🏢 NVIDIA Just Made a Billion-Dollar Bet... on Light!? Something Is Failing?

I. Introduction

Something strange happened this month.

NVIDIA, the company synonymous with AI chips and GPU dominance, quietly moved $4 billion into 2 companies most people had never heard of.

Instead of investing in a flashy AI startup or a new chip designer, they invested in 2 companies that build components made of light.

Coherent and Lumentum each received $2 billion from NVIDIA, backed by multi-year purchase commitments.

At the same time, Marvell Technology made its own moves, closing a $3.25 billion acquisition of an optical chiplet startup and adding another $540 million deal for a networking company, bringing Marvell's total photonics spend close to $3.8 billion in a single quarter.

That is over $7 billion flowing into one narrow corner of the supply chain in just a few weeks.

We all know that money doesn’t move like that by accident. When it does, it usually points to a deeper shift. This time, the shift is about physics, not just software anymore.

II. Why is Copper Failing in AI Data Centers?

Copper struggles to handle rising data speeds in AI systems. At 1.6 Tbps, it creates too much heat, uses too much power and loses signal quality. These issues are not fixable with software. There are limits of physics.

Key takeaways

• Heat increases with higher data speeds.

• Power usage rises sharply, not linearly.

• Signal quality degrades at high speeds.

• Affects all major AI data centers.

Every AI data center is essentially a city of chips. Thousands of GPUs and processors are constantly exchanging data, every second, at a massive scale. The system depends on fast connections and today those connections still rely on copper cables.

Right now, speeds are already at 800 gigabits per second (Gbps). The next step pushes that to 1.6 terabits per second (Tbps).

At that level, copper starts to break down, not in one way but in 3 simultaneous ways.

The Three-Threat Failure

• The first issue is heat.

Copper resists electrical flow and as data speeds increase, that resistance generates more heat. At 1.6 Tbps, the heat isn’t a small side effect anymore. It becomes a core engineering challenge that affects stability and cooling.

• The second issue is power.

This is where it gets worse. As speeds increase, power usage doesn’t rise smoothly; it explodes.

AI data centers already consume huge amounts of electricity, sometimes comparable to a small city (over 100 megawatts). When data rates increase over copper, energy demand grows faster than expected, making both cost and cooling harder to manage.

• The third issue is signal quality.

At these speeds, even short copper connections start to lose clarity. Signals degrade, errors increase and systems need to resend data. That slows everything down, which defeats the purpose of faster hardware.

This is not a bug you can fix with a software patch. It’s a hard limit from physics and every major tech company building large-scale AI infrastructure (Microsoft, Google, Amazon, Meta) is hitting it right now.

Simple version: chips are fast enough. The connections between them are not.

III. What Is Silicon Photonics?

Silicon photonics replaces electrical signals with light. Data moves through waveguides instead of copper wires. This reduces heat, lowers power use and improves signal quality. It allows much higher data transfer at scale.

Key takeaways

• Uses light instead of electricity.

• Improves efficiency by ~3x per link.

• Works over longer distances with less loss.

• Built using existing chip manufacturing processes.

1. Why are the Numbers Hard to Ignore

Silicon photonics delivers around 3 to 3.5x better power efficiency per link compared with copper. That might sound like a modest improvement until you scale it.

A large AI data center has 40,000 to 50,000 interconnect links. Apply a 3x efficiency improvement across every single one of them and you are not talking about shaving a few percentage points off an electricity bill.

You are talking about a structural transformation of how data centers are built and operated.

2. The Market Opportunity

In 2025, the silicon photonics market sits at roughly $2.65 billion. It is projected to reach $9.65 billion by 2030, growing at nearly 30% annually. More aggressive forecasts push it toward $28 billion by 2034.

TSMC, the world's most important chip manufacturer, announced a dedicated production line for photonic components in 2026. When TSMC commits manufacturing capacity to a technology, that is the most credible validation signal the industry can send.

