Almost 25 Trillion Tokens in 72… | The AI Corner Summary

1. Key Themes

Compute Is the Binding Constraint, and Capacity Planning Is a Long-Fuse Forecast

Pichai ties today's capacity to a decision made two years ago and plans 2–3 years forward, with the whole industry short on supply.

"The compute we have today is a function of decisions I made in 2024."

Google builds its own compute and data centers, so lead times run long. It plans for what it needs 2 to 3 years out, and "some decisions may look suboptimal 2 years out, depending on where the state of the art lands."
"Everyone is short today, he notes, wishing for 2x what they planned."
Launch demand shows the scale: "In the first 72 hours since we launched, it's almost up to 25 trillion tokens used, which is about 3x what Gemini 3.0 had in the first 72 hours." Google remains capacity constrained while serving the traffic.

Frontier Research Gets Compute First, and the Cost Is Named

Google allocates compute by strict priority: frontier AI development first, everything else planning around it. Pichai openly says cloud revenue is the trade-off.

"The first principles we operate on is to making sure we have enough compute to invest in the frontier. And that's non-negotiable."

"I think our cloud business is, at any given moment in time, foregoing short-term revenues because we are making compute trade-offs to support frontier AI development."
He traces the rule to 2015, when he set the company to be AI first. Research and model development get compute first, and each business then plans from "its own long-range projection, backed by demand signals and committed contracts."
Constraints also have a research upside: "Some constraints are also good. That's what inspires creativity." They push toward more efficient algorithms, "maybe ones that use less memory."

Agents Are Absorbing Chat, and the Interface Is Collapsing Into One Input

Gemini is removing the chat-versus-agent distinction as agentic tasks take a growing share of usage.

"You don't need to think about whether you're asking a chat or a, agentic task, it'll just kind of automatically do the right thing."

A steadily growing share of what arrives in the Gemini app is agentic tasks, "where last year most requests were chats."
Personal agents will work in the Gemini app, Google search and Gmail, "running persistently in the background," and Google wants to get it right first.

AI Code Quality Is Measured by Rollbacks, Not Volume

Google has retired the "percent of code written by AI" metric in favor of outcome measures.

"That number in our world today is like meaningless in the sense that every part of code we now touch is influenced by AI deeply."

"In the 4 weeks since launch, he reports a 1/3 decrease in rollbacks, and he reads that as people writing code more productively and better."
Per his account, "people pointed the model at complex projects, including some touching millions of lines of code."

Trust and Safety Will Lean on Independent Evaluators and Standards

Pichai points to the financial-controls model as a template for AI oversight.

"If you look at areas like financial controls, we have like accounting standards, we have independent auditors, et cetera, right?"

The next steps, he says, are "better standards and independent evaluators, and regulation will end up playing a role."

2. Contrarian Perspectives

"This Is Not Fiber": Rebutting the Bubble Comparison

The common bear case likens the AI buildout to the dot-com fiber overbuild, with idle "dark" capacity. Pichai rejects the analogy on utilization grounds.

"This is not like fiber. It's not dark. We are using it."

Evidence: Gemini 4 Argon drew almost 25T tokens in 72 hours, about 3x Gemini 3.0's launch window, while Google remained capacity constrained.
His planning inputs: "demand signals, committed contracts, internal evals and model trajectories," described as "a disciplined approach that is also confident and aggressive."
The author flags a caveat: stress-test the checklist before using it as a template, since "demand moves quickly."

Confident on Overbuilding, but Only After Naming the Failure Modes

Asked directly whether the industry is overbuilding, Pichai says yes, it is possible, and still backs Google's plan.

"Am I confident that the compute we are planning for, we will have uses for it given enough time? Absolutely, yes."

Risks he lists first: constrained components can cause bottlenecks, and external economic shocks count as macro risks.
Notably, he also says some of Google's decisions "may look suboptimal 2 years out," conceding forecast error is built in.

The Headline Adoption Metric Is "Meaningless"

Against the industry habit of touting AI-generated code share, Pichai calls the figure meaningless once AI touches everything. The author sharpens the point: "AI-generated code can look fine while churn climbs," and "volume flatters everyone."

3. Companies Identified

Google / Alphabet

Description: Parent of Gemini, Google Cloud, Search, Gmail, Waymo and TPU hardware.
Why mentioned: Main subject. Launched Gemini 4 Argon, is capacity constrained, and prioritizes frontier compute over cloud revenue.
Quotes: "Google took a bit longer with this release than with its other models, since it wanted the testing and deployment done with care." / "Google builds its own compute and its own data centers, he says, so lead times run long."

