Sundar Pichai on Gemini 4, scarce compute, and Google's agent plans

In the first 72 hours after Gemini 4 Argon's wider release, it processed nearly 25 trillion tokens, about 3 times the volume recorded during Gemini 3.0's equivalent launch period. After sharing those internal figures, Google and Alphabet CEO Sundar Pichai acknowledged that the company remained constrained by compute.

The numbers measure model usage. They cannot be converted directly into users, revenue, or a quality ranking. They do, however, illustrate Google's planning problem: model capabilities and demand can change quickly, while the data centers, power, and chips supporting them must be arranged years ahead.

In his conversation with Sources host Alex Heath, Pichai kept returning to that coordination problem. As AI takes on more work, compute, products, organizational responsibilities, and operating safeguards have to keep pace.

Why Gemini 3.5 Pro did not arrive as planned

Pichai described Argon as strong across knowledge work, agentic tasks, and coding, while acknowledging room for improvement in programming. Harder assignments demand better code, clearer review, and stronger safety. Hallucinations remain an ongoing problem.

That context helps explain what happened to Gemini 3.5 Pro.

According to Pichai, Google had been developing Flash and Pro together when it recognized gaps in agentic capabilities, particularly coding. The team concentrated its experiments on the smaller, more efficient Flash model. It moved through 3.5, 3.6, 3.7, and 3.8 Flash in roughly 3 months.

By the time those lessons could be applied to Pro, Argon's pretraining was complete. Releasing 3.5 Pro at that point would have given it a short useful life. The team put its findings into Argon's post-training instead, directing limited compute and time toward the next generation.

Pichai presented this as a resource and product decision. He did not describe 3.5 Pro as Argon under another name, or promise a release date for a more efficient model.

Argon itself is large. At the time of the interview, access included the API, Gemini's Pro and Ultra subscriptions, and an expanding rollout in Antigravity. Google was also working on distillation to preserve much of the capability at a lower serving cost and broaden access.

Google's official release account describes a phased rollout. A released model, availability to some paying users, and ample access for everyone remain different milestones.

Agents are taking on work with long chains of dependencies

Heath raised the Minecraft experiment to ask where sustained agent capability might lead. A system that can work for more than a day raises questions well beyond games.

The research paper provides important boundaries: 4 Argon agents worked together on Easy difficulty, completing the task in 27 hours and 14 minutes by operating the game through code. Researchers could pause or stop them. The paper describes a single run, which cannot establish a reliable success rate.

Heath mentioned spending below $2,000. The paper estimates spending using API list prices; introductory versus standard pricing, and counting only the winning run versus the entire session, produce different totals. Those figures do not establish the model's actual serving cost or the full cost of a real engineering project.

Pichai focused on Google's internal work. He described a Chromium feature that lets a tab move into another experience while remaining live, without reloading. The visible interaction sounds small, but implementing it touches multiple layers of a vast codebase. In his account, Argon can persist through those dependencies and advance the work.

The value of a long-running agent depends on maintaining the goal, managing dependencies, recovering from failure, and producing changes people can review. The game provides an observable example. Production systems still need their own validation.

These capabilities are also entering model research. Pichai said researchers work with models on training recipes and the evaluation of subsequent checkpoints. Some stages are more automated than others. Recursive self-improvement remains a shifting continuum with substantial human participation.

He expects that collaboration to accelerate progress over the next 12 months and sees 2027 as a potential step forward in Google's productivity. Those are forecasts, rather than results already delivered.

Frontier model development has become an all-consuming job

Building frontier models now requires coordinating data, compute, algorithms, people, and agents at considerable scale. Pichai described it as work that consumes nearly all of a leader's attention.

That was his central explanation for changes at Google DeepMind. Koray Kavukcuoglu took on its principal management responsibilities, while Demis Hassabis devoted more time to AGI safety, scientific applications, and drug discovery at Isomorphic Labs.

Google's formal announcement identifies Hassabis as Chair of Google DeepMind and Chief Scientist of Alphabet, while continuing to lead Isomorphic Labs. Kavukcuoglu is SVP of Google DeepMind and retains his position as Google's Chief AI Architect.

The discussion also covered Jeff Dean's departure to start Discovery Loop. Pichai recalled roughly 27 years of contributions and characterized the move as an opportunity Dean wanted to pursue. The interview supplied no evidence for speculation about internal conflict.

Pichai's account suggests a clearer division of attention between the daily operation of model development and longer-term scientific and safety work. He also stressed that DeepMind's integration with Google's products had already been developing for years.

Today's compute reflects decisions made years earlier

Observers assess demand with today's information. Much of the supply available to management reflects decisions made about 2 years earlier. Land and power planning can extend 3 years ahead; the interview was already discussing 2028 and 2029.

That lag makes current model rankings an inadequate guide to infrastructure. Agentic coding can create a surge in demand much faster than new capacity can be built.

Heath framed a question around free cash flow turning negative over the summer and asked why Google was willing to keep increasing investment. Pichai clarified that the measure under discussion was free cash flow, then explained his confidence.

He described several signals: sustained use by consumers and enterprises; practical benefits from AI inside Google; progress in internal evaluations and research; and long-term customer contracts with committed revenue.

These signals check one another. Better capabilities alone do not establish demand, while a brief burst of usage is a weak basis for years of construction. Pichai considers the technology trajectory, practical applications, and contracts together.

