In the race to build the infrastructure that powers artificial intelligence, Alphabet Inc.’s Google has an enviable position: The company has a healthy cloud computing business, makes its own chips, and has struck deals to share them with companies like Anthropic PBC and Meta Platforms Inc.
Google’s success has made its computing resources so valuable, though, that its own AI researchers have to get in line.
Last summer, Andrew Dai, then a researcher in Google’s AI lab, discovered a blind spot in Gemini, the company’s flagship AI model. While playing a board game, Dai took pictures of the board and asked Gemini a simple question: who’s winning? To his surprise, Gemini was stumped, as were models from rivals. He became convinced of the need to build AI that could better understand what was happening in images.
Dai discussed his idea with some of his colleagues, but he quickly concluded that he wouldn’t be able to secure enough computing power to tackle the problem within Google, he said in an interview. He had to leave the company if he wanted to do it.
Dai is among current and former employees who say Google’s leadership in AI development has turned computing power into a precious resource, accessible mostly to people with high-priority projects, like improving Gemini.
AI researchers sometimes feel like they are losing out on computing power to paying customers, the people said. Google’s search and cloud computing units are also jockeying to use the company’s chips, known as tensor-processing units, or TPUs. Within the AI lab Google DeepMind, access to computing power influences the projects that researchers pursue, the leaders they align themselves with and the pace at which they work.
“Inside Google, every TPU has three suitors,” said Oren Etzioni, a veteran AI researcher who is a professor emeritus at the University of Washington. “If you find yourself in the uncomfortable position where you have a pie-in-the-sky project and you are competing with a revenue-yielding customer, that’s a tough position to be in.”
Google said in a statement that the company has a “rigorous, ongoing process that ensures our compute resources are allocated to the most important priorities, balancing today’s customer and user needs along with our long-term investments to advance research and innovation.” Alphabet Chief Executive Sundar Pichai has said that when deciding where to devote computing power, company leaders are focused on making sure that Google DeepMind has the resources that it needs to build cutting-edge AI models, “because it’s a foundation for everything we do.”
Alphabet said Google Cloud’s backlog — the measure of contracted work that hasn’t been recorded as revenue yet — nearly doubled from the prior quarter to over $460 billion. “We are compute constrained in the near term,” Pichai said. “We are working through that moment and investing.” Google will unveil its latest suite of product advancements at its annual developer conference in Mountain View on Tuesday.
AI researchers once regarded Google as a place where they could have the freedom to pursue intellectual passions, almost like in academia, but with better pay and more resources. Researchers at the company have long angled for more computing power, but until relatively recently, the models were small enough that they didn’t need as much to run a meaningful project, former employees said. But in 2022 the launch of OpenAI’s popular chatbot ChatGPT prompted Google to invest in large language models, AI programs that can spin up a professional-sounding cover letter or term paper in seconds. Now Google is focusing on models that write computer code, which competitors have shown can be a hit product and generate revenue.
Under the strategy followed by top AI labs, “you have to build the world’s best coding model, because ultimately no one wants to be second to AGI,” Dai said, referring to the widely held Silicon Valley ambition of building AI that can perform on par with humans. That makes the idea of pouring resources into other projects, especially experimental ones that may not generate revenue, harder for Google to justify.
Dai left Google to found Elorian, an AI startup that recently exited stealth mode and specializes in visual reasoning, which Dai says is key to bringing AI to industries such as architecture, automotives and robotics. He is one of several former Google AI researchers who say they have had better access to computing power as startup founders. The researchers said that founding companies gives them the freedom to seek computing power from multiple sources — and they can use the chips they secure as they wish, without navigating Google’s bureaucracy, or worrying access could disappear if company priorities shift.
Former Google DeepMind researcher Ioannis Antonoglou said he had access to ample computing power while working on AlphaGo, an AI model designed to play the strategy game Go, which made waves by beating one of the world’s best players. Later, he was part of the push to build Gemini, one of Google’s most important strategic initiatives. But he felt that the company wasn’t devoting enough computing power to post-training, a stage in which models are fine-tuned with data related to specific fields, such as legal documents or computer code.
“Both myself and my cofounder, we believed in reinforcement learning as being the next frontier,” said Antonoglou, who left with fellow DeepMind researcher Misha Laskin in 2024 to found ReflectionAI, a startup dedicated to building AI models in the open. “It wasn’t clear that Google or DeepMind would take this path back then.”
When AI researchers are poised to defect, access to computing power is a lever that companies can pull. Former DeepMind researcher Anna Goldie said the company offered her more computing power to try to dissuade her from leaving to launch a startup. She ultimately departed anyway, founding a company called Ricursive Intelligence with fellow DeepMind researcher Azalia Mirhoseini that launched in late 2025.
Goldie said she has been pleasantly surprised by how much computing power she has been able to find on the outside, from a range of sources. She declined to say how much computing power the company has obtained after raising $335 million, but she said it is on par with what she had been offered to stay at Google.
“I don’t need to ask like 10 layers above me for permission,” she said. “I can just make a decision with my cofounder to do what’s best for the company. I can listen to my employees and hear their ideas.”
At top AI labs, some researchers work on language models because it’s the priority, even if their true interests lie elsewhere, said Tom McGrath, a researcher who left Google in 2023.
“There’s the carrot of compute and promo and generally being part of the glory of the big training run,” said McGrath, who is chief scientist at Goodfire, a startup that aims to better understand the inner workings of AI models. “There’s also the stick that you won’t have any accelerators if you don’t.”
It’s a new way of life for some researchers at Google. To catch up in the AI race, Google in 2023 merged two AI labs: London-based DeepMind, which had a more top-down structure, and Google Brain, where researchers pursued passion projects with minimal supervision.
Researchers at Brain each received credits to buy chips in an internal system where price fluctuated based on demand, similar to the stock market, Dai and Goldie said. Some researchers made the most of what they had by pooling resources and then using the credits of their teammates while they were on vacation or sleeping, Goldie added. “That was a powerful way that you could bond together and make something happen,” Goldie said.
Google still has a pool of computing power for individual researchers, but supply is constrained when the company is training large AI models, Dai said. This means researchers are effectively competing for slices of a smaller pie.
Now, researchers who want more computing power often focus on short-term research questions that might yield something that could be incorporated into the next version of Gemini, Dai said. “Then it makes leadership believe it makes more sense.”
Researchers can’t always bank on receiving the computing power they are promised. In 2024, a large training run prompted Google to pause some research projects for about a quarter, Dai said. Some people abandoned their work as a result.
Startups offer an “element of control over your own destiny — being much clearer that if you pay for this much compute over the next year, you’re going to get it,” Dai said. “No one’s going to take it away from you.”
To make the most of the computing power he has as Elorian ramps up, Dai said he has focused on hiring researchers who have experience with limited resources.
“The game of AI has always been twofold,” Antonoglou said. “One is, who has the most compute. And the second is, who can actually use it better.”
Love writes for Bloomberg.
The post Inside the AI compute crunch driving Google researchers to quit appeared first on Los Angeles Times.
