All the Thoughts of Google CEO Sundar Pichai: Learning Management Decisions in the AI Era, from the Secret Story of Transformer's Birth to Capital Allocation and Organizational Transformation
Sundar Pichai has been CEO of Google for 10 years. Alphabet (Google's parent company) plans to invest $175–185 billion (approximately 26–28 trillion yen) in capital expenditures in 2026, clearly signaling its stance as a frontrunner in the AI race. In conversations with Stripe CEO Patrick Collison and Elad Gil, Pichai spoke candidly about the origins of Transformer, how they were beaten to the punch by ChatGPT, the future of search, the internal logic of capital allocation, and organizational transformation. This article summarizes the key points and provides an analysis for business professionals and investors.
1. The Birth of Transformer—It Wasn't Just About "Research"
1-1. Technology Born from Translation Improvements and TPUs
Pichai disputes the common belief that "Transformer was born from pure research."
“Transformer was created to solve specific product needs. The team was thinking about how to improve translation. In the case of TPUs, we were trying to solve the inference cost problem of not having enough chips to provide speech recognition to 2 billion people.”
Transformer was immediately deployed into Google Search, contributing to improvements in language understanding quality as BERT and MUM. According to Pichai, the biggest factor that allowed Google Search to pull away from competitors during this period was precisely these Transformer-based technical improvements.
“People underestimate the jump in search quality that BERT and MUM brought. We measure search quality very strictly, and it was precisely because of these technologies that search pulled away from everything else during that time.”
1-2. "The Product Version of ChatGPT Was Already In-House"
Pichai goes even further with his comments.
“We had conceived the exact same product as ChatGPT in-house. That was LaMDA. You remember the commotion when an internal engineer felt LaMDA was 'sentient.' What he was talking about was something like an early version of ChatGPT.”
At Google I/O in 2022, they released a limited version of LaMDA as "AI Test Kitchen," but the internal version had not yet applied RLHF (Reinforcement Learning from Human Feedback) and produced a lot of toxic output, so it was not in a state to be released as is. In addition, high standards for search quality raised the hurdle for shipping.
“As a company with a search quality bias, our standards for acceptable product quality were probably high. We were in the middle of figuring out how to release it.”
1-3. "Surprise" is Inevitable in Consumer Internet
Pichai says we shouldn't view this experience too dramatically.
“In consumer internet, three people build a prototype and put something out into the world for millions. We had Google Video, but YouTube came along. We had Facebook, but Instagram came along. No one builds and ships an iPhone in a garage, but that is the world of consumer internet. You have to be aware of that and internalize it.”
However, he admits there were signals they missed.
“On the coding side, we probably should have seen a bigger leap. The progress from GPT-2 to 3, and then to 4, might have been much more pronounced if we had been using it for coding than if it were just language. There is a possibility that there was a signal there that we missed.”
2. Obsession with Speed—Latency Budgets in Milliseconds
2-1. Improving Search Latency by 30% in 5 Years
Google has historically used "speed" as a differentiator. Query processing time displayed in search results, high-speed Gmail searches, Chrome rendering speed—it is a philosophy common to all of them.
Pichai specifically revealed how the current search team operates.
“Each sub-team has a latency budget measured in milliseconds. If 3 milliseconds are saved, 1.5 milliseconds are returned to the team's own budget, and the remaining 1.5 milliseconds are returned to the user. There is a strict review process for the budget.”
Despite the continuous increase in features, search latency has improved by 30% over the past five years.
2-2. The Capability-to-Speed Trade-off between Flash and Pro Models
Gemini also adheres to this philosophy. The Flash model offers about 90% of the capabilities of the Pro model while being delivered much faster and more efficiently. Vertical integration with TPUs supports this.
“We are constantly thinking about the balance between the capability and speed frontiers. We want to keep latency low, but the capability frontier is also advancing simultaneously. Finding that balance is the difficult part.”
As another form of "speed," Pichai also mentioned the speed of shipping and iteration. It is the recognition that both product latency and the speed of the organization's release cycle are important.
3. The Future of Search: Evolution into an Agent Manager
3-1. "An Expansive Moment, Not a Zero-Sum Game"
To the question, "Is search dying?" Pichai's answer is clear.
