OCP Global Summit 2026 | Ciena | Spencer Giacalone, Helen Xenos | The Rise of Scale Across

3 minutes ago

8 min read

1. Ciena's definition: scale-across runs on a dedicated coherent network

The talk was split between Spencer Giacalone, Ciena's Senior Director of Cloud AI Architecture, on architecture, and Helen Xenos, Senior Director of Product Marketing, on products. Giacalone drew the boundary up front: this week everyone has been talking about congestion control, load balancing and collective optimization, but one slide carried a heavy judgment: host and protocol fixes manage contention; they do not create the capacity that is required.

Ciena defines scale-across as extending lossless connectivity between data center campuses or regions to train a single AI model. It has two layers:

Packet layer: congestion management, fairness and load balancing at the IP layer and above
Optical layer: a dedicated coherent optical network with enough capacity to absorb large, bursty, loss-sensitive traffic

The slide states it plainly: scaled coherent DWDM is the main differentiator between scale-out and scale-across. It also names the uses: pipeline parallelism, and the fault-tolerant HSDP that Meta and others ran on 100,000 GPUs.

Why go across sites at all? Ciena cites the 2026 ScaleAcross Explorer paper from Harvard and Meta: gigawatt-scale power and grid limits keep a single location from hosting hundreds of thousands of synchronous GPUs, and the distance between buildings is set by power, grid and cooling. Giacalone added that the power quoted for a campus is its end-state nominal figure, not what is available on day one.

2. Breaking down capacity: 14x, 8:1, and the fiber as the unit of bandwidth

Ciena gave two capacity signals:

14x: in larger projects, the capacity deployed between two training campuses is 14 times what was deployed between all data centers in an entire hyperscaler front-end region around 2020
More than 8:1: citing the same paper, the slide puts scale-across oversubscription above 8:1

The split between IMDD and coherent was just as direct:

Item	O-band IMDD (campus)	C+L coherent DWDM (inter-campus)
Reach	Up to ~10 km (DR/FR/LR)	~120 km per span, 1,000 km+ with ILAs
Capacity per fiber	Single-digit Tb/s	Tens of Tb/s
Modulation	2 bits/symbol	DP-16QAM 8 bits/symbol + WDM + C+L
Fiber growth	One port per fiber, linear growth	Many router ports per fiber

The reason is that you cannot get that much fiber between campuses: a campus is one contiguous piece of land where you can pull your own fiber, while a link between campuses is not. Giacalone's conclusion: coherent is not an option for scale-across; it is a requirement. The slide also concedes that coherent costs more, which limits its use inside campuses, and that Coherent-Lite is meant to close that gap.

3. Near and far: the 150 km line

Ciena splits designs into two kinds:

Near scale-across: an extended campus within 150 km, capacity pushed to the limit, no in-line amplifiers (ILAs) needed
Far scale-across: beyond 150 km, a truly distributed cluster, high but not extreme capacity, ILAs required

▲ Ciena draws the line at 150 km, splitting scale-across into near designs that need no ILA and far designs that do; both deploy in whole fiber pairs across the full C+L band (Image source: Ciena | 2026 OCP Global Summit presentation)

Both share the same pattern: point-to-point, no add/drop, and the full C+L band lit on day one. The slide shows three implementations. Near scale-across uses ZR+ plugs in the router with an optical protection switch for N+1, so when a fiber is cut the optical layer fails over before the packet layer notices. Far scale-across is the same design plus ILAs. The third is active-active: the router emits only grey optics, and external transponders handle the coherent modulation.

The trend is what matters. Giacalone said all of Ciena's near scale-across customers are looking at or moving to far scale-across, mostly because of power. Asked about the distance limit, he said customers will go to 500 and 800 km and ultimately about 1,000 km, still for synchronous pre-training.

Ciena's 2027 trends slide adds two more signals: training-only routes will grow into an optical AI backbone that carries inference, training and front-end traffic together, and the practice of using optical protection to save router ports will extend to regional networks, DCI and even OCS inside the data center. For where OCS stands inside the data hall, see ECOC2026 | 95% of OCS shipments are Google's business.

4. Breaking down the amplifier huts: 22 down to 1

Once distances grow, the bottleneck moves to the ILA sites. Xenos explained that a traditional amplifier hut is a small building with modest capacity needs: one chassis amplifies one fiber pair, with 4 rails per rack and 16 rails per hut.

▲ A 20 Pb/s AI network needs 22 amplifier sites with a conventional line system, but only 1 with RLS Hyper-Rail at 128 rails per rack, a 32x density gain (Image source: Ciena | 2026 OCP Global Summit presentation)

A 20 Pb/s network built the conventional way needs 22 huts. Ciena's RLS Hyper-Rail puts 16 C+L fiber pairs in a single chassis; Xenos said orally that this is 16 times the previous density at 75% lower power, and a full rack reaches 128 rails, so one hut is enough. She said the product is "close to generally available" and described line systems as having seen no real innovation in about 30 years.

Notably, Cisco's multi-rail OLS at the same summit (Open Transport 3000) also reaches 128 rails per rack, within a 12 kW rack. Two coherent system vendors have converged on the same density, which suggests multi-rail is now table stakes for scale-across rather than a differentiator. The real separation comes in the next section, at the transponder.

