Optical Networking

Where Copper Stops Scaling

AI clusters are outgrowing copper. A single-rack system with 72 GPUs can still rely on copper backplanes at sub-meter distances (NVIDIA's NVLink does exactly this). But the industry is moving to multi-rack deployments with hundreds or thousands of GPUs, and at 800G and 1.6T signaling rates, copper degrades beyond about one meter of reach.¹ Once a cluster scales past the rack, the network shifts to optics. Not as an upgrade, but as the only option that maintains bandwidth at distance.

Reach is only half the problem. Power is the other. In a fully loaded 32-port 800G switch using traditional DSP-based pluggable optics, the transceivers alone can consume more than 500 watts, more than the switching ASIC itself. At cluster scale, that adds up quickly: each GPU requires roughly six optical transceivers, and across one million GPUs, transceiver power alone approaches 180MW.² Networking already accounts for about 5% of total data center energy use, and that share is expected to rise as bandwidth scales.

Optics solve both the reach and power problem more cleanly than copper. Signals carried over fiber maintain bandwidth across much longer distances, without the resistive loss that makes high-speed copper links hard to extend beyond the rack. Fiber can also carry multiple wavelengths at once, pushing far more bandwidth through a much smaller physical medium.¹ The result is a network that scales farther, denser, and more efficiently than copper can at the same speeds.

The AI-Specific Demand Multiplier

The shift from traditional server workloads to AI training and inference clusters fundamentally changes optical demand. In a conventional data center, one server requires one optical connection. In a modern AI cluster, each GPU requires approximately six high-speed optical transceivers in a three-tier fabric for scale-out networking.³ This ~6x attach rate multiplier is one of the most important demand drivers in the sector.

The transition is already well advanced: over 80% of hyperscale data center rack-to-rack links now use optical solutions.¹ As AI clusters grow from single-rack configurations (72 GPUs with NVL72) to multi-rack deployments (576+ GPUs), more links exceed copper's practical reach. The question is not whether optics will matter more (that is already happening) but how far inward optical will push, and on what timeline.

The Technology Roadmap

In prior optical cycles, one speed generation ramped, plateaued, and was then replaced by the next. This cycle is more compressed: 800G, 1.6T, and 3.2T are ramping on top of one another rather than in sequence. That overlap is what makes the current setup unusual. It sustains demand across more of the supply chain and delays the inventory digestion that ended prior cycles. Whether that holds through 2027-2028 is the central question for the sector.

800G: The Current Workhorse

800G pluggable modules in QSFP-DD800 and OSFP form factors are now the default for new AI data center buildouts. Over 24 million 800G+ transceiver units shipped in 2025, with 63 million projected for 2026, a 2.6x increase.⁴ Pricing has settled at $360-450 per module depending on reach, with continued room to decline as manufacturing scales.

The 800G ramp is the largest generational deployment in optical transceiver history, driven by hyperscaler AI cluster buildouts that require thousands of transceivers per facility.

1.6T: The Next Wave

2026 is the year 1.6T enters volume production. The enabling technology is the 200G-per-lane EML laser chip, which doubles the lane rate from the 100G-per-lane architecture used in 800G modules. Cignal AI projects over 5 million 1.6T units will ship in 2026, with Nomura forecasting 20 million by year end.⁵ Current pricing sits at $1,300-1,500 per module, expected to decline to approximately $1,100 within two years.

The IEEE 802.3dj standard covering 200G/lane PAM4 signaling is expected to be finalized by mid-2026, which will further catalyze the ecosystem.⁶ Google and Amazon are qualifying 1.6T modules for their next-generation AI clusters, and AAOI secured a $200M+ volume order for 1.6T transceivers in March 2026.⁷

3.2T and Beyond

3.2T transceivers require 448G PAM4 SerDes, expected to become available in 2027 with volume manufacturing in 2028. Broadcom's Taurus DSP platform and Coherent's 400G/lane demonstrations at OFC 2026 point to a viable path.⁸ At 3.2T speeds, pluggable modules approach thermal limits, which is when co-packaged optics becomes essential rather than optional.

