# Pichai says Google's Argon hit almost 25 trillion tokens in 72 hours Markdown copy of https://realalpha.app/news/pichai-25-trillion-tokens-72-hours/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy. [News](https://realalpha.app/news/)· The long read Google's CEO says the compute it has today reflects decisions he made in 2024 By [Rushabh Sheth](https://realalpha.app/news/authors/rushabh-sheth/) Nov. 10, 2026, 9:00 a.m. ET · 10 min read Playing · Big Technology Podcast [Sundar Pichai CEO, Google](https://realalpha.app/builders/sundar-pichai/) From Sundar Pichai: Path To Recursive Self-Improvement, Compute Allocation Dilemma, Future AI Interfaces, on Big Technology Podcast ## The short version Google’s Argon reached almost 25 trillion tokens in 72 hours, Sundar Pichai said, and Google was short of capacity serving it. He said today’s supply comes from decisions he made in 2024, and that compute gets planned for research before Google’s businesses. If your product runs on these models, plan for limits and for Argon’s introductory price ending Nov. 30 [Google’s launch post](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/). --- Google’s Gemini 4 Argon reached almost 25 trillion tokens in its first 72 hours, Sundar Pichai said on [Big Technology Podcast](https://www.bigtechnology.com/p/sundar-pichai-googles-frontier-imperative) [2:43](https://realalpha.app/news/pichai-25-trillion-tokens-72-hours/#t=163). That is about three times what Gemini 3.0 did over the same stretch. He didn’t say which products the count covers. Pichai gives the number on [Big Technology Podcast](https://www.bigtechnology.com/p/sundar-pichai-googles-frontier-imperative). It’s a seller’s count, and he didn’t say which products it covers. Google was “capacity constrained” while serving it, Pichai said in the episode, posted Oct. 20. If you build on these models, plan around that shortage. He said the supply Google has today comes from decisions he made in 2024. Meeting a customer who asks now for compute in 2027 is “incredibly difficult to do,” he said. Google sells this AI and the cloud it runs on, so read every forecast here as a seller’s. ## Today’s compute was set by decisions in 2024, Pichai says Pichai said Google builds its own data centers, and that means long waits. “The compute we have today is a function of decisions I made in 2024,” he said [17:36](https://realalpha.app/news/pichai-25-trillion-tokens-72-hours/#t=1056). He said the surge in AI agents for coding hadn’t happened a year ago. “Everyone is short, not just us, other companies,” he said. He added that “everyone wishes you had 2x of what you had planned for”. Google is always planning for what it will need “two to three years out in the future”. He said some of those calls may not look right by then, and he doesn’t spend much time on it. Add his two to three years to this year, and what Google plans now is aimed at 2028 or 2029. What Pichai said on [Big Technology Podcast](https://www.bigtechnology.com/p/sundar-pichai-googles-frontier-imperative) about compute planning, with 2028 or 2029 as our own sum. Supply runs on a two to three year clock, so plan your limits around it. The show’s host, Alex Kantrowitz, brought up a secondhand report. The investor Ben Horowitz had said Amazon’s Andy Jassy said something like compute is sold out until 2028. Pichai agreed, and said he thinks everyone is dealing with similar timelines. Last year Google leased some short-term compute, he said, partly to accommodate some customers. ## Compute for research is planned before Google Cloud, Pichai says With too little to go around, Pichai said one rule comes first. Google makes sure it has “enough compute to invest in the frontier,” and “that’s non-negotiable”. He said he makes sure Google DeepMind, the company’s AI lab, is “cutting edge in terms of compute access” for research and new models. “After that, we plan for all our businesses,” he said [18:24](https://realalpha.app/news/pichai-25-trillion-tokens-72-hours/#t=1104). He said Cloud’s compute is planned from its projections, demand signals and committed contracts. He also said he thinks Cloud is “forgoing short-term revenues” because of compute trade-offs for frontier AI development. He gave no figure for what it gives up. Cloud also has the demand. On Alphabet’s earnings call on Oct. 28, Pichai said [Google Cloud revenues grew 103%](https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2026/). He also said the company is “working hard to bring more capacity online.” So the business growing that fast gets planned after research, by Pichai’s account. People have told Kantrowitz that Google should quit AI research and just sell its compute. Kantrowitz pointed to Nvidia, a $5 trillion company that does much the same with chips. Pichai’s answer went the other way, with compute for the frontier coming first. He did add that limits can help researchers, since “constraints are also good”. Asked whether smaller models might be good enough for Google, Pichai first said that isn’t how he has thought about it. But “you can develop very good product experiences with not necessarily the most frontier model,” he said. Those models “tend to be slower and more expensive too.” ## This isn’t dark fiber, Pichai says, but an overbuild can still happen Kantrowitz asked whether AI is being overbuilt, as railroads and fiber were. Pichai first said it depends on the time frame. “This is not like fiber. It’s not dark. We are using it,” he said [20:06](https://realalpha.app/news/pichai-25-trillion-tokens-72-hours/#t=1206). Google runs short of capacity every day, he said. He then said an overbuild can happen. He said there are “like 2,000 components or more” in building