He also acknowledged that genuine demand can coexist with industry overinvestment or outside shocks. Google's roughly 6 to 8 years of aggressive investment in Cloud served as an analogy for his long-term approach. It offers no guarantee that AI investment will follow the same path.

Real demand and investment risk can exist together. His answer supported the former without excluding the latter.

Pichai's own agents help him understand what is happening

Pichai's personal examples were concrete.

Google has numerous discussions covering pretraining, mid-training, post-training, and data. Catching up on their latest developments could previously take about 1 hour a day. Agents now summarize that material, while he investigates the areas that need closer attention.

Other agents read compute telemetry and provide information about efficiency, utilization, and allocation across products. Important compute decisions remain with people, including Pichai himself.

The approach also helps explain Google's personal-agent ambitions. Pichai said enterprise and consumer agents can share the model, tool-execution framework, and runtime. Their different contexts account for much of the distinction.

The enterprise version emphasizes isolated execution, permissions, governance, and auditability, with tasks able to run in individual small virtual machines. Google wants to extend those capabilities to consumers, bringing conversation, persistent work, scheduled tasks, and notifications into a coherent entry point.

This remains a product direction unfolding in stages. The interview does not establish that every account already has all of those capabilities.

A shared agent could connect Search, Gmail, and Gemini

Heath asked whether agents clicking through browsers were an awkward transitional arrangement. Today's web was built for people, and imitating their actions can be cumbersome.

Pichai expects both browser interaction and protocol-based connections to develop. Browsers preserve an interface people recognize and can supervise. Protocols such as MCP and UCP offer a way for services to exchange information and cooperate directly, with greater potential for scale.

Adoption requires agreement across the ecosystem. Merchants, platforms, and users must find the arrangements worthwhile. Technical capability may advance faster than that commercial coordination.

The same reasoning applies inside Google. Asked whether Search and Gemini would eventually merge, Pichai described an agent layer operating across products.

An agent can retain a task's context while running in the cloud. Work started in Search could continue in Gmail or Gemini. His answer did not settle whether the products would formally combine; the intended continuity concerns the task experience.

An implication of that discussion is that personal-agent competition includes the ability to carry information, permissions, and progress across entry points.

Space compute, Waymo, and quantum run on different timelines

Project Suncatcher takes the compute question much further out. Heath brought up an early pilot involving Planet and SpaceX. Pichai said it was going well and discussed the project on a 10 to 20 year horizon.

His reasoning included the energy available in sun-synchronous orbit and the pressure that long-term computing demand could place on land. Google's project account identifies unresolved challenges including radiation, vibration, and cooling. The present work tests hardware and technical feasibility; it does not establish a commercially operational orbital data center.

Waymo faces the more immediate work of expanding an existing service. Pichai said paid rides had exceeded 500,000 per week, with 10 new cities that year. Over the following 12 months, priorities included vehicles, customer support, operations, international markets, weather, and highways.

He cited a 95% crash reduction. Under Waymo's safety metric, that figure concerns serious injury or worse crashes, compared with a human-driving benchmark over comparable distance in its operating areas. It is a specific company-reported safety comparison, not a 95% reduction in every category of collision.

Partnerships with other transport platforms remain part of Pichai's outlook. Selling autonomous cars directly to consumers is not what the Waymo team is currently working on, he said, although he did not rule it out for the future.

Wing is pursuing delivery growth through partnerships and lightweight drones that are easier to manufacture and deploy. Heath referred to potential future coverage of forty million US households. Pichai expects delivery volume and categories to expand in 2027. The household figure is a prospective coverage goal, not a count of existing users.

Quantum computing has a less certain commercial timetable. Pichai said the team had met or exceeded internal milestones over roughly 10 years. He sees the possibility of useful commercial applications within 5 years, including Google Cloud negotiating quantum deployments at some scale. That remains an expectation, without a confirmed product delivery date.

More autonomous work requires stronger operating controls

Early in the interview, Pichai discussed external security research, the Fairwind program, US government testing, and extensive internal use before Argon's release. Later, he connected safety and governance to competitive strength.

Once models execute tasks, they also need sandboxes, real-time monitoring, access controls, and activity records. Systems must detect unexpected behavior early and allow people to intervene. Those safeguards consume compute themselves and belong in deployment cost calculations.

On policy, Pichai favors more consistent evaluation standards and independent assessment resembling financial audits. He prefers uniform rules to a state-by-state patchwork. Heath also raised regulatory capture and room for startups to compete. These are policy positions, rather than a description of a settled regulatory regime.

Taken together, the interview exposes Google's coordination challenge. Model releases can accelerate month by month, infrastructure must be planned over years, and cooperation across products depends on permissions, standards, and commercial relationships. Pichai wants stronger models to take on more work while people remain responsible for resource choices and consequential decisions. Google still has to show that these activities, moving at different speeds, can advance together.

Original source: Sundar Pichai on Gemini 4 and Google’s biggest bets

Platform and channel: YouTube, Sources Podcast. Host: Alex Heath. Guest: Sundar Pichai, CEO of Google and Alphabet. Published 2026-10-21. Duration: 59:09.

Based on English auto-generated captions, with spoken repetition and sponsorship removed. The linked primary sources were used to check terminology and statistical scope. Operating figures, internal examples, and forecasts remain attributed to their speakers. This article reorganizes and interprets the discussion rather than reproducing it verbatim.