Almost 25 trillion tokens hit Gemini 4 Argon in its first 72 hours, and Sundar Pichai says Google is capacity constrained while it serves the demand.
He frames the whole moment this way: “We are at early stages of the most consequential technology ever.”
The Alphabet CEO takes questions on Big Technology Podcast about compute, agents, AI-written code, safety standards and data centers in space, and several answers carry a number or a decision you can copy.
I watched the full interview and pulled out the 10 answers that matter most.
Sponsored by GrowthSchool
Argon hit almost 25 trillion tokens in 72 hours. The people who win 2027 are learning to build with Gemini and its rivals now.
GrowthSchool’s 7-Day AI Mastery Challenge is usually $299, and it’s free this Black Friday with code AICONOV1.
Build: 10+ hands-on projects, from AI automations to your own AI agent
Learn: 20+ tools, including ChatGPT, Claude, Gemini and n8n, live for the first 3 days
Get guided: by experts who’ve worked at Microsoft, Google, Amazon and NVIDIA
1. Almost 25 Trillion Tokens Hit Gemini 4 Argon in 72 Hours and Google Stayed Capacity Constrained
Launch-week usage settles the argument before any benchmark chart does.
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.
Usage numbers end arguments that benchmark charts keep alive.
Google took a bit longer with this release than with its other models, since it wanted the testing and deployment done with care, he says. Even so, Google is capacity constrained as it serves the traffic.
My first check on any new model is tokens per day, because agent costs climb with usage and AI costs keep rising even as tokens get cheaper.
Plan capacity for launch week, since that week sets your ceiling.
“We took our time with it.”
2. Google Retired the AI-Written Code Percentage (Here Is What It Tracks Instead)
Volume alone flatters every team that ships a lot of code.
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.
When AI touches everything, the share stops telling you anything.
The host asked him to update a 25% figure, and he answers that Google now tracks newer metrics. 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.
By his account, people pointed the model at complex projects, including some touching millions of lines of code.
A falling rollback rate outranks any share-of-code figure for me, whether the diff comes from Claude Code or a person, because AI-generated code can look fine while churn climbs. Teams chasing the output half can start with context engineering.
Report output and rollbacks as one pair.
3. Gemini Merges Chat and Agent Tasks Into One Input as Agents Take a Growing Share
The mode switch between chatting and delegating disappears from the interface.
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.
Hiding the chat-versus-agent choice removes friction early.
The Gemini app already shows chat and task, he says, and a steadily growing share of what arrives 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.
Actual Gemini traffic is what sells me on the merge, so route each goal to whichever agent fits. Some jobs still fit a plain workflow, so check whether you need an agent.
Offer 1 entry point and route work behind it.
To build the background-agent version of this today, start here:
Google just made agents 3x cheaper to run. Here is the playbook
Anthropic just solved the hardest part of building AI agents
4. Today’s Compute Came From a 2024 Decision (and Pichai Plans 2 to 3 Years Out)
Every capacity order is a forecast with a long fuse.
The compute we have today is a function of decisions I made in 2024.
One 2024 forecast still sets the ceiling today.
Google builds its own compute and its own data centers, he says, 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. Dylan Patel did the math on the world’s compute by 2028, a second forecast to check against.
A capacity plan passes my test when each figure carries a logged error estimate, so the review in 2 years has something to grade.
Grade last year’s capacity call before you make this year’s.
Planning a build on that horizon? These 3 sit closest to your decision:
5. Frontier Compute Comes First at Google (and Every Other Business Plans Around It)
One rule decides the order of every compute request at Google.
The first principles we operate on is to making sure we have enough compute to invest in the frontier. And that’s non-negotiable.
A non-negotiable line settles every allocation fight in advance.
He traces the rule to 2015, when he set the company to be AI first, so AI now runs through every business from search to Waymo. Research and model development get compute first, and each independent business then plans from its own long-range projection, backed by demand signals and committed contracts.
To me, the order of operations is the part worth copying: settle the top priority first and let every team forecast around it. Pichai’s I/O 2026 keynote is the companion read.
Protect the 1 input your strategy depends on.
6. Google’s Cloud Business Gives Up Revenue So Research Gets Compute (and Some Constraints Help)
Every allocation has a loser, and Pichai names his.
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.
Naming the loser out loud takes discipline.
That trade is already happening in some form, he says, and Google commits to investing what it takes to stay at the frontier. Some constraints help researchers too, he notes, since they push toward more efficient algorithms, maybe ones that use less memory.
I read that as a priority list with a price attached. Investors weighing the same trade can start with compute economics.
Name what you give up before you choose.
“Some constraints are also good. That’s what inspires creativity.”
7. Pichai Calls the AI Buildout Different From Fiber (Here Is the Evidence He Cites)
A bubble comparison meets a 1-line rebuttal about idle capacity.
