Replay Available

Meet the Speakers

Sanjai Parthasarathi – Chief Marketing Officer – Coherent Corp.

Lorenzo Ghioni – VP of Product Management & Technical Marketing – Cisco

Kyle Hollasch – Lead Analyst Transport Hardware – Cignal AI

Presentation Summary

Compiled with AI assistance from the session transcript.

Key takeaways

• Scale-across spending is mostly still ahead. Cloud drove record optical spending last quarter with almost no scale-across contribution yet – Cignal AI models roughly $9 billion of scale-across spend by 2030.
• Back-end bandwidth dwarfs traditional DCI. Kyle’s hedged range is 5-20x traditional DCI; Cisco’s internal engineering estimate puts scale-across traffic at 14x front-end datacenter traffic. Latency and error tolerances tighten at the same time.
• Line systems take the majority of scale-across spending by 2030. Pluggables lead in 2026-27, then longer routes and 800ZR-to-1600ZR price-per-bit compression flip the mix toward Layer 0, where prices are holding or rising.
• Multi-rail amplification answers the hut wall. Hundreds of parallel fibers hit power and space limits in roadside huts; both vendors framed pooled, arrayed amplification as the fix, with Cisco quantifying the chassis math and Coherent detailing the component set.
• The technology trickles down. The same investment accelerates IPoDWDM, open line systems, thin transponders and day-one L-band – operators without scale-across requirements still benefit.

What is scale across?

• Definition (Kyle) – one training cluster spread over many sites, made to behave like a single site. Frontier models have outgrown what one building can physically hold and power.
• The technology – coherent pluggables directly in switches, carried over line systems and many parallel fibers. Kyle declined to call it IPoDWDM since little IP runs on the back end: “pluggables over OLS, lots of bandwidth, lots of fibers.”
• Where it sits (Lorenzo) – Cisco’s normalized traffic ladder: WAN traffic in and out of the datacenter = 1, scale-up (within the rack) roughly 500x, scale-out (spine-leaf fabrics scaling to ~50,000 GPUs as one system – roughly GPT-5 training scale, and about the practical single-site maximum) 56x, and scale across at 14x front-end traffic between sites.
• Metro first, then regional – the shortest routes light up first; Kyle cited a common consensus that routes around 500 km will become typical as nearby datacenter pairings are exhausted.

Market context (Cignal AI data)

• Record cloud spend: $2 billion in the quarter, up 63% year over year. Cloud is now 38% of worldwide optical spend, up from 28% a year ago, and pluggables were nearly half of that growth.
• North America dominates – nearly half of worldwide spend, up from a third a year ago; within North America, cloud is almost three-quarters of spend. Roughly four companies, with a small tail, outspend hundreds of service providers.
• Cloud demand is understated – MOFN network builds appear as service provider revenue, but the demand behind them is entirely hyperscale.
• Long haul grew 20% – AI is the tide lifting all boats, but today’s long haul (transponders, ROADMs) has little to do with scale across. The scale-across wave is still arriving.
• 800ZRx today: the vast majority of shipments go into scale across. Cisco Acacia is by far the leader, with Ciena and Nokia ramping quickly. Kyle tied Marvell’s miss on early volumes to its decision not to implement standardized OpenROADM, while expecting them to catch up.
• The crossover: scale-across bandwidth passes front-end DCI in late 2027 or early 2028 – and both keep growing. By 2030, 1600ZRx optics, which do not ship today, become the single largest slice of scale-across spend.
• Supply is the Layer 0 story: line systems have fewer suppliers than pluggables, component lead times run 12-18 months if supply exists at any price, and the bottleneck traces back to indium phosphide wafers for pump lasers. Amplifier and line-system pricing is holding or rising while merchant pluggable ASPs keep falling.
• Line system landscape: Ciena was sole supplier of the early scale-across metro systems (predominantly Meta) and announced the first hyperscaler multi-rail purchase order for its RLS HyperRail. Cisco pairs mature DCI- and RON-oriented line systems with its Open Transport 3000 multi-rail and announced a hybrid RON-MOFN win. Nokia is regaining line-system share with a hyperscaler multi-rail design win for its 1830 GX. Pluggables and line systems are independent races – there is no bundling anymore.

Beyond hyperscale: the trickle-down

• Thin transponders take off – pluggables are displacing embedded transponders, starting at 400G and accelerating in the 800G generation, pushing embedded designs to ultra-long-haul and subsea (Kyle).
• L-band for everyone – the 120 GBaud generation (800G pluggables and 1.2T+ embedded interfaces) ships with L-band from day one, and integrated C+L ROADMs are ramping. L-band is becoming standard in nearly any long-haul network.
• Rural operators are building hyperscale networks – MOFN demand is pushing hyperscaler-grade builds to operators that historically bought traditional gear, as datacenters land far from major markets.

