Demis Hassabis is no longer running a laboratory whose breakthroughs can be judged mainly by papers, games or scientific demonstrations. Google DeepMind now supplies the intelligence for products used by hundreds of millions of people, for cloud services sold to large enterprises and for experimental systems that can generate hypotheses, manipulate software and act across the web. Its chief executive has acquired influence over Alphabet’s growth, capital spending and risk profile that few research leaders have ever held.
The shift became explicit at Google’s May developer conference. Gemini 3.5 was presented as a family designed not only to answer questions but to complete agentic work. Gemini Omni joined reasoning with video creation and conversational editing. Computer-use capabilities, coding agents and a new enterprise agent platform moved the model closer to taking consequential action. At the same time, Google DeepMind introduced Gemini for Science and published work on a multi-agent Co-Scientist that can generate, challenge and refine research hypotheses.
These releases express a coherent ambition: one family of models should support consumer assistance, enterprise automation, creative production and scientific discovery. They also create a difficult operating test. The evaluation standard for a helpful search feature is not the same as the standard for a biological hypothesis, a security action or an autonomous workflow. Hassabis must make one research organisation move at product speed without allowing distribution pressure to outrun evidence.
From frontier lab to Alphabet operating system
Google DeepMind was formed by combining DeepMind with Google Brain in 2023. The merger gave Hassabis a concentrated pool of researchers, engineers and compute, but it also tied the group more directly to Google’s commercial priorities. Gemini now sits inside Search, Workspace, Android, Cloud and the standalone Gemini application. The laboratory’s performance is therefore visible through user growth, advertising economics and cloud contracts as well as benchmark results.
The scale is already striking. Google said in May that the Gemini application had passed 900 million monthly active users, more than twice the level reported a year earlier. Its model interfaces were processing roughly 19 billion tokens a minute, while 8.5 million developers were building with Google models each month. Across Google’s surfaces, monthly token processing had risen sevenfold in a year to more than 3.2 quadrillion. These are measures of activity rather than profit, but they show how quickly a research decision can become an infrastructure obligation.
Alphabet is financing that obligation aggressively. It spent $91.4 billion on capital expenditure in 2025 and expects $175 billion to $185 billion in 2026, mostly for servers, data centres and networking. Management has said just over half of its machine-learning compute this year is expected to serve Cloud, leaving a substantial allocation for DeepMind research and Google’s own products. Hassabis must compete internally for that capacity while ensuring the resulting models improve the economics of the businesses paying for it.
The financial evidence is encouraging but incomplete. Alphabet’s annual revenue exceeded $400 billion in 2025. In the final quarter, Cloud revenue grew 48 per cent and its backlog reached $240 billion, while products built on generative models recorded rapid growth. Gemini Enterprise sold more than eight million paid seats within months of launch. Yet the group does not disclose a standalone profit and loss account for DeepMind or for Gemini. Investors can see the capital bill and broad product momentum, but not the return on each layer of model development.
That opacity matters because frontier training, inference and safety work consume different kinds of resources. A model that lifts engagement in Search may protect an enormous advertising franchise. A scientific agent may require years of validation before it produces revenue. Open-weight Gemma models can strengthen developer adoption while giving away some immediate monetisation. Hassabis has to maintain a portfolio in which the most commercially legible work does not crowd out research that could create a much larger advantage later.
Science is the differentiator and the liability
Hassabis has long argued that advancing science is one of the strongest justifications for building more capable AI. AlphaFold gave that claim credibility by making predicted protein structures broadly available and useful to researchers. Google DeepMind is now trying to extend the approach beyond prediction. Co-Scientist uses multiple agents to propose and debate hypotheses. Gemini for Science offers tools for literature analysis, algorithm design and experimental reasoning. Other systems address weather, genomics and biological function.
The commercial logic is attractive. Scientific and industrial customers value accuracy, proprietary data and workflow integration more than entertaining conversation. If Google can make frontier models dependable in laboratories, engineering groups and healthcare systems, it can sell higher-value cloud services and establish deep institutional relationships. The work can also generate discoveries for Isomorphic Labs, Alphabet’s drug-design company, and reinforce Google’s claim that its AI spending creates benefits beyond advertising.
The validation problem is harder than the model demonstration. A plausible scientific hypothesis is not a finding. It must survive experiments, replication and peer review, often over years. As agents produce more candidate ideas, the bottleneck may move from generation to verification. That could increase demand for laboratories, high-quality datasets and expert judgement rather than reduce it. It may also create a flood of convincing but weak work that strains already limited review systems.
