Google DeepMind's Research Map Now Extends Beyond Its Payroll
On August 5, Google published a new map of the company’s research leadership. Demis Hassabis would move from Google DeepMind chief executive to chairman, add the role of Alphabet chief scientist, and continue leading the drug-discovery company Isomorphic Labs. Koray Kavukcuoglu would become senior vice president of Google DeepMind, reporting to Alphabet CEO Sundar Pichai.
Google’s published employee messages said Jeff Dean and Sanjay Ghemawat were launching an independent public benefit corporation. Google would remain involved as a founding investor and Cloud partner, collaborating on a research framework for machine-learning systems and related infrastructure. Axios reported that Oriol Vinyals and Quoc Le were also leaving, and identified the new company as Discovery Loop.
Work once concentrated inside one organization now sits in three settings. One leader will set scientific direction across Alphabet. Another will run the lab and ship Gemini. Four researchers are expected to work outside Google’s payroll, with Google retaining a financial and computing relationship with their new company.
A paper can become a model capability, a cloud product, or a company. That range turns the announcement into an employment story as well as an executive succession. Google can offer salary, equity, colleagues, vast compute, and distribution. An outside founder can claim more control over the research agenda and a larger share of whatever the work creates. Discovery Loop appears to put both offers around the same table.
Several important terms remain private. Google has disclosed no investment amount, ownership percentage, board rights, cloud credits, intellectual-property agreement, hiring plan, or customer contract for Discovery Loop. Public reporting also leaves each researcher’s motive unknown and establishes no causal link between the leadership change and the announced exits.
Five decisions provide an operating test: who selects the research problem, who allocates compute, who controls a product deadline, who owns the resulting asset, and who receives the upside. Titles describe the new map. Those decisions will reveal how it works.
Research authority moves inside and outside Google
Hassabis had carried two jobs under one title. He represented the scientific identity of DeepMind and ran the organization that turns frontier work into Gemini products. The August change divides those responsibilities without removing him from either field.
As Alphabet chief scientist, Hassabis can shape research across the parent company. As Google DeepMind chairman, he retains a formal connection to the lab. His continued leadership of Isomorphic Labs adds a third setting where research is expected to become a commercial system.
Kavukcuoglu receives the operational side. Google’s message says he will oversee Gemini model development, frontier AI research, and the Gemini app and developer teams. A roadmap assigns scarce things: accelerator time, engineering teams, evaluation capacity, release windows, and executive attention. It turns a research option into a dated product commitment. Reporting directly to Pichai puts that work close to Alphabet’s capital allocation and product priorities.
This is a succession from inside the lab. Hassabis wrote that he and Kavukcuoglu have worked together for more than 13 years, dating to DeepMind’s early days. Pichai pointed to Kavukcuoglu’s work on WaveNet and DQN and said the Gemini models had already been in his hands with other leads. The public case for continuity rests on that shared history, not on a newcomer arriving with a turnaround mandate.
Discovery Loop is planned as a third locus of authority. Its founders are leaving the employee hierarchy. The organization is described as independent, yet Google will remain both an investor and a Cloud partner. Its founders seek room to select projects and build a company. Google keeps a route to the people, workloads, and possible discoveries without keeping them as employees.
Here is the operating map supported by current disclosure:
| Research setting | Disclosed leader or founders | Decision right indicated publicly | Economic tie | Material term still missing |
|---|---|---|---|---|
| Alphabet scientific direction | Demis Hassabis | Company-wide scientific leadership | Alphabet executive role | Budget, veto power, staff, and relationship to product groups |
| Google DeepMind operations | Koray Kavukcuoglu | Gemini models, frontier research, Gemini app, and developer teams | Alphabet employment and product economics | Budget, hiring, publication, and release authority |
| Google DeepMind oversight | Demis Hassabis | Chairman role | Alphabet employment | Scope of approval, escalation, and performance accountability |
| Discovery Loop | Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le, according to Axios | Independent company formation | Google as founding investor and Cloud partner; joint research framework announced for systems and infrastructure | Ownership, board rights, compute terms, IP, hiring, customers, and detailed benefit purpose |
This map prevents several easy overstatements. An investment does not establish control. A Cloud partnership does not prove subsidized compute. A chairman title does not reveal which product decisions require approval. Independence does not reveal whether prior Google inventions, future joint work, or employee recruiting are covered by special agreements.
