OCP Global Summit 2026 | Nokia | Clayton Wagar | Multi-Dimensional Scalability for AI Network Infrastructures

1. Why a router company talked about the "fourth wave" of traffic at OCP

Wagar opened by saying this was not a product launch but an invitation. His argument rests on internet history: voice, data and video each arrived as a traffic wave larger than the last, and each forced a new control plane, from TDM and MPLS to CDNs. The slides cast AI as the fourth wave, with the token as its unit of sale and traffic that is bidirectional, latency-critical and machine-driven.

The scale is already large. Nokia cites 1.3 trillion AI sessions a year and 236 EB of traffic a month, and Google alone processes 3.2 quadrillion tokens a month. Wagar stressed that all of this is still a human typing into a box and waiting; the machine-to-machine layer has barely started.

▲ AI supercycle figures cited by Nokia: global data center spending of more than US$5.5 trillion in 2026–2030, intra-DC capacity growing at a 65% CAGR, and WAN AI traffic at 23% (Image source: Nokia | 2026 OCP Global Summit presentation (p. 2))

The most telling part of this slide is the gap between two growth rates: intra-DC capacity at a 65% CAGR versus WAN AI traffic at 23% (both 2025-2034). By STT's calculation, that compounds to roughly 90x versus 6.4x over nine years, a gap of about 14x. Optical interfaces show the same pattern: intra-DC ports grow from 50 million in 2025 to 215 million in 2030 (about 4.3x), coherent ports from 970,000 to 3 million (about 3.1x).

2. Breaking down the three scales: all inside the same fence

The shared vocabulary at OCP this week is scale-up, scale-out and scale-across. Nokia lined up the purpose and the bottleneck of each in one table:

Dimension	Purpose	Bottleneck
Scale-up	Bigger system	Compute density
Scale-out	Bigger clusters	Cluster capacity
Scale-across	Bigger facility	Building size

The line under the table is the real point: all three scale capability within a single operator domain. Wagar said Nokia internally calls the next step scale-beyond (no product launched yet): taking AI traffic out of the data center and across different operators.

▲ Nokia's distributed AI grid: scale-beyond uses IP routing and optical transport to connect devices, base stations, enterprises and AI factories (Image source: Nokia | 2026 OCP Global Summit presentation (p. 8))

Wagar gave three reasons inference will spread out: sovereignty, regulation and data privacy. Inference itself is also being disaggregated into GPUs, data platforms, prefill and decode, KV cache and CPUs, and every plus sign between them is a network. Nokia also works with NVIDIA on AI-RAN, which lets third-party workloads run on a GPU at the base station, such as an RTX 6000. In other words, the places where inference can happen now stretch from the AI factory all the way to the cell site.

3. Breaking down CIR: only one of four tiers fell behind

The weightiest slide of the session is the Compute Intensity Ratio (CIR) from Lightstorm Chief Product Officer Prasanna C: how many bytes each interconnect tier can move per FLOP, in millibytes per FLOP (mB/F), with the H100's 989 TFLOPS of dense BF16 as the denominator.

Tier	Interconnect	CIR (mB/F)	vs NVLink
Memory	HBM	3.39	3.7x above
Scale-up	NVLink	0.910	Baseline
Scale-out	InfiniBand NDR / 400G Ethernet	0.006-0.050	18-152x below
Beyond	Conventional WAN (1-10 GbE)	0.0000158-0.000158	5,760-57,600x below

▲ CIR of the four interconnect tiers on a log scale: the first three roughly keep pace with compute, while WAN lags NVLink by four to five orders of magnitude (Image source: Simple Tech Trend original, data from Nokia/Lightstorm 2026 OCP Global Summit presentation)

The argument of this table lies in its shape, not its values. The three tiers the industry funded have tracked compute: the B200 generation holds on-package CIR at about 3.56 mB/F, close to the H100's 3.39. The WAN is the one tier that has not kept up, and it is the tier every distributed workload has to cross. The ranges reflect link class and oversubscription; the method is in arXiv:2608.14967.

4. 429x is the ceiling; 58 Gbps is the more honest starting point

The second key slide compares two workloads on the same link:

• Training: 5 Gbps. Gradient exchange is periodic and latency-tolerant, so the WAN can absorb it. The slide notes that this is an assumed rate, not a measured one.

• Inference: 2,147 Gbps. A 70B model with multi-head attention (MHA) and 4K context must move 10.74 GB of KV cache inside a 40 ms time-to-first-token (TTFT) budget.

▲ Cross-site demand for training is about 5 Gbps, while the ceiling for moving KV cache in inference is 2,147 Gbps, a 429x gap; GQA brings it down to 268 Gbps and latent attention to about 58 Gbps (Image source: Lightstorm/Nokia | 2026 OCP Global Summit presentation (p. 11))

What makes this slide unusual is that it marks its own number down: grouped-query attention (GQA, 1/8 KV heads) needs 268 Gbps and latent attention about 58 Gbps. Measured against one 800G coherent wavelength (STT calculation), MHA needs about 2.7 wavelengths, GQA about a third of one, and latent attention less than a tenth.

