Alphabet CEO Sundar Pichai recently marked his tenth year leading the company by participating in a joint interview with Stripe co-founder John Collison and tech angel investor Elad Gil. Pichai reflected on Alphabet's journey in AI, from initially trailing to achieving a leadership position. He addressed the perception that Alphabet missed an opportunity with the Transformer architecture, which originated within the company but was later leveraged by OpenAI to launch ChatGPT, a product that disrupted the search industry. Pichai clarified that this view is somewhat misunderstood, explaining that Transformer was developed to solve the practical problem of improving translation quality, not as a theoretical exercise. The delay in public release was partly due to Alphabet's higher standards for search quality and concerns that early internal versions were too unreliable to launch.
Regarding the current AI competition, Pichai believes the market is far from a zero-sum game, describing the value growth curve as "extremely steep." He revealed that he personally approves computing resource allocations for at least one hour each week, calling it "the most critical task at present." Pichai views Alphabet's full-stack vertical integration—spanning from its seventh-generation TPUs to models and applications—as a core advantage. He disclosed that capital expenditures are projected to reach $175 billion to $185 billion by 2026. On resource constraints, he identified wafer production capacity as the "fundamental bottleneck," warning that 2026 will be a "year of supply tightness." He emphasized that the U.S. must learn to "build physical infrastructure ten times faster." Pichai also confirmed that Alphabet is exploring space-based data centers, comparing the initiative to "Waymo in 2010"—a seemingly distant concept that has started with a small team and budget.
Pichai firmly believes that search functionality will not disappear but will evolve into an "agent manager," where users issue commands and AI agents complete tasks. He even predicted that by 2027, Alphabet's internal business forecasting will be fully automated by AI, eliminating the need for human intervention.
The following is an edited version of the interview:
**01 “We Were Not Slow; Our Standards Were High”** Q: Many point out that Transformer was invented at Alphabet but became the foundation for ChatGPT. How do you look back on this? Pichai: There's a misunderstanding. Transformer didn't appear out of nowhere; it was driven by our need to improve translation. The same applies to TPUs: we had speech recognition technology, but to serve two billion users, we first had to solve inference efficiency with existing chips. Q: So Transformer was product-focused from the start? Pichai: Yes, our research team aimed to solve real-world problems. We immediately applied Transformer to search, followed by BERT and MUM, leading to significant quality leaps. We also developed something similar to LaMDA internally but weren't first to market. Q: In other words, you conducted the research and saw returns but didn't capitalize on it fully. Pichai: It's more than that. We had explored a ChatGPT-like product with LaMDA internally. The engineer who thought LaMDA was conscious was actually working with an early version. We launched AI Test Kitchen at I/O 2022, powered by LaMDA, but imposed restrictions because it lacked RLHF and was too unreliable for public release. Alphabet's high search quality standards also meant a higher bar for product launches. When OpenAI released ChatGPT, its partnership with Microsoft was just finalized. ChatGPT's success wasn't inevitable; OpenAI benefitted from spotting an opportunity in programming via GitHub, a signal we might have missed. Progress in coding from GPT-2 to GPT-4 was more dramatic than in pure language tasks. Multiple factors converged to create the outcome. Q: I recall ChatGPT's launch was low-key during Thanksgiving week; no one expected its impact. Pichai: That's typical in consumer internet—unexpected disruptions happen. We had Google Video Search before YouTube; Facebook acquired Instagram. You don't sense dramatic disruption initially; small teams prototype constantly. The key is to recognize this dynamic and embed it into the organization's DNA.
**02 Search Is Evolving, Not Dying** Q: Alphabet is known for speed, from early search response times to Gmail and Chrome. Gemini on TPUs remains incredibly fast. Is this deliberate strategy or more complex? Pichai: Speed has two aspects: user-perceived latency and iteration speed for new features. Both matter. The search team operates on a millisecond-level latency budget; saving 3 milliseconds means 1.5 ms for user experience and 1.5 ms for new features. Over the past five years, we've reduced search latency by 30% while adding features. Gemini Flash offers 90% of Pro's capability but is faster and cheaper, thanks to vertical integration. Q: Will search exist in 10 years? Some say chat is the new interface or everyone will have personal agents. Pichai: Search's capabilities expand with each technological shift. User expectations change, and we adapt. Many queries will become agent-based tasks. Search will become an agent manager. I already use multiple agents in Antigravity. Q: Will the traditional keyword-and-links format persist? Pichai: AI-powered search already enables deep research. We'll see more long-running, asynchronous tasks. Q: So the search box might remain but become peripheral? Pichai: Device forms and input methods will evolve. Predicting ten years ahead is challenging; we're excited about the next year given the steep improvement curve. This is an expansive moment, not zero-sum. YouTube thrives alongside TikTok and Instagram. Innovation prevents zero-sum dynamics. Running both search and Gemini is beneficial despite overlaps.
