Stripe

The history and future of AI at Google, with Sundar Pichai

Key Themes

Google AI historyfuture of SearchGemini and multimodalityAI infrastructure constraintscapital allocationagentic workflowsinternal AI adoption

1h 10mApr 7, 2026

Summary

Sundar Pichai outlines how Google built its AI stack, why Search is evolving rather than being replaced, and where compute, capital, and internal adoption become the real bottlenecks.

In this interview, Sundar Pichai traces Google’s AI journey from early internal product efforts like LaMDA and AI Test Kitchen to Gemini’s multimodal momentum and a broader company-wide reorientation around agents, infrastructure, and long-term bets. He argues that Google’s supposed AI setback was overstated because the company had already spent years building TPUs, data centers, and a full-stack platform. The conversation also highlights the physical limits on AI scale-up, including wafers, memory, power, and permitting, and closes with a look at how Google is diffusing AI internally across teams and workflows.

Google’s AI narrative is increasingly tied to infrastructure depth, not just model quality.

Pichai repeatedly emphasizes TPUs, data centers, and a full-stack approach as the basis for Google’s comeback, suggesting competitive advantage may come from scale and integration as much as from frontier model performance.

Physical supply chains may be the real governor of AI growth over the next two years.

The interview stresses that wafers, memory, power, electricians, and permitting constrain how much compute can actually be deployed, which implies that investors should watch infrastructure vendors and bottlenecks as closely as model releases.

Capital allocation inside AI leaders is becoming more granular and compute-aware.

Pichai says he now spends dedicated time on compute allocation and that TPU scarcity changes how Google evaluates projects, which suggests rising strategic importance for planning discipline and resource control.

Internal AI adoption may create a lagged but meaningful productivity upside for large incumbents.

Pichai describes AI diffusion at Google as uneven, with workflow, permissions, and security barriers slowing rollout; that means the financial benefits of AI may arrive later than the headline model improvements.

Select any chapter text to Deep Dive with AI

01Google’s AI history and the future of Search

Sundar Pichai reflects on Google’s early internal AI product efforts, including LaMDA and AI Test Kitchen, and explains why Google’s slower public rollout was shaped by product quality, safety, and the surprises of consumer internet dynamics. The discussion then turns to Google’s speed culture and ends with a forward-looking view of Search becoming more agentic and task-oriented alongside Gemini.

Google was productizing AI internally before broader public awareness.

AI Test Kitchen used a constrained version of LaMDA because the system was not yet fully safe.

Consumer internet shifts are often surprise-driven and hard to predict.

Search culture has long treated speed and latency as core differentiators.

Many information-seeking queries may become agentic over time.

Search and Gemini are likely to overlap in some use cases but diverge in others.

02Google’s AI comeback and Stripe network intelligence

Pichai argues that the market underestimated Google because it had spent years building vertically integrated AI infrastructure, even if it was behind in frontier LLMs. He highlights Gemini’s multimodal strengths as a key perception shift, discusses AGI as largely a semantic disagreement, and frames AI as a potential macroeconomic expansion force. The chapter ends with Stripe’s pitch that its payments platform functions as network intelligence built from global transaction data.

Investor sentiment around Google was deeply negative before the recent turnaround.

Google’s TPU and data center investments positioned it for the AI wave.

Gemini’s multimodal strength helped change external perceptions.

The AI frontier remains highly competitive and dynamic.

Pichai treats AGI debates as mostly semantic rather than substantive.

Dogfooding and internal product use are essential to staying close to users.

AI could expand the software market and the broader economy even if GDP effects are hard to measure.

Stripe frames payments data as network intelligence for fraud and verification.

03Bottlenecks in AI Scale-Up

Pichai details the physical constraints that limit how quickly AI can scale, including wafer capacity, memory supply, power, and permitting. He argues that compute scarcity may temporarily cap model divergence while also forcing efficiency gains, and he broadens the discussion to Google’s long-horizon bets in quantum, robotics, Wing, and Isomorphic.

CapEx is rising, but physical constraints still limit deployment.

Wafer capacity and memory are major short-term bottlenecks.

Power is solvable faster than permitting and regulatory delay.

Compute scarcity may cap how far any one company can pull ahead.

Constraints can spur efficiency improvements and product innovation.

Google continues to invest in quantum, robotics, Wing, and drug discovery.

04How Google allocates capital across AI, TPU, and Waymo

Pichai explains Google’s capital allocation philosophy as a process of choosing the best use of cash across very different technical bets. He focuses on Waymo, quantum, TPUs, and Google Cloud, emphasizing opportunity cost, milestone-based evaluation, and the need to plan compute much more carefully as AI demand rises.

Capital allocation is about opportunity cost and highest-and-best use of cash.

Google often starts with small commitments and increases funding as milestones are proven.

Waymo is an example of sustained investment through skepticism.

TPU scarcity has made compute planning much more granular.

Cloud commitments require careful forward planning.

AI may act as an orchestration layer over large product surfaces.

Persistent consumer AI tasks are part of the agentic future.

05How Google is diffusing AI internally

Pichai describes AI adoption inside Google as uneven but steadily spreading through the organization. He cites barriers like prompting skill, code sharing, data access, permissions, role redesign, and security, and he suggests that 2027 may be a major inflection point for non-engineering workflows. The chapter closes with his excitement about small experiments like space-based data centers and other high-upside technical projects.

AI adoption at Google is spreading in concentric circles rather than all at once.

Some teams are already operating with agent-manager workflows.

The Search team only recently received access to the tool.

Prompting, permissions, and data access remain key bottlenecks.

2027 may be a major inflection point for broader workflow change.

Younger companies may have an advantage in building AI-native processes.

Small projects can become meaningful if they start with a strong technical insight.