LPO: The Power Fix

Linear Pluggable Optics (LPO) removes the DSP from the transceiver module entirely, replacing it with high-linearity analog components and shifting signal processing to the host switch ASIC. Traditional 800G modules with DSPs consume 14-17W; LPO modules consume 7-8.5W, a 40-50% reduction, while delivering roughly 3x lower latency per hop.⁹ LPO is already shipping in NVIDIA Spectrum-X and Meta AI network deployments. In AI training clusters where collective operations span thousands of GPUs, the per-hop latency savings compound into measurable performance gains.

CPO: The Paradigm Shift

Co-packaged optics (CPO) represents the largest architectural change on the roadmap. Instead of pluggable transceiver modules connected to a switch via front-panel connectors, CPO embeds optical engines directly onto the switch ASIC package. The conversion from electrical to optical signals happens on-chip, and fiber connects directly to the switch.

Meta tested Broadcom's CPO solution for one million link hours without a single link flap, measuring approximately 5.4W per 800G of bandwidth delivered versus 15W for a pluggable equivalent, a 65% reduction.¹⁰ NVIDIA's Spectrum-X Photonics, arriving in the second half of 2026, claims 3.5x power efficiency improvement and 63x better signal integrity.¹¹

CPO penetration remains minimal today, estimated at roughly 0.5% of optical modules in AI data centers in 2026. TrendForce projects 35% penetration by 2030.¹² The near-term debate centers on serviceability: a failed optical engine in a CPO switch may require replacing the entire package, whereas pluggable modules can be hot-swapped individually. Hyperscalers are planning around this by overbuilding port capacity by 5-10%.

The Supply Chain

The optical stack is not equally valuable at every layer. The key question is where scarcity sits, where margins concentrate, and which companies control those choke points.

The EML Bottleneck

One of the clearest binding constraints in the optical supply chain is EML laser capacity. Electro-absorption modulated lasers (EMLs) are the core component in every 800G and 1.6T optical transceiver. They convert electrical data signals into modulated light, and capacity remains well short of projected demand.

McKinsey's June 2025 analysis projected that 800G transceiver production would fall 40-60% short of demand through 2027, with 1.6T shortfalls of 30-40% persisting through 2029.¹³ The root cause is manufacturing concentration: EML production requires extremely high-precision Indium Phosphide fabrication, and only a handful of companies worldwide can produce them at the quality and volume required. Lumentum is currently the only supplier shipping 200G-per-lane EMLs at volume, the critical component for 1.6T transceivers.¹⁴

NVIDIA has responded by pre-allocating laser chip capacity at major suppliers, pushing delivery timelines for competitors past 2027. The $4 billion combined investment in Lumentum and Coherent in March 2026 was explicitly aimed at securing priority access to this scarce capacity.¹⁵ When NVIDIA invests $4B to secure optical supply rather than simply placing purchase orders, it strongly suggests that the company views photonics as a structural gating factor.

Where the Margins Are

The supply chain is not uniformly profitable. Margin pools concentrate at two points: upstream laser chips and switching silicon.

The China Dynamic

Chinese manufacturers dominate module assembly. Innolight holds over 50% of NVIDIA's 800G optical module procurement and generated $3.3 billion in revenue in 2024, up 123% year-over-year, with 20-22% net margins. Eoptolink achieved 179% revenue growth to $1.2 billion and holds the highest net margins in the sector at 33%.¹⁶ Together they control approximately 60% of the global 800G transceiver market.

Their competitive advantage is straightforward: aggressive pricing (20-25% below Western incumbents), fast execution, and deep integration with NVIDIA's qualification process. The vulnerability is equally clear: they depend entirely on US-sourced DSP silicon from Broadcom and Marvell, and they struggle with leading-edge laser technology. Both companies are shifting manufacturing to Thailand and Vietnam to mitigate tariff exposure, but the engineering and supply chain remain anchored in China.