“We are far from a zero-sum game right now. The value of what people can do is also rising sharply along a curve. YouTube is doing well even after TikTok and Instagram. As long as we continue to innovate and the product continues to evolve, it will not be a zero-sum game.”
Search and Gemini will overlap in some areas and diverge significantly in others. Pichai stated that it is important to have both and to embrace both.
3-2. Search Will Become an Agent Manager
Regarding search 10 years from now, Pichai envisions it as follows.
“Search will become an agent manager. Many agents will move, and tasks will be completed. Deep research queries are already being performed in AI mode, and it no longer fits the traditional definition of "ranking results for a single-line query." It will also begin to execute tasks that run asynchronously for long periods.”
However, he does not forget to warn against thinking too far into the future.
“Models change dramatically in a year. Precisely because the curve is so steep, just thinking about a year ahead is exciting. Being on the curve itself is exciting.”
4. Google's AI Revival: The Rebound from a $150 Stock Price
4-1. What Investors Misjudged
From the spring to summer of 2025, market sentiment toward Google was extremely negative. "Search is over," "The core business model is under attack"—the stock price fell to around $150.
Mr. Pichai reflects on this period.
"We were behind on frontier LLMs, but we had all the capabilities internally. The TPU had reached its seventh generation, and at Google I/O in 2016, we had already announced the construction of AI data centers. Looking at the full stack, we had the research teams, the infrastructure teams, and all the platforms in place.
"Suddenly, I felt that one common technology could accelerate everything: Search, YouTube, Cloud, and Waymo. It was a highly leveraged way to proceed. I never thought of this as a zero-sum moment at all. Everything scales by 10x, and there is room for other companies as well.
4-2. Gemini 2.5 and the Advantage of Multimodality
The shift in external perception likely occurred around the time of Gemini 2.5.
"The Gemini model was designed to be multimodal from day one. The upfront fixed costs were high, but its strengths began to show. This is a credit to the Google DeepMind team.
However, Mr. Pichai also acknowledges that the competition at the frontier is extremely dynamic.
"Even if things go well this month, we might fall behind in some areas next month. Two or three labs are pushing each other hard. The situation will likely change again in a few months.
4-3. Rebuttal to the Claim of Not Being 'AGI-pilled'
Regarding comments from researchers at other labs that 'Google lacks the belief that AGI is imminent,' Mr. Pichai offered a rebuttal.
"At one point, Demis Hassabis, Jeff Dean, Ilya Sutskever, and Dario Amodei were all at Google. For anyone who has watched the last 20 years, the claim that Google does not understand this technology makes no sense.
The fact that CapEx has been expanded from $30 billion to approximately $180 billion speaks for itself regarding our view on the curve. Mr. Pichai's view is that the languages used may differ, but there is no difference in the fundamental outlook.
5. The CEO's 'AGI Moments' and Dogfooding
5-1. The First AGI Experience in 2012
The first time Mr. Pichai 'felt AGI' was in 2012, when Jeff Dean demonstrated an early version of Google Brain. It was the moment the neural network recognized a cat. Since then, milestones have continued, such as the self-driving car in the DARPA Challenge and Demis Hassabis's demonstration of an AI model's 'imagination.'
He says that currently, he feels 'AGI' the most when coding.
"When I hand over a complex task and see it completed in the world of agent managers without ever opening an IDE, it feels like magic. I recently did a hobby project, and after everything was working, I had to ask, 'Which language did you use?'
5-2. Using Gemini Live for 30 Minutes at the Gym
Mr. Pichai consciously maintains contact with his products.
"I am constantly dogfooding internal versions. Two weeks ago, while stretching at the gym, I talked about a single topic with Gemini Live for 30 minutes. There are parts that work well and parts that are frustrating, but I learn a lot.
He also utilizes AI agents. Using an internal version called "Antigravity" (internal name "Jet Ski"), he inputs queries like, "Tell me the five worst and five best things people are saying about this launch." Previously, it took more time to get a sense of things, but he says AI has significantly improved efficiency.
6. CapEx Bottlenecks—Why $180 Billion Is Not Enough
6-1. List of Constraints
Pichai stated that even if Google wanted to invest $400 billion, it would be physically impossible.
