Talent & Culture Strategy at Google DeepMind: Startup Speed Inside a Trillion-Dollar Company

This article is part of MokaHR's Talent & Culture Strategy series, which profiles how leading companies build their people strategies.

Introduction

On a January morning in 2026, Demis Hassabis picked up his phone and called Sundar Pichai. Again. This had become routine — the CEO of Google DeepMind and the CEO of Alphabet now speaking every day, sometimes multiple times, about model architecture decisions, compute allocation, and competitive intelligence on OpenAI. It was a rhythm that would have been unimaginable three years earlier, when Hassabis ran a semi-autonomous research lab in London that published papers but shipped nothing, and Pichai ran a search company that happened to own the lab.

The shift is the most consequential organisational design decision made by any AI lab in the last five years. In April 2023, Alphabet merged DeepMind — the London-based research group it had acquired in 2014 for a reported $500 million — with Google Brain, its in-house AI research division and the birthplace of the Transformer architecture that powers almost every modern large language model. The merger was framed publicly as a strategic consolidation. Internally, as multiple accounts have since confirmed, it was an emergency response to ChatGPT. Sergey Brin returned from semi-retirement to code. Hundreds of engineers from both Brain and DeepMind were conscripted into what would become Gemini. Hassabis was named CEO of the combined entity, while Jeff Dean, who had led Brain, was given the title of Chief Scientist — a prestigious but, by many accounts, less operationally central role.

Three years later, the structure appears to be working. Google DeepMind has grown from roughly 1,567 employees in 2022 to around 8,000 by early 2026. Gemini reached approximately 750 million monthly active users by the end of the fourth quarter of 2025. Gemini 3, released in November 2025, prompted what secondary reporting described as an urgent internal "code red" at OpenAI. And DeepMind's two-year employee retention rate sits at 78%, second only to Anthropic in the AI industry and well above OpenAI's 67%.

The question worth studying is how a research lab famous for deliberate, interdisciplinary, long-horizon work reorganised itself to ship products at the cadence of its fastest-moving competitors — without losing the cultural identity that made the research worth shipping in the first place.

How does Google DeepMind attract and hire talent?

Google DeepMind's hiring model is the product of a specific historical accident: two of the best AI labs in the world operated inside the same parent company for nearly a decade before being combined. That history shows up in the current hiring approach, which blends DeepMind's academic, publication-heavy filter with Google's long-standing "Googleyness" emphasis on collaboration at scale.

The publication filter, adapted

Unlike Anthropic's mission-first hiring model, which explicitly treats values alignment as a gate, DeepMind's published guidance frames the bar around systems-level thinking and scientific rigour. The company looks for people who can lead large projects at scale. Recruiter-compiled analyses describe DeepMind as a lab that "values publication history and systems-level thinking" and specifically assesses whether candidates can operate at Google scale. One recruiter guide estimates that machine-learning engineers with publications at NeurIPS or ICML have a 30–40% higher chance of reaching an interview than comparable candidates without them.

That said, DeepMind has been explicit that publication count alone is not the bar. The company's own recruiting material notes that post-PhD research scientists are hired on "abilities rather than publications or academic achievements", with significant weight given to internships, industry experience, side projects, and open-source contributions. The intent is to avoid the failure mode of academic filters — that they reward people who write papers rather than people who produce research that changes the field.

A long, multi-stage interview process

DeepMind's published interview process for research roles typically runs about 45 days end-to-end. Candidates begin with a recruiter call, then a hiring-manager technical conversation, followed by two to three coding interviews, one or more machine-learning rounds covering probability, statistics, and ML theory, and then a research or portfolio deep-dive. The final stages include team-lead interviews, focused on how candidates reason about ambiguous problems, and a people-and-culture interview.

For alignment and safety roles, the process is unusually transparent. A public post from Rohin Shah, a DeepMind research scientist, explained that the company's rough test for a Research Engineer role is whether the candidate can "reproduce a typical ML paper in a few hundred hours" and whether their interests align with the team's. For alignment-specific questions, interviewers are explicit that candidates are not expected to know the right answer — the team is assessing how candidates think about confusing and ill-specified questions, because that is the daily work of alignment research.

