Microsoft lost its way in the AI race. Can Copilot get it back on course?

The crew is engaged on a brand new AI product, one which capabilities as a private assistant, able to doing all the pieces from reserving flights to responding to emails to discovering a superb native plumber. They know competing groups at different firms are engaged on comparable merchandise. As in the event that they wanted a reminder that quite a bit is driving on their work, Satya Nadella drops by. He desires to point out them one thing.

The Microsoft CEO opens a laptop computer and fires up an software. It’s a type of system for instructing and controlling a number of AI brokers. He calls it “Chain of Debate.” As Nadella walks them by the demo, the engineers commerce figuring out appears to be like, the type regulars on the native basketball courtroom trade after they understand a beginner’s bought sport. As a result of Nadella didn’t get somebody to construct this app for him. He created it himself, vibe coding with an AI software. “That set the tone for a way exhausting the crew was going to push,” recollects Jacob Andreou, the manager vp liable for the design of Copilot, Microsoft’s AI assistant. He was within the room with of us, like over their shoulder, there along with his machine out.

Watching the boss get such pleasure out of constructing new issues impressed the crew. It wrapped up its massive push in late February when it rolled out Copilot Duties, the computer-using private assistant AI software. (Nadella’s personal prototype served because the mannequin for a function referred to as the mannequin council in addition to different parts of Copilot.)

However the truth that Nadella is spending a lot time with the groups constructing AI merchandise, even rolling up his sleeves and constructing prototypes himself, says quite a bit about Microsoft’s present predicament. In spite of everything, it is a $3 trillion firm, not some scrappy startup the place the CEO routinely logs on for coding sprints with the builders. Nadella is worried sufficient concerning the firm’s AI technique that final October he introduced he was stepping again from some business duties to concentrate on AI analysis, product innovation, and the build-out of AI information facilities.

There’s definitely purpose for concern. Microsoft’s inventory has endured a troublesome run. After hitting an all-time excessive in October, over the following 5 months Microsoft’s share value fell some 34%, regardless of its cloud-computing platform Azure’s AI-related revenues having greater than doubled prior to now 12 months. Microsoft has been a distinguished sufferer of the SaaSpocalypse, the selloff of software program shares precipitated by the appearance of AI coding brokers. Many buyers are satisfied these merchandise imply companies gained’t purchase AI choices from software-as-a-service (SaaS) distributors reminiscent of Microsoft — that, maybe, they gained’t purchase off-the-shelf software program in any respect.

Gross sales of Microsoft’s enterprise Copilot merchandise have been slower than the corporate would really like. Lower than 4.5% of the 450 million prospects of its Microsoft 365 workplace suite presently pay for Copilot options. Utilization of its consumer-facing Copilot chatbot, in the meantime, lags far behind ChatGPT, Gemini, and Claude. GitHub Copilot, as soon as the main AI coding assistant, has been supplanted—first by AI startup Cursor, then by Claude Code.

Two years in the past, Microsoft seemed to be one of many early winners of the AI period. Because of Nadella’s prescient guess on OpenAI, Microsoft had unique entry to the high-flying AI startup’s fashions, and will use them to create AI options throughout its merchandise. If firms needed to entry OpenAI’s know-how, the one cloud supplier they may use was Microsoft Azure. The corporate even thought OpenAI gave it the perfect probability in years to compete with Google Search. Nadella, a decade into his tenure, had steered Microsoft by one platform shift — desktop to cloud — and seemed poised to repeat the feat. However AI is fast-moving, and two years is a lifetime. That is the story of how Microsoft fumbled its early AI lead, and the way it’s attempting to get it again.

What went flawed

Microsoft was tripped up partially by the very deal that put it on the entrance of the AI pack to start with: its partnership with OpenAI. Microsoft noticed the younger San Francisco firm early, investing its first $1 billion in 2019, and ultimately committing $13 billion to the startup. Microsoft used OpenAI’s tech to launch its Copilot-branded AI merchandise throughout each its client and enterprise software program portfolio.

However OpenAI’s explosive development and hovering ambitions after the launch of ChatGPT in late 2022 quickly strained the partnership. The 2 companions clashed over computing capability (OpenAI continually needed extra); over mental property (Microsoft thought OpenAI was sluggish to honor its contractual obligation to share improvements); over prospects (OpenAI pitched AI fashions on to the identical enterprises Microsoft was promoting Copilot to); and, when OpenAI sought to restructure, over how a lot fairness Microsoft ought to obtain within the new for-profit company.

