Google DeepMind researchers are lining up for TPU access inside a company that built some of the world’s most coveted AI infrastructure. At DeepMind, access to TPU compute is now rationed against profit potential, internal priorities, and, in some cases, paying customers – a reminder that the AI boom has turned silicon into a political asset, not just a technical one.
That tension is becoming a quiet force inside the company. Google still has advantages most rivals would kill for: its own chips, a major cloud business, and deals to share infrastructure with companies such as Anthropic and Meta. But when demand for training and inference explodes, the spoils do not automatically go to the people inventing the models.
Google DeepMind TPU access is a routing problem
The internal bottleneck is blunt. Oren Etzioni said there can be at least three claimants for every TPU inside Google, which means researchers with ambitious but uncertain work can lose out to projects with a clearer path to revenue. That is a very Silicon Valley way to discover scarcity: build the factory, then start arguing over the output.
Google says it uses a strict ongoing process to allocate compute to the most important priorities, balancing current customer demand with long-term research. Sundar Pichai has also said DeepMind gets the resources it needs to build frontier models because that underpins everything else the company does. The problem is that ”everything else” increasingly includes products that can pay for the GPUs and TPUs first.
ChatGPT changed the culture at Google
Before ChatGPT, Google’s AI labs had a reputation for giving researchers something close to academic freedom, just with better pay and far more hardware. After ChatGPT’s arrival in 2022, the company leaned harder into large language models that could ship as products and generate real revenue, while more experimental work lost some oxygen. That shift was not subtle: the dream job started looking more like a scheduling conflict.
Google tried to simplify the internal structure in 2023 by combining DeepMind with Google Brain, which had given researchers more room for personal projects and even an internal chip-credit system. But the merger did not abolish competition for compute; it just moved it into a more centralized machine. Google Cloud’s backlog nearly doubled from the previous quarter and passed $460 billion, which tells you plenty about where the pressure is coming from.
Why researchers keep leaving anyway
The most revealing part of the story is not that Google can’t satisfy everyone. It’s that some of its own researchers now see startup life as the easier path to getting things done. Anna Goldie said she was offered more compute to stay, then left anyway to found Ricursive Intelligence, which has already raised $335 million. Outside the bureaucracy, access to chips can still be hard, but at least the answer is not filtered through 10 layers of management.
That trade-off is becoming familiar across the AI industry. OpenAI and Anthropic have both built their reputations on the same basic formula: get enough compute, turn it into models quickly, then use the models to justify even more compute. Google should understand that better than anyone, but owning the infrastructure does not stop internal scarcity from reshaping behavior.
The race for TPU compute is also a talent race
Former DeepMind researcher Ioannis Antonoglou framed the whole thing neatly: the first question is who has the most compute, and the second is who uses it best. Google still wins the first round more often than not. The harder part is keeping enough of its own talent convinced that the second round is worth playing inside the company rather than outside it.
If compute stays tight, expect more internal prioritization, more frustrated researchers, and more spinouts that claim they can move faster. Google can outspend a lot of the market, but it cannot completely firewall ambition from bureaucracy. That is the part of the AI boom no one likes to put on a keynote slide.