsoft IQ is a clean integration story on a slide; inside an actual company, where mail lives in one system, files in three others and the org chart nobody updated, it is a harder promise to keep.
That is the trade Microsoft is making. It is stepping out of the one place its advantage was uncontested—the operating system, installed on nearly every desk in the enterprise—to compete on a layer, conversational software, where Codex and Claude's Cowork got there first, and where its own products, Word, Excel, PowerPoint, are being folded in as capabilities inside somebody else's chat window rather than destinations of their own. Either the interface Microsoft is betting on turns out to be sharply, durably better than what a faster-moving field has already shipped, or the company has traded a long-standing fortress for a share of open ground. Microsoft may be burning a deep moat to compete on a shallow one.
The Second Intermediation
Microsoft wanted the clear sky too, once. It spent years pushing Outlook toward the browser, presenting Microsoft 365 as a subscription to intelligence that lived in Azure rather than software that lived on a machine. It never quite worked on its own terms: the browser never made the thick client irrelevant, and the mail kept doing what mail does, syncing onto a laptop that, in 2026, also carries an NPU, a GPU and more memory than anyone budgeted for.
That same machine is exactly where a new kind of software wants to live. A small, telling example: laya-mlx, an independent Apple Silicon port of the decision model Laya, answers a typed question—is this urgent, does it need a reply, can it be archived—in about seven milliseconds, entirely offline, on a laptop's own chip. Such a system, built-in in a thick client, could for example process any mail that lands on it, sort, flag or draft a first response without a single token reaching a cloud model, let alone Azure. We could call it a 'cloud-last' architecture: not anti-cloud, but cloud as the top floor, reached by escalation—literally, by taking the stairs up—rather than the ground floor everyone starts on.
Another signal from Google reinforces the pattern: in its latest release, Google's own Antigravity SDK now hands real coding work to a swarm of local Gemma models—not a demo toy: in one recorded run, a cloud model spent 95 tokens on planning while Gemma did the rest on-device, 97.2% of the run's tokens, reproducing and patching a security vulnerability without a line of code leaving the machine. Yet, Google has no thick client to defend and no legacy pulling it toward the cloud—Chrome OS was built, from the start, to erase the client altogether. So when the company with the least reason to defend the client puts its own frontier model family to work there, the shift looks less like a Microsoft problem than a change in where computing itself wants to sit.
There is a second cost folded into the same move. For years the cloud flattened this distinction: a mediocre corporate laptop and a high-end MacBook could reach the same model through the same browser, so memory, GPUs and NPUs barely mattered to what a person could actually do. Local inference reverses that. Microsoft's advantage in this story is the raw number of laptops running its software inside most large companies—but that fleet was built for spreadsheets and email, not for the neural cores that make local inference fast. Apple's machines had been built for different purposes, but happened to be remarkably well prepared for this one, and Apple has spent the past year turning that advantage into an on-device model and runtime of its own. Microsoft owns the installed base. Apple happens to own the machine the moment now rewards.
Installed vs Inhabited
One Microsoft product didn't need a rebrand to become, in practice, the place the workday starts. Inside plenty of large companies, that product is Teams—not because anyone called it an OS, but because the day already opens there. Teams is a different kind of asset than Windows or PowerPoint. Windows is installed, and so is PowerPoint; Teams is inhabited: closer to a social graph of the company than to a piece of software, and every channel is a small vein of context about who trusts whom, who is blocked on what, who actually made the decision the slide deck later took credit for.
That distinction matters more than it sounds. A tenant knows where a document lives; it does not know why it exists. Microsoft can point to Entra, SharePoint, Outlook and Teams and call the sum of it enterprise context—the pitch behind what it now calls Microsoft IQ—but the residue of an organization's work is not the work itself. The document is what's left after the decision; the conversation that produced it, who was asked, whose objection quietly settled it, is the part no repository stores. Teams sits closer to that layer than anything else Microsoft owns.
What Microsoft does with that isn't decided yet, and it doesn't strike me as a small choice. Turn Teams into a directory of agents sitting in every channel, and colleagues stop talking to each other and start performing for an audience that includes software. Inviting an agent into a conversation with a colleague is not the same gesture as opening an app. Nobody has quite worked out what happens to trust once a conversation might be read by something built to summarize it.
Storm Clouds, No Lightning Yet
« Quand le ciel bas et lourd pèse comme un couvercle (...) »
— Charles Baudelaire, Spleen, 1857
None of this shows up yet in a quarterly result. Microsoft 365 Copilot still sells for $30 a user a month, on top of the base license, with Copilot Credits metering anything heavier. Salesforce runs three pricing models on Agentforce at once—$125 a user a month, $2 a conversation, Flex Credits at $500 per 100,000—which looks less like a strategy than a company that hasn't decided what a unit of agentic work costs. Inside Microsoft alone, three businesses want three different things. Azure wants the computation in the cloud. Windows increasingly needs it on the device. Copilot wants to become the layer that decides between them. All three report to the same CEO.
So the weather report, officially, is fine. No earnings call has mentioned a hole where Azure revenue should be, no analyst has downgraded the stock over a mail sorter running on somebody's laptop. But the sky Windows opened onto forty years ago was never actually empty—it only waited long enough to look that way. It is filling back in now: with models, with agents, with clouds gathering at a distance, in the blue that gave Azure its name. Whether the storm passes to one side or breaks directly overhead is exactly the part nobody, including Microsoft, has decided yet.