A researcher considering the move will care more about the columns than the prestige of each title. An internal lab can supply a large team, proprietary infrastructure, established evaluation systems, and immediate product reach. An external company can offer founder equity and a clearer claim on the agenda. A corporate chief-scientist role can range across projects while leaving daily staffing and delivery to someone else.
Managers below the executive layer will feel the division first. The friction will be mundane. A research lead wants another training run. A product lead needs a capability in the next Gemini release while a study needs more time. A scientist has to choose among publication, patent, and product. Each needs to know who can settle the conflict.
Titles and reporting lines are now public. Escalation rules are still internal. Decisions will show whether the open interfaces have owners.
The 2023 unit remains inside a wider network
Google created the current structure in April 2023 by combining DeepMind and the Brain team from Google Research. The DeepMind announcement described the new organization as “a single, focused unit” intended to accelerate progress and bring research into products. Hassabis became chief executive. In a companion Google announcement, Pichai named Dean chief scientist for Google Research and Google DeepMind, reporting to him on research direction.
Combining the teams addressed a real coordination problem. Frontier systems depend on more than a model idea. They need large-scale computing systems, training data, safety work, evaluation, product engineering, distribution, and a budget that can tolerate experiments with uncertain returns. Housing the Brain team and DeepMind together shortened some organizational distances among those resources.
It also concentrated different professional clocks. Research is rewarded when a difficult question yields a result that survives scrutiny. Product work is rewarded when a capability reaches users reliably and on schedule. Infrastructure work is rewarded when many teams can run safely, quickly, and at an acceptable cost. A single unit can align those clocks, or it can make one dominate the others.
Dean’s departure carries institutional weight because his career joined research and infrastructure long before the current model race. His official Google profile says he joined the company in 1999 and co-founded Google Brain in 2011. His work spans machine learning and the systems used to train and serve it. Ghemawat’s Google profile dates his arrival to late 1999 and traces work on the distributed systems behind Google’s scale.
Vinyals and Le connect that systems history to the model era. Google’s May 2026 Gemini 3.5 launch listed Vinyals as a Google DeepMind vice president alongside Kavukcuoglu and Dean. Le’s research profile records work across deep learning, representation learning, and automated machine learning. A byline or profile proves proximity to the work, rather than sole ownership of a model or roadmap.
Their planned departure will remove more than four individual contributors. Long-tenured researchers carry a map of how decisions have worked across reorganizations: which infrastructure teams can solve a bottleneck, which evaluation failed before, which collaboration survives a product deadline, and which ambitious project can earn another year. Some of that knowledge can be documented. Much of it travels through judgment and trust.
That loss is felt below the executive layer. A staff researcher may lose the person who knew how to protect an uncertain project through a budget review. A systems engineer may lose an informal route to a model team. A manager hiring a specialist may need a new senior sponsor. Google can promote other leaders and recruit replacements, but the handoff cost appears in slower introductions, repeated context, and decisions whose precedent left with a colleague.
The quartet can preserve its working relationships at Discovery Loop. Old access to teams, internal data, production feedback, and unpublished context stays behind unless agreements reopen it. Google’s investment and Cloud partnership preserve a commercial bridge. They change the candor and incentives of people who once shared an employer.
Google has kept the 2023 unit and changed its edges. Gemini delivery stays inside Google DeepMind. Scientific leadership moves upward across Alphabet. A separate group of veterans moves outward. The result is a network around the unit that 2023 tried to focus.
Such a network may fit a mature research portfolio. Some projects need close product integration. Some need a company-wide scientific sponsor. Others may attract founders only if they can build outside the corporate hierarchy. Supporting all three requires more negotiation across boundaries.
