# What Is Photonics? Inside NVIDIA's $4 Billion Bet on Light
> We traced NVIDIA's $4 billion photonics bet: why copper hit a wall at AI speeds, what replaces it, and what the market had already priced in.
- Author: Barebone Research, Barebone AI
- Published: 2026-03-13
- Canonical: https://barebone.ai/resources/what-is-photonics-nvidia-4-billion-bet-on-light
- Publisher: Barebone AI (https://barebone.ai)
---
## Two $2 Billion Checks
On March 2, NVIDIA announced two strategic agreements that most people scrolled past. It is investing **$2 billion in Lumentum** and **$2 billion in Coherent** — two companies that make lasers and optical components — with multi-billion-dollar purchase commitments and future capacity rights layered on top of each deal.
The market did not scroll past. Lumentum jumped **+11.7%** that day to $783.25. Coherent jumped **+15.4%** to $298.91. Both closed at the top of their entire twelve-month range.
Which raises the question almost nobody on the receiving end of those headlines could answer: what, exactly, is photonics — and why is the most valuable chip company on Earth writing ten-figure checks to secure it?
We used Barebone AI to pull both press releases off SEC EDGAR, the keynote where Jensen Huang first laid out the math, and twelve months of price history across the optics complex. The short version: **Lumentum +933% in twelve months. Coherent +289%. The S&P 500, +20%.** The market started pricing this shift long before the checks were signed.
Here is what it actually is, why it became the center of the AI-infrastructure trade, and — just as important — what the past two weeks say about how much is already in the price.
## Copper Hit a Wall
A modern AI model doesn't train on a chip. It trains on tens of thousands of chips that have to behave like one — which means GPUs are constantly shipping enormous volumes of data to each other, every microsecond, across the entire data center.
Today most of that traffic moves the way it has for decades: as electrical signals over copper.
Copper has real virtues. A passive copper cable costs almost nothing, draws almost no power, and rarely fails. But it has one defect that gets worse every hardware generation: as signaling speeds rise, the distance a copper link can carry a clean signal collapses. At the data rates of current AI fabrics, passive copper is good for roughly **one to two meters**.
You can see the consequence in NVIDIA's own rack design. The GB200 NVL72 packs 72 GPUs into a single rack lashed together by a copper spine — because copper only works when everything is within arm's reach. The moment traffic leaves the rack, it has to ride light through glass fiber. And frontier clusters now span buildings.
> "The challenge with scaling out GPUs to many hundreds of thousands is the connection of the scale-out." — Jensen Huang, GTC keynote, March 2025
At that keynote, Huang put numbers on the problem. Every GPU in a scaled-out cluster needs roughly **six optical transceivers** — the plug-in modules that convert electrical signals to light and back — at about **30 watts each**. That is **180 watts per GPU** spent on nothing but moving data, before a single token is generated. Scale to a million GPUs and you need six million transceivers burning roughly **180 megawatts** — a power plant's worth of electricity, consumed entirely by the conversion between electrons and photons.
That is the wall. Photonics is the answer to it.
## What Photonics Actually Is
Photonics is to light what electronics is to electrons: the science of generating, routing, and detecting photons to carry information. In a data center, it means a laser pushing bits down a strand of glass fiber instead of electrons pushing through copper.
None of this is new. Photonics already carries essentially all long-distance internet traffic — every transoceanic cable is a photonics product. What is new is *where* the light starts.
In today's architecture, data leaves a switch chip electrically, travels a few centimeters to the faceplate, and only there — inside that 30-watt pluggable transceiver — becomes light. Every transceiver carries its own lasers and its own signal-processing chip. Multiply by millions and you get the 180-megawatt problem.
**Silicon photonics** changes the geography. It uses standard chipmaking processes to build optical components — modulators, waveguides, detectors — directly into silicon. The flagship application is **co-packaged optics (CPO)**: instead of bolting transceivers onto the outside of a switch, the optical engine moves inside the package, millimeters from the switch silicon, and the power-hungry electrical hop disappears.