The pattern from prior generations is instructive: each new speed tier launches at premium ASPs with Western suppliers leading, Chinese suppliers follow 12-18 months later at lower prices, and margins compress. The treadmill continues at the next generation. For 1.6T, Innolight was the first to complete NVIDIA testing, suggesting the gap is narrowing.¹⁶ For Western module makers, that means today's margin window is real but unlikely to stay open indefinitely.

Where the Value Concentrates

Not all optical companies benefit equally. The value in this cycle concentrates at structural choke points, places where physics creates scarcity and substitution is difficult. Below, we group the key players by how they fit in the stack and how durable their position is likely to be.

Choke-Point Winners: Lumentum and Coherent

These two companies sit at the scarcest layer of the supply chain: laser chips. You cannot build an 800G or 1.6T optical transceiver without EML lasers, and there are not enough of them.

Lumentum holds an estimated 50-60% share in high-end EML laser chips and is currently the only supplier shipping 200G-per-lane EMLs at volume, the critical component for 1.6T transceivers.¹⁴ NVIDIA's $2 billion investment at $695.31 per share with multi-billion dollar purchase commitments underscored that position.¹⁵ The company is expanding EML production by 40% annually, targeting $100M in quarterly OCS revenue by December 2026, and was added to the S&P 500 in March 2026.¹⁷

Coherent is broader: the result of II-VI acquiring Finisar (2019) and the original Coherent (2022), it spans InP lasers, silicon photonics, transceivers, and optical circuit switches.¹⁸ Its data center book-to-bill ratio exceeds 4x, with InP lasers sold out through 2027. At OFC 2026 it demonstrated a 6.4T socketed SiPho CPO system. Coherent also received $2B from NVIDIA.¹⁵

Key risks: Customer concentration is rising as NVIDIA becomes a dominant buyer for both. The CPO transition could change the form factors these companies sell, though laser chips remain essential regardless of packaging. Coherent's breadth cuts both ways: executing across all layers simultaneously during hypergrowth is demanding, and its 1.7x debt leverage is higher than peers.

Execution Bets: AAOI and Fabrinet

These companies benefit from the demand cycle but face more competitive pressure. Their position is less structural and more dependent on execution and customer relationships.

AAOI is one of the highest-risk names in the sector. It designs and manufactures its own InP laser chips, which in an era of laser scarcity transformed it from a marginal player into a serious contender.⁷ Revenue is targeting a doubling in 2026, anchored by a $200M+ order for 1.6T transceivers. A new $150M facility in Sugar Land, Texas is dedicated to 1.6T/3.2T production. But the stock trades at roughly $114, well above the $66.80 average analyst target, and five executives sold shares in January 2026.¹⁹ The transformation is real but unproven at scale.

Fabrinet does not design optical products. It manufactures them, producing transceivers and modules for NVIDIA, Cisco, AWS, and Lumentum.²¹ Its first billion-dollar quarter (+36% YoY) and clean balance sheet make it a strong execution story. But as a contract manufacturer, Fabrinet's throughput is constrained by upstream EML laser availability regardless of its own capacity additions.

Defensive and Infrastructure: Corning and Ciena

These companies are less exposed to the transceiver technology cycle and more tied to the physical buildout itself.

Corning is probably the most defensive position in the optical sector. It holds a near-monopoly in high-performance fiber, and fiber demand grows regardless of which optical technology wins: pluggable or CPO, DSP or LPO, Lumentum or Innolight. Meta's Louisiana data center alone requires eight million miles of fiber. The $6B Meta deal is the largest customer commitment in Corning's history, and at OFC 2026 it unveiled multicore fiber offering 4x capacity per strand.²²

Ciena operates at the transport layer: WDM line systems and coherent transponders connecting data centers to each other. Its $7B backlog provides unusual visibility, with most orders scheduled for FY2027. Direct cloud providers are 42% of revenue, and neoscalers (OpenAI, Anthropic) are emerging as new buyers. Ciena operates in a duopoly with Nokia-Infinera and was added to the S&P 500 in February 2026.²⁰ Despite beating Q1 estimates, the stock fell 13% — illustrating how elevated expectations have become.