To make this easier to apply, I’ve organized the full breakdown into this template so you can reuse it to analyze any AI infrastructure shift step by step.

IV. How to Think About This with 3-Layer Framework

Not every company in this space carries the same risk or offers the same kind of upside. To invest intelligently (or just to understand it clearly), it helps to break the photonics ecosystem into three layers stacked on top of each other.

1. Base Layer: Mega-Cap Enablers. 

These are massive, established companies where silicon photonics is only one part of a much bigger business. The exposure is indirect but they provide the core infrastructure that everything else depends on, which makes them more stable.

2. Middle Layer: Foundries and Manufacturers. 

These companies do not design the technology; they manufacture it at scale. Their advantage comes from volume, so they benefit from rising demand without carrying as much technology risk.

3. Top Layer: Pure Play Photonics. 

These companies live and die by how fast the copper-to-photonics transition happens. If adoption accelerates, they benefit the most but if it slows down, they feel it immediately. This layer carries the highest risk but also the highest upside.

The easiest way to see it is like building a house:

• The base layer acts as the foundation, steady and reliable.

• The middle layer forms the structure that makes everything work.

• The top layer is the roof, the most exposed but also where the biggest gains show up when everything comes together

Where you place your focus depends on how much risk you’re willing to take.

V. The Base Layer: Mega-Cap Enablers

Before anything else, you need to understand this: the AI boom isn’t just about models.
It’s about infrastructure and a handful of mega-cap companies are building the entire base layer.

1. NVIDIA: The Architect

NVIDIA is no longer just a chip company. They sell entire rack-scale AI supercomputing systems.

Their latest platform, Vera Rubin (shipping in the second half of 2026), packs 72 GPUs and 36 CPUs into a single rack with 1.3 million components per system.

It has already integrated silicon photonics directly into its systems:

• Spectrum-6 switches deliver ~5x better power efficiency than traditional networking.

• SuperNICs run at exactly 1.6 Tbps, where copper networking starts breaking down.

Of NVIDIA's $215 billion in fiscal 2026 revenue, $197 billion came from data centers alone. That is the market every photonics company in this ecosystem will be selling into.

When NVIDIA adopts a technology into its platform architecture, the entire supply chain follows around it. That’s what happened with CUDA and InfiniBand and the same shift is now happening with photonics.

2. Broadcom: The Co-Packaging Pioneer

Broadcom is pushing something called co-packaged optics.

Instead of plugging optical modules into a networking chip, they embed optics directly onto it. This shortens the data path, cuts power consumption and dramatically improves signal quality.

They’re already on their third generation, running at 200 Gbps per lane, with the next version doubling that.

Right now, photonics is a small slice of their business but as the global data center shift away from copper, this becomes a major growth layer on top of an already dominant business.

3. Cisco: The Enterprise Backbone Play

Most people think of Cisco as old-school networking infrastructure. That reputation actually puts Cisco in a strong position for the photonics transition.

So, when every enterprise data center eventually needs to replace copper with optical interconnects, the company already selling the switches and routers is in the best position to upgrade them.

Cisco acquired Acacia Communications, one of the original silicon photonics pioneers and is already manufacturing 800 Gbps silicon photonic transceivers inside today's data centers. They have demonstrated a co-packaged optics prototype in a 25 Tbps switch and are on track for their strongest revenue year ever.

Photonics is not why you get excited about Cisco today. But it is a big part of why Cisco stays relevant in the AI infrastructure era.

4. Intel: The Controversial Wild Card

Intel is often dismissed in AI conversations but its photonics work is hard to ignore.

They’ve been working on silicon photonics R&D for over 25 years. They’ve shipped more than 8 million photonic chips with over 32 million on-chip lasers, far ahead of most competitors.

Their OCI (Optical Compute Interconnect) chiplet runs at 4 Tbps bidirectionally and consumes just 5 picojoules per bit, compared to 15 for traditional pluggable modules, a 3x power advantage.

Intel is also the only foundry on the planet offering an optics-based manufacturing option for external customers.

