Two Checks, One Map
On March 2, NVIDIA announced two $2 billion investments in a single day. One went to Lumentum, a laser company in San Jose. The other went to Coherent, a laser company headquartered in Saxonburg, Pennsylvania. Neither makes GPUs. Neither trains models. Most investors could not sketch what they actually sell.
NVIDIA can. Both deals came bundled with multi-billion-dollar purchase commitments and rights to future manufacturing capacity - the kind of terms you negotiate when you are worried about supply, not returns.
Why is the world's most valuable chipmaker writing checks into the optics industry? Because of the funnel above it. On their Q4 earnings calls in late January and early February, the four hyperscalers guided 2026 capital expenditures to record levels: Amazon to roughly $200 billion, Alphabet to $175 - 185 billion, Meta to $115 - 135 billion, Microsoft to $110 - 120 billion. Up to $630 billion from four companies, against $388 billion actually spent in 2025 - a 62% step-up in a single year.
Almost every one of those dollars buys the same machine: racks of accelerators that must move data between themselves as light. The path that light travels - GPU package to fiber in the wall - runs through seven supply-chain layers, most controlled by five or fewer companies. Some by one.
We used Barebone AI to trace the chain end to end: filings, capex guides, shipment data, and two weeks of violent tape. This is the map.
The Map at a Glance
If you read our photonics explainer last week, you know the physics: at AI speeds copper stops working beyond a meter or two, so everything that leaves the rack travels as laser light through glass. This piece is about who gets paid along the way.
| Layer |
What it is |
Who controls it |
| 1. Raw minerals |
Gallium, germanium, indium - the refined feedstock |
China (~99% of refined gallium) |
| 2. Substrates |
Indium phosphide crystal wafers |
Sumitomo Electric, JX Advanced Metals, AXT (~80 - 90% combined) |
| 3. Lasers |
The chips that generate and modulate light |
Lumentum, Coherent, Broadcom, Mitsubishi, Sumitomo |
| 4. Foundry / CPO |
Fusing photonics onto logic silicon |
TSMC at the leading edge |
| 5. Transceivers |
Pluggable modules that put light into servers |
Innolight, Coherent, Eoptolink lead |
| 6. Switching |
Directing light across the data center |
Broadcom and NVIDIA silicon |
| 7. Deployment |
AI data centers at hyperscale |
Amazon, Microsoft, Google, Meta |
Walk it from the ground up.
Layers 1 - 2: The Ground Floor
Everything starts as rock. The laser chips at the heart of the system are grown from gallium and indium compounds; germanium shows up in the photodetectors at the far end of the fiber. The refining is even more concentrated than the mining: by the most recent full-year figures, China refined roughly 99% of the world's gallium and about 60% of its germanium.
That concentration has already been weaponized once. In December 2024, Beijing banned exports of gallium and germanium to the United States outright; a US Geological Survey analysis put the potential cost of that ban alone at $3.4 billion, nearly half of it landing on the semiconductor industry. Then, after the October trade truce, China suspended the ban on November 9, 2025 - until November 27, 2026, with export licenses still required and sales to US military end-users still prohibited.
The chokepoint did not disappear. It is on a timer.
Layer 2 is where those elements become something you can build on: single-crystal discs of indium phosphide, sliced into wafers - the blank canvas every high-end photonic chip starts from. Growing these crystals without defects is brutally hard, and industry estimates put 80 - 90% of the InP substrate market in the hands of three companies: Sumitomo Electric (roughly 30% alone), JX Advanced Metals, and AXT. The wrinkle most investors miss: AXT is listed on Nasdaq but grows its crystals through its Beijing subsidiary, Tongmei - even the American name in the trio produces inside China.
Sumitomo has announced plans to expand 6-inch InP wafer capacity about 40% by the end of 2026. The layer is tight, and everyone in it knows.
Layer 3: Five Companies Make the Light
On top of those wafers come the lasers - dies smaller than a grain of sand that generate light and switch it on and off billions of times per second to encode data. The workhorse for AI links is the EML, the electro-absorption modulated laser, and according to TrendForce only five companies produce them at scale: Lumentum, Coherent, Broadcom, Mitsubishi Electric, and Sumitomo Electric.