Nadella knew that staking his firm’s AI technique on an unproven startup was dangerous. These dangers have been underlined in daring in November 2023 when the nonprofit board that managed OpenAI’s for-profit arm fired CEO Sam Altman for not being persistently candid—informing Nadella minutes earlier than saying its resolution. Nadella scrambled to reassure buyers that Microsoft retained entry to OpenAI’s know-how, whereas working with Altman to stress the board to reverse its resolution. Nadella introduced a plan to rent Altman and any OpenAI workers who needed to return with him. The prospect of mass defections pressured the board to cave and reinstate Altman. At OpenAI, the five-day disaster grew to become referred to as the blip. But it surely left Nadella shaken, in response to these conversant in his pondering. He wanted to hedge his bets.

Plan B got here within the type of Mustafa Suleyman, a Google DeepMind cofounder who had left to create his personal AI startup, Inflection. In March 2024, Microsoft employed Suleyman and Inflection’s technical workers and licensed their know-how in a $650 million deal. Suleyman was put in as CEO of a brand new Microsoft AI division—MAI for brief—accountable each for constructing in-house frontier fashions as insurance coverage towards OpenAI, and for rising the person base of Microsoft’s Copilot chatbot.

It didn’t go properly. Microsoft’s partnership with OpenAI prohibited Microsoft from coaching fashions past a sure measurement. “We have been very a lot solely in a position to prepare native fashions, Microsoft fashions, within the SLM, or small language mannequin, measurement,” Suleyman tells Fortune. The primary normal goal language mannequin MAI examined publicly, in August 2025, referred to as MAI-1 preview, ranked properly down the efficiency leaderboards. It was by no means extensively launched.

MAI additionally didn’t flip the Copilot chatbot right into a client hit. A 12 months into Suleyman’s tenure, Copilot utilization had flatlined at about 20 million weekly lively customers, in response to press stories, whereas ChatGPT’s person base rocketed to ever better heights on its solution to 900 million. A serious 2025 improve, designed to make Copilot extra like a private assistant that might carry out duties, did not jump-start development. And the AI-enhanced model of Microsoft’s Bing search engine barely dented Google’s share of the search market.

Plan A was additionally operating into bother. In 2023, OpenAI’s GPT fashions have been head and shoulders above the sector. By early 2025, Anthropic’s Claude routinely topped AI leaderboards, and lots of companies most well-liked it for advanced duties. Google’s Gemini was more and more aggressive on visible duties. However Microsoft’s Copilot choices have been powered completely by GPT. What had as soon as been the engine of Microsoft’s AI technique was beginning to really feel like a millstone.

Judson Althoff, Microsoft’s business CEO, acknowledges a number of missteps. For one factor, calling each its client and its enterprise merchandise Copilot was complicated. “The one factor worse than not having a copilot is having a couple of,” Althoff, who holds a personal pilot’s license, quips. Microsoft additionally incentivized its gross sales reps to push each the freemium model of its enterprise M365 Copilot in addition to the premium model, when solely the premium model delivered the worth companies needed. “We bought that flawed,” he says.

Microsoft was additionally struggling to maintain tempo with the pace at which AI know-how was evolving. A break level got here in 2025, when Anthropic launched Claude Code, which might autonomously write total applications from only a description of what the developer needed. This was now not a copilot, it was an autopilot. Inside six months it had reshaped software program improvement. Then, this January, Anthropic debuted Claude Cowork, an agent that might use software program — together with Microsoft’s productiveness instruments, like Excel and PowerPoint—to autonomously full duties.

Claude Cowork posed a critical problem to M365 Copilot and the AI brokers Microsoft had been pushing prospects to undertake. Actually, it posed a menace to most enterprise software program. That realization sparked the SaaSpocalypse inventory market rout, which might ultimately wipe greater than $2 trillion off the worth of tech shares—together with a one-day plunge that erased $357 billion of Microsoft’s market cap in a single buying and selling session.

The fixes

By the autumn of 2025, Nadella realized it was time for a reset. The corporate’s actions since then have mirrored a tough balancing act, as the corporate strives to innovate at AI pace like a buzzy startup, whereas nonetheless reliably delivering for buyers and enterprise prospects just like the staid Microsoft of outdated.

He handed a lot of his business and each day operational duties to Althoff, a longtime Microsoft exec, in order that he might concentrate on AI product improvement. Althoff says he handles horizon zero and horizon one whereas Nadella takes care of horizons two and three. On the similar time, Nadella started breaking down silos to make Microsoft sooner, flatter, and extra agile than ever earlier than, Althoff says.

