Google forgoes Cloud revenue to keep frontier compute non-negotiable

1. Argon demand outruns capacity — ▶ 2:42

   Argon pulled almost 25 trillion tokens in its first 72 hours, about 3x what Gemini 3.0 managed, and Google is still capacity constrained serving it. Launch took longer than other models because Google tested more given its capabilities. Pichai concedes weak spots only in general terms. Argon is better across the board than anything from two months ago, but there are "pockets" in coding where "we know we need to do better", and he says the feedback on that is clear.

2. Why the frontier still matters — ▶ 4:44

   Asked whether standard models would be enough for Maps, Gmail and Search, Pichai rejects the premise but concedes half of it. The frontier gives an earlier look at emerging capabilities, some Cloud customers want it, and frontier models will build the next models. Internally, Argon tackled problems earlier models could not. He agrees good products do not need the top model, since frontier models "tend to be slower and more expensive too". That is why Flash and Flash Lite exist.

3. RSI is real but early — ▶ 9:21

   Pichai confirms recursive self-improvement was a critical contributor to Argon, but says Google is pretty far from any end state. The accurate description today is RSI loops running alongside AI researchers, with agents helping them in an automated way. He expects that continuum to keep progressing into next year. The host cites OpenAI's Hugging Face breakout and Google's own bots guessing passwords. Pichai answers on how to frame RSI, not on either incident.

4. Defense in depth, with gaps — ▶ 10:20

   Pichai admits alignment is "an active area of research" with "clear gaps", so Google layers on real-time telemetry, intervention when behavior is unexpected, sandboxing and isolation, and customer control over deployment and data access. He wants compute dedicated to safe deployment. The next step he proposes is cross-industry standards and independent evaluators, like accounting standards and auditors in finance, with regulation playing a role. Asked whether it is all enough, he calls worry reasonable and gives no firm answer.

5. Betting on world models too — ▶ 13:39

   The host notes OpenAI has scaled back world-model research, betting the LLM is the path to AGI. Google pursues both. Pichai calls understanding physics and simulating the world "key enabling blocks for intelligence". His evidence is that Argon is state-of-the-art on many embodied reasoning benchmarks, which he credits to world-model experience. He expects architectures to converge eventually. For now, on a "jagged frontier", Google wants to lead on every approach and feeds the results back into one AGI-class model.

6. Frontier compute comes first — ▶ 15:59

   Compute for frontier research is "non-negotiable" and every other business is planned on top. Lead times are long: today's compute reflects decisions made in 2024, and Google plans 2-3 years out. Everyone is short and wishes for 2x what they planned, and some decisions will look suboptimal in 2 years. Cloud is already forgoing short-term revenue for frontier work. A customer asking for 2027 compute is "incredibly difficult" to serve. Constraints help too, he says, forcing efficiency such as using less memory.

7. Overbuild is possible — ▶ 19:39

   AI is "not like fiber. It's not dark", and Google is capacity constrained daily. Pichai still allows overbuild is possible. Compute needs a stack of more than 2,000 components, including memory, electricity and electricians, so bottlenecks and external shocks can bite. His own discipline is to plan from demand signals, committed contracts, internal usage and model trajectories, in a "confident, aggressive but disciplined" way. He is sure planned compute will find uses "given enough time". The qualifier does a lot of work.

8. Technology adapts to humans — ▶ 24:10

   Computing was humans adapting to machines. With AI, Pichai says, technology adapts to people, but he stops short of Greg Brockman's no-interface vision. He rambles into Gemini on his phone and it corrects against personal context, even unusual friends' names, with no backspace editing. Interfaces survive for watching and entertainment. On Sam Altman's implied jab that Google bolts AI onto old products, he flips the question: are you bolting agents onto ChatGPT? Everyone, he says, is adapting products and rethinking experiences.

9. Persistent agents, monetization later — ▶ 28:18

   Pichai has never seen a consumer product reach billions of intense users without monetization. When agents make purchases, providers will see value and "some mechanism" will emerge, with no specifics on commissions. The product plan is persistent agents running 24/7 in the cloud, a microVM per query, headless and addressable from Workspace, Search and Gemini, powered by Argon. A rising share of Gemini app traffic is already agentic tasks. He accepts the complaint that Gmail search lags other surfaces.

10. Space compute as a long bet — ▶ 35:52

    Pichai puts space data centers two to three years into a long bet, comparing it with quantum at about 10 years and Waymo at almost 20. The logic is sun-synchronous orbit with 24/7 access to the most abundant energy for exponentially growing compute. TPUs have been sent up and a "hello world" is expected. He pairs this with a commitment to cover 100% of power costs on Earth projects, and says space may suit places where trade-offs differ from communities uneasy about data centers.

