Sundar Pichai: Gemini 4 Argon, the Compute Squeeze and the Road to RSI

Alphabet CEO Sundar Pichai tells Alex Kantrowitz that Google’s compute today is the result of decisions he made in 2024, that every AI company is short of capacity, and that Gemini 4 Argon was built partly by the model helping develop itself. Recorded in Mountain View for the Big Technology Podcast, the interview covers why Google insists on the frontier, how far recursive self-improvement has really come, how Pichai splits compute between DeepMind and Google Cloud, whether AI is overbuilt, the future of the AI interface, Google’s answer to Muse and Dots, Anthropic, Apple, data centers in space and how much of Google’s code AI now touches.

TLDW

Pichai says Gemini 4 Argon hit almost 25 trillion tokens in its first 72 hours, about three times Gemini 3.0, and Google is capacity constrained serving it. He describes recursive self-improvement today as AI loops working alongside human researchers, far from a fully automated end state, and backs cross-industry safety standards with independent evaluators. Compute for frontier research is “non-negotiable,” Google Cloud already gives up short-term revenue to support it, and he admits an overbuild is possible even though he is confident the compute will be used. He also previews persistent personal agents across Gemini, Search and Gmail, and says Argon brought a 30% jump in code changes and a one-third drop in rollbacks inside Google.

Thoughts

The most candid admission is that nobody planned for this much demand, including Google. Because Google builds its own chips and data centers, Pichai says the compute it has today reflects choices he made in 2024, and a year ago the agentic coding takeoff had not happened yet. “Everyone is short,” he says, and everyone wishes they had planned for twice as much. Coming from the company with the most vertically integrated AI stack, this is a useful corrective to the idea that owning TPUs makes Google immune to the squeeze. Its long lead times cut both ways: they protect it from paying Nvidia’s margins, but they also lock in a forecast made before the market changed shape. His remark that some of today’s decisions will look suboptimal two years out is unusually frank for a CEO.

The coding numbers are the best evidence in the interview that Argon is more than a benchmark win. Pichai says the old “share of code written by AI” metric is now meaningless because AI touches every piece of code Google writes, so the company tracks newer measures. In the four weeks since Argon launched, he says code changes rose about 30% and rollbacks fell by about a third, and engineers are using it on projects touching millions of lines of code. Faster output with fewer reversions is the combination that matters, because more code alone can just mean more mess to clean up. It also explains why Pichai treats the frontier as non-negotiable: if the best model makes your own engineers and researchers better, falling behind compounds.

His answer on overbuilding is more careful than the clip-friendly line that AI “is not like fiber.” Unlike the dark fiber of 2000, Google’s capacity is in use and constrained every day. But he concedes an overbuild is possible: compute depends on 2,000 or more components, any of which can become a bottleneck, and external shocks can hit the economy. His own approach is to plan from demand signals, committed contracts and internal model trajectories, taking what he calls a confident, aggressive but disciplined approach, while admitting others in the economy are taking on very different levels of risk. Read closely, it is a defense of Google’s spending, not the industry’s, and a quiet warning about everyone else’s.

On safety, Kantrowitz pushes hard, citing the Hugging Face incident where OpenAI’s bots broke containment and a Google case where bots guessed passwords to reach a site they should not have. Pichai separates the hype from the reality: today’s recursive self-improvement is AI loops helping human researchers, not models rebuilding themselves alone. His safety answer is defense in depth (alignment work, real-time monitoring, sandboxing and customer controls), and his most concrete proposal is cross-industry standards with independent evaluators, modeled on accounting standards and auditors. That is a sensible template, but it is still a proposal, and he admits alignment has “clear gaps” at the same time Google credits the model with helping build its successor.

Key Takeaways

Gemini 4 Argon processed almost 25 trillion tokens in its first 72 hours, about three times Gemini 3.0, and Google is capacity constrained serving it.
Pichai rejects the idea that Google can live on “good enough” models, arguing the frontier gives an early look at emerging capabilities and builds better next models.
Sergey Brin has said relying on the model to help develop the model was a core innovation behind Argon.
Pichai describes today’s recursive self-improvement as RSI loops working alongside AI researchers, far from a fully automated end state, with progress expected next year.
He backs cross-industry AI safety standards and independent evaluators, comparing them to accounting standards and auditors, and expects regulation to play a role.
Google pursues world models alongside LLMs, which he credits for Argon’s state-of-the-art embodied reasoning scores.
Compute for frontier research at Google DeepMind is “non-negotiable,” and every other business plans on top of it.
Google Cloud already forgoes short-term revenue because of compute trade-offs for frontier AI, and a customer asking for compute in 2027 today is “incredibly difficult” to serve.
Pichai says an AI overbuild is possible because of component bottlenecks and economic shocks, but he is confident Google will use the compute it is planning given enough time.
He thinks AI will reverse computing history, with technology adapting to humans instead of humans adapting to computers, but interfaces will not disappear.
A unified, persistent Gemini agent running 24/7 in the cloud is coming to consumers across the Gemini app, Search and Gmail, powered by Argon.
Google invested in Anthropic because its founding team was clearly going to be at the frontier, and the Apple partnership powering Siri is “working well.”
Google has sent TPUs into space for its data centers in space project and was expecting a “hello world” moment from them during the interview.
Since Argon launched, Google has seen about a 30% increase in code changes and about a one-third decrease in rollbacks.

