What was actually opened
The five models cover different fields. AlphaFold 3 predicts protein and biomolecular structures, including complexes with ligands and nucleic acids. AlphaGenome works on genomic sequences and the effects of variants. WeatherNext is DeepMind's weather forecasting system. AlphaEvolve searches for algorithms and code improvements. AlphaEarth Foundations produces geospatial embeddings for mapping and Earth observation.
The announcement was reported by The Next Web and confirmed by Google DeepMind and Google Cloud. Access is granted to a maximum of 200 European scientific researchers. That is the cap in the announcement, and nobody has said whether it will be raised.
Two hundred is a small number against the size of the European research community working on computational biology, climate modelling and materials. It is a pilot, not a distribution programme.
Two hundred seats and $3 million in credits
Google Cloud is contributing $3 million in compute credits. Spread evenly across the full 200 seats, that is $15,000 per seat, about €13,000 at current rates.
Whether that is a lot depends entirely on what you run. Training a genomics model from scratch is not happening on $15,000 of cloud time. Fine-tuning an already released model, or running large batches of inference on AlphaFold 3 or AlphaEarth Foundations, fits comfortably. The credits buy participation; they do not buy a frontier-scale training run.
RAISE itself is the bigger number. Resource for AI Science in Europe is backed by €107 million from Horizon Europe for the pilot phase. That money is not divided between the 200 seats, since RAISE funds the wider initiative, and most of it is for training and attracting AI researchers and for securing computing time at the planned EU AI gigafactories. Anyone reading the two figures as a per-head budget is reading them wrong.
For contrast, our own AI Arena rig runs local models on a single 16 GB GPU and measures tokens per second at zero marginal cost. That is a completely different budget profile, and it is the one most European labs actually work with day to day.
The AlphaFold team has been reorganised
In July, Google DeepMind broke up the AlphaFold team and reassigned most of its members to other initiatives, according to The Next Web. The reshuffle followed the departure of John Jumper, the Nobel laureate who led the work, a month before the reorganisation.
So the models being handed to European scientists were built by an organisation that no longer keeps that team together. The announcement describes access and does not describe long-term maintenance commitments. A lab planning a three-year project on top of AlphaFold 3 is building on a released artefact, not on a dedicated group.
This is normal in industrial research, and it is also the sort of thing that should be written into a project plan rather than discovered in year two.
Europe already has compute routes
The compute gap RAISE is meant to close is real. European academics have long complained that they compete for GPU hours while well-funded labs treat clusters as a fixed cost.
Other options exist. Google has said it will invest €13 billion (around $14.6 billion) in AI infrastructure in Finland through 2028. That is corporate investment rather than academic compute allocation, but it changes the regional supply of compute.
A researcher who misses the 200-seat cap is not out of options. Open weights on your own hardware is slower, but nobody can revoke it.
Data location and the AI Act
The credits are Google Cloud credits, which means the work runs on Google infrastructure. For a genetics group handling human genomic data, that raises the standard GDPR questions: where the data is processed, what the legal basis is, and which processing agreement covers the transfer. European institutions handle this routinely through standard contractual clauses and adequacy decisions. It is still paperwork a project has to plan for, and a model access grant does not remove it.
On regulation, Chapter V obligations for general-purpose AI models have applied since 2 August 2025, with Article 50 transparency duties and oversight by the European AI Office running from 2 August 2026. Pure research use sits at the lighter end of the scale. A university project that later turns into a commercial product inherits the obligations of whatever it becomes.
Separately, the Council's 2026 Digital Omnibus moved the compliance deadline for standalone Annex III high-risk systems from 2 August 2026 to 2 December 2027. Labs that had been planning around the earlier date can slow down.
What a researcher can do with this
Access is described as running through RAISE and the European Commission, not through a direct Google application. Details on eligibility and selection come from the initiative rather than from DeepMind's blog.
If you do get a seat, the sensible move is to treat the credits as a fixed budget and decide early what consumes them. Inference at scale on AlphaEarth Foundations or AlphaGenome is a reasonable use. Iterative retraining is not.
Do I need to move my data to Google Cloud to use the models?
The compute credits are Google Cloud credits, so work funded through them runs on Google infrastructure. Projects handling personal or clinical data still need a legal basis and a transfer mechanism under GDPR, exactly as with any other cloud provider.
Can I download the models instead of using them through the programme?
The announcement covers access to the models, not redistribution of weights. Some of DeepMind's earlier releases have been published openly with restrictions on commercial use, but the terms differ per model and should be checked individually before you plan around them.
Is the €107 million from Horizon Europe part of the 200 seats?
No. That figure funds the pilot phase of the RAISE initiative as a whole, with most of it for training and attracting AI researchers and for securing computing time at the planned EU AI gigafactories. The 200 seats and the $3 million in Google Cloud credits are the parts attached specifically to the DeepMind model access.
