On March 2, 2026, Jensen Huang wrote two cheques.
Each was for $2 billion. Both went to photonics companies — Lumentum and Coherent. Same day. $4 billion total. And alongside each investment came something even more consequential: a multi-billion dollar purchase commitment for advanced laser components over multiple years.
Jensen Huang does not make sentimental investments. He is not diversifying a personal portfolio. When NVIDIA spends $4 billion on optical companies, it means one thing: optical components are a genuine bottleneck to the AI infrastructure buildout that NVIDIA's entire business depends on, and he needs to make sure that bottleneck gets solved.
This post is about why that bottleneck exists, what Lumentum is doing to solve it, and what the next two to three years look like for the business as a result.
If you've read Parts 1 and 2, you know the business model and the financial history. This is where the forward thesis gets built.
The AI Infrastructure Bottleneck Cascade: A Pattern Every Investor Should Know
The AI buildout has created a predictable sequence of scarcity events. Each year, a different component becomes the critical constraint. And each constraint creates a multi-year growth window for the company that can relieve it.
Read this chart left to right. 2024 was about GPU scarcity — NVIDIA H100 chips had six to twelve month lead times. 2025 was about High Bandwidth Memory (HBM) — the specialised memory chips that sit inside every AI processor. 2026 is the optical year. And 2027–28 will likely be about power and cooling. Each bottleneck creates a concentrated, winner-takes-most opportunity.
Let me explain what "optical bottleneck" actually means at the ground level.
A modern AI training cluster consists of thousands of GPUs — often 10,000 to 100,000 in a single AI factory. These GPUs need to communicate with each other constantly during training. The calculations are massively parallel — different GPUs handle different parts of the model, and they need to exchange results billions of times per second.
For small clusters, copper cables work fine. But as you scale to tens of thousands of GPUs, copper hits a wall — it generates too much heat, it can't maintain signal integrity over distances, and the energy consumption becomes enormous.
The solution is optical fiber. But connecting 10,000 GPUs with optical fiber requires an enormous number of optical transceivers — the devices that convert electrical data signals to light and back again. A cluster of 10,000 NVIDIA H100 GPUs might need 40,000 to 80,000 transceiver ports. A cluster of 100,000 GPUs needs ten times that.
And inside every high-speed transceiver — the heart of the device — is a laser chip. Specifically, an EML (Electro-Absorption Modulated Laser). And the world's primary supplier of the most advanced EML chips, the ones needed for the fastest transceivers? Lumentum.
Why Demand Is Running Ahead of Supply: The 1.6T Story
Here's the data that makes the current moment extraordinary.
The left side of this chart shows the AI optical transceiver market growing from $8.9B in 2022 to an estimated $26B in 2026 — nearly a 3× expansion in four years. The right side shows something even more important: each generation of transceiver roughly doubles the optical content per server rack. A 1.6T transceiver carries 8× the data of the original 100G baseline. Same rack space. Eight times the optical content. Eight times the revenue opportunity per rack.
The current inflection is driven by the transition from 800G to 1.6T transceivers.
800G transceivers use four lanes of 200 gigabits per second each (4 × 200G = 800G total). They're the dominant technology in 2025–26.
1.6T transceivers use eight lanes of 200 gigabits per second (8 × 200G = 1.6T total). They deliver twice the data through the same physical form factor — critical for the GPU-dense clusters that AI training requires.
The key insight: 1.6T transceivers require 200G-per-lane EML chips — precisely the technology where Lumentum holds 50–60% global market share and where demand materially exceeds supply.
And the supply situation? AI data center optical component shortage is pushing 800G transceiver production 40–60% below demand through 2027. Lumentum is investing in expanding its San Jose InP fabrication facility, while analysts expect double-digit price increases on 200G EMLs in 2026 given the absence of viable second-source suppliers.
Think about what "40–60% below demand" means in practical terms. Hyperscalers are ordering optical components, and they're receiving less than half of what they asked for. That doesn't make them cancel orders — it makes them order even more and queue up for future capacity. Which is exactly what the $400M+ OCS backlog and NVIDIA's purchase commitment reflect.
The forecast revisions tell the story most vividly: Earlier this year, industry forecasts called for roughly 10 million 1.6T optical transceivers to ship in CY2026; by mid-year that estimate doubled to 20 million, and it now stands near 35 million units. An estimate that tripled in a single year. This is not analysts being bullish — this is hyperscalers pulling forward demand faster than anyone modelled.
What NVIDIA Actually Did: The $2B Deal Anatomy
Let's be precise about what happened on March 2, 2026, because most coverage has been imprecise.
This chart breaks down the three components of the NVIDIA deal: the $2B equity investment (preferred stock convertible to common), the multi-billion dollar purchase commitment for laser components, and the future capacity access rights at the new US fabrication facility.