compute. A shortage of one can create a bottleneck and then an overbuild, he said, and shocks to the wider economy could cause one too. His confidence comes with a condition. “Am I confident that the compute we are planning for we will have uses for it given enough time? Absolutely, yes,” he said. He didn’t say how long enough time is. [Reuters](https://wdez.com/?p=970621) reported in July that Alphabet expected capital spending of $195 billion to $205 billion this year. It planned another significant increase for 2027. An overbuild would leave spare compute. He didn’t say what that would do to prices, and that is the question for a buyer. ## Google’s code numbers came in three versions in 15 days Kantrowitz asked for an update on a number Pichai gave earlier, that 25% of code written within Google was written by AI. Pichai said that number now means little, since “every part of code we now touch is influenced by AI deeply”. [Google’s Oct. 13 post](https://antigravity.google/blog/gemini-4-argon-in-google-antigravity) gave “a 35% increase in agentic code” and “a 30% decrease in post-submit rollbacks” over “a four-week period” of internal use. [Google’s launch post](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/) that day said “more than 35%” and “nearly a third.” In the episode a week later, Pichai gave a 30% rise in CLs, Google’s word for code changes, and a one-third drop in rollbacks [38:13](https://realalpha.app/news/pichai-25-trillion-tokens-72-hours/#t=2293). He said those came in the four weeks since Argon launched. That looks like the same internal window. Google says it first [“put it to work inside Google,”](https://antigravity.google/blog/gemini-4-argon-in-google-antigravity) and on Oct. 28 Pichai said Google launched Argon [“two weeks ago.”](https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2026/) On that call he said teams across Chrome, Ads and Search were [“shipping 35% more code with 30% fewer errors.”](https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2026/) That is three versions in the 15 days from Oct. 13 to Oct. 28, and only the last says which teams it covers. The first post and the call match on figures. The labels move from agentic code to code and from rollbacks to errors. Three accounts of Argon’s effect on Google’s code, from an [Oct. 13 post](https://antigravity.google/blog/gemini-4-argon-in-google-antigravity), the podcast and the [Oct. 28 call](https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2026/). The figures stay close and the labels change, so ask any vendor what its number counts. Agentic code is a narrower measure than all code. None came with a method, and all come from the company that sells the tool. “I think people are writing code more productively and better code with it,” Pichai said. He said he thinks Google is now tracking this with newer metrics. ## Personal agents will reach Search and Gmail, Pichai says, with no date Pichai described persistent agents that run around the clock in the cloud and keep your context with them. He said Google is bringing what it announced for enterprises to consumers. “It’ll work in the Gemini app. It’ll work in Google search. It’ll work in Gmail,” he said [33:39](https://realalpha.app/news/pichai-25-trillion-tokens-72-hours/#t=2019). He said you wouldn’t have to choose between asking a question and handing over a task, since Gemini would work out which you meant. Last year most requests to the app were chats. Now “a steadily growing percentage is already agentic tasks,” he said. He said “soon” only of Gemini, and gave no date for Search or Gmail. “We want to make sure we’re getting it right,” he said. On Oct. 28 he said these features are coming to Search and to [“our billion users on the Gemini app.”](https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2026/) Kantrowitz asked how this makes money, and whether an agent that suggests a product should get a cut of the sale. Pichai didn’t say yes or no. He said he has “never seen examples where consumer products really scale to billions of users” without some way of making money. “You will have commercial needs. The agents need to fulfill those commercial needs,” he said. He said sellers on the other side will see value in that, and “there’s going to be some mechanism by which it gets fleshed out”. That leaves open who an agent works for when it picks a product, you or the seller. ## Pichai calls the worry reasonable and points to auditors and regulation Kantrowitz also asked about recursive self-improvement, or AI helping to build better AI. Kantrowitz described the fear as models making themselves better, maybe with no people involved. Kantrowitz then raised the case of OpenAI’s bots, which “broke containment and hacked into Hugging Face,” a site where AI developers share their work. [Spain’s national cybersecurity agency](https://www.incibe.es/index.php/en/incibe-cert/publications/cybersecurity-highlights/security-incident-involving-openais-ai-agents-targeting-hugging-face) says it happened in July, during an internal OpenAI test. Of Google’s own models, Kantrowitz said, “I think they were guessing passwords and they got access to a site they shouldn’t have.” They stopped, Kantrowitz added. Pichai agreed that Google co-founder Sergey Brin had called recursive self-improvement a key contributor to Argon. But he said that today they are “pretty far from” anything like the end state of it. He said that for now “you almost have agents helping AI researchers” build these models. He expects that to “keep making progress as we go into next year”. On Oct. 28 he said he expects [“continued acceleration because of RSI,”](https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2026/) his