Box count is a poor test of success. Product teams need clearer decisions, scientists need room for uncertain work, and external founders need a collaboration path with defined control. If every hard choice returns to the same small group, the new map adds titles without adding capacity. Authority that follows the roles could let specialists run different clocks.
Gemini delivery moves under an operator
Kavukcuoglu inherits a product with enormous reported reach. In its second-quarter 2026 earnings exhibit filed with the SEC, Alphabet said the Gemini app had 950 million monthly active users. Gemini models were processing 22 billion API tokens per minute. Nearly 90 percent of the Fortune 100 used Gemini Enterprise, according to the company.
The figures measure three different things. A monthly active user may open the app once or use it daily. Token volume captures processing, not customer value. Fortune 100 presence omits depth, paid seats, workflow completion, and return on investment. Even with those limits, a research decision can reach consumers, developers, and enterprise buyers quickly.
Roadmap control at that scale is an operating job. A release needs a model-readiness call, safety and security blockers, infrastructure capacity, a usable interface, and customer commitments. Those schedules move together. Optimizing one in isolation simply pushes delay or risk into another team.
Google’s scientific portfolio also reaches beyond the Gemini app. A May 2026 Google Research update described Gemini for Science, an AI co-scientist, Earth AI, and Computational Discovery. One disclosed system reviewed 10,000 papers before researchers experimentally tested a proposed biological mechanism. These projects connect model development with scientists, laboratories, and domain-specific evidence.
Breadth produces choices that a generic call for more innovation cannot settle. A scientific assistant may need a slower evaluation path than a consumer feature. An enterprise agent may require controls, connectors, and support work that do not improve a public benchmark. A discovery model may need expensive experiments whose value arrives through knowledge rather than immediate revenue.
Separate roles for Hassabis and Kavukcuoglu could make those choices more legible. Scientific direction can defend a portfolio whose payoff is distant. Day-to-day leadership can decide what the next Gemini release needs and which dependencies are late. Both roles need a common arbitration point before either promises resources that the other must supply.
Pichai’s reporting line provides that point at the top. It also raises the stakes of each escalation. A chief executive should decide a few portfolio conflicts, not serve as the routine scheduler for research and product. The useful evidence will appear lower down: faster approvals, fewer repeated reviews, stable research programs, reliable launches, and clear ownership when a capability crosses teams.
Employees can read the same evidence through their calendars. If the new structure works, researchers spend less time restating a case to several leaders. Engineers receive an answer when a paper result becomes a product dependency. Evaluators know which risk owner can delay a release. Program managers can plan scarce compute and lab access before a public date becomes immovable.
Enterprise customers have a narrower concern. They need API stability, security response, capacity, documented model changes, and a reliable route for escalation. A famous researcher leaving does not establish a service problem. A new operator earns credibility with those customers by keeping commitments while the research organization changes behind the interface.
Outside observers lack a baseline for those measures. Approval time, roadmap churn, cross-team transfer time, researcher regretted attrition, and experiments stopped for product work are undisclosed. Model launches alone will not isolate the effect of a leadership change. A cleaner chain of decisions may be visible internally months before it appears in a benchmark.
Shared research costs stay outside the segments
Alphabet’s accounts show the scale around the reorganization without providing a Google DeepMind budget. Second-quarter research and development expense reached $18.219 billion, up from $13.808 billion a year earlier. Capital spending was $44.924 billion, almost double the prior year’s $22.446 billion. Company headcount reached 198,933, an increase of 11,830 from June 2025.
All three numbers cover Alphabet, rather than DeepMind alone. R&D spans the wider workforce and project portfolio. Capital spending supports servers, data centers, networking equipment, offices, and other assets across products. Headcount leaves the number of researchers, Gemini engineers, evaluators, and departing employees undisclosed.
One accounting line gets closer while remaining broad. Alphabet-level activities recorded a $5.789 billion operating loss in the quarter, compared with $3.372 billion a year earlier. The company said the category primarily reflects shared AI research and development. Its quarterly report also places corporate initiatives, shared functions, severance, and office costs in that line.
