Google details Virgo Network for AI Hypercomputer

Google used its Cloud Next keynote to introduce Virgo Network, a scale‑out data‑center fabric designed to sit under its AI Hypercomputer and link large TPU and GPU clusters for training and serving at hyperscale.

Google calls Virgo a megascale, purpose‑built network fabric that connects accelerator pods across racks and sites to form much larger training and inference domains than a single rack or pod can support. The company presented Virgo as the physical network layer that lets the Hypercomputer scale out.

In marketing materials and technical notes, Google says Virgo can deliver non‑blocking bi‑sectional bandwidth measured in the tens of petabits per second and can stitch together hundreds of thousands — and in some descriptions more than a million — TPU chips into single clusters for distributed training. Those capacity figures appeared in multiple Cloud Next writeups.

The practical payoff Google emphasizes is lower latency for multi‑rack model serving and faster synchronization for very large training jobs. By reducing network bottlenecks between racks, Virgo aims to shrink the time spent exchanging gradients, checkpoints and attention states during distributed training. Google framed the fabric as integral to reducing end‑to‑end latency in both training and inference.

Virgo was announced alongside Google’s eighth‑generation TPUs, which include separate chips tuned for training and inference. Google presented the two as a combined system — chips, storage, software and the Virgo fabric — that together make up the AI Hypercomputer stack. The company said the hardware and fabric will become available to customers later in the year.

Beyond raw bandwidth numbers, Google described software and orchestration changes that let large models span many pods. The company named JAX and Pathways as frameworks that, together with Virgo and improvements in shared storage, let users scale training experiments into vastly larger clusters. Managed Lustre and other storage upgrades were highlighted as part of the same stack.

Google also said Virgo is not limited to its own TPUs: the fabric can be used to link third‑party GPU systems in some configurations, letting customers build heterogeneous superclusters. Several Cloud Next posts noted Virgo’s compatibility with certain large GPU superpod designs in Google’s materials.

Network architects and cloud engineers at Cloud Next argued that a purpose‑built AI fabric answers a basic systems problem: chips are getting faster, but end‑to‑end training and serving speed depends on how quickly data and intermediate results can be moved between accelerators. Virgo is Google’s attempt to make that physical movement less of a limiter for frontier models. This is a systems‑level argument Google has made about its Hypercomputer approach before.

Industry coverage treated Virgo as part of a larger push by hyperscalers to vertically integrate chips, networking and software to control latency and costs. Analysts and reporters pointed out that Google’s claims about linking very large numbers of chips put the company in a direct architectural competition with large GPU‑centered offerings from other vendors. Observers also flagged that real‑world results will depend on software maturity, availability across regions and customer demand.

For customers, the headline is a familiar one: faster training, cheaper inference and a broader set of consumption options if Google’s integration works as described. Google announced partner programs and financing to help enterprise adopters move to the agentic‑AI and Hypercomputer model it promotes, signaling the company expects Virgo to be a commercial plank, not just an internal optimization. Availability timelines in Google’s materials said the new stack will roll out through the year.