Google Cloud Next '26: the shift from building single agents to governing thousands

The framing has moved to agent fleet management

Pichai's central observation is that the customer question has changed. He describes the shift explicitly, and it reframes what Google is selling.

The conversation has gone from "Can we build an agent?" to "How do we manage thousands of them?"Montana Labs

The answer is the new Gemini Enterprise Agent Platform, positioned as a "mission control" to build, scale, govern and optimize agents. This sits on top of Gemini Enterprise, which Google says saw 40% quarter-over-quarter growth in paid monthly active users in Q1. The demand signal Google cites for the underlying models is concrete: first-party models now process more than 16 billion tokens per minute via direct customer API use, up from 10 billion the prior quarter.

That growth is backed by a capital commitment worth noting: Google expects just over half of its overall 2026 machine learning compute investment to go toward the Cloud business. The company is redirecting the majority of its ML infrastructure spend toward external customers rather than internal model training.

Two TPUs for two different jobs

The eighth-generation TPU launch is notable for splitting into two chips with distinct workloads rather than a single general part. The TPU 8t targets training, scaling to 9,600 TPUs and 2 petabytes of shared high-bandwidth memory in one superpod, with three times the processing power of Ironwood and up to 2x more performance per watt.

The TPU 8i targets inference, connecting 1,152 TPUs per pod with 3x more on-chip SRAM to prioritize low latency. Google ties this directly to the agent thesis: the chip is meant to "concurrently run millions of agents cost-effectively." The design choice tracks the argument in the rest of the post — if enterprises are running large fleets of agents, inference latency and throughput become the binding constraint, not just training scale. Google notes these sit alongside a portfolio of NVIDIA GPU instances, so the split TPU strategy is an addition rather than a replacement.

Google as its own proof point

The most specific claims come from Google's "customer zero" section, where it reports on its internal adoption. The company says 75% of all new code at Google is now AI-generated and approved by engineers, up from 50% last fall. It describes a complex code migration completed six times faster than was possible a year earlier, and says the initial MacOS release of the Gemini app was built with its Antigravity platform, going from idea to native Swift prototype in a few days.

On security, Google says its Security Operations Center agents automatically triage tens of thousands of unstructured threat reports each month, cutting threat mitigation time by more than 90%, and that it actively uses Gemini-based agents such as CodeMender to find and fix critical software flaws. On the operations side, marketing teams generated thousands of creative asset variations for the Gemini in Chrome launch, reporting 70% faster turnaround and a 20% increase in conversions.

These are internal metrics, not customer results, and they should be read that way. But they are unusually specific for this kind of keynote, and they double as a marketing argument: the tooling Google is selling is the tooling Google runs on.

What the security pairing signals

Google frames AI as both threat and defense, and pairs its own Threat Intelligence and Security Operations with Wiz's Cloud and AI Security Platform, alongside Wiz's new AI Application Protection Platform covering code to cloud to runtime across multicloud and hybrid environments. The inclusion of Wiz here is the clearest sign that Google intends to sell agent governance and agent security as a single package rather than leave runtime protection to third parties.

The specific implication of Cloud Next '26 is that Google is betting the next unit of enterprise complexity is the agent fleet, and it is building the full stack — a management plane, inference-optimized silicon, and integrated security — around operating that fleet at scale. For teams evaluating the platform, the useful questions are not about model quality but about governance, latency economics, and whether the internal productivity numbers translate outside Google's own codebase. I/O on May 19 is flagged as the next checkpoint.