Wafer Starts: The most fundamental constraint
Memory: "One of the most important components right now. Major memory manufacturers are not expected to significantly increase production in the short term."
Power and Energy: More solvable, but takes time
Permitting and Regulatory Environment: Even if there is land in Texas or Nevada, construction permits cannot keep up
Shortage of Electrical Engineers: Lack of manpower to build physical infrastructure
" I am in awe when I see how fast China can build things. The U.S. needs to switch to a mentality of building 10 times faster in the physical world."
6-2. Efficiency and Creativity Born from Constraints
Pichai also mentioned the positive aspects of bottlenecks.
" Constraints stimulate creativity. It forces compaction cycles and makes things more efficient. We will make these things 30 times more efficient. That is happening simultaneously."
6-3. The "Musical Chairs" of Computing Resources and Oligopoly
As model self-improvement (its own data labeling, its own code generation) progresses, a " game of musical chairs where whoever has the computing resources now wins" is unfolding.
The host asked, "If everyone only gets a proportional share of industry capacity, won't there be a ceiling where no one can get significantly ahead?" Pichai acknowledged that it is a " reasonable framework regarding the inference side," but pointed out the unique nature of AI models.
" You run a data center for months, and the output is a flat file that fits on a USB drive. It's not like a SpaceX rocket. This unique attribute casts doubt on that framework."
6-4. Security as a "Hidden Constraint"
Pichai also mentioned a constraint that is often overlooked.
" These models have the potential to destroy almost all existing software. The black market price for zero-day vulnerabilities is dropping. This is due to the increased supply caused by AI."
In the current situation where coordination on the security front is insufficient, an unexpected 'system shock' could occur. "It could be a sharp moment. It's not something you can just wish away," Pichai warned.
7. Google's "Hidden Gems"—From Space Data Centers to Robotics
Pichai introduced several projects that are being nurtured as long-term investments.
7-1. Space Data Centers
"When you look 20 years ahead, where will you put most of your data centers? It's a really difficult problem. Space data centers are in the very early stages, moving toward their first milestones with a small team and a small budget. It's just like Waymo in 2010."
7-2. Robotics and Waymo
The Google DeepMind team is working on robotics in earnest, and the Gemini Robotics model is achieving SOTA (state-of-the-art) in spatial reasoning. Partnerships with Boston Dynamics and Agility Robotics are also underway.
Regarding Waymo, while acknowledging that the transition to end-to-end deep learning was a major breakthrough, Pichai emphasizes the accumulation of over 15 years of system integration.
"Waymo is a highly integrated system. Like TSMC or SpaceX, the complexity of system integration requires 'craft' that takes time."
7-3. Wing: 40 Million People Within Reach
The drone delivery service Wing is also expanding, and Pichai said, "Within a reasonable timeframe, 40 million Americans will have access to Wing's delivery service. It's not years away."
7-4. Quantum Computing
"Nature is inherently quantum, and to simulate it, you need quantum systems. Weather simulation, reality simulation—quantum computing will have an advantage there."
Citing the historical pattern that once technology reaches a certain scale, creativity on top of it finds applications, he stated, "Mobile phones plus GPS created Uber. No one who was building phones predicted that result. Quantum is the same."
7-5. Isomorphic Labs (AI for Drug Discovery)
Regarding Isomorphic Labs, which follows in the footsteps of AlphaFold, he highly praised it, saying, "A comprehensive approach that increases the probability of success at every step, not just in molecular design but all the way to Phase 3 trials, is the smartest strategy I've seen."
8. Thinking on Capital Allocation—"Bet Small Early, Judge by the Technology Curve"
8-1. CEO Spends 1 Hour a Week on TPU Allocation
The capital allocation issue Pichai is currently feeling most acutely is TPU allocation.
"I spend at least an hour a week thinking about TPU allocation at a fairly detailed level. I track and evaluate compute unit usage by project and by team. Right now, compute resources are the scarce resource, and it is extremely important to ensure they are being used for the most valuable efforts."
In the past, the majority of the R&D budget was 'people walking around the building'—that is, personnel costs. Now, in addition to headcount budgets, TPU budgets have become a critical component of project resourcing.