NVIDIA announced its version at GTC in March 2025 — Quantum-X and Spectrum-X Photonics switches, built with TSMC around a component called a micro ring modulator. Huang's claim on stage: "It is the world's first 1.6 Tb/s CPO, based on a technology called a micro ring modulator." NVIDIA's stated numbers versus pluggable transceivers: **3.5x better power efficiency**, **4x fewer lasers**, **10x better network resiliency**, with InfiniBand versions slated for late 2025 and Ethernet versions for 2026.
Here is what NVIDIA's own keynote arithmetic looks like at AI-factory scale:
At one million GPUs, the transceiver line item alone is 180 megawatts; at NVIDIA's claimed 3.5x efficiency, co-packaged optics cuts it to roughly 51 — about **130 megawatts handed back**. In a buildout where every new campus is power-constrained, a saved megawatt is a sellable megawatt: it becomes more GPUs on the same grid connection. Huang framed the savings as "tens of megawatts" per data center, and that framing is the entire investment case in one sentence.
## The Hyperscalers Were Already There
NVIDIA is the loudest convert, but not the first.
**Google** has spent years routing traffic between its TPU pods through optical circuit switches — its Apollo system uses arrays of microscopic MEMS mirrors to steer beams of light from one fiber to another, with no electrical conversion at all. Google's own paper on the project describes optical switching deployed at datacenter scale.
**Microsoft** bought hollow-core fiber maker Lumenisity in 2022. Light travels about **47% faster** through a hollow, air-filled core than through solid glass, and Microsoft has been threading the technology into Azure's network.
**Oracle** began taking orders in September 2024 for OCI Superclusters scaling to **131,072 Blackwell GPUs** — at that scale, everything rack-to-rack rides optics by physical necessity — and followed in October 2025 with Zettascale10, its next-generation successor. Meta and Amazon are pushing the same direction inside their own networks.
The hyperscalers didn't adopt photonics because it's elegant. They adopted it because at their scale, nothing else closes the math.
## Anatomy of a Land Grab
So why did the March 2 announcement matter, if the direction was already obvious? Because of what NVIDIA bought. Read the two deals side by side:
| Term | Lumentum | Coherent |
|------|----------|----------|
| Direct investment | **$2B** | **$2B** |
| Purchase commitment | Multi-billion | Multi-billion |
| Capacity rights | Advanced laser components | Advanced lasers + optical networking products |
| Manufacturing | New U.S. fab | U.S. manufacturing buildout |
| Exclusivity | Non-exclusive | Non-exclusive |
> "AI has reinvented computing and is driving the largest computing infrastructure buildout in history. Together with Lumentum, NVIDIA is advancing the world's most sophisticated silicon photonics to build the next generation of gigawatt-scale AI factories." — Jensen Huang, March 2, 2026
Notice what the money targets: **lasers**. Silicon photonics has one famous gap — silicon is excellent at steering and detecting light but poor at emitting it, so the laser must be made separately from compound semiconductors and attached. Lasers are the hardest, lowest-yield, most supply-constrained part of every optical link, and Lumentum and Coherent are two of the world's largest merchant laser makers. Both were already named partners in NVIDIA's photonics ecosystem at GTC 2025. Coherent's CEO noted the deal extends a **20-year relationship**.
The timing tells you the rest. Per TrendForce — in a note published two days before this article — CPO will account for only about **0.5%** of the optical transceiver modules used in AI data centers in 2026, rising toward roughly **35% penetration by 2030**, with optical scale-up links arriving alongside NVIDIA's Rubin-generation systems. One widely used industry forecast puts the optical transceiver market at **$14.7 billion in 2025**, growing to **$46.1 billion by 2034** — a 17% compound rate, more than tripling.
In other words: the technology transition is at the steep base of its adoption curve, and the scarcest input is laser capacity. NVIDIA's $4 billion isn't really buying revenue. It is buying a place at the front of the queue — capacity rights — for the moment demand outruns supply, and paying to make sure a chunk of that capacity gets built on U.S. soil. Less a bet than a land grab.
