Sundar Pichai's candid take on AI, Search latency, and Google's $180bn bet
Google CEO Sundar Pichai details millisecond latency budgets, a $180bn CapEx plan, and why Search will become an agent manager within a decade on a podcast.
Search
Sundar Pichai has now spent more than a decade as CEO of Google and Alphabet. Today he appeared in a conversation published on YouTube with John Collison, co-founder of Stripe, and investor Elad Gil on the Cheeky Pint podcast to discuss Google's history with artificial intelligence, the scale of the company's infrastructure commitments, and where Search is heading as agentic systems become more central to the product. The episode had gathered over 17,000 views by the time of this writing and covers ground that matters directly to anyone working in digital marketing.
The video was recorded in what appears to be a pub setting - a contrast to the boardroom or conference stage format usually associated with Pichai's public appearances. That relaxed format produced candid admissions and specific technical detail that quarterly earnings calls rarely offer.
Transformers, LaMDA, and the ChatGPT question
One of the more direct sections of the interview addressed a question the industry has debated since late 2022: why did Google invent the Transformer architecture but not ship the consumer product that made it famous?
According to Pichai, the framing is "a bit misunderstood." He explained that Transformers were developed at Google specifically to solve product problems - translation and speech recognition at scale - not as a pure research exercise. "Transformers were specifically - it was from our research teams, but they were guided by solving product problems," he said. The architecture was applied immediately to Search through BERT and later MUM, which Pichai credited with producing some of the largest quality improvements in Search during that period. "Some of the biggest jumps in search quality in that period where search went ahead of everyone else was because of BERT and MUM," he said.
What Google had internally, though, was an early version of what became the ChatGPT product format. The company built something called LaMDA. Pichai noted that Google "even conceived the product, which is ChatGPT - it was LaMDA." A public-facing version appeared at Google I/O 2022 as AI Test Kitchen. But the internal version Pichai saw was, in his words, "a lot more toxic at a level." The company also maintained what he described as a higher quality bar shaped by years of rigorous Search measurement. "We had a higher bar, maybe, for what we thought was an acceptable product quality to go out," he said.
OpenAI's launch timing also played a role. According to Pichai, ChatGPT launched "the week of Thanksgiving" in 2022 without much fanfare - "it was a little bit of a buried launch." The competitive signal that really mattered, he suggested, came from the coding side. The performance jump between GPT versions was more visible to engineers using models for code than to those using them only for language tasks.
Latency as a product value, measured in milliseconds
The interview's most technically specific section concerned Search latency - an area with direct relevance to how advertisers and publishers experience the platform. Pichai described an internal system of latency budgets operating at the sub-team level within Search.
"They now have for sub-teams, latency budgets in the milliseconds," he explained. The structure works as an incentive mechanism: if a team shaves three milliseconds off an existing process, they earn 1.5 milliseconds of latency budget to spend on new features. Budgets range from 10 to 30 milliseconds depending on the type of work involved, and teams face rigorous reviews against those allocations.
On the same subject, Nick Fox, Google's SVP of Knowledge & Information, posted on LinkedIn today citing the same Cheeky Pint conversation. According to Fox, Google has "reduced latency significantly (by over 35%) in the last five years" while adding more AI capabilities. Pichai's figure in the interview itself was 30% over five years. Fox framed speed as "a reflection of both a product's technical health but also deep respect for our users' time."
The practical threshold matters here: humans perceive latency in the low hundreds of milliseconds. The millisecond-level budgets Google manages internally sit well below conscious perception, but they accumulate into the product-level experience billions of users feel across billions of queries.
Flash models and the capability-speed tradeoff
Adding AI capabilities to Search without degrading its speed is an unsolved engineering problem, not a default outcome. Pichai addressed the tradeoff directly, describing how Gemini Flash models are positioned to manage it. According to him, Flash models operate at "90% the capability of the pro models" while being substantially faster and more efficient to serve. Google's vertical integration between its custom TPUs and its model development makes this tradeoff manageable in ways harder for providers relying on third-party hardware.
This is the logic behind Alphabet's announced 2026 capital expenditure of between $175 billion and $185 billion, the majority directed at AI infrastructure. Pichai confirmed the range in the conversation - "We have said it'll be between 175 and 185" - and noted that Google has scaled its CapEx from $30 billion to approximately $180 billion. He framed this not as speculative AGI investment but as a response to observable demand. "We are supply-constrained. We are seeing the demand across all the surface areas," he said.
The CapEx commitment reflects investments deeper in the stack. Google is currently on its seventh generation of TPUs. Pichai traced the strategy back to Google I/O 2016, when the company first publicly announced TPUs and described its plans to build AI data centers. "We were thinking about... the company was operating in an AI-first way. We had deeply internalized this shift," he said.
Memory, wafer starts, and the physical limits of AI scaling
The conversation turned unusually specific on supply-side constraints shaping 2026 - a level of detail that matters for anyone modelling how quickly AI capabilities can actually be deployed at scale.
