Google Cloud Next ‘26 finished this week and the flagship announcement is the Gemini Enterprise Agent Platform, which is what Google is now calling the thing that used to be Vertex AI. New name, extended remit. The platform bundles model selection, model building, and agent building into one product, with new features aimed at agent integration, DevOps, orchestration, and security. Model Garden inside the platform now offers first-class access to more than 200 models, including Gemini 3.1 Pro, Gemini 3.1 Flash Image, Lyria 3, and Gemma 4, plus a growing list of third-party frontier weights that Google is now perfectly happy to host on Google infrastructure. Underneath: the eighth-generation TPU family, which shipped as two specialized chips, one weighted toward training and one weighted toward the kind of high-concurrency inference that agent workloads actually look like.
The framing is worth reading carefully. Google is not saying “our agent is better than OpenAI’s or Anthropic’s.” Google is saying: your enterprise is going to run hundreds of agents, from many vendors, hitting many data sources, calling many tools, and you need a control plane. The Gemini Enterprise Agent Platform is the control plane. Bain called it the “agentic enterprise control plane” in its wrap-up, which is analyst-speak but is also basically what Google is going for. The A2A (agent-to-agent) protocol Google is pushing alongside the platform is meant to be the thing agents from different vendors use to talk to each other. Every platform vendor with a control-plane pitch also has a protocol they want to be the standard, so read that with the appropriate salt.
The strategic frame is that this is the hyperscaler version of what Microsoft is doing with Frontier and what Amazon is doing with Bedrock AgentCore. Nobody is fighting for “your enterprise picks one AI vendor” anymore. That war is over, and the answer is “your enterprise picks all of them and someone has to govern the mess.” Google’s specific advantage is that it is the only one of the three that owns competitive frontier silicon end-to-end, from TPU-v8 down to the network switch. That is what the whole eighth-generation TPU announcement was there to signal. The pitch to a CIO is: run your agent fleet on the platform whose underlying compute we ourselves designed, and stop worrying about GPU allocations from a supplier who might ration you next quarter.
The product is more polished than it needed to be. Whether IT departments actually adopt “one control plane per hyperscaler” as their organizing principle, or whether some cross-cloud third party ends up owning that layer instead, is the question the next twelve months will settle. Google has now placed its bet in a shape you can look at.