This is not like fiber. It’s not dark. We are using it.
Busy capacity is the strongest answer to a bubble claim.
Over a long enough timeframe, this level of build makes complete sense to him. For the plan itself he weighs demand signals, committed contracts, internal evals and model trajectories, and he calls the result a disciplined approach that is also confident and aggressive.
I’d stress-test that checklist against the biggest risks in the AI thesis before using it as a template for any build decision. Demand moves quickly, as AI spend shows.
Plan from demand signals and committed contracts.
8. Pichai Says Yes to the AI Overbuilding Question (and Still Backs Google’s Plan)
A direct question gets a yes, and the risk list comes before the confident close.
Am I confident that the compute we are planning for, we will have uses for it given enough time? Absolutely, yes.
That confidence carries weight because the risk list came first.
Asked about overbuilding, he starts with a yes: constrained components can cause bottlenecks. External shocks to the economy count too, and he files them under macro risks.
I trust the confidence more because the failure points came first. My pre-build checklist opens with the bottlenecks, then asks what risk sits behind the forecast.
List your bottlenecks before you approve the build.
9. Pichai Points to Auditor-Style Evaluators for AI (and Says Regulation Will Play a Role)
Finance built standards and auditors; he points to the same toolkit for AI.
If you look at areas like financial controls, we have like accounting standards, we have independent auditors, et cetera, right?
The analogy fits because finance already runs on outside checks.
Worry is reasonable, in his view, and the right response is to ask early what more the industry can do. The next steps, he says, are better standards and independent evaluators, and regulation will end up playing a role.
My first question for any AI vendor is which outside party tests their model, since an internal eval grades its own homework. Buyers already hesitate on trust, and our Six AI Trends of 2026 breakdown tracks how fast regulation is landing.
Demand outside testing before you trust any eval.
If you ship AI to customers, read these before your next release:
The US government switched off Anthropic’s most powerful model 3 days after launch
Dario Amodei: AI Is Closer Than You Think (And Society Isn’t Ready)
The AI model that can hack anything, and why you can’t use it
10. Google Dates Its Space Data Center Bet at 2 to 3 Years and Compares It to Waymo
A moonshot gets an honest clock from the CEO behind it.
You’re always, with all these technologies, you’re at a moment in time where it looks a bit like sci-fi, but 10 to 20 years down the line, it’s actually working in a meaningful way.
Honest timelines make ambitious bets easier to fund.
He recalls watching the Stanley car compete at the DARPA Grand Challenge in the desert, the car that became the basis for Waymo. By his clock, Waymo is almost 20 years in and space is 2 to 3, with quantum named as another long bet.
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.
The habit I’d borrow is stating a project’s age out loud, since patience in the world of atoms runs on long clocks.
Date your moonshot, then hold it to milestones.
What This Means for You
Plan capacity years ahead, and judge AI output by what survives rollbacks.
Founders: Pichai ties today’s compute to decisions he made in 2024, so write down the capacity you need in 2 years and the demand signal behind each line. Read why lean AI startups are winning, then place a small first order this quarter.
Investors: Ask each AI company when it ordered the capacity it serves today and what share goes to research first. Test the answers against Altman’s rules for the AI era and Andreessen’s AI moat argument before your next diligence call.
Operators: Pair every agent rollout with a quality metric, the way Pichai reports more productive coding alongside 1/3 fewer rollbacks. Measure each agent’s ROI beyond tokens and time saved, and benchmark your scorecard against Anthropic’s 8 agent metrics.
Everyone else: Hand 1 weekly task to an agent and track what breaks as well as what ships. Before you pay for a plan, see which AI plan is worth paying for, where we priced 32. Try code-free agents if you want a starting point.
The 4 Principles to Steal
Date every forecast. Write the expected miss beside each capacity number and grade it later.
Name your compute trade-off. Say what the frontier costs the rest of the business, the way Pichai does with cloud revenue.
Pair output with rollbacks. Volume flatters everyone, and graph engineering covers what a loop does with the 1 number it is given.
Layer checks around the model. Outside evaluators and standards catch what your own telemetry misses.
Long lead times reward the people who decide early. Make your next capacity call this week.
Watch the full interview on Big Technology Podcast for his answers on Anthropic and Apple too.
Keep reading:
Anthropic just cut the price of frontier intelligence in half
China just open-sourced Opus-level intelligence. Here is the playbook
Google Open-Sourced the AI Agent Framework Behind Its Own Products
Demis Hassabis Says AGI Arrives in 2 to 5 Years. Here Is the Full Picture
Anthropic is closing in on a $1 trillion valuation. Dario won.
If this breakdown saved you the watch time, send it to one founder or investor who needs it.
If those tokens don’t create adequate value, Google will no longer be capacity constrained.