Alphabet keeps the category outside Google Services and Google Cloud segment results. That placement matters when comparing research investment with product revenue. Shared model work can support Search, Cloud, Workspace, Gemini, and future products. Assigning all of the cost to one segment would misstate the portfolio. A shared category also prevents an outside reader from calculating the economics of Google DeepMind or one model family.
New roles change the people who influence this pool while its allocation stays opaque. Hassabis will shape scientific direction across Alphabet. Kavukcuoglu will oversee models, frontier research, the app, and developer teams. Other product and research leaders compete for the same infrastructure and specialized employees, with Pichai above them. The allocation formula remains internal.
Compute is part of a researcher’s employment offer because it determines which ideas get tested. Access depends on total capacity, hardware type, deadline priority, systems support, data permission, and patience for repeated failure. A huge corporate fleet still feels scarce when an experiment loses every scheduling contest. A startup with less capacity can feel freer when its founders control the queue.
Discovery Loop’s Google Cloud relationship puts this issue at the center of its design. Cloud turns an internal resource contest into a commercial service with a price and contract. A priced contract can offer a clearer budget and fewer internal claimants. It also brings capacity limits, obligations, and a bill. Published terms reveal none of that balance yet.
Compensation has the same missing denominator. Alphabet reports employee stock compensation across the company. It does not publish the pay, retention grants, or forgone equity of the four departing researchers. Discovery Loop’s founder ownership has not been disclosed. Any comparison between their Google package and startup upside would be speculation.
Rising shared AI R&D or capital spending can accompany good or bad allocation. A high Gemini user count measures reach, not the return on every research dollar. The operating case links a specific capability to usage, revenue or strategic option value, direct and shared costs, and the time required to learn. The reorganization can change that path before quarterly accounts reveal it.
Employees will see the economic signal in concrete choices. Which projects receive another training run? Which teams can hire? Which researchers can publish? Which product deadline overrides a longer experiment? Money becomes organizational authority through those decisions.
Discovery Loop leaves with a commercial tie
The phrase “independent public benefit corporation” sounds precise while leaving most of the company unknown. It identifies a corporate form. Discovery Loop’s capitalization, board, voting rights, revenue model, and relationship to inventions made at Google remain undisclosed.
TechCrunch, citing the company’s launch release and a New York Times interview, reported a plan to use AI to accelerate scientific research. Discovery Loop wants to run many experiments in parallel, automate parts of the experimental loop, and study systems that improve AI itself. Radical Ventures and Khosla Ventures co-led its funding round, with Kleiner Perkins, Lightspeed, Doerr Capital, and Alphabet also participating, according to the report.
Dean also described the organizational appeal in the New York Times interview summarized by TechCrunch. A public benefit corporation, in his view, gives researchers more latitude to make scientific choices that do not serve only a company’s financial interests. That statement supports a preference for a different decision environment. It does not establish dissatisfaction with Google, and the other founders have not publicly supplied the same motive.
Those statements describe an ambition, rather than a demonstrated system. Discovery Loop has published no operating product, benchmark, customer result, completed experiment count, or technical paper that establishes the promised automation.
Google’s announcement supports two ties to the former employer. Google will be a founding investor. Google Cloud will be a partner. The sides also plan a research framework for machine-learning systems and infrastructure. Those ties create a middle position between employment and a clean break.
The attraction for founders is agenda control with access to industrial computing. They can recruit for a new mission, hold founder equity, and decide how research becomes a product or service. A known investor and infrastructure supplier shortens the route to capital and a computing stack.
Google preserves several options. A promising discovery may become a cloud workload, a partnership, an investment return, or a future product relationship. The company can remain close to researchers after their employment ends without carrying every new project inside its headcount and reporting structure.
Each benefit carries a conflict that contract terms must settle. An investor may want information or commercial access. A cloud provider can see usage patterns and holds leverage through pricing and capacity. Founders need freedom to choose customers and publish work. Google needs protection for confidential information and intellectual property developed before departure.
Future employees face a simpler question: are they joining an independent company or a lab economically dependent on one partner? The answer will sit in revenue concentration, infrastructure portability, voting rights, and the freedom to choose research and customers.