Platform Leverage: The Broader Ecosystem

These companies are not pure-play optical, but they sit at critical nodes in the stack and benefit significantly from the optical buildout.

Broadcom controls over 80% of high-end switching silicon and supplies the PAM4 DSPs in virtually every 800G and 1.6T transceiver. Its Tomahawk 6-Davisson is the first 102.4 Tbps switch with integrated CPO.¹⁰ Marvell's $3.25B acquisition of Celestial AI brought "Photonic Fabric" technology in-house, with AWS backing the deal.²³ Both companies sit at multiple choke points and have visibility into demand across the chain.

The Numbers

The financial case for the optical networking sector rests on three pillars: unprecedented hyperscaler capital expenditure, broad-based earnings acceleration, and order backlogs that provide multi-year revenue visibility.

Hyperscaler Capex: The Demand Engine

Sources: Futurum Group, Goldman Sachs, CreditSights²⁶

Approximately 75% of 2026 capex ($450B+) is directly AI-related: servers, GPUs, data centers, and networking equipment. Capital intensity has reached 45-57% of revenue, historically unprecedented. Goldman Sachs projects cumulative 2025-2027 hyperscaler capex of $1.15 trillion, more than double the $477 billion spent in 2022-2024.²⁷

Earnings Acceleration

Order Backlogs

The sector's order visibility is unusually deep. Ciena's backlog increased by $2 billion in a single quarter to a record $7 billion, with most new orders scheduled for FY2027.²⁰ Coherent's data center book-to-bill ratio exceeds 4x, and its InP lasers are sold out through 2027.¹⁸ Lumentum's OCS backlog exceeds $400 million, with multi-hundred-million dollar CPO orders in hand.¹⁷ Corning is concluding multi-year, multi-billion dollar agreements beyond the $6 billion Meta deal.²² Tower Semiconductor's silicon photonics capacity is over 70% reserved through 2028.²⁵

This backlog depth provides unusual revenue visibility for a hardware sector. Most optical companies can now see 12-24 months of demand with reasonable confidence, which is why guidance has consistently surprised to the upside.

The NVIDIA Signal

On March 2, 2026, NVIDIA committed $2 billion each to Lumentum and Coherent, plus additional investments in Scintil Photonics (a French multiplexing laser startup) and Ayar Labs (optical chiplets). The investments included multi-billion dollar purchase commitments for laser and optical networking products.¹⁵

The Valuation Question

The optical networking sector's stock performance over the past 12 months ranks among the most extreme in recent market history. Whether current prices are justified is the central question for any investor considering the space.

Current Multiples

Lumentum's forward P/E ranges from roughly 50x to 60x depending on the source, well above the hardware industry median of approximately 21x.²⁸ Even accounting for rapid earnings growth, these are multiples that assume years of flawless execution.

The 2000 Comparison

The obvious historical parallel is the 2000 telecom bubble, when optical networking stocks reached extreme valuations before collapsing. JDS Uniphase traded at triple-digit P/E multiples at its peak before crashing 99.8%. The comparison is unavoidable, but the differences matter.

Goldman Sachs assessed in October 2025 that "valuations of the technology sector are becoming stretched, but not yet at levels consistent with historical bubbles. The appreciation has been driven by fundamental growth rather than irrational speculation about future growth."²⁹

The distinction is important. Lumentum's 65% revenue growth is backed by NVIDIA purchase commitments, not the circular vendor financing that inflated 2000-era telecom revenues. Ciena's $7 billion backlog represents real customer orders, not letter-of-intent accounting. But "real demand" and "fairly valued" are different things. A stock can have genuine fundamentals and still be priced for perfection. When expectations are this high, even beating estimates may not be enough, as Ciena demonstrated when it fell 13% after reporting above consensus.²⁰

The 2017-2018 Parallel

The more relevant comparison may be the 2017-2018 data center buildout cycle, which was smaller in scale but followed the same pattern: demand surge from hyperscaler buildouts, aggressive capacity ramps, demand normalization, oversupply, and 50-70% stock corrections. The cycle was driven by 100G transceiver deployment and ended when inventory digestion caught up with production.