**03 Investing $180 Billion Annually in AGI Exploration** Q: Some researchers feel Alphabet is less "AGI-obsessed" than other labs. Does this affect your direction? Pichai: Our capital expenditure surge from $30 billion to $180 billion shows our belief. It's partly semantic; as a large company, our communication differs. But founders like Demis Hassabis, Jeff Dean, Ilya Sutskever, and Dario Amodei, all formerly with Alphabet, were AGI enthusiasts. Perceived differences may stem from geographic factors like San Francisco's concentration of young labs. Fundamentally, we share similar views on the technology curve. Those deploying AI agents daily witness exponential growth firsthand. Q: When did you last feel an AGI moment approaching? Pichai: In 2012, Jeff Dean demonstrated Google Brain recognizing a cat. Later, Larry Page and I saw DARPA's self-driving cars. Demis showed early models with "imagination." Recently, programming advances are striking—agents complete complex tasks without IDEs. That's an AGI moment. Q: How do you stay connected to products beyond reports? Pichai: I use internal versions intensively. Two weeks ago, I used Gemini Live at the gym for 30 minutes on one topic—some experiences were good, others frustrating, but I learned. X provides direct feedback. In Antigravity, I ask AI for top and bottom user comments on new features. AI assists, but I still invest time experiencing products.
**04 Supply Chain Alerts: Memory and Electricians** Q: You mentioned supply constraints defining 2026. With capex around $180 billion, what are the bottlenecks? Pichai: Wafer capacity is the fundamental constraint. Power and energy are easier, but permits and regulations slow progress. The U.S. must learn to build faster, like China. Memory is a critical short-term bottleneck. Q: Will price increases stimulate supply? Pichai: Leading memory makers won't massively expand. Shortages will ease slowly, driving efficiency innovations—we've improved efficiency 30x. Constraints foster creativity. Q: Does this reinforce oligopoly? More compute lets leaders advance further. Pichai: There's some truth, but we released Gemma 4, a strong open-source model. While Gemini 3 is far ahead, the gap isn't immense time-wise. Capital incentives are huge, but balancing wafer capacity, permits, and capital suggests constraints may be less severe overall. Q: But 2026-2027 physical bottlenecks remain, like the Strait of Hormuz analogy. Pichai: Safety and other limits exist. Models are pushing beyond current software limits, possibly already surpassing them unbeknownst to us.
**05 Three "Hidden Gems"** Q: Beyond SpaceX, Anthropic, and Waymo, what undervalued projects does Alphabet have? Pichai: We pursue long-term projects like space data centers in early stages, similar to Waymo in 2010. Quantum computing is advancing steadily. Q: Where will quantum computing impact most? Pichai: It's better suited for simulating nature, which is quantum-mechanical. Classical computers with compression might suffice, but quantum could excel in weather or complex natural phenomena. Once usable, unexpected applications will emerge, like GPS enabling Uber. Pichai: Google DeepMind is advancing robotics; AI was the missing piece earlier. Gemini Robotics leads in spatial reasoning, collaborating with Boston Dynamics and Agile. Wing drone delivery will soon serve 40 million Americans. Isomorphic uses AI for drug discovery, improving success rates despite clinical trial hurdles.
**06 Regret Over Delayed Waymo Investment** Q: How does Alphabet allocate capital across projects with different return curves like YouTube, Waymo, and AI research? Pichai: TPU allocation highlights this issue. AI will soon assist by analyzing unlocked data. Alphabet's advantage is deciding early; long-term projects need modest initial funding. Assessing progress, like quantum computing's error rates, guides continued investment. We bet heavily on TPUs and increased Waymo investment when others retreated during autonomous driving pessimism. Q: Was Waymo's breakthrough switching to end-to-end deep learning with Transformers? Pichai: Waymo is a highly integrated system; timing and accumulated expertise matter. End-to-end learning accelerates progress, but team cultivation is a major advantage.
**07 Weekly Personal Review of Compute Allocation** Q: With compute as a major budget item, how does allocation work? Pichai: Compute is severely constrained. I spend at least an hour weekly reviewing allocations by project and team, ensuring optimal use. Q: How do you balance internal needs with Google Cloud customer commitments? Pichai: Forward planning handles most issues; cloud teams prioritize customer commitments. Q: AI as an orchestration layer simplifies complex platforms like GCP. Pichai: AI as an orchestrator is powerful for enterprises and consumers. We can improve, but the opportunity is huge. Q: When will consumers have persistent, stateful AI agents? Pichai: Reliable, secure long-term tasks are the future. Identity and permissions need resolution, but we're exploring this frontier. Q: How is Alphabet adapting workflows to AI? Pichai: Change spreads concentrically. Teams like Google DeepMind already use agent managers like Jet Ski (Antigravity). Large companies face change management hurdles smaller firms avoid. Q: Challenges include prompt engineering, code collaboration, data permissions, and role redefinition. Pichai: Gemini Enterprise and Antigravity teams are addressing these. Internal use identifies obstacles, leading to robust solutions. We're in the fixed-cost phase now.
**08 AI Taking Over Human Tasks Timeline** Q: When will AI fully automate Alphabet's business forecasting? Pichai: 2027 will be a turning point. Human verification will phase out gradually. Non-engineering processes will also AI-ize by then. Q: What internal projects excite you currently? Pichai: Space data centers—starting small with a tiny team and budget, achieving initial milestones. Big ideas begin modestly.
Comments