2026 OCP APAC Summit | NVIDIA | Gilad Shainer | Scaling the AI Factory: Open Networking for the Gigascale Era

2 days ago

6 min read

On the OCP stage, NVIDIA put its "entire AI factory" — compute boards, NICs, switches and racks, power, and management and orchestration software — into open specifications. It looks generous, but it is really a declaration: in a gigawatt-scale AI factory, the variable that decides what a GPU is worth has moved from "the GPU itself" to "the network." NVIDIA's calculation is to use four separately customized networks (NVLink for scale-up, Spectrum-X open Ethernet for scale-out, Spectrum-XGS for scale-across, plus CPO, which moves the optical engine into the switch) to turn the network into a moat that hangs an "open" sign out front while NVIDIA writes the rules. The real point of this 12-minute keynote isn't any single product — it's the calculus of openness.

1. Why Now: Contributing the Whole Factory to OCP Isn't Generosity — It's Defining the Battlefield

Gilad Shainer opened by pointing at his slide: what you see is the "entire NVIDIA factory" contributed to OCP — compute boards, NICs, switches, racks, power distribution, plus a large body of management, orchestration and security software, all released with open interfaces. The subtext is clear: when AI factories scale to the gigawatt level, no single company can monopolize them with closed specs, so "open the foundation first, then lead on top of it" becomes the only answer.

He highlighted two recent contributions. One is the MGX rack (which he jokingly calls the "magic box"): a high-density rack that flexibly houses GPUs/CPUs, has a built-in NVLink switch backplane and supports high-temperature liquid cooling, with a third generation coming this year. The other is 800V rack power distribution. They may sound dull, but they are the hardest industry signals of the talk: raising rack voltage to 800V and stacking compute density is what lets NVIDIA "keep using copper." Using copper instead of optics inside the rack saves power — and keeps optical module suppliers outside the rack door. Wherever copper holds up, there's no room for optics; we broke down this path fully in After Copper Runs Out for AI: Seven Paths for Scale-Up Optical Interconnect.

An AI supercomputer is essentially a distributed computing machine stitched together from "multiple separately customized networks"; how the network is designed directly determines how much the machine is worth.

2. Four Networks: One Supercomputer Is Really Four Networks, Each Fighting Its Own Battle

This section is the backbone of the talk. Shainer split the AI factory's networking into four layers, each an independent infrastructure "customized for a single purpose," rather than one big network for everything:

Scale-up (NVLink): the in-rack interconnect that links many compute ASICs into "one giant GPU." It demands ultra-low latency and ultra-high bandwidth, with built-in in-network computing that grows generation by generation. His number: NVLink delivers roughly a 2x improvement on workloads. This layer is "building the GPU" itself.

Scale-out (Spectrum-X open Ethernet): scaling rack-level GPUs horizontally to hundreds of thousands, running as one AI factory. Shainer repeatedly stressed "eliminating jitter" — in synchronous workloads like training, the slowest step sets the pace, and with more jitter the GPUs sit idle. Spectrum-X isn't just pitched as "faster" but as deterministic performance where "every step is consistent," delivering more than 40% end-to-end performance gain in training. More notable is the "open" approach: they open-sourced the implementation to SYCL and support running different network operating systems such as SONiC and Cumulus on top. This echoes AMD writing "open Ethernet" into its rack backbone at Advancing AI — a signal we unpacked in AMD Writes "Open Ethernet" into the Rack Backbone.

Scale-across (Spectrum-XGS): when a single data center isn't enough, multiple remote data centers are linked to run the "same" AI workload. He specifically noted that they avoid "deep-buffer switches," because deep buffers create jitter; instead they use extended adaptive routing and congestion control algorithms, yielding 2x network performance across remote data centers.

Storage (BlueField-4): a new storage tier for inference context memory, letting context extend beyond the GPU server to support more users and longer contexts.

3. Enter CPO: Scale-Out Power Eats Nearly 10% of Compute — the Real Reason to Move to CPO

This is the most important section for the optical supply chain. Shainer gave the reason for CPO directly: optical transceiver transmission and power at the scale-out layer are approaching 10% of total compute. In other words, nearly a tenth of the "compute budget" of the GPUs you buy goes to feeding the network's optical-electrical conversion — an astronomical figure at gigawatt scale.