The honest assessment: Intel carries real execution risk given its recent struggles. But the technical foundation is genuinely impressive. If they execute, the upside is much bigger than people expect.

VI. The Middle Layer: Foundries and Manufacturers

This layer is where designs become real products. These companies don’t always get the attention but they do the actual building and without them, nothing ships.

1. Tower Semiconductor: The Specialist Foundry

Tower Semiconductor is a specialty foundry. They do not design chips; they manufacture them for others, especially in silicon photonics. So, when someone creates a photonic integrated circuit, Tower is often the company fabricating it on silicon wafers.

They focus on silicon photonics and run multiple platforms, including one that integrates lasers directly onto silicon.

The CEO has publicly stated that Tower is "by far the leading foundry in silicon photonics". Photonics revenue reached over $220 million in 2025, more than double the year before. They’ve committed $920 million to expand capacity by five times, with over 70% already reserved through 2028 through customer prepayments.

They also have a direct partnership with NVIDIA to build 1.6 Tbps optical modules.

2. GlobalFoundries: The Scale Player

Where Tower is the specialist, GlobalFoundries is the scale player, one of the largest contract chip manufacturers globally, with fabs in the United States, Germany and Singapore.

Their unique advantage is integration: GlobalFoundries can build both the optical chip and the full module in one place, combining packaging and photonics into a single process. That end-to-end capability is hard to replicate at scale.

Their photonics revenue doubled to over $200 million in 2025 and is expected to nearly double again in 2026, with guidance toward a $1 billion run rate by 2028.

Photonics is currently about 3% of total revenue but the growth curve, 3 years running, is what makes it interesting.

3. Fabrinet: The Hidden Beneficiary

Fabrinet sits even further down the chain but it benefits from almost everything happening above it. The company is based in Thailand and focuses purely on precision manufacturing rather than designing products.

They run over 2 million square feet of factory space and specialize in sub-micron alignment, something very few companies can do well.

Here’s the key detail: both Coherent and Lumentum, the 2 companies that received $2 billion each from NVIDIA, manufacture through Fabrinet. So when money from NVIDIA flows into those companies, it goes directly to Fabrinet’s factories.

Instead of betting on which design wins, Fabrinet benefits from the entire ecosystem growing.

Their book value has doubled in just five years. For investors who want exposure to the silicon photonics buildout without the volatility of pure-play stocks, Fabrinet is one of the simplest ways to ride the growth without taking on the risk of a single design bet.

VII. The Top Layer: Pure Play Photonics

This layer is where the core hardware gets built. The companies here don’t just support the system; they are the system.

1. Coherent: The Vertically Integrated Giant

Coherent is the world's largest vertically integrated photonics company. While most companies focus on one piece of the supply chain, Coherent builds everything from the laser chip to the final transceiver module in-house. It’s also one of only two companies producing both indium phosphide and silicon photonics at scale.

The numbers reflect it:

• Revenue grew roughly 5x over 6 years, from $1.1 billion in 2018 to $6.3 billion on a trailing 12-month basis.

• Their most recent quarter at $1.7 billion was the best in company history.

• Gross margins have been recovering, moving from 30% back toward 36%.

Then came the big signal: the NVIDIA $2 billion strategic investment and multi-year purchase commitment landed in March 2026, the largest single deal in the company's history.

PEG ratio: ~1.1

2. Lumentum: The Laser Supplier to Everyone

Lumentum makes the lasers that power the entire photonics ecosystem.

Silicon can’t generate light on its own. Every silicon photonics chip needs an external laser source and Lumentum's lasers sit inside other companies’ products, even competitors.

That means no matter which design wins, Lumentum still gets paid.

The financial turnaround has been sharp.

• In under two years, Lumentum went from a $547 million net loss to $252 million in trailing net income.

• NVIDIA announced the same $2 billion deal structure, the largest laser supply commitment in Lumentum's history.