This contrasts sharply with OpenAI's interview model, which runs shorter four-to-six hour final loops and decides within a week. DeepMind's longer, more deliberative process reflects the company's origin as a research lab and the parent company's risk tolerance: Alphabet can afford a 45-day hiring cycle in a way that venture-funded startups cannot.

The 20% project as a pipeline

A particularly interesting hiring channel is the internal 20% project pathway. Google's long-standing tradition of allowing engineers to spend part of their time on projects outside their core team is used by DeepMind as a de facto audition process for team transfers. Public accounts from current research engineers describe joining alignment teams after running 20% projects through the summer, getting familiar with the codebase, and then converting to full-time roles on the target team. For internal candidates, the 20% pathway functions as a low-risk trial period; for the hiring team, it produces a signal far richer than any interview loop can generate.

Compensation competitive but not Meta-tier

DeepMind compensation ranks at the top of the industry but, like Anthropic, does not match the nine-figure packages Meta deployed through 2025. According to The Wall Street Journal coverage of the talent war, Microsoft's AI division hired more than 20 researchers and engineers from Google (including former VP of Engineering Amar Subramanya), and Meta's Superintelligence Labs successfully recruited DeepMind researcher Jack Rae. The losses are real but bounded — DeepMind's 78% retention suggests that mission, compute access, and the AlphaFold halo effect retain most staff against external pressure.

📄 2025 AI Recruitment Casebook DeepMind's ability to hire at a less than 1% acceptance rate while retaining 78% of staff against aggressive poaching illustrates the value of a structured, multi-stage hiring process. To see how companies across 10 industries are using AI-driven assessment to raise hiring bars without slowing pipelines, download MokaHR's full AI Recruitment Report. Download the free report →

Trusted by: Tesla · NVIDIA · McDonald's · Nestlé · Schneider Electric

How does Google DeepMind develop and manage employee performance?

DeepMind's post-merger performance and development model is shaped by one overriding constraint: the lab needed to become dramatically faster without losing the research depth that had produced AlphaFold and the Transformer. The response has been a deliberate attempt to re-adopt what Hassabis has repeatedly called a "startup or entrepreneurial" operating model inside a $3.9 trillion company.

Startup practices, rediscovered

In an April 2026 interview on the 20VC podcast, Hassabis described the two years since the merger as a period of deliberate acceleration. "We've caught up" in the last two to three years, he said, by "aligning talent from around the company, sort of pushing in one direction" and gaining access to the compute infrastructure DeepMind had previously lacked at scale. The combination — DeepMind's research culture plus Google Brain's engineering infrastructure plus Alphabet's compute — is what Hassabis has called the real "secret" behind Gemini 3: world-class research, engineering, and infrastructure "all working closely together with relentless focus and intensity".

Operationally, the shift translates into shorter development cycles between model research and production. For ML engineers, that means faster access to new model capabilities and APIs, and a higher cadence of changes to the internal library, model, and infrastructure stack. For researchers, the cultural shift increases pressure to deliver replicable, production-ready results — with greater collaboration with software engineering and product teams than was historically typical at DeepMind.

The integrated career path

DeepMind's career model, like Anthropic's, deliberately integrates research and engineering rather than separating them into parallel tracks. The organisation structure runs a "Core" research team that handles more pure-research projects (AlphaGo, AlphaFold, AlphaZero) and an "Applied" team that focuses on production deployment, but both sit under a common mission and staff move between them. Koray Kavukcuoglu, the CTO, is explicitly positioned as the bridge between pure research and product — a role that exists precisely because the lab needs people who can translate between the two modes.

A similar pattern of integrating technical career tracks has worked well at Microsoft's growth-mindset transformation, where the elimination of rigid boundaries between research, engineering, and product was a precondition for the company's AI-era comeback.

