The History and Future of AI at Google, with Sundar Pichai

ELI5/TLDR

Sundar Pichai sits down with Stripe’s founders (John and Patrick Collison) and Elad Gil to talk about Google’s AI journey, where the money goes, and what comes next. Google invented Transformers and had its own ChatGPT internally (LaMDA) but was slower to ship it — not because they missed the idea, but because they had a higher bar for product quality and a more toxic early model. The real bottleneck on AI progress right now is not money or ambition but physical stuff: memory chips, wafer capacity, permitting for data centers, and electricians. Google is spending roughly $180 billion in CapEx in 2026, and Pichai personally spends an hour a week deciding which teams get scarce compute.

The Full Story

The Transformer Origin Story Everyone Gets Wrong

The standard narrative goes: Google invented Transformers, then fumbled the ball while OpenAI ran it into the end zone. Pichai pushes back on this, and his version is more interesting than the usual “big company slow, startup fast” story.

Transformers were not some abstract research project. They were built to solve specific product problems — making translation better, making speech recognition scale to two billion users when Google did not have enough chips. Once built, they were immediately deployed in Search via BERT and MUM, producing some of the largest search quality jumps Google had measured. This was not research sitting on a shelf.

Google even had the chatbot product. LaMDA was, in Pichai’s telling, essentially an early ChatGPT. The engineer who thought it was sentient was talking to what amounted to a proto-ChatGPT inside Google. But the internal version was, to use Pichai’s word, “toxic” — not RLHF’d, not safe enough to ship. Google’s search quality culture meant a higher bar for what counted as “acceptable.”

“We even had the product version of it in the multiverse, somewhere else. Google probably shipped that nine months later or something like that.”

Pichai frames this with a surprisingly relaxed analogy: Google had Google Video Search, then YouTube came along, and Google just bought YouTube. Facebook had its social network, Instagram came along, Facebook bought Instagram. In consumer internet, three people in a room can always prototype something surprising. That is the nature of the game.

“I don’t think people wake up in a garage and ship a better iPhone. That’s not going to happen. But that’s not how consumer internet is.”

The signal Google may have missed, he admits, was on the coding side. The capability jumps between GPT-2, 3, and 4 were more pronounced in code than in pure language. If you were watching the coding use case, the exponential was steeper.

Speed as Religion

Google has long been obsessed with latency. The original Search page famously displayed query time in the results. That obsession continues in the AI era.

Search teams now operate with “latency budgets” measured in single-digit milliseconds. If you ship an improvement that saves 3 milliseconds, you earn 1.5 milliseconds of latency budget for future features; the other 1.5 goes directly to the user. Despite adding enormous AI functionality, Google Search has gotten 30% faster over the past five years.

This is the logic behind Gemini Flash models: roughly 90% of the capability of the Pro models, but substantially faster and cheaper to serve. The vertical integration — Google making its own TPUs, running its own data centers — is what makes that tradeoff possible.

Search Becomes an Agent Manager

Asked whether Search will exist in 10 years, Pichai gives an answer that is both honest and corporate-safe: it will keep evolving. But his description of what it evolves into is genuinely interesting.

“Search would be an agent manager in which you’re doing a lot of things.”

He describes a future where what we now call “search queries” become long-running, asynchronous tasks. You will have multiple agent threads doing things for you simultaneously. The one-line prompt returning ten blue links is already morphing — people are doing “deep research queries” in AI mode that do not fit the old definition at all.

He refuses to get paralyzed by 10-year predictions, arguing the curve is steep enough that thinking one year ahead is plenty:

“The models are going to be dramatically different in a year’s time. I think riding the curve itself is exciting.”

The Market Got Google Wrong

A year ago, in spring/summer 2025, the prevailing sentiment was that Google Search was “cooked.” The stock sat around $150. Pichai says he found this baffling. Google had been operating AI-first since at least 2016, was on its seventh version of TPUs, and had invested intentionally across the full stack — research, infrastructure, platforms, applications.