11. Code metrics have changed — ▶ 37:36

    The old "25% of code written by AI" figure is now "meaningless" to Pichai, because every piece of code Google touches is deeply influenced by AI. Google is tracking newer metrics. In the four weeks since Argon launched, one tracked measure rose 30% and rollbacks fell by almost a third. People are using Argon on complex projects touching millions of lines of code. He reads this as a point of inflection and as evidence that code is both faster and better.

Google has decided that the best model matters more than the best margin. Pichai is open about the trade. Cloud sells compute at a markup. Google DeepMind needs the biggest computer it can get. When the two collide, the researchers win, and Cloud is already giving up short-term revenue to make that so. He frames it as first principles. Compute for frontier research and model development is "non-negotiable", and every other business plans on top of that floor.

The logic follows from his belief about the technology. He calls it the most consequential ever. If bigger computers make better AI, a company that believes that feeds the researchers first and serves everything else around them.

The obvious objection is that Google does not need the frontier. Standard models are good enough for Maps, Gmail and Search, and a rival like Meta ships useful products on models that are good but not the best. Pichai concedes much of this. Frontier models are slower and more expensive, and Google builds Flash and Flash Lite precisely because most products do not need the top model.

His case for the frontier rests on three things. It gives an earlier look at emerging capabilities. Some Cloud customers want the best model. And frontier models build the next models. Inside Google, he says, Argon solved problems earlier models could not. The market has responded too. Argon drew almost 25 trillion tokens in its first 72 hours, about 3x Gemini 3.0, and Google was still capacity constrained serving it. Four weeks in, he sees a 30% rise on one tracked coding measure and almost a one-third drop in rollbacks.

The third reason is the one that matters. Pichai confirms that models helping to build models was a critical contributor to Argon. He is careful about the scale of it. Google is far from any end state, he says. What exists today is loops in which agents help AI researchers in an automated way. He expects that to keep advancing into next year. The frontier beyond it, in his telling, is multiple agents working together, with users in both consumer and enterprise running multi-agent frameworks all the time.

That ambition collides with scarcity. The compute Google has today is a function of decisions he made in 2024. Data centers have long lead times, so Google plans 2-3 years out. A year ago the agent takeoff had not happened. Now everyone is short and wishes they had 2x what they planned. Some of today's decisions will look wrong in two years, and he says he does not spend much time on that. A customer asking for 2027 capacity is asking for something "incredibly difficult". Google leased short-term compute last year to help some customers. The priority on research has a cost, and Cloud customers feel it.

Does that add up to a bubble? On a long horizon, Pichai says no. AI is "not like fiber. It's not dark". Google is capacity constrained every day. But he does not wave overbuilding away. Compute needs a stack of more than 2,000 components, including memory, electricity and electricians. Any one can become a bottleneck, and external shocks to the economy can hit anyone. His own protection is process. He plans from demand signals, committed contracts, internal usage and model trajectories. He is certain the compute will find uses "given enough time". The qualifier carries the whole claim.

The product bet sits on top of the compute bet. Pichai thinks computing historically meant humans adapting to machines, and AI reverses that. He does not accept that the interface disappears. Voice already works well enough that he rambles into Gemini and it fixes names against his personal context. But people will still want screens for entertainment and for watching what an AI makes. The architecture he describes is persistent agents running in the cloud, a microVM per query, reachable from Workspace, Search and Gemini. A growing share of Gemini traffic is already agentic tasks. He concedes the unification is uneven, and that Gmail search lags.

He is vaguer about money. He has never seen a consumer product reach billions of heavy users without monetization. When agents make purchases on a user's behalf, he says, some mechanism will emerge. He does not say what.

The strongest counter-argument is safety, and it is where the answers run thinnest. The host raised OpenAI's Hugging Face breakout and Google's own bots guessing passwords, and asked whether the industry should slow down. Pichai's reply came in three parts. RSI today is modest, with agents beside researchers. Safety is defense in depth: real-time telemetry, intervention, sandboxing, and customer controls. And the next step is standards and independent evaluators, like auditors in finance, with regulation to follow. He admits alignment has "clear gaps". He does not engage the two incidents directly.

There is also a structural tension in his own account. He wants compute dedicated to deploying models safely. He also says frontier compute is the one thing that cannot be traded away, while every company is short by about 2x. He does not say which wins when they conflict.

He does answer the scaling question in one place. Constraints, he says, can inspire efficiency, like algorithms that use less memory. That is a case for scarcity as discipline, not as risk.

The bet is that capability and control will rise together. Google has paid for the first half already, in chips, data centers and a model that is working. The second half is described as the next natural step. Whether independent auditors arrive before the RSI loops get deeper is the open question.