On Wednesday, at the end of the day in London, Google CEO Sundar Pichai and Google DeepMind CEO Demis Hassabis had some news for employees.
Hassabis, DeepMind’s co-founder and the lab’s longtime steward, was giving up the CEO title and moving to a chairman role, they said. Minutes later, the news broke that long-time chief scientist Jeff Dean and several other prominent researchers were leaving the company.
Staff in DeepMind’s London office were shocked. They had questions. Was the move a promotion or a demotion? Would Hassabis retain real influence over DeepMind’s direction? And was Google finalizing a power shift from the London lab to its Mountain View headquarters?
One DeepMind engineer familiar with internal reaction told Fortune employees were “very mixed” on what the reshuffle meant, with some wondering whether the chairman role would ultimately end with Hassabis phased entirely out of the company.
In reality, the shock exit was just the tip of the iceberg of what has been a difficult year for Google DeepMind. Within the lab, three current and one former employee told Fortune tensions were already simmering. Delayed models, high-profile exits of talent, and a very public fight with its own staff over controversial defense contracts had left insiders wondering if Google had lost its way in the AI race.
Once a frontrunner, the company has recently ceded ground to rivals, especially OpenAI and Anthropic. This was most evident in delays around the release of the company’s Gemini 3.5 Pro model, which has missed three release deadlines in recent months. Three of the DeepMind engineers who talked to Fortune blamed the delays on the company failing to prioritize AI coding abilities.
And now, for the second time in just four years, the company is again playing catch-up.
Google DeepMind is struggling to retain its technical edge, keep hold of top researchers, and contain an employee revolt over its defense work. Pichai’s reshuffle—which employees say shifts power back to Mountain View and pulls DeepMind closer to Google proper—arrives at a critical moment for the company. But it’s also a gamble. Will new leadership restore Google’s AI momentum, or worsen internal tensions at a crucial moment in the AI race?
A leadership shakeup
In his note to staff on Wednesday, Hassabis said that he was handing over day-to-day operational responsibilities for the lab to allow him the “time and space to focus on the big picture and help influence what is to come to the best of my ability.”
Instead of selecting a new CEO for DeepMind, however, Pichai announced that DeepMind’s chief technology officer, Koray Kavukcuoglu, was taking over day-to-day operations as senior vice president, reporting directly to him rather than holding a standalone DeepMind CEO title. Kavukcuoglu, unlike London-based Hassabis, works 6,000 miles away, in the California offices in Mountain View. A Google spokesperson said that Hassabis was moving into an expanded role to focus on advancing science and long-term strategy.
Three DeepMind engineers, who asked for anonymity to discuss company matters, told Fortune this represents a broader structural change and a shifting of power to Mountain View, where Google is headquartered, and away from London, where DeepMind was originally founded.
Losing Hassabis as a CEO and not getting a new one, they said, also means the lab is being somewhat absorbed into the rest of the company and losing some of its independence. The first engineer called it a “small step, but gradually continuing the trend.”
“A lot of people are quite upset by it,” they said. “Both Demis exiting, and losing some of the separation from Alphabet and identity as DeepMind.”
For months, there had been rumors of a significant power shift to California, and concerns about the message that change would send to the London-based staff. Workers describe an ongoing “background pressure” of things moving to Mountain View and several prominent London-based engineers, including AlphaFold veterans John Jumper, Jonas Adler, and Alexander Pritzel, recently exited the company to join Anthropic.
The company had also been slowly moving reporting lines away from Hassabis for a while, according to the second engineer.
A Google spokesperson disputed the characterization that power was shifting to Mountain View and said Google DeepMind remained committed to its London presence. The spokesperson said Hassabis, along with much of Google DeepMind’s leadership, would remain based in the London office. They added that DeepMind would retain separate leadership and research autonomy under Kavukcuoglu.
From a commercial perspective, there may be an argument that things need to change, especially if the company is having issues shipping models in an increasingly competitive environment. Mountain View may well have run out of patience and want a tighter grip on its AI crown jewel.
“It’s very clear that they’re having a big internal debacle,” the second engineer, who spoke on the condition of anonymity, said. “This might have also contributed to speeding up some of these power shifts. Generally in Google, when something is significantly underperforming or not working, their solution is to just do a huge reorg, a reallocation of compute and priorities to try and fix it.”
DeepMind’s journey
This isn’t the first time Google has hit a stumbling block on its AI journey.
