Gemini 4 Argon processed almost 25 trillion tokens in its first 72 hours, about three times what Gemini 3.0 handled over the same period, according to Google and Alphabet chief executive Sundar Pichai, who said Google remains capacity constrained as it works through the launch. Pichai made the remarks in a 58-minute interview with Alex Heath on the Sources podcast, published on YouTube on October 21, 2026, roughly a week after Argon's wider release. The conversation covered model access, the DeepMind leadership changes, capital spending, browser agents and the web, Waymo, Wing, space-based data centers and quantum computing.

In Short

Google's boss said its newest AI model was used three times as much in its first three days as the previous big model, and that Google still does not have enough computing power to give it to everyone. That matters to marketers because the same model and its agents are being wired into Search, Gmail and Gemini, which changes how people reach websites and ads. Pichai offered no numbers on what agents will do to Google's advertising business, saying only that the answer has to put users first and still work for the businesses involved.

The launch numbers

The headline figure arrived in the first two minutes. Asked about early reception, Pichai said: "in the first 72 hours since Argon launched you know we've had about almost 25 trillion tokens which is about 3x what 3.0 had in the first 72 hours."

On that ratio, Gemini 3.0's first three days would have come to roughly 8 trillion tokens. Pichai did not say whether the count covers the Gemini app, the API, AI Mode in Search, Google Antigravity or all of them combined, and he gave no split between consumer and enterprise use. Gemini 3 was released in November 2025, when Google priced Gemini 3 Pro in preview at $2 per million input tokens and $12 per million output tokens, so the comparison spans nearly a year of growth in Google's overall user base.

A rough sense of scale is available from Alphabet's own disclosures. In June, Alphabet said it was processing 3.2 quadrillion tokens a month across its products. Spread evenly, that is about 107 trillion tokens a day. Argon's reported 25 trillion over three days works out to around 8.3 trillion a day, or close to 8% of that June daily average, by PPC Land's calculation. The two figures cover different periods and possibly different scopes, so the comparison is indicative only.

Pichai described Argon as "a world-class frontier model" and pictured its strengths as a heat map across agentic tasks, knowledge work and coding. He was less specific about weaknesses. Coding remains the area Google wants to improve, he said, and "behavioral refinements" based on user feedback are under way. Hallucination, he said, is a problem that never ends: "It's an inherent part of the current state-of-the-art technology."

Who can use it, and why

Argon is a large model and expensive to serve, Heath noted, and it is currently available through the API and to Google AI Pro and Ultra subscribers. Pichai did not dispute that description. He said that at launch Google made Argon available to all Ultra tiers and the Pro tier in the Gemini app, followed by the Pro tier in the Antigravity app, and that Google is building smaller versions to reach more people.

The method is distillation, the same approach Google used with the 3.0 generation. According to Pichai, Google can produce "versions which are nearly as capable but much more efficient", and he said "we'll we'll continuously expand access from now on." He gave no date and no indication of whether the free tier of the Gemini app will receive any Argon-derived model.

That gating matters because of how Google has restructured consumer access in recent weeks. PPC Land reported that from October 9, free Gemini users lost access to two of three models, keeping only Flash-Lite. When Argon arrived on October 13, Google priced it at an introductory $2 per million input tokens and $10 per million output tokens, with consumer access limited to paid tiers. The same day, the Gemini app gained a toggle that switches chat into a task mode that keeps working while users are away, for subscribers paying $19.99 a month and up.

Asked whether rumors about a "carbon" model referred to a coming release, Pichai said only that Google always has internal versions of its frontier models under code names, and that this is "different from us getting a more efficient version of Argon available more broadly."

Safety testing and monitoring

Heath suggested that Google had been more methodical with safety testing for Argon than for previous models. "Uh definitely," Pichai replied. He tied the extra work to the model's agentic and coding capabilities and to Google's enterprise customers.

He described external testing that included the US government and early access for cyber researchers, plus wide internal testing. The heavier investment, by his account, went into infrastructure around the model rather than the model itself: sandboxing, real-time monitoring and telemetry. "That real time monitoring as these models are deployed I think ends up being critical," he said, so that Google gets early signals "if something is not working as intended and you can intervene." For customers, he said, the aim is that "you exactly know where it's running what access it has, uh, all the telemetry you need."

Pichai returned to the government point later in the interview. "Even though it's voluntary we are submitting our models to the US government for testing," he said, describing it as "a change from a year ago." A June executive order created a voluntary federal review under which developers can give the government access to covered frontier models for up to 30 days before release.