The solution is integrating the optical engine into the network device: Co-Packaged Optics (CPO). The tech stack he named is all the hard stuff of optical communications: microring modulators, high-power lasers (feeding more fibers with fewer lasers), fiber arrays and advanced packaging. The stated benefit: CPO cuts the power of this infrastructure layer by up to 5x, and — more importantly, in his view — improves reliability by about 10x: fewer interruptions, more usable compute, more tokens. And he said explicitly: "A lot of (the supply chain) is right here in Taiwan."

We covered CPO's turn from "crying wolf" to finally reaching commercial deployment in CPO Is Finally No Longer "Crying Wolf"; and for why AI interconnect ultimately converges on CPO and what the pJ/bit battle looks like at each layer, OIF has drawn an official map — see OIF Draws an Official Map for AI Interconnect.

4. The Calculus of Openness: NVLink Fusion Is the Spear, "Open Ethernet" Is the Shield

Shainer closed on NVLink Fusion: customers can use their own XPUs and custom compute engines, connect to the NVLink network via NVLink chiplets, connect to supporting CPUs via NVLink C2C, and plug the whole system into NVLink switches, NVLink backplanes and MGX racks — all the way into NVIDIA's full AI infrastructure. The ecosystem includes CPU partners, optical interconnect partners and custom silicon partners.

Put this together with the earlier sections and you get NVIDIA's full calculation: externally, scale-out uses "open Ethernet + open-source SYCL" to ease customer concerns about closed InfiniBand, so neoclouds, sovereign AI and hyperscalers are willing to commit; internally, scale-up NVLink, racks, backplanes and CPO remain highly co-designed domains that NVIDIA controls. What's opened is the most competitive layer, the one hardest to monopolize (Ethernet); what's held tight is the most profitable, stickiest layer (NVLink and packaging). This isn't generosity — it's drawing the battlefield where it suits NVIDIA best.

5. Timing and the Counter-Case: Three Question Marks Behind the Pretty Numbers

Line up Shainer's numbers and the sense of timing is strong: scale-up +2x, Spectrum-X training +40%, Spectrum-XGS cross-data-center +2x, CPO power –80% / reliability ×10, and rack voltage jumping to 800V. But readers should look at three things calmly.

First, nearly all these numbers are NVIDIA's own benchmarks; the baselines and test conditions aren't necessarily public, so "+40%" relative to what deserves a question mark. Second, "keep using copper" cuts both ways — 800V and high density let copper last another generation inside the rack, but that also pushes back in-rack optical interconnect adoption, which is bad news for suppliers hoping to sell scale-up optical modules. Third, CPO's ×10 reliability is a "spec promise"; real yield and serviceability (how do you replace a failed optical engine?) will only be proven in volume production — the last mile of CPO commercialization has always been packaging and repair, not vision.

6. What It Means for Taiwan's Supply Chain

Shainer's line that "a lot of it is right here in Taiwan" wasn't a pleasantry. The keynote's significance for Taiwan breaks into three threads.

Advanced packaging and CPO: microring modulators, laser integration, fiber arrays and optical engine co-packaging are all links where Taiwanese companies (foundries, OSATs, passive optical components, FAU) can claim a position. Once CPO moves from samples to volume production, the biggest winners will be packaging players that can do optoelectronic co-packaging — not traditional pluggable module makers. Copper's extended life: 800V plus high density gives in-rack copper cable, connector and power module suppliers another stretch of upside, but it also warns Taiwanese companies working on scale-up optical interconnect that adoption will be slower than expected. The open Ethernet opportunity: Spectrum-X going with open Ethernet and supporting SONiC opens a door for Taiwan's white-box switch and networking ODMs — the battle shifts from "selling NVIDIA-specified parts" to "building systems and thermal solutions on open specs," and cooling 51.2T-class switches is exactly where Taiwanese companies can compete.

Conclusion

In this 12-minute keynote, NVIDIA really wanted to say just one thing: in gigawatt-scale AI factories, the network is the main variable determining GPU utilization, token economics and time-to-first-token, and NVIDIA wants to be both the "contributor of the open foundation" and the "definer of the rules on top of it." For Taiwan's supply chain, what to track isn't yet another pretty performance number but three milestones: when CPO moves from samples to volume production, how long 800V racks extend copper's life, and whose open Ethernet ecosystem (Spectrum-X vs the AMD camp) takes shape first. Whoever understands the calculus of openness first will know which layer to stand on.

This article is for technology and industry trend analysis only and does not constitute investment advice.