The Light Between The GPUs: Why Optical Networking Could Be The Next Big AI Investment Theme

Every conversation you have with an AI, every search, every summary, every generated image, travels through light. Not metaphorically. Literally. The moment an AI data center gets large enough, copper wires physically can't carry the data fast enough between chips without melting down or losing signal. The solution is photons: light pulses moving through thin glass fibers at extraordinary speeds. This is the story most AI investors are entirely missing. While the world debates which next model will outperform the last, a quieter but structurally more durable opportunity is building in the plumbing that connects it all; optical components and high-speed networking. And in March 2026, Nvidia made a $4 billion bet that confirms it.

Let's break down why optical networking sits at the heart of the AI hardware stack, which companies are cashing in, what the numbers actually say, and companies investors should watch.

Optical Networking: The AI Bottleneck Worth Understanding

Here's a mental model worth keeping. Think of AI training like cooking a massive feast at a wedding. Nvidia's GPUs are the world-class chefs. Everyone knows about the chefs. But if the kitchen has narrow doorways and slow conveyor belts moving ingredients between stations, even the best chefs will be standing around waiting. That "conveyor belt", the interconnect layer, is optical networking.

A modern AI training cluster doesn't just use one GPU. It uses tens of thousands of them, all talking to each other constantly. The communication between GPUs during training can account for 30-40% of total training time. When you scale from 10,000 GPUs to 100,000 GPUs, which hyperscalers are now doing, the sheer volume of data shuttling between chips every second becomes astronomical. Copper-based electrical cables, the traditional workhorse of data centers, have three fatal flaws at this scale: they degrade over distance, they consume enormous power, and they simply can't carry enough bandwidth per unit of space.

Optical fiber solves all three. Data travels as light. No heat buildup from resistance, no signal degradation over tens of meters, and bandwidth capacity that scales with the speed of photons, which, last anyone checked, is 299,792,458 meters per second.

This isn't new technology. What's new is the urgency.

AI Optical Networking $30B Market Opportunity

Datacom optical component revenue crossed $19 billion in full-year 2025; a more than 70% increase over 2024, according to market research firm Cignal AI. That isn't gradual adoption. That's a structural inflection.

Cignal AI expects total datacom optical component revenue to grow at a 20%+ CAGR from 2024 through 2029, reaching nearly $29 billion by end of that period.

The Demand Engine: $700 Billion in Capex and Rising

The five largest US cloud and AI infrastructure providers namely, Microsoft, Alphabet (Google), Amazon, Meta, and Oracle, have collectively committed to spending between $660 billion and $690 billion on capital expenditure in 2026, nearly doubling 2025 levels.

Roughly 75%, or approximately $450 billion, of that spend is directly tied to AI infrastructure: servers, GPUs, data centers, and networking equipment.

Every dollar poured into this infrastructure creates downstream demand for optical transceivers, fiber cables, photonic chips, and high-speed networking switches. The link between AI compute spend and optical demand is direct and growing. Cignal AI data shows datacom optical revenue more than doubled in two years, from roughly $11 billion in 2024 to over $19 billion in 2025, even as overall AI capex grew at a slower pace, suggesting optical spend is outgrowing compute spend as cluster sizes scale. 

According to Dataintelo's market analysis, every dollar invested in AI compute hardware drives approximately $0.18 of incremental optical interconnect spending, up from $0.09 in 2020,  though this figure represents the firm's own internal estimate and has not been independently verified by a third-party research house. 

Nvidia’s $4 Billion Bet On Optical Infrastructure

On March 2, 2026, Nvidia announced two separate multiyear strategic partnerships with Coherent Corp. and Lumentum Holdings, investing $2 billion in each, alongside multibillion-dollar purchase commitments and future capacity access rights for advanced laser and optical networking products.

"AI has reinvented computing and is driving the largest computing infrastructure buildout in history," said Jensen Huang, founder and CEO of Nvidia. "Together with Lumentum, Nvidia is advancing the world's most sophisticated silicon photonics to build the next generation of gigawatt-scale AI factories."

This isn't a vendor relationship. This is Nvidia locking up supply. When the most valuable chip company on Earth spends $4 billion to secure optical component capacity, including pre-paying for manufacturing it doesn't yet need, it tells you exactly where the next bottleneck could be.

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. Nvidia is essentially buying insurance against this. For investors, this is a signal worth noting.

Companies To Watch In Optical Netwroking: Where the Opportunity Actually Sits

Coherent Corp (NYSE: COHR)

The most direct play. Coherent reported Q3 FY2026 revenue of $1.81 billion, up 21% year-on-year, with datacenter and communications segment revenue reaching $1.4 billion, up from $1.0 billion a year earlier. CEO Jim Anderson stated the company delivered "exceptionally strong demand across our datacenter and communications businesses." Beyond the Nvidia deal, Coherent's InP (Indium Phosphide) lasers are reportedly sold out through 2027.

Lumentum Holdings (NASDAQ: LITE)

The other beneficiary of Nvidia's $4B optical bet. In Q3 FY2026 (reported May 5, 2026), Lumentum posted revenue of $808.4 million, up 90% year-over-year, with Q4 FY2026 guided at $960 million-$1.01 billion, which if achieved would mark another record. Full-year FY2026 revenue is tracking toward approximately $2.8-$3.0 billion. CEO Michael Hurlston said: "This multiyear strategic agreement reflects our shared commitment to advancing the optics technologies that will power the next generation of AI infrastructure." Lumentum is currently the only supplier shipping 200G-per-lane EMLs, the critical laser components inside 1.6T transceivers, at meaningful volume. 