NVIDIA's non-exclusive agreement includes a multi-billion purchase commitment and future capacity access rights for advanced laser components, alongside a $2B investment to support R&D, future capacity, and operations as Lumentum builds out US-based manufacturing.
Break that down into three distinct signals:
Signal 1: The $2B Investment — Lumentum issued 2,876,415 shares of new Series A Convertible Preferred Stock to NVIDIA at $695.31 per share. This converts to common stock at a 1:1 ratio at NVIDIA's option. The $695 conversion price implies NVIDIA's internal valuation at the time of investment — and they clearly believed the stock would be worth more over time.
Signal 2: The Purchase Commitment — This is arguably more important than the equity check. A purchase commitment means NVIDIA is contractually agreeing to buy specific quantities of laser components over a defined period. This de-risks Lumentum's capex decisions — when you build a billion-dollar fab, you want to know you have customers lined up before the first shovel breaks ground.
Signal 3: Capacity Access Rights — NVIDIA reserved production capacity at the new Greensboro, North Carolina fabrication facility that Lumentum is building. They're not just buying components that exist today — they're pre-allocating production from a fab that isn't operational yet. That is the strongest possible forward demand signal.
One critical nuance worth stating clearly: this deal is non-exclusive. NVIDIA made the identical investment in Coherent on the same day. Jensen Huang is not picking a winner — he is securing supply across multiple capable vendors because he cannot afford a single point of failure in his optical component supply chain. This is rational capital allocation on NVIDIA's part, but it means Lumentum does not have an exclusive NVIDIA partnership. Investors should not overweight that angle.
What the deal definitively validates: optical interconnects are not a temporary architectural choice. They are a permanent, scaling bottleneck in AI infrastructure — and NVIDIA has now committed that publicly, in capital, not just in words.
The Three Growth Engines: EML, OCS, and CPO
Lumentum is not a one-product story. The business has three distinct revenue engines operating at different stages of maturity. Understanding all three is essential for modelling the forward revenue trajectory.
The stacked bar chart at the bottom of this visual shows how the revenue mix evolves. EML chips and transceivers carry the business through 2026. OCS revenue begins scaling meaningfully from mid-2026. CPO revenue appears in 2027 and becomes significant by 2028. Each layer represents a new wave of growth that extends the runway.
Engine 1: EML Chips — The Core Business (Now)
EML chips are Lumentum's foundational product — the laser source inside every high-speed transceiver. Lumentum is the only company shipping 200G-per-lane EMLs at volume, and demand exceeds supply by 25–30%. As the industry transitions from 800G to 1.6T transceivers over the next two years, every transceiver maker needs Lumentum's chips. This supply-demand imbalance enables pricing power and supports the margin expansion trajectory through FY2027.
The 40% annual capacity expansion (committed for both 2025 and 2026) is Lumentum racing to close the supply gap. Until they do, they have pricing power — the ability to raise prices because demand exceeds supply and there's no viable alternative. Analysts expect double-digit EML price increases in 2026. That's the rarest thing in hardware — a manufacturer that can charge more simply because no one else can fill the order.
Engine 2: Optical Circuit Switches — The Scaling Business (Now → 2027)
OCS is the most exciting near-term growth vector, and it barely existed 18 months ago.
Let me explain what an Optical Circuit Switch does, simply. In a conventional data center, when data moves between servers, it gets converted from light to electrical signals at each "hop" — each time it passes through a router or switch. This conversion adds latency (delay) and consumes significant power.
An OCS eliminates these conversions entirely. It routes optical signals in the light domain — the data stays as light from source to destination, switching dynamically between different fiber paths without ever becoming electricity. Google has demonstrated that OCS-based architectures deliver 40% power savings in data center fabric — a number that matters enormously when you're building a 1-gigawatt AI factory.
The demand signal is unambiguous: "In OCS, we are scaling rapidly to meet extraordinary customer demand that has already driven our backlog well beyond $400 million."
Lumentum's OCS business cleared $10 million per quarter three months ahead of internal targets. The order backlog has surged past $400 million, with the majority slated for delivery in H2 CY2026. Management is targeting $100 million per quarter in OCS revenue by the end of calendar 2026.
And OCS customer diversification is expanding: OCS demand diversification expanded to three customers, broadening revenue sources and reducing customer risk concentration.
This is the product line that could be transformative. If OCS scales to $400M/year by FY2027 — which the backlog and customer expansion suggest is plausible — it adds roughly 10–15% to total revenue entirely from a product that barely registered in FY2025.
Engine 3: Co-Packaged Optics — The Pipeline Business (2027–28)
CPO is the most technically advanced and furthest from volume revenue, but it's worth understanding because it represents the long-term direction of the industry.
In current architectures, optical transceivers are pluggable — they slot into a switch or server as a separate module. This works well, but there's an energy penalty every time the signal crosses the boundary between the module and the chip.