Pichai identified memory as one of the most critical constrained components in the near term. "There is no way that the leading memory companies are going to dramatically improve their capacity" quickly, he said. The constraint is expected to ease as supply responds to price signals over time. Wafer starts - the number of semiconductor wafers entering fabrication - represent what he called a deeper constraint, a "fundamental" ground truth that spending alone cannot resolve in the short term.
Power and permitting were identified as more solvable, though not trivial. Even in pro-growth states like Texas, Nevada, and Montana, permit velocity is a real limitation. Pichai expressed concern about construction pace relative to China and called it a strategic issue: "I really think we need to learn to build things much faster." Data center moratoriums in some jurisdictions compound the problem.
The dynamic creates what Pichai described as a musical chairs situation for compute. "Who has the compute right now and how much can you actually scale relative to overall industry capacity?" he said, acknowledging that this places a ceiling on how far ahead any single lab can pull relative to competitors. Everyone is working within roughly the same constraint envelope.
Security is a less-discussed dimension of the same picture. Pichai noted that AI models are "definitely really going to break pretty much all software out there" and that the black market price of zero-days is falling because AI is increasing exploit supply. "Somebody was telling me the black market price of zero-days is dropping because the supply is growing due to AI, which I thought was a really interesting market metric," he said. He predicted a moment of sharp coordination requirements that are "not happening today."
The future of Search: agentic, not terminal
The section most directly relevant to the marketing community concerned Search's trajectory. Pichai rejected the zero-sum framing that dominated coverage around spring and summer 2025, when Alphabet was trading near $150 a share and the prevailing view, as he put it, was that "Search is cooked."
His description of where Search is heading is closer to expansion than replacement. "A lot of what are just information-seeking queries will be agentic in Search," he said. Users will complete tasks rather than submit queries. Many threads will run simultaneously in the background. The search box itself may not survive as the primary interface in ten years as device form factors and input methods change - but the underlying function connecting people to information and actions will persist and expand.
Gemini 2.5 was identified as the model where external observers began recognizing Google's frontier capabilities, particularly around multimodality. Pichai credited Google DeepMind teams and explained that the Gemini architecture was designed to be multimodal from day one. "We paid a bit more of a fixed cost upfront, but we designed the Gemini models to be very multimodal from day one," he said. That upfront cost is now generating returns visible in competitive benchmarks.
Capital allocation in a TPU-constrained world
The capital allocation section was unusually candid. Pichai described spending "a dedicated hour a week" reviewing compute allocation at a project and team level. "I will know by projects and by teams, the compute units they are using," he said. This is a significant operational detail: the CEO of one of the world's largest companies measures strategic priority in compute units rather than headcount.
The constraint has made TPU allocation the practical expression of strategy in a way that headcount planning once was. Waymo, for instance, competes with AI model training for the same scarce resources. Pichai described an internal tool called "Antigravity" - an agent manager platform he uses to query user sentiment about product launches without manually reviewing feedback threads.
On Waymo, Pichai said he now rides it to work every day when possible. He expressed that he would have invested capital in the project faster in hindsight. The breakthrough - moving from hand-mapped heuristics to end-to-end deep learning - arrived with the broader Transformer wave, and the years of prior continuous investment allowed Google to capitalize on it rather than starting from scratch.
Gemma 4, open source, and the model as flat file
One of the more striking passages involved Pichai's description of what a trained model actually is as a physical artifact. He had just shipped Gemma 4, an open-source model based on the Gemini 3 architecture, described as competitive outside China. "You're talking about a set of weights which can fit on a USB stick," he said. "I'm always shocked that you run a data center for months, and then your output is a flat file. It's like having a Word doc or something, and that's your model."
Gemma 4's release comes as Google's AI model stack has reshaped Google Marketing Platform, with the same Gemini architecture powering advertising tools presented at NewFront 2026 in March.
Long-term bets: space, robotics, quantum, drug discovery
The final section covered projects at earlier stages. Pichai confirmed that Google has a small team working on data centers in space - "literally a few people with a small budget to go to the first milestone" - positioned as a twenty-year infrastructure problem. The constraint-inspires-creativity logic applies: with physical land, power, and permitting all limited on Earth, space becomes a long-range planning option.
On robotics, Pichai said the Gemini Robotics models are state-of-the-art on spatial reasoning, and the company is partnering with Boston Dynamics, Agile, and others. He acknowledged that previous robotics efforts came too early. "It turned out AI was the missing ingredient for a lot of ideas maybe 15 years or 10 years ago," he said.
Quantum computing was addressed more speculatively. Pichai's instinct is that quantum will have an advantage in simulating physical phenomena - weather, molecular behavior, the Haber process for fertilizer - but he acknowledged that classical computing and deep learning may narrow that gap in unexpected ways. Isomorphic Labs, focused on drug discovery using AI models across the full pipeline beyond just molecular design, was described as "really exciting."