The compressed upgrade timelines (800G, 1.6T, and 3.2T ramping simultaneously, with CPO creating a fourth demand layer) could prevent the usual inventory correction by sustaining demand across more of the stack. Or the pattern repeats at larger scale, with the correction proportional to the size of the run-up. Both readings are defensible. The sector's track record favors the second.

What Could Go Wrong

The risks that matter most

Cycle and valuation. Every prior optical cycle has ended in correction. The 2017-2018 cycle saw 50-70% drawdowns after inventory digestion. The compressed upgrade path (800G, 1.6T, 3.2T ramping simultaneously) could sustain demand longer, or it could accelerate commoditization at each tier. At current multiples, the stocks price in near-perfect execution for years. If hyperscaler capex guidance disappoints in any quarter, or if the 1.6T ramp is slower than expected, corrections could be severe. The market has priced in the best case; it has not priced in the base case.

CPO disruption. If optical engines move onto the switch ASIC package, the pluggable transceiver market where most current revenue sits could shrink structurally. CPO is at least two to three years from volume, and pluggables will coexist for longer-reach applications. But the market is forward-looking, and CPO announcements could pressure valuations well before volume arrives.

Structural risks

Customer concentration. The sector is dangerously dependent on a small number of buyers. Ciena derives 67% of revenue from three customers. Fabrinet's largest customer is believed to represent over 30% of revenue. AAOI depends on limited hyperscalers for its transformative 1.6T orders. If any single hyperscaler pauses its buildout in any quarter, the impact ripples through the entire chain.

Chinese competition. Innolight and Eoptolink already manufacture 60%+ of 800G modules at 20-25% lower prices. As they climb the technology ladder (Innolight was first to complete NVIDIA 1.6T testing), they compress margins for Western module makers.¹⁶ The pattern from 100G and 400G is consistent: each generation starts at premium pricing, Chinese suppliers follow 12-18 months later, and margins compress.

Execution. Several companies are scaling at rates that leave little room for error. AAOI is targeting a revenue doubling. Lumentum is expanding EML production by 40% annually. Fabrinet is building a new facility to double 1.6T capacity. Any manufacturing stumble, yield issue, or qualification delay gets magnified at these growth rates.

Why This Matters

Optical networking is the physical infrastructure layer that connects AI compute. GPUs dominate the narrative, but beyond the rack, every link between GPUs runs on light. Every data center requires millions of miles of fiber. The $660 billion hyperscaler capex cycle includes a large optical component that the market largely ignored until 2025.

The supply picture is what elevated the sector from background to foreground. EML laser chips are 40-60% undersupplied through 2027 according to McKinsey. InP wafer capacity is near its limits. NVIDIA's $4 billion investment in optical suppliers is a strong signal that the company views photonics as an emerging gating factor for AI cluster deployment. Whether that makes optical networking "the new bottleneck" or simply "an important constraint alongside several others" is still an open question, but it is clearly no longer background infrastructure.

What the market is pricing reflects this shift in narrative, from compute bottleneck to interconnect bottleneck. Whether current valuations prove justified depends on two questions. First, does the demand cycle persist through 2027-2028, or does it normalize as first-wave buildouts complete? The compressed upgrade cycles (800G to 1.6T to 3.2T) and continued hyperscaler capex commitments suggest persistence, but no hardware cycle has ever avoided correction entirely. Second, can the pure-play optical companies maintain margins as Chinese competitors and CPO both apply pressure from different directions?

Our base case is that demand persists through at least 2027, supported by simultaneous generational upgrades and the structural EML laser shortage. The supply constraints provide margin protection for the next 12-18 months at minimum. Beyond that, the trajectory depends on whether CPO accelerates (favoring Broadcom, NVIDIA, and vertically integrated players) or pluggables persist longer (favoring Lumentum, Coherent, and Fabrinet).