OCP Global Summit 2026 | Cisco | Rakesh Chopra, Tom Wiliams | Breaking the Walls: Open Scale Across for the AI Era

1. What scale-across is: not just faster DCI

Cisco Fellow Rakesh Chopra opened the session, Vipul Deokar, VP of Core Network Infrastructure at Meta, took the stage mid-way to describe operations at scale, and Tom Wiliams, Cisco VP of Product Management, closed with the optics section.

Cisco's definition is clear: AI infrastructure scales up inside a rack, scales out across racks, and finally scales across data centers. Traditional DCI connects loosely coupled applications, CPU to CPU, to storage and to users, with many mixed, bursty, asynchronous flows where a network event affects only individual services. Scale-across is GPU to GPU: fewer, sustained, synchronized high-bandwidth flows, and when one slows down the whole training job stalls.

▲ Cisco splits cross-site networking into two tiers, traditional DCI and scale-across, with the latter sitting directly on the critical path of distributed AI workloads (Image source: Cisco | 2026 OCP Global Summit presentation (13:01))

The driver is quantified too: GPT-2 used 100+ GPUs, xAI Colossus trained Grok on 200K+ GPUs, and training compute doubles roughly every 5 months (Stanford HAI 2025 AI Index). The session abstract is blunter: frontier clusters are approaching 100 MW, a single campus can no longer hold them, and GPUs must span multiple data centers tens of kilometers apart.

When the network sits in the critical path of training, every network hiccup becomes idle GPU cost.

2. Meta's breakdown: from EBB to Rings and RBB

Meta's segment was the most substantive part of the talk. Vipul first showed the footprint: 23 data center regions worldwide (19 in North America, 3 in Europe, 1 in Asia) connected by EBB (Express Backbone). The Prometheus cluster region used for its latest models mixes H-shaped halls, newer single-building designs, tents and third-party cloud capacity, and targets 1 GW by year end.

Then came two new concepts:

• Rings: overlay the fiber topology on the build plan and group neighboring regions into rings by round-trip time bands of 1, 5 and 10 ms. In the slide's words, latency follows fiber, not geography.

• RBB (Regional Backbone): regions within a ring share one service stack (database, cache, queue, warm storage) while others host only compute, so new regions come online faster. The cost is that local reads become backbone traffic: intra-ring capacity is 10x inter-ring, and the first trial pair already freed several MW of baseline power.

A detail worth noting: the RBB is labeled shallow buffer on the slide. In other words, Meta runs general regional traffic on shallow-buffer FBOSS white boxes, and deep buffers are really needed in the scale-across tier covered next.

3. The reach breakdown: three thresholds at 3 km, 10 km and 120 km

Meta offered a very practical reach ladder:

• Within 3 km: FR modules are enough

• Within 10 km: LR modules

• Beyond 10 km: ZR coherent modules plus DWDM are required

Beyond 10 km, Meta connects 10,000s of 800G ZR on each side from the frontend and backend switches, through optical protection switches and muxes, over a working path and a protection path. Without in-line amplifiers (ILAs) the link reaches 120-150 km, and with them 500 km+. Building protection into the optical layer means ZR and IP ports do not have to be duplicated for redundancy.

▲ Meta's beyond-10 km scale-across architecture: tens of thousands of 800G ZR modules at each end, through optical protection switches and multiplexers, reaching 120–150 km without ILAs (Image source: Meta | 2026 OCP Global Summit presentation (09:45))

▲ Three reach thresholds for scale-across: 3 km FR, 10 km LR, and beyond 10 km a switch to ZR plus DWDM, reaching 120–150 km without ILAs and 500 km or more with ILAs (Image source: Simple Tech Trend original, data from Cisco's 2026 OCP Global Summit presentation)

The industry meaning of this ladder: past 10 km, the optical bill of materials switches from IM-DD to coherent. DSPs, narrow-linewidth lasers, coherent receivers and C+L band amplifiers form an entirely different value chain. Cisco also ran a live scale-across demo over 80 km of fiber at its booth.

4. The routing silicon breakdown: why deep buffers are back

Rakesh tied Cisco's system lineup to one chip: the P4-programmable Silicon One P200, stressing that programmability, deep buffering, congestion response, security and telemetry only deliver value together. The same P200 comes in two form factors:

• Fixed (Cisco 8223, Nexus 9000): 51.2T per system, smaller failure domain, independent node increments

• Modular (Cisco 8800, Nexus 9800): 28.8T per line card, maximum radix, fewest devices to manage

An STT estimate explains why deep buffers matter: light travels about 5 µs per km one way in fiber, so 100 km adds roughly 1 ms of round-trip time, hundreds of times more than inside a campus. For a 51.2T system, a 1 ms bandwidth-delay product means about 6.4 GB in flight (STT estimate, not a slide figure). A shallow-buffer chip facing congestion at that distance can only drop packets, and a single dropped packet in synchronized training traffic can hold up the entire job.

That is Cisco's commercial thesis for this session: scale-out belongs to shallow-buffer Ethernet switch silicon, but scale-across gives deep-buffer routing silicon a new ticket in.

5. The optics breakdown: ZR+ eats transponders, multi-rail OLS eats racks

Tom's optics section carried two sets of hard numbers. The first is coherent pluggables replacing gray optics plus transponder shelves: the router takes ZR+ directly and the transponder in the middle disappears.