Public labels leave those conflicts to contracts. Board representation, a right of first refusal, exclusivity, preferred access to discoveries, and special hiring terms are undisclosed. So are Discovery Loop’s freedom to move workloads across providers, its access to Google data, and any negotiated recruiting protocol.
The label “spinout” would imply a transfer that has not been disclosed. Public reporting identifies no Google business, employees, patents, or assets moving into the new entity. “Outsourced research” would imply a research contract. Google has announced an investment, a cloud relationship, and a collaborative framework instead.
What is known is unusual enough. Four researchers are reported to be leaving the hierarchy to become founders. Google will become an investor, supplier, and research collaborator. That change replaces managerial authority with negotiated rights.
It also changes the talent proposition for future hires. A Google employee joins a large institution with liquid public-company equity, internal mobility, established systems, and a broad product surface. A Discovery Loop employee may receive concentrated equity, proximity to founders, and a narrower mission, while accepting startup risk and a less certain path to market. Cloud access reduces one startup barrier. It does not supply customers, research breakthroughs, or organizational stability.
Lab leaders elsewhere can study the structure without copying it. The relevant decision is which relationship fits a project and its people. Work tightly coupled to a live product may belong inside. Work that needs an independent agenda may fit a venture. A senior scientist whose value spans many groups may fit a corporate research role. Forcing all three into one employment contract can make departure the only route to authority.
Shazeer’s return lasted two years
Google has already tried an expensive route for bringing a prominent researcher and team back inside. In 2024, Character.AI granted Google a non-exclusive technology license while co-founders Noam Shazeer and Daniel De Freitas and some colleagues joined Google. Axios later reported a price of about $2.7 billion for the broader license and talent arrangement. That figure was not personal compensation to Shazeer.
In June 2026, Shazeer left for OpenAI, according to Axios reporting on the move. His return to Google lasted roughly two years. Google received technology rights and employee work during that period. The agreement had value and still left Shazeer free to make another career decision.
Discovery Loop adds another model. Google can retain a relationship after employment ends, rather than pay to bring every researcher back. Whether that proves cheaper or more productive depends on undisclosed terms and future work. An equity investment could become valuable. Cloud usage could generate revenue. The researchers could build something outside Google’s product priorities. Any of those paths may occur together.
Frontier research talent is often discussed as a compensation auction. Money matters, especially when competing companies can offer equity tied to different valuations. The complete offer has more parts:
| Researcher choice | Value offered | Cost or risk carried by the researcher | Evidence a lab can monitor |
|---|---|---|---|
| Stay as an employee | Salary, liquid equity, colleagues, compute, distribution, internal mobility | Possible limits on agenda control, product deadlines, diffuse ownership | Regretted attrition, project continuity, internal moves, accepted offers |
| Take a company-wide science role | Broad mandate, executive access, ability to connect fields | Distance from daily experiments, influence dependent on delegated rights | Budget decisions, cross-company projects, time to resolve conflicts |
| Found an external company | Founder equity, mission control, hiring authority, customer choice | Financing risk, infrastructure bills, recruiting burden, uncertain demand | Team formation, capital terms, publications, products, customer work |
| Keep a commercial tie to the former employer | Faster infrastructure access, known partner, possible route to market | Dependence, negotiation overhead, possible limits on strategic freedom | Cloud concentration, contract renewals, partner revenue, governance rights |
Different people will price these rows differently. A systems researcher may value access to huge production workloads. A scientist may value the freedom to pursue a result that has no product date. A founder may accept less liquid wealth for concentrated ownership. A research manager may prefer a broad institution because recruiting, legal support, and evaluation already exist.
Prominent exits are an incomplete retention measure. Hassabis told Semafor in June that Google still had the broadest bench and won its share of contested hires. That is a management judgment without a denominator.
SignalFire’s 2025 talent report estimated two-year retention at 78 percent for DeepMind, compared with 80 percent at Anthropic and 67 percent at OpenAI. Its proprietary career-profile data covered an earlier cohort and the wider companies, rather than 2026 frontier researchers. The figures provide context, not a current attrition rate.
