OCP Global Summit 2026 | NVIDIA | Gilad Shainer | Gigawatt AI Factory Networking: Recovery Time Beats pJ/bit

4 days ago

7 min read

1. Five Networks, One Theme: Failure

Shainer opened with NVIDIA's definition of an AI factory: a factory is a "single unit of computing" containing five purpose-built networks, not one general-purpose fabric.

Network	NVIDIA product	Slide claim
Scale-in	BlueField-4 DPU	18x AI factory services bandwidth, 2x storage acceleration
Scale-up	NVLink	10x packet rate, 3x lower latency
Scale-out	Spectrum-X Ethernet	1.6x RDMA bandwidth, 1.3x lower jitter
Scale-across	Spectrum-XGS Ethernet	1.9x multi-site performance
Context scale	Vera BlueField-4 Storage Processor	5x token throughput, 5x power efficiency

▲ The five networks of an NVIDIA AI factory and the multiplier claimed for each; the slide states no baseline (Image source: NVIDIA | 2026 OCP Global Summit presentation)

None of these multipliers states a baseline, so read them as direction only. The two slides with a complete baseline that can actually be checked are in the middle of the talk: the multiplane failure timeline and training step time under multiple jobs. Both make the same point: performance in a gigawatt factory depends on how much you lose when things break or congest, not how fast you run at best.

2. Breaking Down Multiplane Resiliency: 1,080 ms vs 2.68 ms

Multiplane means one NIC connects to several independent switch planes at once. NVIDIA says Spectrum-X multiplane with CPO can reach 512K GPUs on a flat two-tier network with 40% fewer switches.

The question is what happens when a plane fails. The traditional approach does load balancing in software: detect first, then redistribute traffic to the remaining planes, and during that window RDMA bandwidth drops to zero. NVIDIA moved the hardware adaptive-routing IP from its switch into the SuperNIC, so the NIC detects and rebalances by itself. The slide compares:

Phase	OTS Ethernet multiplane	Spectrum-X multiplane
Fault detection	1,080 ms at 0% bandwidth	2.68 ms (slide: 400X)
After detection	90% bandwidth	90% bandwidth
Back to 100% after link repair	1,080 ms	100 ms (slide: 11X)

▲ The same plane failure: off-the-shelf Ethernet bandwidth drops to zero for 1,080 ms, while Spectrum-X falls to 90% within 2.68 ms and keeps running (Image source: Simple Tech Trend original; data from NVIDIA 2026 OCP Global Summit presentation)

Put simply: a plane failure on off-the-shelf Ethernet is an outage; on Spectrum-X it is a 10% slowdown. For synchronous training, a second of zero bandwidth can stall an entire collective, so the cost is more than that one second. NVIDIA titled this slide 1.6x higher AI-factory goodput. Both multipliers can be checked from the timings: 1,080 / 2.68 ≈ 403 and 1,080 / 100 ≈ 11; the preceding product slide rounds detection to 2.6 ms.

3. Breaking Down the Value: 8% Power Saved, or 1.6x Goodput Kept

This is our core argument. Shainer said the scale-out optical network draws power equal to about 10% of compute, CPO reduces power by up to 5x and raises MTBI 10x, and NVIDIA has accumulated "hundreds of millions of hours" running CPO.

Doing the math (STT estimate: 10% × (1 − 1/5)), full CPO adoption in scale-out frees at most about 8% of compute power. The estimate rests on three assumptions: the 10% is Shainer's spoken "almost 10%"; the 5x compares the optics themselves and is applied here to the whole optical network, so it is an upper bound; and all scale-out optics move to CPO. That is not small, but it is clearly an order smaller than the 1.6x goodput claimed for multiplane and the 1.9x training gap under multiple jobs.

▲ DeepSeek V3 multi-workload training: once multiple workloads start, off-the-shelf Ethernet step time stretches, with a 1.9x gap marked on the chart, while Spectrum-X stays flat (Image source: NVIDIA | 2026 OCP Global Summit presentation)

The training curve deserves a closer look. With a single job the two curves nearly overlap, and the 1.2x label marks only one jitter spike. Once multiple jobs run together, OTS Ethernet step time jumps from about 650 ms to around 1,100 ms; the labeled gap is 1.9x (about 1.7x on average, STT chart reading), and Shainer said "almost 2x". In other words, NVIDIA's advantage concentrates in noisy conditions: multi-tenancy, multiple jobs, plane failures. That is exactly the normal state of a gigawatt factory.

So in NVIDIA's story CPO and multiplane are a pair: CPO cuts how often failures happen through 10x MTBI, multiplane cuts what each failure costs through millisecond failover. In OCP APAC Summit 2026 | Broadcom | CPO's Third-Generation Report Card we saw Broadcom present its CPO reliability results; NVIDIA goes one step further, extending reliability from the component to the NIC and the topology.

4. Breaking Down Scale-Across: Removing Deep Buffers

The fifth network is scale-across, joining distant data centers into a single workload. The traditional approach uses deep-buffer switches to absorb unpredictable congestion along the path, at the cost of latency.