Cisco: make deployment match AI’s pace

• AI traffic is not a bubble – Lorenzo’s view is that inference growth plus agent-to-agent traffic will keep compounding demand well beyond training.
• One solution spans both – the same platforms should handle traditional DCI and scale across despite the differences in distance, so buyers are not forced into parallel architectures.
• The hut math: legacy sites built to a 3 kW budget support 4-5 C+L line amplifier sites today; multi-rail integration lifts that to 16 parallel C+L amplifier sites in one chassis at the same 3 kW, evolving toward 128 rails of C+L within the 12 kW envelope used for new builds.
• Automation is the real unlock. The arc runs from ROADMs (removed route planning pain) to tunability (removed wavelength ordering pain) to OTDR, dynamic gain and channel monitoring – and now fiber-type detection, letting systems read the fiber plant and automatically set per-channel power, spectral density and gain profiles across open, multi-vendor line systems and DCOs.
• Deskilling is deliberate – optical experts are scarce, so systems must deploy and auto-regulate without one on site.
• Supply pressure is structural – a pump laser now ships in every DCO to reach 0 dBm launch power, multiplying pump demand by an order of magnitude versus the line-system-only era. Cisco has locked in capacity through long-term agreements.
• Technology is only half the problem – rights of way, permits, NIMBY resistance, fiber availability and licensing rules mean MOFNs and local operators are essential. Hyperscalers are now buying entire line systems as managed services, not just wavelengths.

Coherent: multi-rail is resource pooling, not replication

• Everything goes optical – rack-to-rack connectivity is already fully optical, within-rack is starting to convert, and chip-to-chip is coming: “it’s hard to fight the physics.”
• Pushback on the pluggables-get-the-glory narrative – Sanjai respectfully disagreed with Kyle on where innovation lives, citing Coherent’s transport lineage: the merchant EDFA business acquired from Corning in 2003, optical channel monitors, pioneering embedded OTDR, dual-chip pumps and a pluggable optical line system. Kyle conceded the point in closing – the engineering is so good it turns invisible at the systems level.
• The multi-rail case: scaling a hut’s capacity 4x with conventional gear means 4x the equipment, power and space – which the site may simply not have, before regulatory hurdles even enter. Multi-rail pools resources instead of replicating them.
• The arrayed component set, introduced over the past 6-9 months: arrayed amplifiers handling 8 fibers (4 bidirectional pairs) in a single amplifier’s footprint with sublinear power scaling, plus multi-port DGE, multi-port channel monitors and multi-port OTDR to match.
• Pump innovation: first to introduce the dual-chip pump, now four chips in a single uncooled package – part of a roughly 40x industry gain in optical power density on GaAs platforms over the past 14 years.
• C+L is not just a wider band – coatings, insertion loss and isolation all must be re-optimized across the full range, demanding materials-level work. Coherent’s March merchant releases (Nano ITLA and IC-TROSA) target C+L and extended-C applications.
• Vertical integration is the moat – optimizing from materials up through modules is what makes arrayed, C+L-capable multi-rail components possible.
• Next: hollow-core and multi-core fiber, with more to come in the coming months.

Themes from the Q&A

• Delivery beats specs right now. Asked what customers push hardest on beyond cost and power, Lorenzo answered delivery – bandwidth where and when it’s needed. Sanjai pointed to ramping every product line at scales the industry has never seen, with reliability treated as a given.
• Inference doesn’t kill training. Kyle argued the whole pie grows: bigger models make tokens more valuable, and hundreds-of-megawatt datacenters will always outnumber multi-gigawatt ones, so they will need connecting. Lorenzo agreed on growth but sees tokens becoming “the next currency,” with workloads routed across models and agents to optimize token cost, citing Cisco’s Jeetu Patel.
• C+L is a continuum. Sanjai sees some deployments arriving C+L-capable on day one while others extend the C-band first and add L when the fiber runs out – a big change from when L-band was scarce and costly.
• Giant datacenters don’t end scale across. Lorenzo: gigawatt-class sites face power reliability, cooling and concentration risk – and their sheer in-and-out traffic becomes scale-across-like anyway. Remote siting only adds transport demand.
• Distance limits are dissolving. Lorenzo offered no hard number but reported that hyperscalers say protocol evolution and fast failure signaling mean distance is no longer the constraint on synchronous training. Andrew noted reports out of China of roughly 100 km deployments.
• Line systems near DCI in dollar terms. Sanjai noted that Kyle’s own market models put line-system spend almost on par with DCI – fitting for a market he has spent most of his career in.

What to watch

• The 1600ZR transition around 2028 – a product that doesn’t ship today becomes the largest slice of scale-across spend by 2030.
• The bandwidth crossover in late 2027 or early 2028, when scale across passes front-end DCI – with both still growing.
• Raman everywhere? If hyperscalers deploy Raman uniformly for ease of deployment rather than selectively per span, Kyle sees potential for another fifth of spend – and more pump laser strain.
• Pump laser and InP supply – 12-18 month lead times and long-term supply agreements will shape who ships.
• Hollow-core and multi-core fiber – Coherent hinted at announcements in the coming months.