Back in late 2022, the company famously declared a “code-red” following the launch of OpenAI’s ChatGPT, fearing the impact on its core search business. Since then, its race to catch up has been somewhat successful. Its Gemini 3 model outperformed all its rivals and pushed OpenAI to declare its own “code-red” in late 2025, as Gemini’s benchmark gains and rapid user growth put pressure on ChatGPT’s lead.
But since then, things appear to have gone in the wrong direction.
Google’s models are currently not competitive with the bleeding edge of the U.S. frontier on intelligence and coding benchmarks. Even some open-source Chinese efforts are also outshining the tech giant.
Gemini 3.5 Pro was supposed to be Google’s next big answer to advanced models such as Anthropic’s Mythos and OpenAI’s GPT 5.6. Unveiled at the company’s flagship I/O conference in May with a promise to ship “next month,” it missed that June target and another mid-July date. The company headed into August with the model still unreleased, even as Google shipped a trio of smaller Flash models around it and claimed it had already begun training Gemini 4.
Micah Hill-Smith, co-founder and CEO of independent benchmarking company Artificial Analysis, told Fortune that Gemini 3.6 Flash, the strongest model Google has shipped, now sits behind models from Anthropic, OpenAI, one or two leading Chinese labs, xAI, and Meta on his firm’s intelligence index, which measures overall performance of AI models.
“For a brief moment, we actually did have [Google’s] model on top as the smartest model in the world. Right now, it’s fair to say that [model] from Google is less compelling,” he said.
Many of the mistakes that led to this started last summer, when the company failed to prioritize the development of Gemini’s coding abilities, according to a third engineer. Google, they said, was preoccupied with protecting its search traffic and beating OpenAI on chat, while rivals were running ahead on agentic and coding capabilities. Google models now significantly lag rivals on coding agent benchmarks, according to Hill-Smith.
A Google spokesperson said Gemini 3.5 Pro was currently in testing and that the DeepMind team was already working on its next generation of models. The spokesperson acknowledged that the company needed to improve in coding and agentic capabilities, but said its teams were focused on that work.
Google perhaps failed to predict that coding abilities would play such a major role in consumer products just a year later. Developer tools such as Claude Code or OpenAI’s Codex, are now underpinning a new wave of consumer and general-use products. They include Anthropic’s Cowork—an agentic tool built on the same underlying technology as Claude Code that lets non-developers hand off multi-step knowledge work, like organizing files or drafting documents, rather than writing code—and OpenAI’s superapp, positioned as a one-stop entry point for AI tools.
“They dropped the ball a little bit [on coding],” the third DeepMind engineer said. “By the time you realize something like that…a number of months have passed, and now you have to play catch-up. Then it also becomes political.” Coding is now in “big focus” inside the lab, the second engineer added.
However, not winning on pure model performance doesn’t mean Google is down and out of the AI race. The company has other advantages such as vertical integration—at the application, cloud, and chip layers—that are structural advantages that can offset weaker performance at the cutting edge of raw model intelligence.
“For OpenAI and Anthropic, because the model layer is so important to what they do, it’s pretty existential for them to win there,” Hill-Smith said. “Google could end up doing very well in the AI race overall, even if they’re not at the frontier of the model layer,” he said, adding that the company’s unusually deep vertical integration is a structural advantage that could well offset weaker performance at the cutting edge of raw model intelligence.
Falling behind—and jumping ship
The technical difficulties and feeling of playing catch-up twice in less than five years have been a knock to morale, all the current and former DeepMind workers who spoke to Fortune say.
That internal frustration has coincided with one of the most damaging stretches of departures in DeepMind’s history. In a single week in June, Google lost Gemini co-lead Noam Shazeer to OpenAI and Nobel laureate and AlphaFold co-inventor John Jumper to Anthropic.
The employees who spoke to Fortune describe the exits as driven by overlapping forces: aggressive poaching from cash-flush rivals, frustration at being behind in the AI race, and engineers racing to get into a pre-IPO lab before they go public, hoping the equity compensation will make them fabulously rich. “There’s a lot of just hopping over to pre-IPO labs,” one employee said.
A Google spokesperson said talent moved regularly between technology companies and that Google remained focused on attracting people aligned with DeepMind’s mission.
Google has tried to combat some of the exodus, with the second engineer describing big retention and equity bonuses being deployed for some staff, and a prior round of enhanced compensation during a period when Meta was making aggressive counteroffers to AI researchers.