Gemini (Gemini 4 Argon / Gemini 3.0 / Gemini app)

Description: Google's frontier AI model family and consumer app.
Why mentioned: Launch-usage benchmark and the merged chat/agent interface.
Quotes: "In the first 72 hours since we launched, it's almost up to 25 trillion tokens used, which is about 3x what Gemini 3.0 had in the first 72 hours." / "The Gemini app already shows chat and task, he says."

Google Cloud

Description: Google's cloud business.
Why mentioned: The named "loser" in compute allocation.
Quote: "I think our cloud business is, at any given moment in time, foregoing short-term revenues because we are making compute trade-offs to support frontier AI development."

Waymo

Description: Google/Alphabet autonomous driving company.
Why mentioned: Used as the long-horizon precedent for space data centers.
Quote: "By his clock, Waymo is almost 20 years in and space is 2 to 3, with quantum named as another long bet."

Big Technology Podcast

Description: Podcast hosting the Pichai interview.
Why mentioned: Source of the interview.
Quote: "The Alphabet CEO takes questions on Big Technology Podcast about compute, agents, AI-written code, safety standards and data centers in space."

GrowthSchool (sponsor)

Description: AI training provider.
Why mentioned: Sponsored promotion of its 7-Day AI Mastery Challenge.
Quote: "GrowthSchool's 7-Day AI Mastery Challenge is usually $299, and it's free this Black Friday with code AICOLNOV1."

Anthropic, OpenAI, Coinbase, Nvidia (peripheral, via linked reading)

Description: Frontier labs, a crypto exchange, and a chip maker referenced in related-article links.
Why mentioned: Context links on agents, compute and valuation, not discussed in the body.
Quotes: "Brian Armstrong Runs 1,200 AI Agents at Coinbase." / "Anthropic is closing in on a $1 trillion valuation. Dario won." / "OpenAI's $122B masterclass: 10 takeaways from Sarah Friar."

4. People Identified

Sundar Pichai

Description: CEO of Google and Alphabet.
Why mentioned: Interview subject.
Quotes: "We are at early stages of the most consequential technology ever." / "The compute we have today is a function of decisions I made in 2024."

Ruben Dominguez

Description: Author of the newsletter post.
Why mentioned: Watched the interview and distilled 10 takeaways.
Quote: "I watched the full interview and pulled out the 10 answers that matter most."

Dylan Patel

Description: Semiconductor and compute analyst (referenced).
Why mentioned: Forecast of world compute by 2028, a second forecast to check against.
Quote: "Dylan Patel did the math on the world's compute by 2028, a second forecast to check against."

Sam Altman, Sarah Friar, Jensen Huang, Dario Amodei, Demis Hassabis, Brian Armstrong, Marc Andreessen (peripheral)

Description: Leaders referenced only through related-article links.
Why mentioned: Companion reading, not discussed in the body.
Quotes: "Demis Hassabis Says AGI Arrives in 2 to 5 Years." / "Test the answers against Altman's rules for the AI era and Andreessen's AI moat argument."

5. Operating Insights

Pair Every Output Metric With a Quality Metric

Report volume and rollbacks together. Pichai reports more productive coding alongside 1/3 fewer rollbacks.

"Pair every agent rollout with a quality metric, the way Pichai reports more productive coding alongside 1/3 fewer rollbacks."

Settle the Top Priority First, Then Let Every Team Forecast Around It

The allocation rule is copyable. Protect the one input the strategy depends on and name what you give up.

"To me, the order of operations is the part worth copying: settle the top priority first and let every team forecast around it."

Date and Grade Your Forecasts

Log an expected error next to every capacity number so the review has something to score.

"A capacity plan passes my test when each figure carries a logged error estimate, so the review in 2 years has something to grade."

6. Overlooked Insights

Space-Based TPUs Are Already in Orbit, With a "Hello World" Imminent

Beyond the long-dated framing, the interview contains a concrete near-term milestone and an energy rationale.

"Sun-synchronous orbits offer 24-7 access to the most abundant form of energy, he says, and Google expects a hello world moment from the TPUs it sent up while he speaks."

Launch Was Deliberately Delayed to Prioritize Testing

Google traded speed for deployment care on a model that then drew record demand. This is a quiet counterexample to the "ship fastest" ethos.

"Google took a bit longer with this release than with its other models, since it wanted the testing and deployment done with care."