Marvell Technology (NASDAQ: MRVL)

Marvell isn't a pure-play optical company, it's better described as the "optical brain." Its PAM4 DSP (Digital Signal Processor) chips are the intelligence inside high-speed optical transceivers, converting electrical signals to optical ones and vice versa. "Marvell delivered record fiscal 2026 revenue of $8.195 billion, growing 42% year-over-year, driven by robust AI demand," said CEO Matt Murphy. Management expects optics to outgrow cloud capital expenditures, noting that optics growth is tied to AI accelerator proliferation and bandwidth intensity, not just aggregate capex. In April 2026, Marvell also acquired Polariton Technologies, a silicon photonics and advanced modulation startup, deepening its optical stack.

Ciena Corporation (NYSE: CIEN)

The coherent optical systems player that connects data centers across cities and regions. In Q1 FY2026 (reported March 5, 2026), Ciena posted revenue of $1.43 billion, up 33% year-over-year, and raised its full-year FY2026 guidance to $5.9-6.3 billion. CEO Gary Smith described demand as "unprecedented, broad-based" and said the company is enabling customers to "monetize their AI investments." Ciena's backlog increased by $2 billion in a single quarter to a record $7 billion, with most new orders scheduled for FY2027. Its DCOM solution was co-developed with Meta. 

Arista Networks (NYSE: ANET)

The networking switch layer; essentially the traffic controller for optical signals inside and between AI clusters. Arista raised its full-year 2026 revenue forecast to $11.5 billion and lifted its AI revenue target to $3.5 billion, effectively doubling its AI sales guidance. CEO Jayshree Ullal said, "Our demand is actually the best I've ever seen in my Arista tenure." At the 2026 Optical Fiber Conference, Arista unveiled XPO (extended pluggable optics), now endorsed by over 100 vendors.

Corning Inc. (NYSE: GLW)

The fiber itself. Corning signed an up to $6 billion multi-year deal with Meta to supply fiber-optic cables for AI-focused data centers. The company's shares had risen about 26% year-to-date as of January 2026, following an 84% gain in 2025. Often overlooked because it makes "glass," Corning is the literal backbone of AI interconnect infrastructure.

The AI Optical Tax

Here's a mental model we're calling the AI Optical Tax.

Every GPU that goes into an AI cluster generates a predictable, growing bill for optical components. Think of it the way India taxes petrol consumption: more vehicles on the road means more fuel tax, regardless of which car company wins. In AI, as model size grows and cluster size grows, the optical spend per GPU grows too, because larger clusters need faster interconnects between more nodes.

In 2020, this “optical tax" was roughly $0.09 per dollar of compute hardware spend. By 2025, it had risen to $0.18. As clusters scale to 1 million GPUs, a threshold Jensen Huang has discussed publicly, this ratio likely moves toward $0.25 or higher. Larger clusters create longer average interconnect distances, which require more fiber, more transceivers, and more sophisticated switching.

This means optical component growth isn't just correlated to AI capex, it accelerates faster than AI capex.

Risks Investors Should Watch

This section matters the most as any serious investment thesis needs to survive its own stress test.

1. Co-Packaged Optics (CPO) Timeline Risk 

The next generation of optical technology: CPO, which integrates optical components directly inside the chip package, promises to dramatically lower costs per bit. If CPO commercializes faster than expected, it could compress margins for pluggable transceiver companies like Coherent and Lumentum, even as volumes rise. CPO is currently expected to see early deployment in hyperscale AI data centers between 2026-2027, with mainstream adoption potentially beyond 2030. But technology timelines are notoriously hard to predict.

2. Hyperscaler Capex Discipline Snapping 

Big tech companies issued $100 billion of bonds in 2026 to fund AI capex, with investors demanding record protection via Credit Default Swaps. If AI ROI disappoints, if enterprise customers don't buy enough AI cloud services to justify the infrastructure cost, hyperscalers could cut capex sharply, as they did in 2022-2023. Optical companies would feel this acutely, given their long-lead-time manufacturing.

3. Chinese Competition 

A significant portion of optical transceiver manufacturing is concentrated in China, particularly companies like Innolight and Eoptolink that dominate 800G module shipments. If US-China trade tensions escalate to include optical components, it could disrupt supply chains in both directions and if Chinese suppliers are blacklisted but then flood emerging markets at lower costs, it creates pricing pressure for US-listed companies.

4. Customer Concentration Risk 

Coherent, Lumentum, Marvell, and Ciena all depend heavily on a handful of hyperscalers for a disproportionate share of revenue. If one major customer (say, Google) shifts architecture, delays builds, or develops in-house optical components (as Google has done with TPUs), the revenue impact is outsized.

5. Overcapacity Cycle 

The same Nvidia supply lock-in that looks bullish today. pre-buying 2028 and 2029 capacity, could create oversupply later if AI cluster demand moderates. Optical companies would then face the same dynamic that crushed solar panel makers who overbuilt in 2012: too much capacity, prices collapse.

Our View On The AI Optical Infrastructure Opportunity

Optical networking isn't a trade. It's a structural position in the physical layer of AI infrastructure, one with real moats (supply scarcity, technology complexity, long design-win cycles), real revenue, and validation from the most credible buyer of technology on Earth: Nvidia.

The near-term opportunity looks strongest in companies with supply constraints like Coherent and Lumentum, because pricing power is highest when demand exceeds supply. Marvell has the broadest exposure across the optical stack, from DSPs to custom silicon. Arista offers the cleanest way to play AI networking switching without direct optical manufacturing risk. Corning is the most defensive as fiber isn't going obsolete. But none of this is a recommendation to buy any specific stock. Valuations matter, and several of these names already trade at significant AI premiums. 

But if you're asking where the next decade of AI infrastructure spending ends up, a meaningful percentage of it could end up as photons, traveling through glass, made by companies that most retail investors couldn't name today.

That's usually where interesting things happen.