What happened to Gemini 3.5 Pro

One of the more concrete disclosures concerned a model that never shipped. Heath pointed out that Google discussed Gemini 3.5 Pro at I/O in May and asked what became of it. Gemini 3.5 and Antigravity 2.0 headlined Google I/O 2026, and PPC Land reported in July that pre-training of Gemini 4 had begun while 3.5 Pro was still with testing partners.

According to Pichai, Google identified gaps in agentic capabilities and coding during the middle of the year. The team judged the fastest way to close them was to iterate on the smaller Flash model, "so you can really get uh a lot more iterative loops there." That produced the run from 3.5 to 3.6 to 3.7 to 3.8 Flash, which Pichai said took about three months. "By the time we got that done, we had a chance to apply it to uh Pro, but Argon pre-training was done."

The conclusion, as Pichai put it, was that "3.5 Pro would have had a very short shelf life", so Google applied what it had learned to Argon's post-training instead. He framed the faster release cycle as a deliberate change for the year: "we needed to pick up the cadence at which we are shipping things." Argon is Google's largest external model, he said, and more 4.0 models will follow.

Agents inside Google

Heath cited a claim that Argon completed Minecraft on its own in a little over a day for less than 2,000 in compute costs, a figure Pichai did not confirm or repeat. Pichai answered with an internal example instead: pointing Argon at Chromium's codebase, which runs to millions of lines, to build a feature that lets a browser tab be dragged out and moved into an AI experience without reloading. Such work historically "cuts through all layers of Chromium", he said, but Argon "can persistently go and work through it and get it done."

He described 2027 as "a golden age" for that productivity frontier and spoke openly about recursive self-improvement, the use of models to accelerate model development. Researchers are "co-working with the models in a deep way", he said, with models suggesting recipes and evaluating checkpoints, while "a lot of humans in the loop" remain for safety reasons. "I would expect the next 12 months to see continued acceleration because of RSI," Pichai said.

He also described his own use. Following Google's pre-training, mid-training and post-training chat groups "could easily take me an hour a day to kind of catch up what's happened in the last 24 hours", he said. Agents now summarize that activity for him, and other agents report on compute usage, efficiency and utilization by product area. Compute allocation decisions, he said, are still "being made by uh humans in the loop including me."

The DeepMind leadership changes

Heath asked about the summer reshuffle at Google DeepMind: Demis Hassabis stepping back from running the unit to become chief scientist and spend more time on Isomorphic Labs, Koray Kavukcuoglu taking over day-to-day leadership of DeepMind, and Jeff Dean leaving after 27 years to start Discovery Loop. Heath said he had not heard Pichai or anyone named in the changes speak about them since.

Pichai's explanation centered on how the job has changed. "Frontier model development is all consuming. It's moving. It's picking up pace," he said, describing a "very scaled operation" running from data and compute to algorithms and team management. Hassabis, he said, wants to spend more time applying AI to science and medicine, citing AlphaFold and Isomorphic, and remains involved as chair of DeepMind. Kavukcuoglu, who wrote Google's Argon announcement on October 13 as chief AI architect, has been at DeepMind for well over a decade and moved to the Bay Area about 18 months ago, according to Pichai.

On Dean, Pichai said he saw "an opportunity to go do something which he was excited by with a set of people," and that Google is excited to see what he does with Discovery Loop. Asked whether the changes signal DeepMind folding closer into Google, he said the coupling had been under way for years, pointing out that the Gemini app is built inside DeepMind. Externally, he suggested, people noticed it as a moment; internally it was a continuum.

Capital spending and the negative cash flow quarter

The financial exchange will interest anyone tracking how much of Google's advertising profit is being redirected into infrastructure. Heath noted that Alphabet went free-cash-flow negative over the summer for the first time since its IPO. Alphabet's second-quarter results, released July 22, showed negative free cash flow of $5.9 billion, capital expenditures of $44.9 billion in the quarter and 2026 capex guidance raised to $195-205 billion, while Search revenue grew 17% to $63.3 billion. Heath said Alphabet had "guided up" capex for 2027. The July results gave no 2027 figure; the chief financial officer said spending would increase significantly again.

Asked what gives him the resolve to keep spending, Pichai, who counted Google among "the top cloud hyperscaler companies in the world," answered in one line: "The strength and durability of the demand signals we are seeing." He listed four sources of confidence: demand across consumer and enterprise products, Google's own internal adoption, internal evaluations that let Google "assess the capability frontier" independently of demand, and contracts. "We also have durable long-term contracts with committed revenue which we are signing. So it's it's not just hope," he said.

He added that research gives Google a view of the next generation: asked internally whether Gemini 5.0 will be much better than Argon, "the answer uh you know is a yes." He gave no timing for Gemini 5.