Safety as a distinct function

One structural choice DeepMind has made that differentiates it from both OpenAI and Anthropic is maintaining a clearly delineated technical AGI Safety team, organised into sub-teams including Alignment and Scalable Alignment, with their own public hiring materials, rubrics, and research agendas. The team publishes its evaluation criteria openly and has been unusually transparent about what it expects from candidates. In a labour market where safety researchers at OpenAI have departed citing resource starvation and safety researchers at Anthropic have resigned warning about the industry's trajectory, DeepMind's decision to keep safety as a named, resourced, and publicly-facing function has become a retention asset.

Leadership and continuity

All three DeepMind co-founders — Demis Hassabis, Shane Legg, and Mustafa Suleyman — started together in 2010, and Hassabis and Legg remain at the company more than 15 years later (Suleyman left and now runs Microsoft AI). That founding continuity is unusual in the AI lab world and materially shapes retention: when the scientific conscience of the organisation (Legg, as Chief AGI Scientist) and the operational leader (Hassabis) have been together for 15 years, the cultural signal to staff is that this is a place where long research careers happen.

What can HR leaders learn from Google DeepMind's approach?

Most companies cannot replicate DeepMind's compute, its Nobel Prize, or its 15-year founding continuity. But the structural choices the lab has made in the post-merger period translate well to any organisation trying to move faster without losing institutional depth.

Merge for speed, but preserve the research identity. The DeepMind-Brain merger is the best-documented case in modern technology of two internal organisations being combined under unified leadership to eliminate duplication and accelerate delivery. The critical design choice was naming the research leader (Hassabis) as CEO rather than the infrastructure leader (Dean) — a bet that in AI, research vision is the scarcer input. For HR leaders handling post-merger integration in any industry, the underlying lesson is that the choice of who leads the combined entity is more important than the combined org chart, because it signals which culture will dominate in contested decisions. Workforce analytics platforms can help HR teams model the talent-retention implications of different integration structures before the decisions are locked in.

Use internal mobility as a recruiting pipeline. DeepMind's 20% project pathway is one of the most effective internal-to-internal hiring mechanisms in the technology sector, because it generates months of real collaboration data before any team-change decision is made. Most large organisations have the raw ingredients — slack time, multi-team projects, internal job posting — but fail to systematise the conversion pathway. The companies that treat internal mobility as a core recruitment channel rather than an afterthought consistently outperform on senior retention. Structured recruitment workflows can formalise the internal-to-internal pathway so that 20%-style trial arrangements convert predictably rather than informally.

Resource your safety or trust function as a named organisation. DeepMind's decision to maintain a public, named, well-resourced Technical AGI Safety team has become a retention asset precisely because the alternative — what happened at OpenAI, where the superalignment team was disbanded and more than 25 senior safety staff departed — demonstrates the cost of under-resourcing the function. The lesson generalises beyond AI: whenever a company's public legitimacy depends on an internal function (safety, compliance, ethics, trust), under-resourcing that function creates a retention cliff. Staff in those roles leave publicly, and the reputational damage is disproportionate to the headcount. AI-powered screening tools can help HR teams identify candidates whose motivations align with specialised trust, safety, or compliance functions before offers are made.

Preserve founder continuity where possible. Anthropic has kept all seven co-founders. DeepMind has kept two of three for 15 years. Neither company achieved this by accident — both have designed equity structures, decision rights, and leadership titles to keep founders engaged over very long time horizons. For HR leaders in scaling companies, the implication is that founder-retention economics deserve explicit design attention, because the cultural signal of long founder tenure is almost impossible to replicate through other means.

What is it like to work at Google DeepMind?

Working at Google DeepMind in 2026 means working at an organisation that has been fundamentally reshaped by the merger. The London-based DeepMind culture — interdisciplinary, neuroscience-inflected, deliberately contrarian about AGI — has been fused with the more engineering-forward, product-integrated Google Brain culture. Employees report that the integration has been genuine rather than cosmetic, but also that the pace has been demanding.