▲ ZR+ plugged directly into routers replaces gray optics plus transponders; Cisco cites 66% CAPEX, 90% energy and over 65% space savings (Image source: Cisco | 2026 OCP Global Summit presentation (16:25))

Cisco cites ACG research showing 40-60% TCO savings depending on place in the network. The second set concerns line systems: bandwidth has grown so much that power and space at amplifier huts have become the bottleneck, and single-rail OLS can no longer keep up.

Metric	NCS 1010 (single-rail)	Open Transport 3000 (multi-rail)
Power per C+L rail	350 W	about 88 W (350 W per 4 rails)
Rails per 12 kW rack	32	128
Chassis density	1 rail per 3RU	16 rails per 5RU

▲ In the same 12 kW rack, a multi-rail OLS fits 128 C+L rails, 4x a single-rail system, with 75% lower power per rail (Image source: Simple Tech Trend original, data from Cisco's 2026 OCP Global Summit presentation)

Tom summed it up precisely: traffic has grown so large that operators no longer turn up one wavelength at a time but a complete line at a time. For the supply chain, that means the purchasing unit for coherent modules, C+L amplifiers and WSS scales up from wavelengths to rails.

6. Open standards: from 400ZR to 1.6T ZR+

The other thread of the session was openness. Cisco stressed that customers need not buy every layer from one vendor; every layer exposes open interfaces, with the NOS supporting SONiC, FBOSS, NX-OS and IOS XR, and the silicon exposed through SAI.

On coherent optics, Cisco laid out the standards progression as follows:

Type	400G	800G	1.6T
OTN	oFEC / OpenROADM	oFEC + PCS / OpenROADM	—
ZR+	oFEC / OpenZR+	oFEC + PCS / OpenROADM	oFEC + PCS / OIF
ZR	cFEC / OIF	oFEC / OIF	oFEC / OIF

Tom singled out Meta: for the 800G generation, Meta asked the industry to move to pluggables four or five years ago, before the AI rush, driving an interoperable PCS solution. Meta also confirmed that 1.6T ZR is now progressing at OIF. In other words, control over the 1.6T ZR / ZR+ spec sits with hyperscalers, not equipment vendors. That matches the DCI demand direction we saw in Earnings Highlights: Fabrinet (FN) Q4 FY2026.

7. Risks and counterarguments

• "10,000s of 800G ZR" has no timeline: the slide shows an order of magnitude on an architecture diagram, not shipment numbers, and does not say whether it covers one site pair or all of them.

• Deep buffers are not the only answer: Meta's own RBB is shallow-buffer, and congestion can also be handled at endpoints and in scheduling. Deep-buffer silicon costs more power and money, and whether customers buy it at every tier remains to be seen.

• The baseline behind the savings: 66% and 90% compare against gray optics plus transponders, and the ACG TCO figure is research cited by Cisco; against a network already fully on pluggables, the increment is much smaller.

• Fiber itself is the hard bottleneck: Meta showed construction photos of trenching and buried conduit. Right-of-way and construction for 100 km-class fiber are harder to accelerate than modules. This echoes the fiber capacity theme we covered in Earnings Highlights: Corning (GLW) FY2026 Q2.

• ASR versus slides: the auto transcript mishears several names and numbers, so every hard number in this article follows the slides.

8. What it means for Taiwan's supply chain

• The slides name no Taiwanese company. Cisco's coherent modules come from its own Acacia, and the slides call Cisco the leading supplier of pluggable modules; systems, silicon and optics are all vertically integrated, leaving no direct slot for Taiwanese vendors in this layer.

• The demand signal is certain; the beneficiaries sit upstream. 10,000s of 800G ZR, dual C+L band and multi-rail OLS translate into narrow-linewidth tunable lasers, coherent receiver components, C+L band amplifiers and WSS. Taiwan's footprint in these coherent components is still limited, which is both a challenge and an open gap.

• Within 10 km, IM-DD still rules. Meta says FR covers 3 km and LR covers 10 km, so the large volume of FR / LR 1.6T modules inside campuses is the battlefield Taiwanese optics makers know, and Cisco's live 1.6T scale-out demo sits in this tier too.

• Open NOS keeps the white-box path alive. Meta's RBB runs FBOSS on OCP SAI, and whether that white-box path extends into scale-across is worth tracking.

For the broader scale-up and scale-out context, see our earlier piece on Marvell Investor Day 2026.

9. Summary

The significance of this talk is not the P200 or any single module, but that Cisco and Meta together promoted cross-data-center networking from DCI to a tier of the AI fabric. Once the network enters the training path, three things change at once: routing silicon needs deep buffers, optics switch from IM-DD to coherent, and line systems move from single-rail to multi-rail.

Timing needs to be read separately. 800G ZR is deploying now, in the tens of thousands; 1.6T ZR / ZR+ is still progressing through OIF and OpenROADM; multi-rail OLS is a shipping Cisco product. Over the next 12 months, watch three things: when the OIF 1.6T ZR spec is finalized, whether Meta's next-generation model really trains across two regions, and whether deep-buffer routing silicon beyond Cisco follows into scale-across.

The cluster bottleneck has moved from the rack to the campus and from the campus to the fiber, and coherent optics volumes must now be counted in AI units.

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