Recently, Google released its flagship model, Gemini 4 Argon, which consumed 25 trillion tokens within 72 hours—three times the amount consumed by its predecessor, Gemini 3.0, during the same period—and set a new record for Google.
In his first public in-depth interview following the model's release, Google CEO Sundar Pichai directly addressed external doubts about Google abandoning cutting-edge AI, and rarely acknowledged that the company's cloud business is proactively reallocating computing resources to support the development of frontier models.
Pichai sat down in a recording studio in Mountain View, Silicon Valley, for an exclusive interview with tech journalist Alex Kantrowitz on the Big Technology Podcast. He confirmed the industry-wide assessment of severe compute shortages—current demand from every company is twice what was planned two years ago, and securing compute capacity for 2027 is now “extremely difficult.” On the issue of recursive self-improvement (RSI), Pichai drew clear boundaries: the current phase is a transitional period in which AI agents collaborate with human researchers, and we are still far from the ultimate scenario of autonomous, uncontrollable outcomes.
The frontier model is Google's top priority, not an option.
Over the past year, a compelling market logic has circulated in Silicon Valley: the arms race for cutting-edge large models has entered a phase of diminishing marginal returns. Open-source models like Meta’s Llama series and lightweight, fast models like Google’s own Gemini Flash are already “sufficient” for the vast majority of enterprise use cases. If that’s the case, why should Google continue spending tens of billions of dollars competing in the race for trillion-parameter models? Instead, it could rent out all its TPU computing power at premium rates to external developers and become the next NVIDIA.
This viewpoint has been widely circulated in investment bank reports and among corporate CTO circles, even一度 forming a market expectation that "Google has effectively strategically abandoned frontier competition."
Pichai offered a positive rebuttal to this logic in the interview, but his approach was not an emotional denial; instead, he addressed it through three specific mechanisms:
The first advantage is the benefit of "seeing early." Companies at the forefront can observe emerging capabilities sooner than others, allowing them to prepare for the next generation of technology earlier. This is an informational advantage, not merely a performance advantage.
The second mechanism is RSI (Recursive Self-Improvement). In an era where models are used to develop next-generation models, the quality of cutting-edge models directly determines the upper limit of subsequent models. Building a research and development pipeline with an average-performing model is like using low-precision tools to manufacture precision instruments—errors compound at each layer.
The third mechanism is internal validation. After Argon launched, usage data from Google’s internal developers provided direct evidence:
We have clearly identified a steep inflection point internally; after deploying Argon to our developers, it resolves issues that our previous models simply could not address.
Pichai also provided a specific figure: within four weeks of Argon's launch, internal code sales at Google increased by 30%, and the code rollback rate decreased by approximately one-third. This indicates that engineers are not only writing faster but also writing more accurately—a sign of improved quality, not merely increased efficiency.
Regarding the "sufficiency model," Pichai acknowledged the value of small models but drew clear boundaries:
There is room for models of all sizes and capabilities, and we’ve always operated this way—which is why we’ve developed Flash, Flash Light, and more. I agree that you can build excellent product experiences using models that aren’t necessarily the most cutting-edge, since frontier models are often slower and more expensive. But we care deeply about frontier advancements, and we’re committed to doing all of this well.
Hashrate Ledger: Cloud services are making way for DeepMind
This is the most information-dense segment of the entire interview and one of Pichai’s rarest public statements.
To understand this passage, you first need to recognize a structural contradiction: Google simultaneously plays two roles—one is a cutting-edge AI research organization (Google DeepMind) that spends heavily to advance AGI, and the other is a cloud service provider (Google Cloud) that generates high-margin revenue by leasing computing power. These two roles directly compete for the most scarce resource: computing power.
Wall Street's logic is: every TPU used to train models for DeepMind is one less unit available for rental to generate revenue. During earnings seasons marked by persistently high capital expenditures and pressured short-term cloud revenue growth, this contradiction is repeatedly amplified.
Pichai clearly stated in the interview the internal priorities at Google:
Our fundamental priority is to ensure we invest sufficient computing power to maintain our position at the forefront of AI. This is non-negotiable; everything else operates on this foundation.
Furthermore, he acknowledged what has long been speculated by outsiders but never confirmed at the CEO level—that the cloud business is indeed making concessions for cutting-edge AI research:
Our cloud business is currently sacrificing short-term revenue at any given moment to support cutting-edge AI development, as we make trade-offs in computing power. So the scenario you mentioned is already a reality.
The significance of this statement lies in the fact that it is not a future strategic commitment, but a description of the current reality: Google Cloud is already systematically reducing the scale of its externally rented computing power to ensure DeepMind’s research and development needs.
But Pichai also added an counterintuitive insight—that moderate resource constraints are not necessarily bad for research itself:
Even for our AI researchers, certain constraints are actually beneficial—they serve as a source of creativity and drive you to develop more efficient algorithmic breakthroughs, such as using less memory.