The planning horizon is long because Google builds its own chips. "At any given moment in time, you're living with all the compute decisions you made 2 years ago," Pichai said, adding that a year earlier the "liftoff" in agentic coding had not yet happened. Land and power planning, he agreed, already reaches into 2028 and 2029.

That spending has been financed partly with outside money. In June, Alphabet raised about $85 billion in equity, and PPC Land has reported on analyst estimates that OpenAI and Anthropic could account for more than 48% of Google Cloud revenue in 2027. Pichai did not address customer concentration.

Is it a bubble?

Heath raised CoreWeave and broader fears about neoclouds and AI financing. Pichai separated Google's position from the wider market. In a moment where "the opportunity is uh uh you know massive and the rewards are high and you have a lot of players coming in you are going to have dynamism and volatility associated with it," he said.

He declined to call the market overheated or not. Business cycles and shocks are "inevitable parts of how the world works," he said, but Google focuses on signals it controls: "contracts backlogs and so on." He used Google Cloud as precedent, describing an aggressive investment posture held "for a period of 6 to 8 years" after Google entered the cloud market late. When Heath asked whether he could see a point where Google would not be compute constrained, Pichai said he hesitated to answer.

Agents, the web and the ads business

For the advertising industry, the most consequential section began at 38:37, under the chapter "How AI agents change the web." Heath described agents such as Meta's Muse using browsers like humans to complete tasks, comparing it to fitting a triangle into a square, and asked whether the web needs to change.

Pichai said agents working at scale means "many many different parts of the economy interoperating with each other," and pointed to momentum behind MCP and payment protocols "like what we did with UCP along with a set of companies." Google introduced the Universal Commerce Protocol in January, co-developed with Shopify, Etsy, Wayfair, Target and Walmart. "There's going to be a chaotic evolution there," Pichai said, "hopefully working towards more interoperable standards."

He did not dismiss browser agents. Browser use and computer use are "a very elegant way to work too because that's what you as a user would do," he said, adding that they make it easier to build controls that keep users in the loop. Protocols, in his view, are "a far more scalable approach" but "will take longer time." Google runs both tracks itself: since July, Chrome auto browse has let AI agents book travel and fill forms for AI Pro and Ultra subscribers in the US.

Heath then put the direct question. Amazon had blocked Muse, he noted, and Google has "a very large ads business that requires engagement. So what do agents do to those kinds of interfaces?" Amazon blocked Meta's Muse from shopping on its site in September, after Meta declined to exclude Amazon voluntarily. DataDome later measured Muse at 33% of agentic browser requests in September.

Pichai's answer did not mention advertising. "The right answer longer term is you put users first," he said. "But ultimately, it has to work for all the businesses who are providing value." The market, he said, will "reach equilibrium and that equilibrium will shift," with new players treating each moment as a disruption. He offered real estate as an analogy: when the internet arrived, many expected home buying to transform overnight, yet "even today most of it happens the way it used to happen." AI agents will be "even more disruptive," he said, but "there'll be micro equilibriums which which will emerge."

That leaves open how Google intends to monetize, or protect, the attention that agents might route around. Neither Heath nor Pichai raised ads in AI Mode, ads in Gemini, or what happens to Shopping ads when an agent completes a purchase through UCP.

One agent layer instead of a merged product

Heath asked whether Google Search, AI Overviews and Gemini will ever fully merge, noting he had discussed this with Liz Reid. Reid had previously described informational queries going to Search or AI Mode and creative or productivity tasks going to the Gemini app.

Pichai's answer shifted the frame from products to agents. He cited personal intelligence, which can be turned on in both Gemini and Search and "has your context and so kind of works in a uh unified way." Then: "I expect our agents to provide that unifying experience across these products." Agents running in the cloud around the clock with a user's context "can pick up things where you left off," he said, so a task started in one Google product can continue in another. "The agent layer offers us a chance to bring a unified product experience better than ever before."

The consumer roadmap follows the same logic. Pichai said the goal for Gemini is "a personal proactive and powerful assistant", and that Google's enterprise and consumer agents use "the same model same harness same runtime." At the enterprise event he referred to, Google Cloud described Gemini agents with their own attested identities, sandboxed execution and audit trails. Each task, Pichai said, gets "essentially a microVM to go complete the task running securely 24/7 with realtime monitoring", and Google is bringing all of that to consumers.

The near-term vision, he said, is "one box", in which the line between a chat and a task disappears. Users will be able to talk to Gemini outside the app, "maybe from Gmail or you can call," he said. He reported that the share of Gemini queries that are agentic "has been steadily increasing over time" but gave no percentage. These features launched mainly for Ultra and Pro users and will become more widely available "as we go through the next few months."