## The Market Got There First
Now the part that should give you pause. Here is the twelve months leading into the announcement:
| | Mar 13, 2025 | Mar 13, 2026 | 12-month change |
|---|---|---|---|
| Lumentum (LITE) | $60.29 | $622.50 | **+933%** |
| Coherent (COHR) | $62.40 | $242.76 | **+289%** |
| S&P 500 | — | — | **+20%** |
Lumentum was a $60 stock a year ago. From its April 2025 low of $49.56 to the March 2 close, it rose nearly **16x**. The repricing didn't wait for NVIDIA's press release — it tracked the CPO roadmap, the transceiver order books, and the AI capex announcements through the back half of 2025. By the time the checks arrived, the market had spent three quarters front-running them.
See the pattern? The $4 billion didn't start the photonics trade. It confirmed a trade that was already crowded.
## The Uncomfortable Part
Which is exactly how the last two weeks have behaved. The announcement-day closes — $783.25 and $298.91 — remain the highest closes either stock has ever posted. Since then:
| | Feb 27 (pre-deal) | Mar 2 (deal day) | Mar 13 (today) | vs deal day | vs pre-deal |
|---|---|---|---|---|---|
| Lumentum | $700.91 | $783.25 | $622.50 | **-20.5%** | **-11.2%** |
| Coherent | $258.93 | $298.91 | $242.76 | **-18.8%** | **-6.2%** |
Lumentum closed back below its pre-announcement price the very next session. Coherent lasted three. As of tonight's close, both trade *below* where they did before NVIDIA wrote the checks — a $4 billion endorsement from the most important company in the AI trade, fully digested within days. That is what "priced in" looks like in the wild.
The other honest caveats:
**Penetration is still 0.5%.** Everything between here and the 35%-by-2030 forecast is execution — new packaging, new failure modes, new manufacturing yield curves. Optical technology transitions have slipped before, and every quarter CPO slips is a quarter the incumbent 30-watt pluggable module keeps selling.
**Copper isn't dead.** Inside the rack, copper remains the right answer on cost and power — NVL72's spine is copper by design, and nothing announced on March 2 changes that. Both agreements are explicitly **non-exclusive**. NVIDIA kept its options open; your analysis should too.
**The supply chain has a chokepoint.** The lasers at the heart of every optical link are built on compound semiconductors — and the raw material gallium is about as concentrated as commodities get: China produced roughly **98%** of the world's low-purity gallium in 2024, per the USGS, banned exports to the U.S. in December 2024, then suspended the ban in late 2025 under a trade truce. A single policy reversal reaches every laser fab on the planet. That chokepoint deserves its own article — it's next in this series.
## What This Means
The AI buildout keeps migrating its bottleneck. In 2023 the constraint was GPUs. By 2025 it was power and land. The next constraint — the one NVIDIA just spent $4 billion to get ahead of — is the connective tissue: moving data between chips fast enough, and efficiently enough, that a million GPUs can act like one machine.
Photonics is the infrastructure layer's answer, and it follows the classic picks-and-shovels logic: whichever model wins, whichever lab wins, the light has to be generated by someone's laser.
What we're watching from here:
1. **GTC, next week.** Huang takes the stage in San Jose. Keynote minutes spent on photonics will be a live read on how central this is to the Rubin generation.
2. **Shipping in volume.** Photonic InfiniBand switches were slated for late 2025 and Ethernet versions for 2026. Volume shipments, not demos, are the proof.
3. **The penetration curve.** CPO moving off 0.5% faster — or slower — than the 2030 forecast implies.
4. **Who locks capacity next.** NVIDIA just showed that laser supply is worth paying billions to reserve. Watch whether the hyperscalers follow.
The question photonics asks of investors isn't whether light replaces copper at AI scale — the physics has already voted. It's whether you understand who gets paid along the way, and how much of that payment the market collected before you arrived. The last two weeks suggest: more than the headlines admit.
---
*Data: Barebone | Sources: NVIDIA–Lumentum and NVIDIA–Coherent press releases (SEC EDGAR, March 2, 2026), NVIDIA GTC 2025 keynote and Spectrum-X/Quantum-X Photonics announcement (March 2025), TrendForce (March 11, 2026), USGS Mineral Commodity Summaries, Google "Mission Apollo" paper | Prices as of March 13, 2026 close*