Why this matters for marketing professionals
Several of the technical disclosures in this conversation carry operational implications.
The latency budget system explains why Search has maintained speed even as AI features have been added. For advertisers, faster Search translates directly into auction dynamics and user behavior. The broader shift toward agentic search is transforming how Google functions as a platform for content and commerce.
The agentic Search trajectory described by Pichai aligns with what Google's own teams have been demonstrating in product releases throughout 2025 and early 2026, including AI Mode, Deep Search, and automated transaction completion. Google Cloud has projected the agentic AI market could reach $1 trillion by 2040, and the internal workflow shifts Pichai described suggest that transition is already underway inside Google itself.
The memory and wafer supply constraints, meanwhile, create a situation where not all demand for AI capability can be met in 2026 and 2027. That compression affects everyone buying AI services, including those using Google Cloud and Google Ads products that depend on the same infrastructure. Pichai's own words make clear that supply cannot grow as fast as demand in the near term, even at $180 billion in annual CapEx.
Pichai expects 2027 to be a significant inflection point for non-engineering functions adopting agentic workflows. He named financial forecasting as a specific example, suggesting that by 2027, Google's forecasting processes could involve AI generating the initial figures with humans reviewing rather than producing them. For marketing teams, this points toward a similar transition in campaign planning, reporting, and budget allocation - a shift that Google's advertising tools are already anticipating with products like Ads Advisor in DV360.
Timeline
- 2012: Jeff Dean demonstrates earliest Google Brain results, neural networks recognizing a cat; Pichai describes it as his first "feeling the AGI moment"
- 2016: Google announces TPUs publicly at Google I/O and begins building AI data centers; company declares itself AI-first
- 2017: BERT applied to Google Search, producing the largest quality improvements of that period; first-generation TPUs deployed at scale
- 2022 (May): Google launches AI Test Kitchen at Google I/O, based on LaMDA
- 2022 (November): OpenAI launches ChatGPT during Thanksgiving week; described by Pichai as "a little bit of a buried launch"
- 2024 (May): Google launches AI Overviews in Search, now reaching 1.5 billion users in 150+ countries
- 2025 (Spring/Summer): Gemini 2.5 identified as the turning point where Google reached the frontier on multimodality; Alphabet stock near $150
- 2025 (March): Google launches AI Mode as experimental feature in Search Labs - covered by PPC Land
- 2025 (July): Google introduces Gemini 2.5 Pro and Deep Search in AI Mode - covered by PPC Land
- 2025 (November): Gemini 3 launched; Gemma 4 based on Gemini 3 architecture shipped shortly after; Google AI Mode begins experimenting with agentic features - covered by PPC Land
- 2026 (January 11): Google announces Universal Commerce Protocol with Shopify, Walmart, Target, and others - covered by PPC Land
- 2026 (January 27): Gemini 3 made default model for AI Overviews globally - covered by PPC Land
- 2026 (February 4): Alphabet reports Q4 2025 revenues of $113.8 billion; 2026 CapEx guidance set at $175-185 billion - covered by PPC Land
- 2026 (March 23): Google NewFront 2026 introduces Gemini-powered advertising tools across Google Marketing Platform - covered by PPC Land
- 2026 (April 7): Pichai sits down with John Collison and Elad Gil on the Cheeky Pint podcast; Nick Fox posts LinkedIn summary citing over 35% Search latency improvement over five years
Summary
Who: Sundar Pichai, CEO of Google and Alphabet, in conversation with John Collison (Stripe co-founder) and Elad Gil (investor) on the Cheeky Pint podcast.
What: A wide-ranging conversation covering Google's AI history from Transformers through LaMDA to Gemini, Search latency management through millisecond-level internal budgets, the $175-185 billion 2026 CapEx plan, near-term supply constraints in memory and wafer capacity, the agentic future of Search, and long-term bets including space-based data centers, robotics, quantum computing, and drug discovery.
When: Published today, April 7, 2026, with the conversation recorded shortly before publication. Key figures referenced span from 2012 through a projected 2027 inflection point for agentic enterprise workflows.
Where: Published on YouTube by Stripe, the payments company co-founded by John Collison, and amplified on LinkedIn by Nick Fox, Google's SVP of Knowledge & Information. The podcast is called Cheeky Pint.
Why: The interview matters for the marketing community because it clarifies the technical and supply-side logic behind decisions that directly affect how advertising and search products operate - from the millisecond latency budgets that shape how fast Search loads to the memory constraints limiting how quickly new AI capabilities can be deployed at scale. Pichai's framing of Search becoming an "agent manager" over the next decade defines the competitive context in which advertising on Google will evolve, and his 2027 projection for agentic enterprise workflows sets a timeline for when non-engineering marketing functions may face the same automation pressure that engineering teams are already experiencing.