“To me, we were behind in terms of frontier LLM models, but we had all the capabilities internally, and we had to execute to meet the moment.”

The turning point in outside perception was probably Gemini 2.5, where multimodal capabilities started to show. But Pichai is clear-eyed that the frontier is intensely competitive, with two to three labs pushing each other hard month by month.

The AGI Question and “AI Psychosis”

When asked whether Google is less “AGI-pilled” than competitors, Pichai offers a pointed response: Google has scaled CapEx from $30 billion to roughly $180 billion. You do not do that unless you believe in the curve. He calls the perception gap “largely semantics” — Google is a bigger company that talks about things differently.

“At one point, Demis, Jeff, Ilya, Dario were all there.”

That is a tidy way of saying: the people who built the other leading AI labs were at Google first.

His own “feel the AGI” moments: watching Jeff Dean demo Google Brain recognizing a cat in 2012, seeing early Waymo, and more recently, using coding agents where he did a hobby project and only afterward thought to ask what programming language it had used.

“That was a detail that I needed to ask it about after everything was up and running.”

The Physical Bottlenecks

This is where the conversation gets most concrete. Google wants to spend $175-185 billion in CapEx in 2026. But even Google cannot spend $400 billion if it wanted to. The constraints are physical:

Wafer starts. The fundamental bottleneck. There is only so much semiconductor fabrication capacity in the world.

Memory. Currently the most critical component. The leading memory companies cannot dramatically improve capacity in the short term. No amount of money fixes 2026-2027 memory supply.

Permitting and construction. Even in pro-growth states like Texas, Nevada, or Montana, building fast enough is hard. Pichai explicitly admires China’s construction speed and says the US needs to learn to build things much faster.

Power and energy. More solvable than the others, but still a constraint.

Electricians. Yes, literally. There are not enough skilled tradespeople.

John Collison draws a sharp analogy: it is like the Strait of Hormuz. You can set whatever price you want for oil, but if 20 million barrels a day disappear from the system, someone is not getting their oil. Memory is similar — someone has to not get the memory they want.

Pichai agrees this creates something like an oligopoly. If every major player gets roughly pro rata access to global compute capacity, there is a ceiling on how far ahead anyone can pull. Though he notes a strange wrinkle: you train a model for months in a massive data center, and the output is a flat file that fits on a USB stick. That is a very different kind of asset than a SpaceX rocket.

Security as a Hidden Constraint

Tucked into the infrastructure discussion is a genuinely alarming aside. Pichai says these models are “definitely really going to break pretty much all software out there. Maybe already, we don’t know.”

He clarifies he means regular software, large platforms, zero-days — not that SSH is suddenly broken. Someone told him the black market price of zero-days is dropping because AI is increasing the supply. He was “not at all surprised.”

“We are going to need more coordination, which is not happening today. There will be a moment of… It could be a sharp moment.”

Google’s Hidden Gems

Beyond the core AI business, Pichai runs through the portfolio:

Data centers in space. A small team, early stages, framed as today’s version of what Waymo was in 2010. The logic: if you take a 20-year view, where do you physically put all these data centers?

Quantum computing. Deeply committed. Pichai thinks its biggest impact will be simulating nature — weather, reality, molecular processes. We still do not fully understand the Haber process for making fertilizer. Classical computing may not be enough.

Robotics. Google was too early the first time. AI was the missing ingredient. Now the Gemini Robotics models are state-of-the-art on spatial reasoning. They are partnering with Boston Dynamics and Agility, in what Pichai calls an “ironic” return.

Wing drone delivery. Scaling to where 40 million Americans will have access to a Wing delivery service in a “reasonable time period” — explicitly not years out.

Isomorphic Labs. Using AI models across every step of drug discovery, not just molecular design. Pichai calls it “definitely the smartest approach I’ve seen.”

Capital Allocation: How Google Decides

John Collison asks the finance nerd’s dream question: how do you compare giving YouTube more money for its recommender versus giving Waymo more money to scale versus investing in a speculative AI approach? The projects are wildly different in nature and payoff shape.