Five suppliers, one dominant customer. By December 2025, TrendForce reported that NVIDIA had "pre-allocated a large portion of EML capacity," producing a worldwide shortage with lead times stretching beyond 2027.
Then came March 2. The $2 billion checks to Lumentum and Coherent were not passive stakes: both companies are putting the money into new US-based fabs, and NVIDIA locked in purchase commitments alongside the equity.
When the most informed buyer in the market stops placing orders and starts buying the capacity itself, the shortage is no longer a rumor. It is on the balance sheet.
Layers 4 - 5: Where Light Meets the Module
A laser is useless until it is married to electronics. Layer 4 is that marriage - and at the leading edge it increasingly happens inside one company. TSMC's COUPE platform, co-developed with NVIDIA and detailed at SEMICON Taiwan in September 2025, stacks a photonic chip directly beneath an electronic chip using TSMC's wafer bonding. NVIDIA's co-packaged optics parts are the first products built on it; Broadcom has reportedly adopted COUPE for its own roadmap. GlobalFoundries and Tower run silicon-photonics platforms of their own - this layer is not literally a monopoly - but the highest-profile co-packaged parts in the industry all route through the same foundry that already makes the GPUs.
Layer 5 is the one piece of this chain most investors have actually heard of: the transceiver - the pluggable module that converts electrical signals to optical and back, where light physically enters the AI server. Jensen Huang did the arithmetic on stage at GTC 2025: six pluggable transceivers per GPU, 30 watts each - 180 watts and roughly $6,000 of optics per GPU before it computes anything. Scale that to a million-GPU factory and the transceivers alone draw 180 megawatts.
The market followed the math. LightCounting's December 2025 update put 2025 optical transceiver sales above $23 billion, up roughly 50% in a year, with Ethernet modules for AI clusters at about $17 billion of that. The vendor table is top-heavy: Innolight leads (2024 revenue of roughly $3.3 billion, up 123%), Coherent sits second, and Eoptolink jumped from seventh to third (revenue of about $1.2 billion, roughly tripled). By industry estimates, Innolight and Eoptolink together took about 60% of NVIDIA's incremental 800G orders for 2025, and Innolight is projected to hold 50 - 60% of the new 1.6T module market. A chokepoint hides inside the chokepoint: the digital signal processor in most modules comes from Broadcom or Marvell.
TrendForce expects worldwide shipments of 800G-and-above transceivers to hit 24 million units in 2025, then jump 2.6x to nearly 63 million in 2026. That is the demand cliff every layer below is straining against.
Layer 6: Two Ways to Switch the Light
Once light is inside the data center, it has to be directed - and the silicon that does the directing is effectively a two-vendor market. Broadcom's Tomahawk switch chips power the merchant ecosystem: Arista's systems and the white boxes hyperscalers design themselves. NVIDIA sells the network as a system - InfiniBand and Spectrum-X Ethernet - and at GTC 2025 it announced the next turn of the screw: switches with the optics co-packaged directly onto the silicon (Quantum-X Photonics first, announced for late 2025, with Spectrum-X Photonics versions due in 2026), which NVIDIA says use 4x fewer lasers and deliver 3.5x better power efficiency than pluggable designs.
Layer 7 - deployment - is the $630 billion we started with. Four buyers, writing the largest capex checks in corporate history. Every dollar of it depends on the six layers underneath.
The Two-Week Stress Test
Maps are theories. The first two weeks of March tested this one in public.
On March 2, the day NVIDIA's investments hit the wire, the chain repriced instantly - and the further down the chain, the bigger the move. AXT, the substrate maker, gained +22.2% in a single session. Coherent gained +15.4%, Lumentum +11.7%. TSMC - too big for photonics to move - slipped 1.5%.