However, employees say, it hasn’t been enough to offset a deeper problem of burnout.
“I’ve been pushing 60-hour weeks for quite a long time because there is so much to do,” the third engineer said. “People are pushing themselves really hard because it is a race, and you have this feeling that you need to catch up… [but] telling people to work 60-hour weeks constantly and that it is constant crunch time usually doesn’t work that well, and people do burn out.”
Google said AI was an intensely competitive field and pointed to researchers who had publicly expressed excitement about their work and the company’s direction.
Employees had issues with leadership too. Even before his exit, workers described Hassabis as absent, especially compared to Google co-founder Sergey Brin, who came out of retirement to help the company’s AI efforts.
“You will see Sergey being incredibly hands-on, working hard on stuff and being involved,” the third DeepMind engineer said. “I can’t say that I’ve ever seen Demis walk around the office in the Gemini area.”
“Demis isn’t a very inspiring leader,” the former employee added. “Certainly not the way people talk about Dario [Amodei] … Demis is more of a guy who is known but not seen.”
A union battle and contentious defense deals
Frustration over the leadership and talent exodus has collided with a separate, more politically charged fight over Google’s defense work.
In April 2026, Google signed a deal with the Pentagon allowing the Department of Defense to use Gemini models on classified networks for any lawful government purpose. More than 580 Google employees, including senior DeepMind researchers, later signed an open letter urging Pichai to reconsider the deal, citing concerns around autonomous weapons and mass surveillance.
Google was not alone in making the controversial deal: OpenAI, xAI, Microsoft, Amazon, and others signed similar classified-use deals. Anthropic was the only outlier, declining to drop its own restrictions on autonomous weapons and mass surveillance—a stance that led the Pentagon to designate the company a supply-chain risk and order the military to stop using its products.
At Google, the Pentagon deal reopened a longer-running fight over the company’s defense and intelligence contracts. Employees have also taken issue with Project Nimbus, a roughly $1.2 billion cloud-computing agreement Google and Amazon struck with the Israeli government in 2021.
For many employees, the Pentagon deal slipped largely under the radar. But for others it was existential.
Alex Turner, who spent more than two years on DeepMind’s AI safety team, resigned in June citingthe Pentagon deal as the trigger.
In a public post, he said that his decision followed months of internal advocacy that went nowhere: he even tried to enlist Google chief scientist Dean as an ally. Andreas Kirsch, another DeepMind research scientist, was also a vocal public critic of the deal. Kirsch posted publicly at the time that the contract’s restrictions amounted to little more than “meaningless weasel words” for PR purposes, and said he felt “incredibly ashamed” to be associated with the lab over the decision.
The Pentagon deal was the final straw for a group of London-based DeepMinders, who launched a union bid in May in an attempt to win back some of their waning leverage. The action is the first formal attempt by a frontier AI lab to unionize.
The union drive really traces back to February 2025, when Google removed language from its public AI Principles committing not to build AI for weapons or surveillance that violates international norms. The decision angered many engineers who felt Google was backtracking on its ethical assurances. The company adopted that pledge in 2018 after the intense backlash over Project Maven—a separate defense contract.
Google has declined to voluntarily recognize the union, instead offering talks through the UK’s conciliation service, ACAS. The two sides have already clashed over basic questions, like which staff should be included in the bargaining unit, a person familiar with the situation said. Now employees worry that American leadership may be less inclined to play nice in negotiations.
A Google spokesperson said the company was still in the early stages of the union process and would continue to engage in negotiations in good faith.
DeepMind’s uncertain future
With Hassabis stepping back and Dean gone, DeepMind appears to be entering a new phase at precisely the moment it can least afford further disruption.
While the lab still has advantages few rivals can match—distribution, cloud business, custom chips, vast research bench, and the ability to embed AI across products used by billions—the reorganization seems to remove a layer between DeepMind and a parent company increasingly focused on turning AI research into commercial—and government—deployments.
For employees, the question is whether new leadership and tighter integration will bring the urgency and clearer direction DeepMind needs to close the gap in coding and agents, or deepen the problems already driving people out: burnout, weakened autonomy, and concerns over how the company will use its technology. Just four years after Google’s first AI “code red,” DeepMind is again trying to prove it can move fast enough to keep pace. But this time, it will be doing so without the two figures most closely associated with its scientific identity.
The post How stalled models, missed deadlines and staff burnout led to the unraveling of Google’s DeepMind appeared first on Fortune.
