Acquired's Ben Gilbert and David Rosenthal live from Google Cloud Next

In this live Google Cloud Next '26 session, Jason Davenport from Google Cloud hosts Ben Gilbert and David Rosenthal, hosts of the Acquired podcast, for a wide-ranging conversation about the evolving AI landscape and Google Cloud's transformation. The duo, who have produced over 11–12 hours of content studying Google across three episodes, share observations from the event and offer their broader analysis of the AI industry.

The conversation opens with reflection on Google's AI comeback following the perceived shock of ChatGPT's 2022 launch. Despite Google having invented the transformer architecture in 2017 and employing most of the leading AI researchers, it appeared caught off guard. Gilbert and Rosenthal describe Google's subsequent resurgence with Gemini and the merger of DeepMind with core Google research as a defining arc — and suggest today's announcements may warrant a 'Part 4' of their Google series.

A central theme is the shift from training to inference as the dominant AI compute workload. Rosenthal highlights the significance of Google's TPU v8 announcement, specifically the split into dedicated training and inference chips. He notes that inference compute loads have now eclipsed training — a signal that AI is being used at massive scale in production, not just development. He draws a precise analogy to the early web: training is like Google's original one-time web crawl, while inference is the ongoing query-serving that eventually dominated infrastructure. He notes this transition took 15 years on the web but is happening in roughly four years with AI. The TPU v8 inference chip specifically offers a 5x speedup over its predecessor, which Rosenthal argues will shift many AI tasks from feeling asynchronous to feeling synchronous — an underappreciated unlock with unknown downstream effects.

The group also explores barriers to broader AI adoption. Rosenthal argues that despite inference eclipsing training in volume (confirming real usage), most people — even AI enthusiasts — still do not reflexively reach for AI tools for everyday tasks. The missing piece is not capability but habit and UI. Gilbert reinforces this by noting that agent-based workflows still largely require a terminal, effectively excluding the vast majority of users. The panel frames this as a 'UI for agents' problem that the industry has yet to solve.

On the 'efficiency phase' of AI, Rosenthal draws an analogy to newspapers digitizing their print editions before discovering the scroll-feed. He argues the industry is currently digitizing existing workflows rather than discovering genuinely native AI business models. He is cautiously optimistic about what comes next but acknowledges that no one has yet identified the 'killer native app' of the agentic era.

Google Cloud's transformation under Thomas Kurian is discussed at length. The hosts cite a Harvard Business School case study in which Kurian surveyed the organization and determined it needed 1,000 times more salespeople to be enterprise-ready — and then actually executed that hiring. They contrast this with Google Cloud's origins: App Engine was the platform, Snapchat became its largest customer with no formal sales relationship, a situation described as 'complete insanity.' Today, Fortune 500 logos fill the keynote stage. The panel identifies Google's integrated stack — custom chips, cloud infrastructure, and frontier models — as a structural advantage no other hyperscaler fully replicates.

Other topics include YouTube as an underappreciated AI research platform (Rosenthal describes using Gemini's video chat feature to search across 50–100 research videos per episode), real-time voice AI as a major unlock especially in markets like India, sovereign cloud and edge deployment of Gemini, and the enterprise cost-efficiency case that is driving faster AI adoption than the internet achieved. The session closes with a preview of Gilbert and Rosenthal's upcoming fireside chat with Google's Jeff Dean and Amin Vahdat on TPU evolution.