▲ Cross-site bandwidth demand for the same 70B model falls from 2,147 Gbps to about 58 Gbps depending on the attention architecture, measured against an 800G coherent wavelength (Image source: Simple Tech Trend original, data from Lightstorm/Nokia 2026 OCP Global Summit presentation)

So the slide's real claim is that cross-site state transfer stops being exotic, not that everyone needs heroic pipes. As model architectures move to GQA and latent attention, scale-beyond bandwidth needs drop directly; but as inference spreads across more sites, total volume still rises. The flip point between moving KV cache and recomputing it is exactly what the field test is meant to answer.

5. Where Nokia sits: selling IP and optics together

Wagar called this his "one product slide", and the point is the combination:

• IP routing: the 7750 service router is a workhorse of carrier backbones; data center switching uses Broadcom merchant silicon (stated in the talk) and has reached hyperscalers and neoclouds.

• High-performance platforms: 800 Gb/s coherent routing and high-density 800GE ports.

• Optics: Nokia combined with Infinera last year and offers coherent pluggables, full-band transponders, Super-C + Super-L expanded-spectrum line systems and direct-attach hyperscale OLS.

• Automation and security: a K8s intent-based policy platform, quantum-safe networking and DDoS mitigation.

This is also why Nokia is pushing scale-beyond: DCI and backbone business happens in the IP and optical layers at the same time. We covered the supply-chain side of DCI demand in Earnings Highlights: Fabrinet (FN) Q4 FY2026.

6. A three-site test in India: turning modelled numbers into measurements

Lightstorm's motive is practical: India uses a lot of AI, but most of the compute sits outside the country, so the load falls on the fiber and IP networks coming in. The plan is a new deep-buried production corridor in India with three sites and multi-vendor open line systems under interoperability test, carrying real RDMA-class traffic and measuring ten metrics split across three parties:

• The operator: provisioning and capacity-step time, delay variation, protection percentiles, multi-tenant isolation and unit cost

• The vendors: how line systems and packet fabrics behave under RDMA-class traffic

• The inference ecosystem: measured TTFT, GPU utilization and the transfer-versus-recompute flip point

The slide says the pilot window opens in late 2026; in the talk Wagar said the proof of concept should come online within a week or two, with other global tier-one operators taking part. Results will be published either way.

This follows the same approach as OIF Draws an Official Map for AI Interconnect: agree on the measurement method and a common language first, then talk products.

7. Risks and counterarguments

• Every number is a model: 5 Gbps is assumed, and 2,147 Gbps is the MHA bound with propagation excluded; until measured, both are analytic values.

• The CIR baseline widens the gap: comparing 1-10 GbE with an H100 represents a conventional enterprise WAN, not the 400ZR or 800ZR DCI hyperscalers already run. With coherent DCI the gap would be far smaller, and the slide does not give that figure.

• Model evolution may solve part of the problem: latent attention already brings demand down to about 58 Gbps, and with KV cache compression, scale-beyond may stay "needed but not urgent" for a long time.

• Cross-domain standards take years: the Metro Ethernet Forum took years to standardize its services, and there is no timeline for an AI equivalent.

• Likely unit error: the product slide reads "11.1Ghz Super-C + Super-L", which by convention should be spectrum on the order of THz; check against Nokia's official material before publishing.

8. What it means for Taiwan's supply chain

• Volume stays intra-DC. The slide's 2030 forecast is 215 million intra-DC optical ports versus 3 million coherent ports, a gap of about 72x. Taiwan's optical supply chain competes mainly in intra-DC optical modules and components, and scale-beyond does not change that ratio.

• Taiwan has little visibility in the coherent layer. The value of scale-beyond sits in routers, coherent DSPs, tunable lasers and line systems; after combining with Infinera, Nokia has its own InP photonic integrated circuits in a vertically integrated model. Neither the slides nor the talk mentioned any Taiwanese company.

• What Taiwan's carriers should consider is Lightstorm's question itself. Like India, Taiwan has a large share of AI inference served from abroad; if sovereign AI and local inference take shape, cross-site bandwidth, delay variation and multi-tenant isolation will become procurement specifications. For now this is a direction to watch, not orders that have happened.

Further reading: ECOC 2026: No Tech Revolution in the Next Decade, Only a Scaling War.

9. Conclusion

The session launched no new product; its value is that it pushes the AI network bottleneck out by one tier. For two years the industry has focused on scale-up inside the rack and scale-out inside the building. Nokia and Lightstorm use CIR to show that the tier lagging by four to five orders of magnitude is the WAN, and that inference, not training, is what drives it.

But the timing needs to be separated. 429x is a ceiling; GQA and latent attention already bring demand down to 58-268 Gbps, and the India test only opens in late 2026. Over the next 12 months, watch three things: the pilot's measured TTFT and transfer-versus-recompute flip point, whether an MEF-style cross-operator AI service specification emerges, and whether coherent port growth can catch up with intra-DC.

Bandwidth inside the data center has caught up with compute; what needs to catch up next is the fiber between data centers.

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