Google DeepMind Opens Science AI Models to 200 EU Researchers

What Google announced

Google DeepMind and Google Cloud say they will give up to 200 European scientific researchers direct access to their frontier science models. The offer also includes a shared pool of $3 million in computing credits.8 Google announced the programme on October 14, 2026, around an AI summit in Dublin. Its statement said the aim was to speed up European innovation and strengthen the continent's scientific competitiveness.413

The package names five systems: AlphaFold 3, WeatherNext, AlphaGenome, AlphaEvolve and AlphaEarth Foundations.8 Each covers a different area. AlphaFold 3 predicts the structures of proteins and other molecules. WeatherNext produces weather forecasts. AlphaGenome is used for DNA research. AlphaEvolve works on improving computer algorithms. AlphaEarth Foundations supports research about the planet.15 Google's examples of what the tools could help with included predicting extreme weather and mapping molecular structures.8

The programme runs through the European Union's Resource for AI Science in Europe, known as RAISE.1 According to one account, the announcement came from Maureen Costello, Google Cloud's vice president for the UK, Ireland and Sub-Saharan Africa.1

Who decides who gets in

The reports disagree on one practical point: who will choose the researchers. Reuters-based and AFP-based coverage says the European Commission will pick participants under RAISE.1113 Greek City Times, citing AFP, says Google told the agency that participants would come from among RAISE's existing beneficiaries. Several outlets present this as a split in roles, with Google providing the models and compute and Brussels deciding who uses them.210

Another report says Google did not explain how researchers will be chosen, when access begins or how the credits will be divided.6 These accounts can both be true. The Commission may own the selection without anyone having published the criteria. Most of the coverage agrees on what is still unknown: there is no published application process, eligibility rules, timeline or length of access.93

In my view, the Commission's role as gatekeeper is the most important part of the design. A public body choosing participants makes this look more like a public research programme than a corporate giveaway. It also gives Google some protection from claims that it is handpicking European scientists who are friendly to its products.

The open-source question: access is not openness

For anyone following AI research through the lens of open science, the key fact is what the programme does not include. Getting access to a frontier model does not mean getting its weights, training code or datasets. Google has said nothing about releasing these five systems as open-weight models through this programme.1 Another analysis agrees. It found nothing in the announcement about publishing weights, code or data, and said this kind of access helps research output but does not give the reproducibility that open models offer.9

This matters for how the research will work in practice. If a paper depends on a model that other labs cannot inspect or run themselves, peer reviewers and other scientists may struggle to check the results. That is especially true once the credits run out. One analysis advises researchers to record which results depend on proprietary models, so they can repeat key steps with open alternatives later.9 The same analysis notes that some of the named systems, including AlphaFold 3, already have separate public tools or terms. Researchers should check each model before assuming what the programme adds.9

PolicyWire's description is telling. It calls the initiative part of Google's mission to widen access to its proprietary AI technologies.7 That is accurate. This programme gives more people access to closed tools. It does not make those tools open, and the two should not be treated as the same thing.

Is $3 million meaningful?

The coverage agrees on the numbers but differs on what they mean. If the $3 million were split evenly across all 200 places, each researcher would get about $15,000. Analysts who did that calculation point out that Google has confirmed no such allocation.19 One analysis says that amount could pay for a lot of model use and some fine-tuning, but nowhere near enough to build a foundation model from scratch.9 Another says the sum matters to individual labs but is small compared with European research spending overall.2

Others describe it more generously. TechBuzz calls it Google's biggest European research bet so far, and says the credits remove a major barrier for European institutions short on resources.14 Analytics Insight is more cautious. It says the real benefit depends on how the credits are split, how long they last and which services they cover. It also warns that short-term access may not be enough for projects that take years.15

I side with the cautious view. Measured against Google's other spending in the region, the credits are small. Alphabet has set aside about €13 billion for AI infrastructure in Finland through 2028.16 RAISE's own pilot has €107 million in EU funding, most of it for training and attracting AI researchers.6 Next to those figures, Google's credits are a gesture of support, not a new source of research funding. Their real value is access to specialised models that are otherwise restricted, bundled with enough compute that a lab doesn't need to find money in its budget before it starts.9

How RAISE fits in

The European Commission launched RAISE in November 2025 as a virtual institute.6 It pools computing resources, scientific data, funding and expertise from across Europe, so scientists can use them without everything sitting in one physical institution.3 It is part of the Commission's wider European AI in Science Strategy and is funded through Horizon Europe.3 RAISE is also meant to secure computing time for scientists at the EU's planned AI gigafactories. A €20 billion fund was set up in February 2025 to build up to five of these large computing centres.6

Google's contribution should therefore be seen as one private add-on to a public framework, not as funding for RAISE itself. One explainer makes this point directly: Google's $3 million is separate from RAISE's budget, with different funding sources, governance and intended recipients.1

A template Google has used before

Europe is not the first place Google has tried this. It has announced similar arrangements with the UK government and the US Department of Energy, giving access to specialist models for scientific work.3 In July 2026, Google committed $40 million in AI tokens and cloud credits to researchers in the Energy Department's Genesis Mission, including access to AlphaEvolve, AlphaFold 3 and AlphaGenome.5

The US commitment was more than ten times larger than the European one. That gap shows the European programme is a smaller version of an existing template, not a new strategy built around Europe. Other AI developers have run similar programmes, such as Anthropic's credits for US federal science agencies and NVIDIA's funding for UK safety computing. Commentators note that these programmes win goodwill, early users and feedback on how models perform in real research.9

The dependency argument

The sharpest disagreement in the commentary is over what Google gains. TechBuzz argues that the programme trains a pipeline of scientists on Google's AI approach. In its view, the resulting papers, students and course material spread Google's influence well beyond the $3 million.14 BigGo Finance makes a similar point, describing the move as positioning Google's models as scientific infrastructure and a way to shape which tools institutions later pay for.16 More neutral accounts raise the same issue more gently, telling institutions to weigh the benefits against data protection, intellectual property, research independence and the long-term cost of relying on commercial platforms.18

I find the dependency argument convincing, though less as a hidden plan than as a predictable result. Credits run out, and researchers may then need to keep using the same provider.1 Data rules add another complication. Hosted models can cause problems for sensitive data such as patients' genomes, and where the data is processed matters for EU data-protection compliance.9

The backdrop: calls to contain frontier AI

The announcement came amid growing demands for tighter AI regulation after a series of security incidents.13 The day before, about 400 researchers from 34 countries, including Geoffrey Hinton and Yoshua Bengio, called for an international body to set containment standards and track security incidents involving advanced models.11 Their open letter says frontier models should at first be cut off from external networks and placed under independent oversight. It also argues that voluntary commitments from companies are not enough.11

Google's programme involves specialised science models rather than general-purpose agents, so it does not directly address those worries. Still, the timing points to a trade-off that will shape AI research in Europe: wider access for scientists on one side, and pressure for tighter controls on frontier systems on the other.

The bottom line

This is a real but small programme. It gives selected European scientists useful tools, and having the Commission choose participants is sensible. But it is controlled access to proprietary models, not open science, and its value depends on rules Google has not yet published.9 Advocates of open research should welcome the access, push for clear rules on whether results can be published and reproduced, and keep backing Europe's public computing efforts. In the long run, those public efforts are what would give European researchers alternatives to models they can only use on a company's terms.