Google Writes Its Answer Into 134,000 TPUs: What Virgo Network Really Reveals Is a Win for OCS
While everyone was still arguing over which year CPO (Co-Packaged Optics) will ramp, Google laid an entire generation's answer on the table at Cloud Next 2026. It is called Virgo Network, a fabric that glues 134,000 TPUs into "one machine." And what actually holds it up is not some exotic packaging, but something Google has quietly been building for six years: the Optical Circuit Switch (OCS). Virgo is the scale-out data center fabric Google launched alongside its eighth-generation TPU 8t. A single fabric connects 134,000 TPUs with 47 Pb/s of non-blocking bisection bandwidth, and Pathways stitches multiple fabrics into a logical training cluster of more than 1 million TPUs. Its most important signal for the optical supply chain: the winners are "massive volumes of pluggable optical modules plus in-house OCS." CPO's turn has not come yet.
1. What Really Limits Million-Chip Training Was Never Compute
When training a frontier model, the bottleneck stopped being single-chip speed long ago. Once you need to keep tens of thousands of accelerators fed and computing in lockstep as if they were one chip, the problem becomes a plain engineering question: how many switch tiers and hops does data have to cross between chips, will it fit, and will one failed chip drag down the whole job?
That is what scale-out is meant to solve. It is different from scale-up, although the market often conflates the two: scale-up uses ultra-high bandwidth to bind nearby chips inside a pod into one large memory domain (ICI for TPUs, NVLink for GPUs); scale-out links pods and whole data halls horizontally into one logical cluster. Virgo handles the latter.
Google built a new network for TPU 8t because the new chip's appetite is too big for the old one. Virgo raises per-accelerator data center network (DCN) bandwidth to 4x the previous generation. Without a network upgrade, even the strongest chip just sits starved.
2. What Virgo Is: A Fabric That Flattens the Network
In one sentence: Virgo is a flat, two-tier, non-blocking network.
A traditional data center network is like a tall building where packets climb several flights of stairs to reach the other side. Virgo switches to high-radix switches, which simply means each switch has more ports and can connect to more peers, so the whole network needs fewer floors. Fewer floors means lower latency: Virgo's unloaded fabric latency is 40% lower than the previous generation. For training jobs that synchronize tens of thousands of chips, predictable latency is worth more than peak bandwidth.
It is also a multi-planar design that splits the network into several independent control domains. Put simply, don't keep all your eggs in one basket: if one plane fails, it does not take the whole network down with it.
3. Three Numbers Worth Memorizing
134,000 TPUs per fabric. A single Virgo fabric connects more than 134,000 TPU 8t chips. (A common slip worth correcting: it is 134K, not 13,440, an order of magnitude apart.) Beyond that, the Pathways/JAX software stack can stitch multiple fabrics across data centers into a single logical cluster of more than 1 million TPUs.
47 Pb/s of non-blocking bisection bandwidth. "Non-blocking" is the key: in theory, any half of the chips can talk to the other half at full speed without the network choking or contending. This is the hard metric for whether a training network is up to the job.
7 hops (down from 16). Using OCS to link up to 36 groups (each with up to 1,024 active chips) into a full pod, any two chips can communicate in at most 7 hops, 56% fewer than the 16 hops of the old architecture. Hop count is a direct source of latency and failure risk.
One number to handle with care: compute specs are reported inconsistently (some say 1.6M ExaFlops, others 1.7K ExaFlops), most likely because of different precision (FP8 vs BF16) or a full-fabric vs single-pod basis. Before quoting it, state the precision clearly rather than citing a bare number.
4. The Real Undercurrent Is Apollo OCS
Virgo's press release puts the spotlight on bandwidth and topology, but the thread worth watching runs underneath: optical circuit switching.
OCS does not convert light to electricity to switch it. It uses MEMS micromirrors to steer a beam of light directly to another output port: light in, light out, never touching the electrical domain in between. Its value is twofold. First, it eliminates the power and latency of large amounts of optical-electrical conversion. Second, when a chip or rack fails, OCS can automatically reconfigure optical paths around the fault with no human intervention. For a training job running for weeks across a million chips, "routing around failures on its own" means rack-level faults are absorbed at the network layer, so effective compute is not dragged down by tail latency.
The key point is that this OCS (internal codename Apollo) is built in-house, not purchased. Since TPU v4 in 2020, OCS has been the structural difference in Google's data centers that is hardest to match for the NVIDIA/Broadcom ecosystem (NVLink plus Ethernet/InfiniBand). Virgo is not a turn on this road but its extension and amplification.
5. For the Optical Supply Chain: Modules and Upstream Win, Not CPO
Many people hear "optical switching, a million chips" and instinctively think CPO. Virgo's architecture is the opposite: the high-radix flat topology has to be fed by massive volumes of 800G/1.6T pluggable optical modules, plus Google's in-house OCS (MEMS micromirrors, circulators and wavelength-selective switches, WSS). In other words, what really gets volume this generation is the optical module itself and its upstream: EML/DML lasers, DSPs and assorted optical components. CPO is still standing outside the door.
And because Apollo is built in-house, what it offers third-party OCS vendors is a "demonstration effect" rather than direct orders. It proves optical switching works at data center scale, but the orders may not land with outside suppliers. That distinction determines where you should look when tracking beneficiaries.
Put another way: the market keeps waiting for CPO to take the stage as the lead, and Google is using Virgo to say that on the road to a million chips, pluggable optical modules combined with in-house OCS are already good enough today, and more controllable.
6. Risks and Counterarguments
Don't read this as "Virgo is invincible." A few open questions are worth keeping:
First, the ICI layer is still unclear. Virgo is scale-out, but Google said little this time about chip-to-chip interconnect (ICI) inside the pod for scale-up. If 8t's in-pod interconnect moves to LPO/LRO, the impact on optical component volumes would be a completely different story.
Second, 47 Pb/s and a million chips are two different levels. Most sources describe "a single fabric within one data center," while the cross-DC million-chip cluster is a higher layer of software stitching. Mixing the two inflates the specs.
Third, in-house means a moat, but also a one-off that is hard to copy. Google can run OCS because it is vertically integrated from chips to network to software. This path is of limited replicability for other cloud providers (or outside suppliers), and it cannot be extrapolated directly into an industry-wide trend.
Conclusion
What Virgo Network really announces is not that "Google has another, bigger network," but that a choice of path has been validated: in the million-chip era, flattening the network, blanketing it with pluggable optical modules, and using in-house OCS to route around faults automatically is a combination that runs today. CPO is still the future, but Virgo shows it is not yet today's answer.
If you are tracking optical communications beneficiaries, the action items are clear: first watch the order pace of 800G/1.6T optical modules and their upstream (EML, DSP, optical components), treat OCS as "confirmation of direction" rather than "a source of orders," and put TPU 8t's ICI interconnect at the top of your next to-check list. Whoever keeps Virgo fed stands on the right side of this round of scale-out.
Comments