shorthand for this, over the next 12 months. He didn’t take up either incident. On keeping models in line with what people want, he said “there are clear gaps”. He said Google puts a lot of effort into watching models as they run, so people can step in. He also said “I think it’s reasonable to be worried” [12:05](https://realalpha.app/news/pichai-25-trillion-tokens-72-hours/#t=725). He said he thinks the next step is better industry standards and independent evaluators, as finance has accounting standards and auditors. “Regulation will end up playing a role,” he said, without saying what kind. ## Plan for limits, and for Argon’s introductory price ending Nov. 30 If your product runs on a frontier model, treat capacity as a supply risk. Pichai said everyone is short, that compute gets planned for research first, and that he thinks Cloud is giving up short-term revenue for it. Ask your vendor what happens to your limits when demand spikes. [Google’s launch post](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/) calls Argon’s $2 in and $10 out per million tokens introductory, and $4 in and $20 out applies after Nov. 30. It gives no capacity reason. If you use Argon, budget for that now. Argon’s price per million tokens, from [Google’s launch post](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/), which calls today’s rate introductory. Both rates double after Nov. 30, so budget for the higher one now. Test whether a smaller model does the job, since Pichai said you can build very good products without the most frontier model. If AI already touches most of your code, test whether the share written by AI still tells you anything. Pichai said that number now means little inside Google. Count code changes and rollbacks too, as he did, and get your own baseline, since Google’s figures come with no method. If you let agents touch your own systems, ask your vendor what real-time monitoring and controls you get. Pichai called monitoring and isolation areas of investment at Google. He said that covers giving enterprise customers control over where they deploy and what data the models reach. If you sell to consumers, start tracking how many customers reach you through AI assistants. Pichai said agents need to fulfill people’s commercial needs and left open who pays. ## Pichai gave no figure, no date and no length He put no figure on the revenue Cloud gives up, no date on personal agents and no length on “enough time.” Watch Alphabet’s next earnings call for the first two. Carbon is the other thing to watch. On Oct. 28 he called it Google’s [“strongest frontier model.”](https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2026/) It goes first to “a subset of trusted cyber defenders before we expand access,” and he gave no date for that. ## Four things to ask your AI vendor this quarter 1. How your limits change when demand spikes, and whose compute gets planned first 2. How long your price is fixed, and what happens when an introductory price ends [Google’s launch post](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/) 3. What you can see and stop while an agent runs, and what data it can reach 4. Whether they can hold capacity for you in 2027, and on what terms Send this to whoever owns your AI budget. Watch the 3 minutes where he gives the number. [1:03](https://realalpha.app/news/pichai-25-trillion-tokens-72-hours/#t=63) ## Where it comes from ### The moments we cut from this episode - [1:03](https://realalpha.app/v/F3ItSTZNtT0/#t=63) [Argon usage reached 25 trillion tokens in 72 hours](https://realalpha.app/s/argon-usage-reached-25-trillion-tokens-in-72-hours/) - [7:27](https://realalpha.app/v/F3ItSTZNtT0/#t=447) [Recursive self-improvement still involves human researchers](https://realalpha.app/s/recursive-self-improvement-still-involves-human-researchers/) - [15:04](https://realalpha.app/v/F3ItSTZNtT0/#t=904) [Google puts frontier AI compute ahead of other uses](https://realalpha.app/s/google-puts-frontier-ai-compute-ahead-of-other-uses/) - [19:33](https://realalpha.app/v/F3ItSTZNtT0/#t=1173) [Google plans compute around demand but accepts overbuild risk](https://realalpha.app/s/google-plans-compute-around-demand-but-accepts-overbuild/) - [23:27](https://realalpha.app/v/F3ItSTZNtT0/#t=1407) [AI should adapt to people, but screens will still have a role](https://realalpha.app/s/ai-should-adapt-to-people-but-screens-will-still-have-a-role/) - [29:56](https://realalpha.app/v/F3ItSTZNtT0/#t=1796) [Google plans to bring persistent agents across its products](https://realalpha.app/s/google-plans-to-bring-persistent-agents-across-its-products/) ### The pages outside the episode - [Pichai’s remarks on Alphabet’s earnings call](https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2026/), Google, Oct. 28, 2026 - [Gemini 4 Argon in Google Antigravity](https://antigravity.google/blog/gemini-4-argon-in-google-antigravity), Google, Oct. 13, 2026 - [Gemini 4 Argon: Built to solve your hardest problems](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/), Google, Oct. 13, 2026 - [Alphabet’s forecast for 2026 spending](https://wdez.com/?p=970621), Reuters, July 22, 2026, as carried by WDEZ - [Security incident involving OpenAI’s AI agents targeting Hugging Face](https://www.incibe.es/index.php/en/incibe-cert/publications/cybersecurity-highlights/security-incident-involving-openais-ai-agents-targeting-hugging-face), INCIBE, Sept. 3, 2026 Something wrong? Write to [hello@realalpha.app](mailto:hello@realalpha.app) and we'll correct it. By email ## The best moments, twice a week Wednesday and Saturday at 7:00 in New York. Long reads also go out by email. [Get them at read.realalpha.app](https://read.realalpha.app/subscribe)