In March, Nadella merged the patron and enterprise Copilot groups. Suleyman was stripped of his duty for client AI merchandise and put in control of a rebranded mannequin improvement effort: the Superintelligence crew. (Suleyman says the identify displays its ambitions and can assist it recruit high researchers.) Andreou, who joined Microsoft in 2025 after stints at Snap and enterprise capital agency Greylock, was put in control of Copilot Expertise throughout each client and enterprise, reporting on to Nadella. Rounding out the Copilot Management Crew alongside Suleyman and Andreou are three veteran Microsoft govt vice presidents: Charles Lamanna, who oversees Copilot, AI brokers, and platform; Ryan Roslansky, liable for Microsoft Workplace and Microsoft-owned LinkedIn; and Perry Clarke, chief know-how officer for software programs. “We wish it to be one backend, one mind, that powers each client and work,” Lamanna says. Nadella himself joins the Copilot Management Crew’s weekly stand-up conferences and participates in an ongoing Groups channel on Copilot improvement.

Andreou factors to 2 new merchandise, Copilot Duties for shoppers—the product Nadella had performed round with prototyping in January—and Copilot Cowork for enterprise prospects, as proof the unified Copilot crew is working as Nadella supposed. “Each of these issues have been mainly frontier-grade experiences, each for shoppers and enterprise customers,” he says. “And each are issues our groups have been in a position to pull collectively and construct in a matter of weeks.”

Microsoft has additionally agreed to OpenAI’s long-pending restructuring on considerably much less restrictive phrases. The software program large secured a 27% fairness stake — potential upside if OpenAI goes public, as is extensively anticipated. However the outdated deal’s exclusivity has been deserted: OpenAI can now lower offers with different cloud suppliers, and Microsoft can work with different AI firms’ fashions. Suleyman says the settlement permits Microsoft to lastly construct bigger, extra succesful frontier AI fashions that can ultimately permit it to be self-sufficient — however provides that it’s going to take two to 3 years for Microsoft to meet up with the highest labs.

The revamped partnership allowed Microsoft to embrace OpenAI’s archrival, Anthropic. The corporate dedicated to investing as much as $5 billion in Anthropic in November and started providing its fashions on Azure. The flexibility to make use of Claude to energy Copilot has been common with enterprises and enabled Microsoft to construct Copilot Cowork.

However Microsoft shouldn’t be merely swapping reliance on one loss-making AI startup for dependence on one other. The Anthropic funding is indicative of a distinct guess: that AI fashions will more and more commoditize, and that worth will accrue—for enterprises, anyway—not within the AI mind, however in all of the physique elements and connective tissue round it. That’s the place Microsoft thinks it could actually win. It already owns lots of these organs and sinew: software program instruments, safety, information warehouses, and cloud computing. It has additionally constructed IQ-branded merchandise that permit enterprises to create personalized workflows, collect their information, after which construct, deploy, and monitor brokers operating these workflows—utilizing any AI mannequin from any vendor.

“We don’t suppose enterprises will swap out their info work platform, their [developer] atmosphere, their safety atmosphere, each time a brand new mannequin drops,” Althoff says.

The strategic shift comes with a brand new enterprise mannequin. Microsoft has historically priced its merchandise with per-user licenses—$30 per person per 30 days for Copilot, for instance. Clients prefer it as a result of it makes budgeting simple. But when the AI brokers inside these merchandise use fashions Microsoft doesn’t personal, Microsoft has to pay the AI vendor for the tokens consumed. So Microsoft has began transferring towards hybrid pricing: a per-user license with a restricted token allowance, plus per-token prices past it — to maintain the model-agnostic technique from torching its margins.

In one other cost-conscious transfer, Microsoft has seemed to slim its workforce. In April, it introduced its first-ever worker buyout supply, geared toward its most long-tenured staff. The corporate says about 7% of its U.S. workforce, or about 8,750 staff, qualify for the buyout, at an anticipated price of $900 million.

There’s proof the revamped enterprise technique is working. On the finish of March, Azure revenues have been rising 40% 12 months over 12 months, and Microsoft’s complete AI enterprise was on tempo for $37 billion in annual gross sales, up 123% from the 12 months earlier than. Of the 20 million M365 customers now paying for Copilot, 1 / 4 have been added within the first 4 months of 2026, and Althoff says adoption is accelerating.

UBS analyst Karl Keirstead says extra Microsoft prospects are telling him they see Copilot’s worth. However the general person base remains to be disappointing. “I don’t suppose they’ve reached a penetration fee to fulfill Wall Road,” he says.

Microsoft’s model-agnostic technique has a attainable flaw: What if the buzzy AI startups begin constructing the boring-old-Microsoft-style connective tissue, too? That’s now not a hypothetical. In February, OpenAI launched its Frontier enterprise platform, providing lots of the capabilities Microsoft is constructing into its new instruments. Anthropic is transferring on this route too, providing a Claude Managed Brokers service.