Biology adds a sharper risk. On 16 July, Google DeepMind and Isomorphic Labs set out a joint bioresilience programme built around prevention, detection and response. They said they had advanced more than 15 partnerships with governments, research groups and biosecurity organisations over the preceding year, while applying threat modelling, evaluations, mitigations and monitoring to powerful models. The initiative recognises the dual-use problem: the same system that helps design vaccines or analyse pathogens may lower barriers for malicious actors.
Hassabis therefore has to turn safety from a set of principles into an operating constraint that survives commercial urgency. Google DeepMind maintains internal responsibility and AGI safety councils and a Frontier Safety Framework. In June it joined partners in offering up to $10 million for external research into the behaviour of interacting agents. Such measures are useful, but the decisive evidence will come from deployment choices: what capabilities receive restricted access, how incidents are monitored and whether product teams accept delay when evaluations reveal material danger.
Asia is a laboratory for useful scale
Asia gives Google DeepMind a chance to demonstrate that its models can create value outside the US technology market. It also exposes weaknesses that English-language benchmarks can conceal. The region combines advanced research centres, huge multilingual populations, manufacturing strength and public services operating at exceptional scale. Models must adapt to local languages, cultural context, regulation and data constraints while running at costs that governments and institutions can sustain.
Singapore has become a base for that effort. Google DeepMind opened a research laboratory there and in May announced national programmes spanning healthcare, life sciences, education, workforce development and climate technology. The partnerships include work on AI support for clinicians, pandemic preparedness, tools for visually impaired athletes and training researchers to use agentic science systems. The group is also collaborating on multilingual and multimodal safety benchmarks, an important corrective to evaluation regimes dominated by Western datasets.
India offers a different test of reach and affordability. In July, Google DeepMind and the government’s Atal Innovation Mission began piloting ATL Saathi, a Gemini-powered assistant for teachers in tinkering laboratories. The initial deployment covers 100 schools, supports eight languages and helps educators create curriculum-aligned robotics and coding projects. The broader laboratory network reaches more than 11 million students. A successful pilot would show that a frontier model can be grounded in national materials and used through teachers rather than deployed as an unaccountable substitute for them.
These programmes have strategic value even when near-term revenue is modest. They create local feedback, institutional relationships and demand for Google infrastructure. They can also distinguish DeepMind from competitors that approach Asia mainly as a source of users. But public partnerships raise the standard of accountability. Errors affecting education or healthcare cannot be treated as ordinary product friction, and governments will increasingly demand data residency, auditability and domestic capability alongside access to the best model.
Competition compresses the research timetable
Google retains advantages that an independent AI laboratory cannot easily reproduce: custom Tensor Processing Units, a global cloud, vast distribution and cash flow from mature businesses. It also faces competitors unconstrained by the need to protect a dominant search franchise. OpenAI has set the pace for consumer adoption and agents. Anthropic has built a strong enterprise position. Meta distributes open models. Chinese groups have shown that capable systems can be trained and served with different cost structures, intensifying price and efficiency pressure.
Hassabis’s answer is breadth allied to integration. Gemini spans text, audio, images, video and action; Gemma provides an open route; DeepMind’s science work supplies a distinctive mission. Yet breadth can dilute attention. Product users expect reliability and speed. Cloud customers expect predictable economics. researchers expect openness and methodological rigour. Safety teams need authority. A single organisation can serve all four constituencies only if priorities and evaluation gates are unusually clear.
The deeper governance question is the division of responsibility between Hassabis and Sundar Pichai. DeepMind develops the core models, while Google’s product and infrastructure groups decide how they are distributed and monetised. That arrangement can combine scientific depth with operational scale. It can also blur accountability when a model problem emerges in a product, or when a commercial deadline influences a research release. Alphabet needs the interface between laboratory and business to be as carefully engineered as the models themselves.
Hassabis has already moved DeepMind farther into daily economic life than its early history suggested. The next phase will not be proved by another benchmark lead. It will be proved when agents complete valuable work without creating unacceptable errors, when scientific systems produce validated discoveries rather than abundant conjecture, and when Alphabet can show that unprecedented infrastructure spending earns durable returns. Maintaining all three outcomes at once is the real frontier now facing Google DeepMind.
