Speaking from the sidelines of the NeurIPS 2025 conference, Google Chief Scientist Jeff Dean offered a detailed look into the company’s hardware trajectory while issuing a broad call for new “moonshot” funding models to sustain academic innovation.
In an exclusive interview with Andy Konwinski for the Laude Institute, Dean discussed the capabilities of Google’s seventh-generation Tensor Processing Unit (TPU), the company’s evolving strategy on open research versus proprietary secrets, and the critical need to apply AI to societal challenges like healthcare.
The Evolution of the TPU
Dean, who serves as the co-technical lead of Google’s Gemini project, confirmed details regarding the company’s latest silicon, the TPU v7. While noting that each generation offers improvements over the last, Dean highlighted specific architectural shifts designed to accommodate the massive training requirements of modern large language models.
“It’s connected together into these very large configurations that we call pods,” Dean said, estimating the scale at “9,216 chips… per pod.”
He emphasized that the new hardware is optimized for efficiency, specifically targeting “much higher performance, especially for lower precision floating point formats, like FP4,” which he noted will be “really useful for training large models.”
Dean recounted the origins of the TPU program, which began as a “thought experiment” in 2013 and resulted in deployed hardware by 2015. The impetus was Google’s realization that running advanced speech models on standard CPUs at scale would be financially and logistically impossible.
“We would actually need to double the number of computers Google had overall in order to just roll out this improved speech model,” Dean told Konwinski. This necessity drove Google to design specialized hardware that was “30 to 70 times more energy efficient than contemporary CPUs or GPUs.”
Advocating for “Moonshot” Research
Beyond hardware, a significant portion of the conversation focused on the state of scientific research. Dean, along with Konwinski and other prominent figures in the AI field, serves on the board of the Laude Institute, a non-profit dedicated to funding ambitious research projects.
Dean argued that while industry labs drive significant progress, a healthy academic sector is vital for long-term innovation. “I advocate that we should have a vibrant academic funding model for academic research because the returns are quite large to society,” Dean stated.
He specifically championed a “moonshot” approach—funding three-to-five-year research horizons that sit between immediate product development and decades-long theoretical inquiry. Dean suggested this timeframe is the “delightful time range” because it is “not so distant that it won’t have impact, but it’s not so short a time period that you can’t conceive of doing something ambitious.”
Healthcare as the Next Frontier
When asked about specific areas ripe for this type of research, Dean pointed to healthcare as a primary candidate for AI integration, despite the regulatory and privacy hurdles.
“How can we as society use every past decision that’s been made in health to inform every future decision?” Dean asked. He acknowledged that achieving this is a “super hard goal” due to privacy concerns and regulatory fragmentation, but suggested that technical solutions like “privacy preserving machine learning or federated learning” could bridge the gap.
Balancing Openness and Competition
The interview also touched on the shifting culture of openness in AI research. With Google facing intense competition in the generative AI space, Dean acknowledged a strategic pivot regarding what the company chooses to publish.
“In this current competitive dynamic we tend to not publish the secret sauce inside our architecture of our Gemini model,” Dean explained. However, he insisted that Google remains committed to the scientific community. “We do publish a lot of stuff in the sort of earlier stage research aspects… of, ‘here are interesting new kinds of model architectures that we haven’t proven out but we’ve experimented with at small scale.’”
Dean cited Google’s computational photography work on the Pixel phone as an example of their current publishing cadence: releasing the product feature first, then publishing the research “a little bit of a delay” later.
The conversation underscored Google’s continued reliance on a vertically integrated stack—from custom silicon to high-level applications—while highlighting Dean’s personal focus on ensuring that the AI revolution extends beyond chatbots and into fundamental scientific and societal progress.