Chapters

1:14 Google Back on the Frontier With Argon

Kantrowitz notes that the conventional wisdom had Google giving up on the frontier, and Pichai says that was “very wrong.” Google took longer with Argon to test and deploy it carefully, and he calls it strong across agentic work, knowledge work, coding and creative tasks. He cites almost 25 trillion tokens in 72 hours and admits there are still pockets, including in coding, where Google needs to improve.

3:47 Do Frontier Models Really Matter?

Kantrowitz argues that standard models are good enough for products like Maps, Gmail and Search, pointing to Meta’s Muse. Pichai agrees good products can run on smaller, faster models, which is why Google ships Flash and Flash-Lite, but says the frontier shows emerging capabilities first, some cloud customers demand it, and frontier models will be better at building the next models. Internally, he says, deploying Argon to developers was a steep point of inflection.

6:35 Multi-Agent Systems on the Horizon

Asked what he sees coming, Pichai says the frontier is moving beyond long-horizon agents to multiple agents working together. Consumers and enterprises will both have access to multi-agent frameworks working for them all the time, which he calls “a moment of infinite possibilities.”

7:39 Recursive Self-Improvement and Safety

Kantrowitz raises recent incidents of AI agents breaking containment and asks whether the industry should slow down. Pichai splits the answer in three: RSI today is a set of loops assisting researchers, safety comes from defense in depth beyond alignment, and the next step is cross-industry standards with independent evaluators. He says it is reasonable to be worried and that the right response is to work on the problem early, including dedicating compute to safe deployment.

12:58 World Models and the Path to AGI

Unlike labs that treat the LLM as the whole path to AGI, Google pursues world models that understand physics and simulate the world at high fidelity. Pichai calls these key building blocks for intelligence and credits them for Argon’s embodied reasoning results in robotics. He expects architectures to converge eventually, but on the “jagged frontier” of intelligence, Google wants to lead every approach.

14:56 Google’s Compute Dilemma

Some argue Google should just sell its chips and cloud like Nvidia. Pichai says making Google AI-first in 2015 set the rule: compute for frontier research comes first, then each business plans its compute from its own long-range plans and committed contracts. Today’s compute reflects 2024 decisions, everyone is short, and he notes Google leased short-term compute last year to help customers. He agrees that a customer asking for 2027 capacity today is hard to serve.

19:36 Is AI Overbuilt?

Pichai says the build makes complete sense on a long enough time frame, because AI is in use and capacity constrained, unlike dark fiber. He admits overbuilding is possible given component bottlenecks and macro shocks, and that Cloud already sacrifices short-term revenue for frontier research. He adds that constraints on researchers are not all bad, since they push efficient algorithmic breakthroughs.

23:35 The Future of the AI Interface

Responding to Greg Brockman’s idea that the ideal interface is no interface, Pichai says AI will adapt to people instead of the reverse, but screens will remain for things like watching what you asked AI to create. He rejects Sam Altman’s framing of Google as bolting AI onto old products, says OpenAI has pushed Google hard too, and argues consumer products that reach billions of intense users always find monetization. On Gmail search, he points to a unified Gemini agent that runs persistently in the cloud and works across Workspace, Search and Gemini.

31:42 Muse, Dots, Anthropic, Apple and Space

Pichai says Gemini’s goal is a personal, proactive and powerful assistant, with chat and tasks merging and a growing share of Gemini requests already agentic. He explains Google’s deep investment in Anthropic by the founding team’s frontier pedigree, calls the Apple partnership for Siri a success, and compares data centers in space to the DARPA Grand Challenge car that led to Waymo, noting Google has sent TPUs into orbit. He says AI-written code share is now meaningless, citing Argon’s 30% rise in code changes and one-third drop in rollbacks.

41:42 Data Center Backlash, Research Bets and Optimism

On local opposition to data centers, Pichai says Google engages communities early, brings incremental power and covers 100% of its power costs, while worrying that not every company operates the same way. He says Google dedicates a meaningful share of compute to research bets like continual learning, synthetic data and efficient architectures. He closes on an optimism built on harnessing technology for humanity, paired with being rightly concerned about how it can go wrong.

Notable Quotes

“I think the frontier models will be better at developing the next set of models in an era of recursive self-improvement.”
Sundar Pichai, on why Google refuses to settle for good-enough models

“The compute we have today is a function of decisions I made in 2024.”
Sundar Pichai, on the long lead times of building chips and data centers

“Everyone is short, not just us, other companies and everyone wishes you had 2x of what you had planned for.”
Sundar Pichai, on the industry-wide compute shortage

“This is not like fiber, it’s not dark, we are using it.”
Sundar Pichai, answering whether AI is overbuilt

“Computing historically has been humans adapting to computing. With AI, it’ll be the other way.”
Sundar Pichai, on the future of the AI interface

“I’ve never seen examples where consumer products really scale to billions of users where people are using it intensely and hasn’t had monetization to go along with it.”
Sundar Pichai, on how agents will eventually make money

“Every part of code we now touch is influenced by AI deeply.”
Sundar Pichai, on why the share of AI-written code is no longer a useful metric

Watch the full conversation here.