Shainer's numbers: fiber propagation itself costs about 5 ns per meter, and the deep-buffer latency penalty is 2 to 3 times the distance itself. At 100 km (STT estimate), one-way propagation is about 0.5 ms, and deep buffers add another 1.0 to 1.5 ms.

▲ At an estimated 100 km, deep buffers add 2 to 3 times the latency of fiber propagation; Spectrum-XGS claims to pay only the latency of distance itself (Image source: Simple Tech Trend original; data from NVIDIA 2026 OCP Global Summit presentation)

NVIDIA's approach: when a job starts, it sends telemetry probes to measure the distance and latency of every NIC and every path, then adjusts the data-distribution algorithm by distance instead of relying on deep buffers. The claimed result is about 1.9x multi-site NCCL performance.

▲ Spectrum-XGS links multiple AI factories with distance-aware load balancing and congestion control, for 1.9x multi-site performance (Image source: NVIDIA | 2026 OCP Global Summit presentation)

On timing, Shainer said this is already deployed across tens of kilometers, and by year-end NVIDIA will start spanning hundreds of kilometers. It is the same logic as the previous section: replace buffering and retries in the switch with hardware intelligence in the NIC.

5. Pluggables Aren't Going Away: USFP Notes from the Q&A

USFP is a pluggable form factor jointly defined by NVIDIA and Cisco; the slide lists 200G/400G SerDes per lane, up to 32 electrical lanes, 6.4T initial capacity, and Amphenol and TE as ecosystem partners. The spec was covered in full in another NVIDIA optics talk, so here we note only what Shainer added:

No separate liquid-cooling circuit required (a design goal stated in the talk)
Up to 100 W, slightly smaller than OSFP (Q&A)
From today's 1.6T to 6.4T, later 12.8T; demos "not long from now", production next year (Q&A)

Asked about NPO, Shainer said NVIDIA will "do NPO if we need to" when designs are too dense and beachfront runs out, but CPO remains the main path for scale-out and scale-up. That matches our view in NPO Is Not a Stepping Stone to CPO: NVIDIA does not reject NPO, it just does not treat it as the main line.

6. Risks and Counterarguments

The 30x and 400x figures use different baselines: the 400X on the multiplane timeline is detection versus detection (1,080 / 2.68 ≈ 403), consistent with Shainer's "almost one second vs under 3 ms"; the Meta Spectrum-6 slide says "30x faster than traditional software-based fault detection" without a baseline time. Back-calculating from 2.68 ms puts that baseline near 80 ms (STT estimate), meaning the 30x compares against a faster software scheme. They do not contradict each other, but any citation should name which baseline it uses.
1.6x goodput depends on failure frequency: millisecond failover only buys that multiple if plane failures are frequent enough; the slide gives no failure-rate assumption.
"OTS Ethernet" is unnamed: every baseline is labeled off-the-shelf Ethernet with no vendor or configuration, and other NIC vendors are also building hardware multipath.
8% is a ceiling: CPO's 5x is stated as "up to", and scale-out optics will not all move to CPO at once.
Hundreds of kilometers for scale-across has no customer or data yet: only tens of kilometers are public today.

7. What It Means for Taiwan's Supply Chain

Three of the four companies in NVIDIA's photonics production photos are Taiwanese. The "NVIDIA Photonics in Production" slide labels its four photos TSMC, SPIL, Lumentum and Foxconn, and Shainer said the optical engines use TSMC's COUPE packaging process. TSMC, SPIL and Foxconn appearing together in NVIDIA's own CPO production imagery shows Taiwanese firms are already in the volume chain; which stage each handles and their share of shipments are not stated beyond the company names.
Multiplane changes the cabling structure. The slide claims 40% fewer switches and a flat two-tier network reaching 512K GPUs; with each NIC connected to multiple planes, how fiber and cabling demand changes was not quantified, so this is an STT inference that needs deployment data to verify.
No Taiwanese firm is on the USFP partner list. The slide describes a "jointly defined form factor and multi-source specification", and the public names are Cisco, Amphenol and TE; whether Taiwanese connector and module makers join the multi-source list before production next year is the window to get in.

8. Conclusion

This talk shifted NVIDIA's networking pitch from "faster and more efficient" to "loses less when things break". CPO reduces failures with 10x MTBI, multiplane turns a failure into a 10% slowdown within 2.68 ms, and scale-across replaces deep buffers with distance awareness; all three point the same way: in a gigawatt factory, the network's value is keeping expensive GPUs from stopping.

Over the next 12 months, watch three things: the real failure rate and goodput once Meta deploys multiplane, whether other NIC vendors can match millisecond hardware failover, and the supplier list when USFP enters production next year. If competitors catch up on millisecond failover before 2027, NVIDIA's scale-out differentiation falls back to CPO itself.

Saving 8% of power is good, but a gigawatt factory that loses only 10% of bandwidth when something fails is what NVIDIA really wants to sell.

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