Every year Google Cloud Next arrives with a stack of announcements. Most are incremental. Occasionally, one resets the frame entirely. The April 2026 event in Las Vegas was the latter — not because of any single product launch, but because of what the full picture reveals about where enterprise AI infrastructure is actually headed.
The short version: the GPU era is giving way to the AI factory era. Compute is no longer a commodity you rent by the hour. It is becoming a full-stack, purpose-engineered system — custom silicon, custom networking, custom storage, and orchestration layers designed from the ground up to run millions of agents at once. Google and NVIDIA are building it together, and the implications for every organization running serious AI workloads are significant.
The Hardware Story: Google Splits Its TPU Line
The most technically significant announcement at Next ’26 was the eighth generation of Google’s Tensor Processing Units — and notably, the fact that there are now two of them. For the first time in the TPU’s history, Google has split training and inference into purpose-built chips rather than trying to optimize one architecture for both jobs.
The TPU 8t is built for training at extreme scale. A single superpod connects 9,600 chips with over two petabytes of shared high-bandwidth memory, delivering 121 exaflops of compute — roughly three times the throughput of the previous generation Ironwood chip. The TPU 8i, by contrast, is purpose-engineered for inference and reinforcement learning. It triples on-chip SRAM to 384 MB, increases high-bandwidth memory to 288 GB, and introduces a dedicated Collectives Acceleration Engine that cuts on-chip latency by up to five times. Google claims 80% better performance per dollar for inference versus the prior generation.
The bifurcation matters because it is an acknowledgment of something the industry has been dancing around: training a frontier model and serving millions of concurrent agent queries are fundamentally different computational problems. Optimizing one chip for both leads to compromise on both. By splitting the line, Google is betting that the era of the general-purpose AI accelerator is ending.
The NVIDIA Partnership: Building the Full-Stack AI Factory
If Google’s TPU announcements were the headline for practitioners, the deepened NVIDIA partnership was the signal for the enterprise market overall. The two companies used Next ’26 to formalize what analysts described as a full-stack “AI factory” — an integrated architecture spanning Google’s AI Hypercomputer infrastructure, NVIDIA’s latest accelerators, and shared networking and software layers.
Google announced the upcoming A5X instance family, a new class of bare-metal compute based on NVIDIA’s Vera Rubin NVL72 platform — 72 Rubin GPUs per rack. When clustered, Google and NVIDIA are pointing toward deployments approaching 960,000 GPUs across multiple data centers. That is not cloud compute in the traditional sense. That is national-scale AI infrastructure available through a cloud API.
This resolves an apparent tension the market has watched carefully: Google competes with NVIDIA on custom silicon through its TPU program, yet also needs to offer NVIDIA hardware because enterprise AI software is overwhelmingly built on the CUDA ecosystem. The resolution at Next ’26 was pragmatic. Google leads with TPUs for internal products and select Vertex AI offerings, but the Vera Rubin partnership lets it claim the broadest possible accelerator support for enterprise customers — from open-source models to proprietary workloads, from training to inference, from cloud to edge via Google Distributed Cloud on Blackwell.
The Network and Storage Layer: Where the Real Bottleneck Lives
Hardware announcements tend to capture the headlines, but the infrastructure practitioners who attended Next ’26 were paying close attention to networking and storage — because at the scale of AI factories, those layers are increasingly where workloads are bottlenecked.
Google unveiled the Virgo Network, a custom-built, AI-optimized fabric designed to connect either NVIDIA Vera Rubin NVL72 systems or TPU 8t superpods into massive supercomputers with hundreds of thousands of accelerators. On the storage side, Managed Lustre with TPUDirect and RDMA support allows data to bypass the host processor entirely, moving directly to accelerators at 10 terabytes per second of throughput. For organizations running large-scale training jobs, this addresses one of the most persistent pain points in production AI operations: storage I/O forcing expensive accelerators to sit idle waiting for data.
The Agentic Platform: Infrastructure Is Only Half the Story
All of this compute and networking serves a strategic purpose that Google made explicit throughout the event: the world is moving from AI models to AI agents, and that transition requires infrastructure at a fundamentally different scale and latency profile than what came before.
The new Gemini Enterprise Agent Platform is Google’s answer to the orchestration layer — a complete workspace for building, governing, and scaling AI agents across enterprise environments. It addresses what Google identified as the central challenge facing enterprise AI teams right now: not “can we build an agent?” but “how do we manage thousands of them?”
Google Kubernetes Engine received new capabilities that deserve attention from anyone running inference workloads: dramatically faster cold starts, scale-out improvements for AI inference, and new agent sandboxes capable of deploying 300 sandboxes per second per cluster with sub-second time to first instruction. These are not marketing metrics. They are the numbers that determine whether an agent-based product is viable in production.
What This Means for Organizations Buying Compute
The announcements from Next ’26 have practical implications for any organization that is serious about AI infrastructure, whether it is building on top of cloud or operating its own compute.
Training and inference are no longer the same problem. Organizations still running unified clusters for both should evaluate whether workload-specific hardware would meaningfully improve their economics. The bifurcation Google has made in silicon will likely accelerate a similar bifurcation in how enterprises design their infrastructure. Storage and networking are the new differentiators — when compute becomes accessible at scale, the bottleneck shifts, and the organizations that get the most out of next-generation accelerators are those that have invested in low-latency storage fabrics and high-bandwidth, AI-optimized networking.
Redeployed infrastructure also has a longer runway than the market assumes. The GPU generations arriving in 2026 and 2027 are extraordinarily capable, but they are also extraordinarily expensive at hyperscaler price points. Enterprise AI workloads — particularly inference-heavy, agentic applications — can run effectively on proven prior-generation hardware when it is properly clustered, networked, and operated. The infrastructure lifecycle is longer than the hype cycle.
The Bigger Picture
Google’s AI systems now process more than 16 billion tokens per minute via direct API use — up from 10 billion just last quarter. That growth curve is what is driving the investment in AI factory-scale infrastructure. When token demand doubles in a quarter, the compute requirements do not scale linearly; the entire architecture must be re-engineered to handle the load without degrading the user experience.
The NVIDIA-Google partnership illustrates something that is easy to miss when covering individual product announcements: the leaders in this space are not competing on a single dimension. They are building interlocking ecosystems where hardware, software, networking, storage, and orchestration are co-designed. For organizations evaluating their infrastructure strategy in 2026, the question is not which accelerator wins. It is whether your architecture is flexible enough to take advantage of rapidly improving price-performance across multiple hardware generations — and whether the infrastructure you are operating today is being fully utilized before you commit to the next wave of capital expenditure.
Comment section