Full episode · 59 min
Sundar Pichai on Gemini 4 and the AI compute crunch
Sundar Pichai says Google is committing to AI compute because demand, model progress and long-term contracts support continued investment.
Sources · Alex Heath with Sundar Pichai
The best minutes
- 22:47Google justifies AI spending with demand and capability signals7 min
- 8:30Pichai expects a productivity leap from AI agents in 20274 min
- 47:26Waymo is scaling through partnerships and broader deployments3 min
Also worth watching
- 1:44Google pairs model access limits with runtime safeguards4 min
- 15:32DeepMind's frontier work has become an all-consuming operation4 min
- 35:30Google wants Gemini to combine chat, tasks and scheduled work3 min
- 38:31Pichai sees protocols as a more scalable path for agents5 min
- 43:47Orbital data centers could address future energy and space constraints3 min
Skip the sponsor reads, the model-race discussion, and the closing regulation and quantum discussion.
Summary · 4 min read
Sundar Pichai is CEO of Google and Alphabet. He discusses Google's Gemini models and its investments in compute, Waymo, Wing, Project Suncatcher and quantum computing.
Pichai says Google is investing in AI compute based on demand across consumer and enterprise, internal capability gains and durable customer contracts, while acknowledging industry volatility and possible economic shocks. He describes Gemini 4 Argon as a strong frontier model with rapid early usage, but says capacity, coding quality, safety monitoring and broader access remain active work. He expects agents to handle more persistent tasks across Gemini, Search and other services, with protocols and user-focused design shaping how they interact with the web. He also discusses DeepMind leadership, Project Suncatcher, Waymo, Wing, AI safeguards and possible commercial applications for quantum computing within five years.
Minute by minute
- 0:00
Argon faces capacity limits. Sundar Pichai said Gemini 4 Argon drew almost 25 trillion tokens in its first 72 hours, about 3x what 3.0 had in the same period. He said Google is capacity-constrained and plans to widen access with more efficient distilled models. He described external and internal safety testing, sandboxing and real-time monitoring for deployments.
- 6:13
Google is speeding up releases. Pichai said he follows external reactions but relies on internal evaluations to track model capabilities. He said Google is committed to the AI frontier over the long term and needed to increase its release cadence, pointing to 3.5, 3.6, 3.7 and 3.8 Flash shipping over about three months.
- 11:24
Models are joining research work. Pichai said Google researchers work closely with models, which can suggest training recipes and evaluate checkpoints, though people remain in the loop. He expects AI-assisted recursive self-improvement to accelerate progress over the next 12 months. He said Google carried coding gains from the 3.5 Flash iterations into Argon post-training after 3.5 Pro pretraining was already complete.
- 15:32
Frontier development reshapes leadership. Pichai described frontier-model development as an all-consuming, scaled operation. He said Demis remains involved as DeepMind chair while focusing more on AI safety and applying AI to science and healthcare. Jeff Dean left after 27 years to pursue Discovery Loop, and a longtime DeepMind colleague stepped up to lead model development.
- 19:47
Compute planning starts years ahead. Pichai said Google plans compute for frontier research and model development, then uses business forecasts, including Cloud demand, to plan capacity. He noted that long lead times mean today's compute decisions reflect plans made about two years earlier, while land and power planning extends further out.
- 25:12
Demand supports disciplined spending. Asked about negative free cash flow and concerns around AI infrastructure financing, Pichai pointed to demand from consumer and enterprise users, model capability gains and long-term contracts with committed revenue. He said the wider market could be volatile or overshoot, but Google should plan for durable demand and invest with discipline, as it did when expanding Cloud over six to eight years.
- 33:13
Agents help monitor Google. Pichai said humans, including him, retain control over compute allocation decisions. He uses internal agents to summarize model-development discussions and provide insights into compute usage, efficiency and utilization. He said AI already informs allocation decisions, as it had helped optimize data-center power use years earlier.
- 35:30
Gemini is becoming a personal agent. Pichai said Gemini's goal is to be a personal, proactive and powerful assistant. Google plans to bring enterprise agent capabilities, including secure execution, monitoring and controls, to consumers, and wants chat and persistent or scheduled tasks to work through one interface. He said Google will expand access beyond Pro and Ultra users.
- 38:31
Protocols could make agents interoperable. Pichai said browser and computer use lets agents work through familiar interfaces with users involved, while protocols such as MCP and Google's UCP could provide a more scalable, interoperable path. He said the agent layer could connect services such as Gemini and Search through shared user context, rather than requiring the products themselves to fully merge.
- 43:47
Suncatcher looks beyond Earth. Pichai said Project Suncatcher's pilot with Planet Labs and SpaceX remains in orbit and is doing well. He sees space data centers as a long-term compute option, citing access to 24/7 energy in sun-synchronous orbits and constraints on Earth's physical space. He compared the effort to early Waymo as a moonshot that can push technology forward.
- 47:26
Waymo is focused on scaling. Pichai said Waymo's opportunity is not zero-sum and expects it to work with partners while the team focuses on scaling its service and operations. He said Waymo had crossed half a million paid rides per week and that a Toronto launch was coming. He said selling Waymo vehicles directly to consumers is not a current focus.
- 50:23
Wing plans broader delivery growth. Pichai praised Wing's lightweight product design as easy to make and deploy. He said the company plans to meaningfully increase deliveries in 2027 and expand into new categories, using partnerships as its approach.
- 51:33
Pichai favors uniform AI rules. Pichai backed more standardized testing and evaluations, and said uniform regulation is preferable to a state-by-state patchwork. He said frontier labs are already submitting models to the U.S. government for voluntary testing and argued that companies investing in safe development can gain an advantage, while rules must keep pace with changing technology.
- 55:49
Quantum could become a business. Pichai said Google's quantum team has tracked or exceeded internal milestones across nearly 10 years. He expects practical commercial applications within five years and imagines Google Cloud negotiating quantum deployment deals at some scale by then.
Published Oct 21, 2026.


