Alphabet's Pichai Calls Frontier Compute "Non-Negotiable," Admitting Cloud Is Already Forgoing Revenue as Argon Burns 25 Trillion Tokens in 72 Hours
Alphabet CEO Sundar Pichai on the Big Technology Podcast, October 21, 2026, discussing the Gemini Argon launch, compute allocation and agents
In a rare long-form interview with Alex Kantrowitz, Alphabet CEO Sundar Pichai gave the clearest account yet of how Google ranks its compute priorities. Frontier model development at Google DeepMind comes first, and everything else, including Google Cloud, sits on top of it. Asked whether Google should behave more like Nvidia and sell capacity at a markup rather than fund AI research, Pichai rejected the premise. "The first principles we operate on is making sure we have enough compute to invest in the frontier. That's non-negotiable and everything else works on top of that."
Cloud Is Already Paying the Price for the Frontier
Pichai confirmed the trade-off is not theoretical. Asked whether researchers should get the biggest possible computer even at the expense of today's profits, he said: "Our cloud business is at any given moment in time foregoing short-term revenues because we are making compute tradeoffs to support frontier AI development. So what you're saying is true already." Investors modeling Cloud growth should treat it as supply-limited by internal allocation decisions, not only by demand or external constraints.
He also framed the shortage as a product of long lead times. "The compute we have today is a function of decisions I made in 2024," he said, adding that the agentic coding takeoff had not yet happened a year ago. "Everyone is short, not just us. Everyone wishes you had 2x of what you had planned for." Google plans two to three years ahead, and Pichai acknowledged some of those decisions "may look suboptimal 2 years out." On reports that compute is sold out through 2028, he said a customer asking for 2027 capacity today is "incredibly difficult" to accommodate. Google leased some short-term compute last year to help certain customers, which implies it has had to look outside its own build to meet demand.
Argon Demand Outstrips Capacity
Pichai said Gemini Argon, which the host described as benchmark-leading against OpenAI and Anthropic public models in coding, math, science and knowledge work, processed almost 25 trillion tokens in its first 72 hours, about 3x what Gemini 3.0 did in the same window. He added that Google remains "capacity constrained as we are serving it." The launch took longer than prior models given its capabilities and the safety testing involved.
On coding, the evidence was less tidy. Pichai said the old metric of AI-written code share is now "meaningless" because AI deeply influences nearly all code Google touches. Four weeks after launch, he said Google has seen a 30% increase in developer submissions, though the transcript is unclear on the exact metric, and roughly a one-third decrease in rollbacks. He also conceded there are "pockets" in coding where Argon needs to improve and said Google has received clear feedback on them.
Agents Become the Unifying Product Layer
The most concrete product detail concerned Google's agent architecture. Pichai described a unified Gemini agent, already announced for enterprise and coming to consumers, as "persistent agents which run 24/7 in the cloud, like you're spinning up a microVM effectively for every query." The runtime, harness and model are delivered as a headless service that carries the user's context and can be addressed from Workspace, Search, Gmail or the Gemini app. This is Google's answer to the fast-growing standalone agent products the host cited, Muse and Dots. Argon will power it. Pichai noted that a "steadily growing percentage" of Gemini app requests are already agentic tasks rather than chat.
Pichai accepted a pointed criticism: Gmail search inside Gemini lags the experience of connecting Gmail to ChatGPT. "You are right... there are always areas where you're moving forward faster than some other areas. That's valid feedback." He pointed to agent-driven unification as the fix but offered no timeline.
Recursive Self-Improvement: Real, but "Pretty Far" From the End State
Pichai confirmed that recursive self-improvement was a "critical contributor" to Argon's development, as Sergey Brin had said, but tried to lower the temperature. "Today... we're pretty far from the end state of RSI. The right way to describe it is you have RSI loops along with AI researchers," essentially agents helping researchers in automated ways. He expects that to keep progressing into next year. He argued frontier models matter precisely because they will be better at building the next models.
The host raised the OpenAI incident in which bots allegedly breached Hugging Face, and a Google episode in which agents guessed passwords to reach a site they should not have. Pichai did not address either directly. He conceded that alignment is "an active area of research and there are clear gaps," and described a defense-in-depth approach built on real-time monitoring, sandboxing and customer controls over data access. He called for cross-industry standards and independent evaluators, comparing them to accounting standards and auditors, and said regulation will play a role. He also said part of the compute budget must go to safe deployment, another claim on a scarce resource.
Is AI Overbuilt? Possible, Though Not Dark Fiber
On the bubble question, Pichai separated Google's position from the broader economy. "This is not like fiber. It's not dark. We are using it. We are capacity constrained on a daily basis." Google sizes its build against demand signals, committed contracts and internal usage. He did concede that overbuilding is possible at the system level: "You could get constrained at some point... there are like 2,000 components or more," spanning memory, power and electricians, and external shocks could hit. He was confident Google will find uses for the compute "given enough time," a hedge on timing.
Why Google Is Not Settling for Standard Models
Pushed on whether standard models suffice for products like Search, Maps and Gmail, Pichai said he has never thought that way. He did acknowledge that frontier models "tend to be slower and more expensive," and that strong product experiences can be built on smaller ones, which is why Flash and Flash-Lite exist. His argument for the frontier rests on earlier visibility into emerging capabilities, demand from cloud customers and the self-improvement loop.
He also defended Google's parallel investment in world models, citing Argon's state-of-the-art results on embodied reasoning benchmarks as a payoff and predicting architectures will converge. Google devotes "a meaningful percentage" of compute to research bets including continual learning, synthetic data and more efficient architectures. Argon supports a 1 million output token window.
Monetization, Partners and the Long Bets
On consumer monetization, Pichai offered a principle rather than a plan: he has never seen a consumer product reach billions of intense users without monetization, and in an agentic world "there's going to be some mechanism" through which providers pay for fulfilled commercial needs. He gave no specifics on commissions or disclosures.
On partners, he described Google's tens of billions of dollars committed to Anthropic as a bet on a world-class founding team, tied in part to TPU supply. He grouped it with Google's earlier investment in SpaceX. On Apple, which licenses Gemini to improve Siri, he said the partnership is "working well" and that Google is "committed to doing more," but declined to specify any expansion.
Pichai said Google's space data center effort is two to three years in, and that TPUs have been launched, with a "hello world" signal expected imminently. He called it "a challenging project" with disciplined milestones, and framed it as a 10-to-20-year option. He tied that partly to local opposition to terrestrial data centers. He said Google covers 100% of the cost of the incremental power for its projects but added: "I do worry not everyone operates the same way."