Project Suncatcher, Waymo and Wing

Google announced Project Suncatcher in November 2025, with a plan for two prototype satellites built with Planet to launch in early 2027. The schedule moved forward: Google said on October 1, 2026 that a Suncatcher prototype satellite had launched on SpaceX's Transporter-18 rideshare mission and was operating as expected. Asked whether it was still in orbit, Pichai said it was "doing great."

He justified the bet on energy grounds. Sun-synchronous orbits offer "access to 24/7 energy," he said, while data centers on Earth "are not just arbitrarily big to how they were like 10 years ago." Extrapolating compute needs 20 years out on the ground would not leave Earth looking good, he suggested. He called Suncatcher a moonshot and the likely future way Google scales compute, but did not say whether it would become a separate Alphabet company.

On Waymo, Pichai said the company had just crossed half a million paid rides per week, and the exchange referred to a coming Toronto launch and 10 new cities this year. He said it drives "with 95% fewer serious crashes" than regular driving. In June, Alphabet had already reported more than 500,000 fully autonomous rides per week across 11 US cities, so the half-million mark is not new; the interview does not say whether "paid rides" is measured differently. Pichai said Waymo will keep working with partners including ride-hailing companies, and that selling Waymo vehicles to consumers is "not what we are working on at this moment", adding "never say never."

On Wing, Heath cited a figure of 40 million American households gaining access. Wing's January announcement with Walmart, covering 150 additional stores, referred to more than 40 million Americans, not households. Pichai said he expects Google "to meaningfully increase the scale of deliveries in 27 and also bring it to newer categories", through a partnerships-led approach.

Regulation and quantum

Pichai's main policy ask was "a more uniform regulation than a patchwork of state-by-state regulations," which he said could set the industry back and make it "inherently unsafer." He compared the next step to financial reporting, with standard evaluation methods and independent evaluators akin to audit firms, and argued that "companies which do that well will have a competitive advantage in the future."

On quantum computing, he said Google's team has tracked or exceeded internal milestones over almost a decade and that he sees "practical commercial applications" within five years. In that time frame, he said, "our Google Cloud team would be negotiating deals on quantum deployments at some scale."

Why this matters for marketers

Three points from the interview bear directly on advertising and search.

First, capacity. Google's most capable model remains rationed by price tier, and Pichai described the constraint as compute rather than policy. The distilled models he promised will determine what free users of Gemini and Search actually see, and when.

Second, interfaces. Pichai's picture of a single agent layer spanning Search, Gemini and Gmail implies fewer distinct surfaces where users type a query and scan results. PPC Land has tracked that direction since Google's I/O, where Search began to be described as an AI agent rather than a tool. If agents carry context across products, attribution and measurement for advertisers become harder to tie to a single session.

Third, the open web. Pichai endorsed both protocols such as MCP and UCP and browser-driven agents, while declining to describe how agent traffic fits with an advertising model that depends on engagement. His answer that equilibrium "will shift" acknowledges that the arrangement between Google, publishers, merchants and advertisers is not settled.

The episode carries sponsor segments for Hyperagent, Granola, Atlassian's Jira, Mercury and Framer, which are paid advertisements and not part of the interview.

Timeline

Summary

Who: Sundar Pichai, chief executive of Google and Alphabet, interviewed by Alex Heath on the Sources podcast. The interview also concerns Google DeepMind leaders Demis Hassabis and Koray Kavukcuoglu, former chief scientist Jeff Dean, and Google's Waymo and Wing units.

What: Pichai said Gemini 4 Argon processed almost 25 trillion tokens in its first 72 hours, about three times Gemini 3.0, that Google is capacity constrained and building distilled versions for wider access, that Gemini 3.5 Pro was dropped in favor of Argon, and that agents rather than a product merger will unify Search and Gemini. He defended record capital spending, described Waymo at half a million paid rides a week, and gave no specifics on how agents affect Google's ads business.

When: The interview was published on YouTube on October 21, 2026, about a week after Argon's wider release on October 13, 2026.

Where: The Sources podcast and its YouTube channel. The products discussed are available through the Gemini app, the Gemini API, Google Antigravity and AI Mode in Search, with consumer Argon access limited to Google AI Pro and Ultra subscribers.

Why: Argon and Google's agents are being connected to Search, Gmail and Gemini, which changes how users reach websites and where ads can appear. Pichai's comments on capacity, access tiers, protocols and browser agents indicate how quickly those changes will reach the wider audience that advertisers pay to reach.