Microsoft’s argument is that a long time of enterprise relationships, its popularity for reliability and safety, and deep integration with the software program prospects already run will give it the sting. Althoff says he welcomes the competitors. “To offer credit score to OpenAI and Anthropic, they’re serving to us run sooner,” he says.

However some query whether or not an organization the scale of Microsoft can match the agility of the AI-native startups. “Microsoft, and albeit all of the software program corporations, are dealing with one thing they haven’t confronted in 10-plus years, which is awfully modern new rivals,” Keirstead of UBS says. “And anticipating any giant incumbent like Microsoft to pivot as quick as OpenAI and Anthropic are in a position to might be too demanding.”

Financial institution of America’s Tal Liani makes the case for Crew Satya. It’s unlikely, he says, that the AI firms will be capable of construct the total product suite that Microsoft gives. Meaning Microsoft doesn’t should win the AI race. It simply has to not lose it. “It is probably not the perfect, but when it’s ok, and also you get nice worth by bundling, that’s really the worth of Microsoft,” he says.

Even the prices of not dropping, nevertheless, usually are not trivial. Microsoft, like different hyperscale cloud suppliers, is expending huge sums on information facilities and specialised chips. The corporate spent $88.2 billion on capital expenditures in fiscal 2025. That was according to friends reminiscent of Google Cloud and Amazon’s AWS. But it surely nonetheless wound up being too conservative. Demand surged, leaving Microsoft with out sufficient compute capability and unable to acknowledge the AI income it had booked on the fee it had hoped to. “I believed we have been going to catch up,” CFO Amy Hood acknowledged on an October earnings name. “We’re not.”

Now the corporate is greater than doubling down: In 2026, Microsoft has forecast it’s prone to spend some $190 billion on capex—greater than 3 times what it spent in 2024. As soon as nervous about such sums, Wall Road appears prepared to tolerate the large spending. But when investor sentiment have been to reverse, Microsoft will likely be extra uncovered than ever.

In November 2025, an unbiased developer named Peter Steinberger launched OpenClaw, a free, open-source system for turning any AI mannequin into an always-on agent that works autonomously over lengthy durations—constructing software program, performing as a digital govt assistant, even operating stock for a web based enterprise. OpenClaw was successful with builders and AI-forward tinkerers. Nadella, reportedly, was amongst them.

However as common as OpenClaw was, it had an issue. It wanted entry to programs, information, fee info, and passwords to be helpful, which made it enormously dangerous. It additionally burned by tokens at an unlimited fee. “I can’t launch OpenClaw at Microsoft,” Nadella instructed a San Francisco tech convention in March. “I don’t have permission to do this, as a result of that will be thought of Microsoft launching a virus. However on the similar time, it’s a implausible innovation.”

Nadella has tasked the unified Copilot crew with constructing a Microsoft model of OpenClaw — similar consumer-grade enjoyable, however with the safety and governance enterprises demand. Andreou frames it because the take a look at of the brand new group: “That’s what a win appears to be like like right here.” Lamanna thinks it may be the factor that ignites Copilot’s development. “The toughest factor has been, how do you assist folks change the best way they work?” he says.

A perpetual AI assistant, if it really works, would make that change simpler. It might additionally imply the unit of AI shifts from mannequin to always-on agent—exactly the type of paradigm shift that assessments whether or not Microsoft’s connective-tissue technique holds when the factor within the center modifications form. Lamanna says an enterprise-grade Microsoft OpenClaw is not very distant.

‘Gigawatt scale’

The week of March 30, Suleyman gathered his new Superintelligence crew—some 500 folks from across the globe—for a three-day offsite in Miami. The aim was to put out a street map for attaining gigawatt scale AI coaching runs—the type that ought to let Microsoft compete instantly with OpenAI, Anthropic, Google DeepMind, Meta, and xAI. Suleyman says it’s important to Microsoft attaining self-sufficiency earlier than 2030; the corporate loses entry to OpenAI’s tech in 2032.

The whole crew gathered in a large ballroom to listen to a keynote handle and ask me something session with Suleyman and Nadella. Nadella, in response to Suleyman, described the second as Microsoft “refounding the corporate” in response to the platform shift to AI. It’s a telling phrase.

After the keynote, the gathering broke up into separate workstreams, every crew gathered at considered one of 40 whiteboards arrayed across the ballroom, as they brainstormed and plotted the following eight-week dash. Quite than depart, Nadella lingered. For the following three hours, he wandered from desk to desk, speaking to researchers, making ideas, providing concepts. If it is a refounding, then Nadella is performing the a part of the startup CEO. He’s not taking something with no consideration. He is aware of that Microsoft has all the pieces to lose — and all the pieces but to play for.