Pichai’s answer is surprisingly personal. He spends a dedicated hour every week looking at compute allocation at a granular level — which projects, which teams, how many compute units. The scarce resource is compute, and the CEO’s job is making sure it goes to the right places.

For long-term bets, Google tries to get in early when the funding amounts are small, then stays committed. The key discipline: evaluate at the technology level, not the business level. With Waymo, they tracked the safety and reliability curve of the “Waymo driver” software. As long as the underlying tech was improving against its goals, they kept investing — even when the rest of the world got pessimistic.

“How do you decide that we’re going to cut Loon but keep Waymo?”

The answer: Loon’s underlying technology was not hitting its milestones. Waymo’s was.

He also reveals that compute allocation has become the new capital allocation. It is not just headcount budgets anymore. Every project gets a headcount budget and a compute budget, and the compute budget is what Pichai personally sweats over.

The Intelligence Overhang

Collison offers a sharp diagnosis of why AI adoption lags AI capability in 2026: prompting skill takes time to develop (both general and company-specific), AI-generated code is hard to collaborate on because the blast radius of changes is so large, data access and permissions need to be rebuilt from scratch, and role definitions (engineer, PM, designer) are blurring.

Pichai says Google faces these same problems internally. The Gemini and Antigravity teams are working through them, using the products internally and running into the same barriers. Identity and access controls are the hardest part. Security requirements add another layer.

He expects 2027 to be a big inflection point for non-engineering processes adopting AI, including things like fully agentic business forecasting.

The Stateful AI Frontier

The conversation closes on consumer AI. Collison notes that current AI apps lack persistence — you cannot tell them “round up my news every morning.” Pichai agrees this is the agentic future and says Google is working on it.

The vision: consumer interfaces will have full coding models underneath, with the ability to persist and run tasks in the cloud. Right now maybe 0.1% of the world is living this future, building things for themselves. Bringing it to mass adoption is the frontier.

Claude’s Take

This is a well-run interview by people who actually understand the business, which makes it more revealing than the average CEO fireside chat. Collison’s capital allocation questions are genuinely sharp, and Pichai gives more concrete answers than he probably planned to.

The Transformer narrative correction is worth taking seriously. Pichai’s version — that Google had the research, had the product prototype, but had a higher quality bar and a more toxic early model — is plausible and internally consistent. The claim that coding was the use case where the capability jump was most visible (and the one Google underweighted) rings true. That said, “we had it first” is the kind of story every incumbent tells. The fact remains that OpenAI shipped first and Google scrambled. Both things can be true.

The physical bottleneck discussion is the most valuable part. Memory, wafer starts, permitting, electricians — these are the actual constraints on AI scaling right now, and hearing a CEO who is spending $180 billion confirm them is useful signal. The Strait of Hormuz analogy for memory supply is apt and grim. If you are trying to understand the AI landscape for the next 18 months, this is the section to pay attention to.

The security comment deserves more attention than it got. Pichai essentially said, with the calm affect of someone discussing weather, that AI models will break most existing software and that we may not know it is already happening. The falling black market price of zero-days as a supply-side indicator is the kind of detail that should keep CISOs up at night.

On the “AGI-pilled” question, Pichai’s response is deft but somewhat evasive. Pointing out that Demis, Jeff Dean, Ilya, and Dario were all at Google is a strong retort. But the question is not whether Google understood AI early — clearly it did — but whether its institutional culture moves with the same urgency as a lab that genuinely believes superintelligence is imminent. A company built to serve two billion users with a “search quality bias” may simply optimize differently than one that thinks it is building God. Whether that is a bug or a feature depends on your timeline.

The data-centers-in-space thing got mentioned twice and both times Pichai compared it to Waymo in 2010. He is clearly seeding this as Google’s next long-term bet. Whether it is visionary or a PowerPoint that escaped the building is impossible to say at this stage. It is a small team with a small budget aiming at a first milestone — which is exactly the right way to run such a project, and also the description of most things that never work out.