8-2. Bet small early, commit for the long term
A characteristic of Google's capital allocation is making decisions at a very early stage of the cycle.
'In the beginning, the capital required is small. However, you commit for the long term and keep checking if there is progress at a deep level. For quantum computing, it's about when the error correction rate of logical qubits will reach its target. For Waymo, it's about how the safety and reliability curves are trending.'
'You think intuitively about the TAM and option value 5 to 10 years out, and make decisions assuming significant growth. TPU investment was exactly that.'
8-3. Contrarian investment in Waymo
The increase in investment in Waymo happened two or three years ago, when the entire industry had become pessimistic.
'We increased our investment when other companies were pulling out. It is a magical experience. I commute by Waymo as much as I can every day now.'
Regarding the criteria that distinguished the fates of Loon and Waymo, Pichai stated the following:
'We follow the progress curve of the underlying technology and evaluate performance against goals. Even if there are periods where things don't move forward, if you are confident in the quality of the team, you believe you can break through. The deeper you can evaluate the technology, the better decisions you can make.'
8-4. 'Was Google under-leveraged?'
Pichai answered frankly to the question, 'Should Google, which has historically maintained a strong net cash position, have utilized more leverage?'
'If Waymo had reached this stage sooner, we would have deployed capital sooner. However, at the time, we were approaching it with safety as the top priority, and the maturity level had not reached the point of expanding investment.'
He positioned the investments in Stripe, SpaceX, and Anthropic as part of 'being a good steward of capital.' He noted that more capital deployment opportunities are emerging in the AI era, and they are taking advantage of them.
9. Organizational penetration of AI—2027 as an 'inflection point'
9-1. 'Intelligence overhang'
Patrick Collison (Stripe CEO) described the gap between AI capabilities and actual organizational utilization as an 'intelligence overhang' and organized the causes.
It takes time to master prompt engineering (general prompts + company-specific prompts)
Collaborating on AI-generated code is difficult (the blast radius of changes is large, and there is a lot of rewriting)
Rebuilding data access and permission management is necessary
Reviewing Role Definitions (Redefining the boundaries between Eng/PM/Design)
9-2. Penetration within Google: From "Jet Ski" to the Search Team
Within Google, the agent manager tool Antigravity (internal name "Jet Ski") is being deployed to leading teams. Pichai stated:
"The curve is shifting dramatically this year. Google DeepMind and some SWE (Software Engineering) groups have truly changed their workflows. We rolled it out to the search team last week."
While acknowledging that change management in a large organization is the biggest challenge for adoption, he expressed his outlook: "We are paying these fixed costs now. Once solved, we will be able to provide it externally in a more robust form. You should be able to see a jump in what people can do."
9-3. 2027: Fully Agentic Performance Forecasting
To Collison's bold question, "How many quarters until Google's first fully agentic performance forecast?", Pichai replied:
"I expect 2027 to be a significant inflection point in several areas. We will check with traditional methods for a while, but a crossover will occur."
Startups naturally take the lead with AI-native team structures, while large companies like Google must promote retraining and transformation. Pichai frankly acknowledged this as a startup advantage.
10. Consumer-facing "Stateful AI" and the Future of Products
10-1. The Arrival of Persistent Agentic AI
When Collison asked about stateful AI (resident AI) like OpenClaw, Pichai responded affirmatively.
"We want to provide users with persistent, long-running tasks in a reliable and secure way. We need to think through identity and access control, but that is the agentic future. Delivering that to consumers is an exciting frontier we are working on."
Coding models, proper harnesses, skills, and persistent execution in the cloud and locally—these primitives are coming together. Pichai stated, "Perhaps 0.1% of the world is living in this future right now. Delivering that to the masses is the next frontier."
10-2. GCP × MCP: AI Becomes the Navigator
Collison highly praised Google Cloud's support for MCP (Model Context Protocol). For GCP, which was difficult to navigate due to its feature richness, AI can now read all API documentation, and one can simply give instructions like "add this feature."
"The wider the product surface, the greater the benefit of the AI orchestration layer. Stripe has the same problem, but it should have a huge effect on GCP."
Pichai said that while "it can still be better," he feels positive about this direction.