PEG ratio: ~0.61 (one of the most attractive valuation multiples in the entire photonics universe by this measure)

3. Marvell Technology: The Data Interpreter and Acquirer

Marvell's digital signal processing chips go inside virtually every 800 Gbps transceiver currently shipping, with a roughly 50% market share. When light exits a photonic chip and needs to be interpreted as digital data, that is Marvell's silicon doing the job.

Marvell also designs custom AI chips for Amazon and Google.

In February 2026, they closed a $3.25 billion acquisition of Celestial AI, an optical chiplet startup building technology that connects computing chips using light instead of electrical traces.

The claimed specs are striking: 25x greater bandwidth and 10x lower latency than current approaches. If that holds at scale, it rewrites how data centers are built.

Quarterly revenue has almost doubled over eight quarters, growing from $1.1 billion to $2.2 billion, while adjusted earnings per share have tripled. There are GAAP losses but they come from acquisition-related amortization and the actual cash flow remains strong.

PEG ratio: approximately 0.64.

VIII. The Investment Breakdown: How to Allocate Across Layers

The idea is simple: don’t treat every company the same.

You put more weight into large, established companies for a stable base and smaller positions into higher-risk players to capture growth without overexposing yourself.

Note: This is a framework for educational purposes only, not financial advice. Always do your own research and talk to a qualified professional before investing.

IX. Key Takeaways: What This All Means

This isn’t just about chips; it determines who controls the next decade of AI infrastructure.

1. The Copper Wall Is A Physics Problem, Not A Software Problem

At ultra-high speeds (1.6 Tbps), copper runs into hard limits, like signal loss, rising power use and heat. You can optimize around it but you can’t break the physics. That’s why the shift is happening.

2. Over $7 Billion In Weeks Is The Market's Loudest Signal

NVIDIA's $4 billion investment in Coherent and Lumentum, combined with Marvell's $3.8 billion acquisition spree, is not betting on a distant future.

They are supply chain lock-ins made by the largest players in AI infrastructure. These companies are buying guaranteed access to the components they cannot build their next generation of products without.

3. TSMC's Production Line Is The Ultimate Validation

Speculation gets real when TSMC builds a production line for it. That moment happened in 2026.

When the world's most important chip manufacturer dedicates capacity to a technology, the transition from "promising" to "inevitable" is complete.

4. The Three-Layer Framework Lets You Calibrate Your Own Risk

Every investor has a different risk tolerance and this framework gives you a map.

• If you want stability, stay close to the biggest AI players where photonics gets embedded.

• If you want leverage, look at manufacturers.

• If you want upside, focus on the transition itself.

The smartest play is probably some combination of all three.

5. Fabrinet Is The Quiet Compounder Worth Watching

It captures spending across the entire ecosystem without needing to pick a winner.

Strong relationships, clean financials and direct exposure to NVIDIA's supply chain commitments, all without the volatility of the pure-play names.

X. Conclusion

Copper is hitting a physics wall, not because of a theory but because it is an engineering constraint that the largest technology companies are already planning around.

Silicon photonics is moving into place as the next foundation of AI infrastructure. It’s becoming the layer that everything else depends on. The companies positioned in this space now (across all three layers) are the ones building, supplying and manufacturing the pipes that future AI systems will run through.

So the question isn’t if this transition happens; it’s who captures the most value as it speeds up.

If you are interested in other topics and how AI is transforming different aspects of our lives or even in making money using AI with more detailed, step-by-step guidance, you can find our other articles here:

• Stop Trusting AI With Your Data. It Doesn't Read, It PRETENDS (Here's the Proof)

• 70% Of Google Searches Get ZERO Clicks (Here's The Fix)*

• Warning: The AI Bubble Burst Could Bankrupt Elon Musk

• ChatGPT vs. Claude vs. Gemini vs. Grok vs. Perplexity: Tips to Make Use of Each One!*

• Kling 3.0 is Crazy and Can Do Anything. Ultimate Guide with Pro Tips & Prompts to Try

*indicates a premium content, if any