One research engineer who joined the AGI Safety team in late 2024 described the post-merger environment as "ferocious" in competitive intensity, with Hassabis's characterisation of the era not rhetorical but structural: "a race in which the resources of incumbents and the ambitions of challengers have converged to a point where institutional inertia is not merely a disadvantage but a disqualifying one". That pressure manifests in shorter research cycles, more frequent product milestones, and a general expectation that research outputs will influence Gemini or another shipping product within a reasonable timeframe.

The compensation sits at the top tier of the AI industry, though not at Meta's nine-figure poaching levels. Equity vests in Alphabet stock, which removes the liquidity uncertainty that affects private-lab equity. Benefits include the full Google benefits package: comprehensive health coverage, generous parental leave, meals, commuter support, and substantial learning stipends. Most staff are distributed across London, Mountain View, Zurich, Paris, and a growing set of other Alphabet research hubs.

The hardest cultural challenge, by multiple public accounts, is managing the talent departures. Microsoft's AI division poached more than 20 engineers and researchers from Google in 2025, including senior engineering leaders. Meta's Superintelligence Labs successfully recruited DeepMind researcher Jack Rae in June 2025. Ex-employees have started competitive labs — Reflection AI, Essential AI, and others — that operate with direct DeepMind lineage. The pattern is the familiar one of a strong research culture exporting senior talent to competitors, in the same way that PayPal exported its early team or Fairchild Semiconductor seeded much of Silicon Valley.

The mitigant is scale. Even with the departures, DeepMind has more than doubled in size since 2022, now employing approximately 8,000 people. The pipeline from top universities, PhD programmes, and Alphabet's internal mobility channels continues to produce replacements. The 2024 Nobel Prize in Chemistry, awarded to Hassabis and John Jumper for AlphaFold, remains the most significant external recognition of any AI lab's work and is a structural asset in the ongoing recruitment competition.

Honest acknowledgment of the challenges — the merger strain, the talent departures, the pace — is part of why studying DeepMind is useful. The company is operating at the scale most HR leaders will never face, but the organisational design choices it has made are unusually well-documented and translate broadly.

Frequently asked questions

How many employees does Google DeepMind have? Google DeepMind has approximately 8,000 employees as of early 2026, up from around 1,567 in 2022 and 5,600 in mid-2024. The unit was formed in April 2023 by merging the original London-based DeepMind with Google Brain, with staff distributed across London, Mountain View, Zurich, Paris, and other Alphabet research hubs.

What is the acceptance rate at Google DeepMind? The acceptance rate for engineering and research roles at Google DeepMind is estimated at less than 1% of applicants, according to recruiter analyses. Candidates with publications at top ML venues such as NeurIPS and ICML are reported to have 30–40% higher odds of reaching interview. Two-year retention sits at 78%, second only to Anthropic in the AI industry.

What is the Google DeepMind interview process? The process typically runs for 45 days end-to-end and includes a recruiter screen, a hiring-manager technical conversation, coding interviews, a machine-learning technical round, a research or portfolio discussion for research roles, team-lead interviews, and a people-and-culture interview. Research scientist candidates are assessed on demonstrated ability rather than publication count alone, with value placed on internships, industry experience, side projects, and open-source contributions.

Is Google DeepMind a research lab or a product team? Both. The 2023 merger of DeepMind and Google Brain created a unified unit responsible for frontier research, the Gemini model family, and AI infrastructure like the Tensor Processing Unit programme. CEO Demis Hassabis has described the post-merger unit as Google's "engine room" — a research-first organisation operating on a product cadence. Core research and applied teams sit under unified leadership, with staff moving between them based on project needs.

See also

• Talent & Culture Strategy at Anthropic: Mission-First Hiring in the AI Race

• Talent & Culture Strategy at OpenAI: Mission, Money, and the AI Talent War

• Talent & Culture Strategy at Microsoft: Growth, Diversity, and Learning Under Nadella

Ready to combine research-grade hiring rigour with product-scale pace? MokaHR's AI-powered ATS helps HR teams run structured, multi-stage interview processes that preserve quality as volumes scale. Book a personalised demo →