Co-Packaged Optics eliminates this boundary. The optical components are packaged directly inside the network switch chip — the laser, detector, and modulator are integrated at the die level. This delivers up to 3.5× energy efficiency improvement and enables substantially higher data densities.
"In CPO, we received an incremental multi-hundred-million-dollar order, deliverable in first half calendar 2027."
That CPO order — described as the largest single purchase commitment in company history — tells you that hyperscalers are already placing bets on this next-generation architecture. The revenue won't show up until 2027, but the order book is being filled today.
Project Vanguard: Why "Made in USA" is a Competitive Moat
The third major catalyst — alongside the NVIDIA deal and the 1.6T transition — is one that doesn't get enough attention in most analysis.
This chart compares the positioning of US and Chinese optical manufacturers side by side, then maps the Greensboro fab construction timeline. The contrast is stark — Chinese optical manufacturers are subject to export controls that prevent them from being in the supply chains of Western hyperscalers building AI infrastructure.
Here's the geopolitical reality. A significant portion of the world's optical component manufacturing capacity sits in China — companies like Accelink and HG Genuine have been low-cost, capable suppliers for years. But Google, Microsoft, Amazon, and Meta are no longer comfortable with their critical AI infrastructure components coming from Chinese suppliers. The reasons are multiple: export control risk, potential supply disruption in a geopolitical crisis, and US government pressure to de-risk strategic supply chains.
Lumentum's Project Vanguard — domestic US manufacturing expansion — positions the company as a primary beneficiary of Western hyperscaler demand for trusted, non-China-sourced optical supply chains.
The new Greensboro, NC fabrication facility is the centrepiece of Project Vanguard. Funded in significant part by the NVIDIA $2B investment, it will be a state-of-the-art InP (Indium Phosphide) wafer fab on US soil. When it reaches full production, Lumentum will be able to offer hyperscalers something no Chinese supplier can: guaranteed, geopolitically stable supply of the world's most advanced optical components, manufactured in the United States.
This is not a marginal advantage. In an era where US-China technology tensions are intensifying, hyperscalers are actively willing to pay a premium for supply chain security. Lumentum's EML chips already command double-digit price premiums over Chinese alternatives — not just because of technical superiority, but because of supply chain reliability. The Greensboro fab deepens that moat permanently.
One additional point for Indian readers: India's own data center build-out — driven by the hyperscalers opening facilities in Mumbai, Pune, and Hyderabad — will eventually require this same generation of optical infrastructure. The optical components powering Indian AI infrastructure will come from the same supply chain that Lumentum is building today. Lumentum has no direct India revenue, but the technology it's developing will be inside the systems that connect India's computational future.
Putting It Together: The Forward Thesis in Three Points
1. Lumentum controls the critical bottleneck of 2026.
The world's only at-volume supplier of 200G EML chips, at a time when demand exceeds supply by 25–30% and the primary driver — 1.6T transceiver adoption — is accelerating faster than any analyst predicted. This is a structural pricing power position, not a cyclical one.
2. The revenue runway has three layers, not one.
EML chips carry the business through 2026. OCS adds a $300–400M revenue layer from mid-2026 as the $400M+ backlog converts. CPO adds another major revenue layer from 2027–28. Each engine extends the growth runway by two to three years beyond what the current headline numbers show.
3. The US manufacturing position is a geopolitical moat that gets stronger over time.
The combination of NVIDIA's $2B investment, the Greensboro fab construction, and Western hyperscaler preference for non-China supply creates a competitive position that is difficult to replicate quickly. The Chinese optical industry is technically capable, but structurally excluded from the most important customer relationships.
The Risk That Keeps Me Honest
All of this is a genuinely compelling thesis. But I would be doing you a disservice if I didn't flag the primary risk clearly: this is a hardware business serving a small number of very large customers, and those customers can change behaviour quickly.
If hyperscaler capex pauses — for macroeconomic reasons, because AI model scaling shows diminishing returns, or simply because they've over-built relative to current AI workload demand — Lumentum's order book can deteriorate rapidly. The 2024 collapse showed exactly how fast this can happen. From peak 2022 margins to trough 2024 losses in under 18 months.
The AI-driven demand cycle appears structurally different from prior telecom cycles. But "appears structurally different" is a sentence that has preceded many painful corrections in the history of technology investing.
We'll deal with this risk explicitly — with numbers and scenarios — in Part 4, which is the valuation deep dive.
Manu | Curious Investing Insights
Applying institutional analytical depth to markets that matter for Indian investors
This is not investment advice. Do your own research. The author may hold positions in securities discussed.
Coming in Part 4: What are you actually paying for this business? Bull case, base case, bear case — with explicit revenue growth assumptions and the DCF math that tells you what the current $940 stock price implies about the future. No hand-waving. Just the numbers.
Tags: Lumentum LITE · NVIDIA Investment · AI Infrastructure · Optical Networking · 1.6T Transceivers · Photonics · OCS · CPO · US Stocks for Indian Investors