Then came the purge: a broad selloff in AI hardware hit that first week - on March 6 alone, Lumentum fell 14.2% and AXT 16.6%. Two weeks after the announcement, at tonight's close:
| Company |
Layer |
Mar 2 (deal day) |
Feb 27 → Mar 16 |
Mar 16 close |
| AXT |
2 - substrates |
+22.2% |
+27.7% |
$48.39 |
| Coherent |
3/5 - lasers, modules |
+15.4% |
-4.5% |
$247.37 |
| Lumentum |
3 - lasers |
+11.7% |
-10.9% |
$624.84 |
| TSMC |
4 - foundry |
-1.5% |
-9.2% |
$340.23 |
Read that table again. The two companies that received $4 billion from NVIDIA ended the fortnight below where they started. The company that received nothing - AXT, one layer further down - held a +27.7% gain through a selloff that took 14.2% out of Lumentum in a day.
That is the market doing what this map predicts. NVIDIA's checks confirmed the bottleneck at layer 3 - and a confirmed bottleneck at layer 3 means laser fabs run flat-out for years, so capital went hunting for the next binding constraint: the substrates those lasers are grown on. The signal traveled down the chain, not up it.
Where the Map Breaks
An honest map marks its own hazards. This one has five.
The architecture is trying to delete layer 5. Co-packaged optics exists to eliminate pluggable transceivers - the very layer that produced 2025's most spectacular revenue growth. If CPO ramps on schedule, the module makers' richest products get designed out of next-generation switches. The transition will take years and pluggables will dominate well into the late 2020s, but the hottest layer on this map is the one the industry's roadmap is aimed at.
Chokepoints attract capital, and capital widens them. NVIDIA's $4 billion explicitly funds new US fabs. Sumitomo is adding 40% to its InP capacity. Today's scarcity premium is funding tomorrow's supply - chokepoint maps have a shelf life of quarters, not decades.
Layer 1 is policy, not geology. China's export ban was imposed in December 2024 and suspended in November 2025. What flexed once can flex again - in either direction - and the current suspension expires November 27, 2026.
The volatility is disqualifying for most portfolios. These names round-trip double-digit moves inside a week. Lumentum lost 14.2% in one session this month after receiving a $2 billion vote of confidence from its largest customer. Concentrated supply is not the same thing as a safe stock; the tape this month is the proof.
The funnel is guidance, not contract. Everything in this article levers off $630 billion of planned spending. Capex guidance gets revised, and a 62% step-up leaves no room for disappointment. If the April earnings season trims those numbers, the entire chain re-rates from the top down.
What This Means
A supply-chain map is not for admiring. It exists to show where scarcity actually binds - because pricing power lives where the supplier count is smallest, and commoditization where it is largest.
Five signals will tell you whether this map still holds through 2026:
- EML lead times. Still quoted beyond 2027? The layer-3 squeeze is intact. Coming inside 12 months? Capacity is catching up.
- CPO ship dates. Quantum-X Photonics volumes and the 2026 Spectrum-X Photonics ramp are the test of whether layer 5 gets redrawn.
- 1.6T qualification wins. Whoever ships 1.6T modules at scale this year inherits the steepest part of the 24-million-to-63-million unit cliff.
- November 27, 2026. The expiry date on China's export-ban suspension is the single most schedulable risk event on this map.
- April capex prints. The first 2026 guidance revisions - up or down - reprice the whole funnel.
The nearest catalyst is days away: GTC, NVIDIA's annual conference, runs this week in San Jose. Last year's keynote handed this chain its co-packaged-optics roadmap; this year's will say whether it is on schedule - and which layers tighten next.
Seven layers, most of them in five or fewer hands. One sits on a political timer, one is in open shortage, one is being deliberately designed out of existence. Putting tickers and valuations against every layer of this map is a separate exercise - but it is the obvious next one.
Data: Barebone | Sources: NVIDIA, Lumentum and Coherent March 2, 2026 announcements (SEC 8-K filings), NVIDIA GTC 2025 keynote, TrendForce (December 2025), LightCounting (2025), USGS, Q4 2025 hyperscaler earnings calls | Prices as of March 16, 2026 close