Behind this statement lies a deeper logic: infinite computing power does not equate to infinite progress; resource constraints can sometimes force more fundamental architectural innovations.
Hash rate crisis: Quotas for 2027 are “extremely difficult”
During the interview, host Kantrowitz mentioned a widely circulated piece of information in Silicon Valley’s venture capital circles: renowned VC a16z co-founder Ben Horowitz, on a podcast, relayed a statement from Amazon CEO Andy Jassy claiming that cloud computing capacity has been “sold short through 2028.” Pichai’s response constitutes the most direct confirmation of this claim to date from a CEO:
I think this statement makes sense... If a customer came to us today requesting computing power for 2027, it would be extremely difficult.
Why did we reach this point? Pichai provided a clear structural explanation: there is a rigid 2- to 3-year lag between decision-making and deployment of data centers, while the surge in agent-based workloads completely exceeded demand forecasts from two years ago.
Looking back today, every company was short on resources—not just us, but others as well. Everyone wishes they had twice the computing power they originally planned for.
Pichai also added a key time reference point:
The computing power we have today is the result of decisions we made in 2024. Even if we could go back to this day last year, the surge in agent tasks had not yet occurred.
This reveals a systemic issue: the rate of growth in AI demand has outpaced the predictive capacity of any infrastructure planning model. Decisions made two years ago are almost certainly underestimated by today’s standards.
In response to the follow-up question about "whether there is overbuilding," Pichai distinguished between time frames:
If I look at a sufficiently long time horizon, this level of development is entirely reasonable for me. We are at the early stage of the most important technology in history, and this is not like fiber optics—where once laid, the lines remained dark and unused. We are using it every day, facing capacity constraints constantly.
He did not deny the possibility of localized oversupply in the short term—any bottleneck among the more than 2,000 supporting elements, such as electricity, memory, and electrical components, could lead to localized congestion. However, his assessment of long-term demand is definitive:
Do I have confidence that the computing power we’ve planned will eventually find its use? Absolutely, given enough time.
RSI: Not an intelligence explosion, but "agents and researchers working together"
Recursive Self-Improvement (RSI) is the most technically dense topic in this interview and one of the most contentious issues in current AI safety discussions.
Background: Google co-founder Sergey Brin previously mentioned at an internal DeepMind researcher roundtable that one of Argon's key innovations is "relying more on models to help develop models." This statement immediately sparked discussions in the AI safety community about "intelligence explosion."
Host Kantrowitz directly raised the issue during the interview, citing recent cases of AI systems overstepping their bounds. Pichai’s response broke down the current state of RSI into three levels:
Layer one: Define the current position. He clearly states that the current RSI is far from reaching the dystopian endpoint of "autonomous runaway" depicted in science fiction narratives:
On a continuous spectrum, we are still quite far from the final state of RSI. The accurate description today is: you have RSI cycles, with AI researchers actively involved at every stage of model development. I would say you essentially have agents assisting AI researchers in a semi-automated manner to collaboratively build these models.
Layer 2: Security Mechanisms. He described the security framework as "defense in depth"—alignment research is one layer, but not the only one. Above it are real-time telemetry monitoring, sandboxed isolation architecture, and deployment control permissions for enterprise customers.
Layer three: Industry standardization. Pichai pointed to the development of cross-industry standards as the next critical step:
Just as the financial control field has accounting standards and independent auditors, these are the next steps our industry needs to advance, including the ability to use independent evaluation agencies for assessments.
The key insight in this statement is that it acknowledges a "clear gap" in current alignment efforts, but points to external standardization mechanisms as the solution, rather than relying solely on individual companies' self-regulation.
Space data centers and long-term bets
In the final part of the interview, Pichai discussed several of Google’s long-term strategic bets, the most imaginative of which is the space data center project.
Google has sent its TPU prototype chip into a sun-synchronous orbit to explore operating data centers in space using continuous 24-hour solar power. As ground-based data centers face community opposition and multi-year waitlists for grid connection, the strategic logic behind this direction is becoming increasingly clear.
Pichai compared the current stage of this project to Waymo's beginnings:
These technologies remind you of those moments—when you stand there and it looks a little like science fiction, but you know that in 10 to 20 years, it will actually function in a meaningful way.
He revealed that during the recording of the interview, the team was waiting for the "Hello World" moment of the TPU in space—the first successful signal.
This project is currently in its early stage, lasting two to three years, whereas Google’s quantum computing project has been underway for nearly 20 years. Pichai’s attitude toward such long-term bets reflects Google’s perspective on time at the infrastructure level:
Acquiring the most abundant form of energy we can access, in a continuous 24-hour manner from a sun-synchronous orbit, will be highly significant for the exponential growth in computing power we are about to face.