Google has slashed DeepMind’s TPU compute allocation to accelerate the Gemini language model, forcing the lab to pause key scientific research. Internal documents show scheduled updates to the GNoME materials discovery project and next-generation structural biology models are delayed until at least 2027 to reallocate hardware to Gemini.
How much did DeepMind’s TPU compute allocation drop?
According to leaked internal Google Cloud scheduling spreadsheets, DeepMind’s science teams saw their dedicated Tensor Processing Unit (TPU) cluster allocation drop from over 30 percent in early 2026 to less than 10 percent by August. TPUs are custom-designed Google hardware chips built specifically to accelerate artificial intelligence calculations.
As reported by Reuters on August 13, senior cloud division leaders expected recent leadership changes to resolve ongoing internal battles over Google’s constrained supply of these chips. The hardware capacity previously reserved for fundamental science was transferred directly to the Gemini training clusters. This shift reflects a sustained internal policy change rather than a temporary server disruption.
Which DeepMind science projects are delayed for Gemini?
Internal project roadmaps show Google delayed planned updates to the GNoME materials discovery database and next-generation structural biology interaction models until at least 2027. GNoME (Graph Networks for Materials Exploration) is a DeepMind AI model that previously identified over 380,000 stable materials for potential use in batteries and superconductors.
Without the required batch sizes on advanced TPU clusters, these non-commercial science teams lack the computing infrastructure to process datasets and validate new models. Quotes from affected DeepMind research scientists working in structural biology confirm their project batch sizes were forcibly reduced to accommodate Gemini.
How are DeepMind scientists reacting to the hardware cuts?
Internal communications reveal that senior science team leads have formally protested the hardware cuts to incoming DeepMind chief Koray Kavukcuoglu. DeepMind is Google’s premier artificial intelligence research laboratory, historically operating with independence in London before an August 2026 restructuring moved operational control to California.
Scientists who joined the lab specifically for its historical independence have contested the new quotas in internal Moma and Slack channels.
Leadership has made the mandate clear. Koray Kavukcuoglu stated internally that commercializing Gemini is an existential necessity to generate the revenue required to fund fundamental science over the next decade.
Why can’t Google Cloud supply enough compute for both?
Alphabet’s data center infrastructure cannot scale rapidly enough to simultaneously train a massive generative language model and maintain DeepMind’s historical scientific research pace. Alphabet’s recent CAPEX filings and standard industry timelines for deploying new data centers demonstrate a strict physical limit on available computing hardware.
This physical infrastructure constraint forces the zero-sum reallocation of resources from science to commercial product development.
How do DeepMind’s delays affect external scientific research?
External biotech and clean-energy academic partners who rely on DeepMind’s open-source database releases have been notified that anticipated Q3 2026 model updates are delayed indefinitely.
The global materials science community heavily utilizes the GNoME database to synthesize new crystals. University labs and startups relying on these foundational weights and datasets must now wait or find alternative models to advance their own work.
Frequently Asked Questions
Why did Google reduce DeepMind’s computing access?
Google reduced DeepMind’s computing access because the company faces a physical shortage of AI infrastructure and requires maximum capacity to train its flagship commercial product, Gemini. Internal Google Cloud allocation logs indicate DeepMind’s science quota dropped from 30 percent to under 10 percent to accommodate this shift.
Which DeepMind science projects are paused?
Google has paused scheduled updates to the GNoME materials discovery model and next-generation structural biology interaction models. Internal roadmaps show these projects will not see major updates until at least 2027 due to the computing reallocation.
Is DeepMind still an independent research lab?
DeepMind no longer operates with full research independence. Following an August 2026 restructuring that placed Koray Kavukcuoglu in operational control of Gemini, decision-making authority shifted from London to Mountain View, and computing resources were tied directly to Google’s corporate commercial interests.