A talent leader also has to look below celebrity hires. Research organizations depend on research engineers, infrastructure specialists, evaluators, program managers, and scientists who may never appear in a launch byline. If founder-style autonomy becomes available only to a handful of stars after they resign, mid-career employees receive a clear signal about the limits of the internal ladder. Internal ventures, project ownership, publication rights, and transparent compute decisions can widen that ladder before an external offer arrives.
Mobility also has a wider economic role. A 2026 OECD working paper on noncompete clauses estimates that about 20 percent of private-sector employees across 15 countries definitely face a noncompete, with probable coverage bringing the share closer to one-third. A 10 percentage point higher incidence was associated with 1.9 percent lower aggregate productivity in the paper’s cross-country model.
OECD’s result is an association from a cross-country model, rather than a randomized causal estimate. It says nothing about the contracts of Dean, Ghemawat, Vinyals, Le, or Shazeer. Public evidence connects none of their moves to a noncompete. The study supplies a broader point: knowledge moves through people, and restrictions that protect one firm can slow diffusion across a market.
Trade-secret protection, confidentiality, and employee mobility require separate treatment. A company can protect confidential work without claiming ownership of every skill an employee developed. A startup can hire experienced researchers without gaining rights to a former employer’s code, data, or unreleased plans. Discovery Loop’s agreements will draw those boundaries, even if the public never sees them.
Retention is an organizational measure as well as a pay measure. A lab can compare cash and equity. It also needs to know how quickly a researcher gets a decision, whether collaborators can be hired, whether compute arrives when promised, whether publication is possible, and whether the person can own a new line of work. A clearer agenda may outweigh a larger package.
Gemini’s next milestone will test the new map
When Google reviews its next Gemini milestone, the discussion will run through a new chain of command. Kavukcuoglu will oversee models, frontier research, the app, and developer teams. Hassabis will hold broader scientific roles. Four veterans who carried decades of Google context are preparing a separate institution.
Early evidence will be mundane. Teams will request capacity, ask to hire, propose publication, delay an evaluation, or move a capability into a product. Each request will reveal whether the new roles shorten a decision or add another approval. Employees will learn the real map before outside observers do.
Several public signals can narrow the uncertainty over the next year:
- Google can describe how the chairman, chief scientist, and DeepMind operator divide research selection, model release, and resource escalation.
- Alphabet can keep separating shared AI R&D from segment economics while adding operational measures that connect cost to delivered capabilities.
- Discovery Loop can disclose its mission, founders’ roles, financing, benefit purpose, infrastructure concentration, and first product or research program.
- Research output can show whether long-horizon programs continue inside Google while Gemini maintains a reliable release cadence.
- Hiring and further departures can indicate whether each organization has an offer that researchers accept, though public moves will remain an incomplete attrition measure.
Causal attribution will remain difficult. Model quality depends on prior work, compute, data, evaluation, engineering, and market timing. A publication may have started under the old structure. A launch delay may be prudent. A departure may reflect family, ambition, money, collaborators, or a project that only works as a company.
Boards and lab leaders should review the organization at the level of contested decisions. Counting papers, launches, or departures alone rewards one clock and ignores the others. Follow the research problem, compute allocation, product transfer, publication, staffing, and ownership from proposal to outcome.
Record the responsible leader, people consulted, time to answer, resource committed, reason for stopping or proceeding, and later result. Patterns will emerge. A product organization may be waiting on evaluations. A science group may be losing months to repeated budget reviews. A founder path may exist only after an employee resigns. Those are design problems that a new title can solve only when authority moves with it.
Google’s August messages create three live comparisons inside one corporate orbit. There is an internal product lab under an operator, a company-wide scientific leader, and an independent founder team linked by capital and cloud. The people, infrastructure, and possible discoveries still overlap. Their contracts and decision rights no longer do.
A breakthrough will attract the headlines. The quieter result will decide whether the structure lasts: where the idea began, who was allowed to pursue it, who supplied the compute, how it reached users, and whether the people who built it chose to stay for the following problem.