5. The full-spectrum transponder: moving coherent out of the router

Today, deployment happens one wavelength at a time. Filling one fiber pair takes 128 x 400G x 2 pluggables, or 64 x 800G x 2. Ciena's 20 Pb/s case involves "hundreds of fiber pairs", about 390 by STT's estimate. With 800G pluggables only, that is 64 per end per fiber pair, so 20 Pb/s (taken as per-direction capacity) means about 25,000 per end and about 50,000 across both ends, each one to be turned up and equalized.

Ciena's answer is the FST (Full Spectrum Transponder): one box lights an entire fiber pair at once through a single line port, while the router goes back to grey optics. The slide lists the benefits: less reliance on DWDM expertise, cookie-cutter deployments, and higher performance without pluggable power limits. With 51.2T FSTs, 20 Pb/s takes about 780 units (STT estimate).

▲ Deployment counts for 20 Pb/s scale-across: 22 amplifier sites vs. 1, and about 50,000 800G pluggables vs. about 780 FSTs (Image source: Simple Tech Trend original; data from Ciena 2026 OCP Global Summit presentation)

Ciena sketched a 51.2T C+L FST concept: four 8-channel service modules, each with eight 1.6T modems, with client fibers coming out of two 6.4T NPOs per module (the previous slide also lists ACC, CPO and grey pluggables as client options).

▲ The 51.2T C+L full-spectrum transponder concept: 4 service modules and 32 1.6T modems, with 6.4T NPO on the client side (Image source: Ciena | 2026 OCP Global Summit presentation)

The diagram says two things. First, the 6.4T NPO drawn on the client side is in the same class as the 6.4T optical engine Ciena pushed at ECOC, which we covered in ECOC2026 | Ciena uses Open CPX to turn 6.4T optical engines back into off-the-shelf parts. Second, Giacalone named "routers moving to CPO" as one case for external transponders. STT's inference is that a CPO router's front panel no longer takes pluggable optics, so there is no slot for ZR+, and coherent has to move outside. Our view: near scale-across remains ZR+ territory, while far scale-across combined with CPO routers is where FST has its opening.

Ciena's roadmap also runs in three steps: first replace ILAs with Hyper-Rail, then remove the terminal ROADM where the link is point-to-point and fully filled, and finally replace the transponders and channel mux with FST.

6. 1e-24: asking for a provably lossless optical layer

The closing call to action carried the most technical weight. Optical networks always have ambient bit errors that FEC corrects, but coherent FEC was designed for loss-tolerant applications, so lossless transport falls back on retransmission. Ciena listed three levers: better fiber (already done), more modem power for stronger FEC, or limiting distance (not desired).

The proposed solution is a provably lossless physical layer, for example 1e-24 post-FEC BER. An STT estimate shows the scale: taking 1e-15 as a reference baseline (an STT assumption; Ciena gave no current figure), 20 Pb/s sees about 20 bit errors per second; at 1e-24, about one every 1.6 years.

▲ At 20 Pb/s, a post-FEC BER of 1e-15 means about 20 errored bits per second, while 1e-24 means about one every 1.6 years (Image source: Simple Tech Trend original; data from Ciena 2026 OCP Global Summit presentation)

Ciena asked OCP to do three things: define a contiguous campus, standardize the definition of scale-across, and standardize the FEC requirement for lossless connectivity. James Harley leads this work at Ciena.

7. Risks and counterarguments

FST is still a concept: the slides call it a "future" product and the 51.2T design "what it could look like", with no timeline.
A bigger failure domain: one FST carries 51.2T, so losing one module takes out 12.8T. Xenos's answer was to split it into four field-replaceable modules or add N+1 inside each module, at a cost in money and space.
The 75% power figure is oral only: the slides give no comparison baseline.
The cost of 1e-24 is not quantified: stronger FEC adds modem power and latency, and shorter distance runs straight against customers heading toward 1,000 km.
Inference may reshape demand: Ciena expects inference to be larger in total than scale-across, but more distributed and meshier, so the fill-the-whole-fiber model may not apply.

8. What it means for Taiwan's supply chain

No Taiwanese companies were named in this talk. Coherent DSPs, C+L amplifiers, Raman and line systems are largely held by system vendors such as Ciena and Cisco plus a few coherent DSP suppliers, and Taiwan's direct share of this segment is limited.
If FST takes hold, value shifts toward grey optics. With routers back on grey optics and FST clients on NPO, part of the coherent pluggable demand turns into IMDD optics and NPO engines, ground that Taiwan's module and component makers know better. That depends on FST actually reaching volume.
Near-term demand is still in coherent pluggables. Ciena said 800G coherent pluggables are already used for scale-across today and now need C+L dual-band versions; we covered the supply-chain side of DCI demand in Earnings Highlights: Fabrinet (FN) | Q4 FY2026. For Taiwanese companies, the thing to watch is whether L-band coherent modules and related passive components show up in revenue commentary, not the FST concept drawing.

9. Conclusion

The value of this talk is that Ciena pulled scale-across back from a protocol problem to the physical layer: capacity comes from filling whole fiber pairs, distance from ILA site density, and reliability from error rates like 1e-24 that metro networks never needed.

Set side by side with Cisco (STT's comparison), the split is clear: Cisco wants coherent inside the router, while Ciena has prepared a path to move it out for far distances and CPO routers. Customers moving to 500 to 1,000 km and routers moving to CPO both favor FST, but today FST is still a concept drawing.

When the fiber becomes the unit of bandwidth, scale-across is won not on per-wavelength speed but by whoever can light the next fiber pair fastest.

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