Fiber Optic Modules Are Manufactured Worldwide

 

The Shenzhen Reality Nobody Puts in Trade Show Presentations

 

Fly into Shenzhen Bao'an, take a taxi to Longhua or Dalang, and you'll pass more optical transceiver factories in forty minutes than exist in all of Europe combined. The density is absurd. I counted seventeen separate SFP manufacturers within a two-kilometer radius of my hotel during one sourcing trip. Seventeen. Some occupy entire industrial parks. Others share a single floor of a nondescript building with an injection molding operation and a company making LED Christmas lights.

This is where the volume happens. Somewhere between 60 and 70 percent of the world's pluggable optical transceivers ship from Guangdong province. The exact number depends on how you count-finished modules versus subassemblies versus bare component kits-but the magnitude isn't controversial.

The factories range from genuinely world-class operations with Class 10K cleanrooms and automated die bonders to shops where I've watched workers hand-place laser diodes using tweezers under a stereo microscope. Both types ship product that passes the same conformance tests. The difference shows up eighteen months later in field failure rates, but by then the module is someone else's problem.

 

Why Japan Still Matters More Than Anyone Admits

 

Here's what the supply chain maps leave out: critical compound semiconductor components-the actual photon-emitting and photon-detecting bits-still flow predominantly from Japan.

Sumitomo Electric. Mitsubishi Electric. Lumentum's Japanese operations. Coherent's legacy II-VI fab in Toyama. These facilities produce the DFB laser chips, EML transmitters, and high-speed photodiodes that Chinese assemblers cannot yet replicate at equivalent performance levels. The 1310nm uncooled EMLs running in your hyperscaler's 400G-DR4 modules? Overwhelmingly Japanese silicon. The high-power pump lasers in your EDFAs? Japanese. The avalanche photodiodes in your OTDR? Japanese.

The material science gap is real. Growing indium phosphide epitaxial layers with the uniformity required for high-yield DFB gratings requires decades of institutional knowledge that doesn't transfer via acquired IP or hired engineers. MOCVD reactor recipes are proprietary down to gas flow rates and temperature profiles. The tacit knowledge lives in the heads of process engineers who've been running the same tools for twenty years.

I spent three days at a Japanese laser fab once. Watched a technician reject an entire wafer because the photoluminescence spectrum showed a 2nm wavelength shift from target. Two nanometers. The wafer probably would have yielded functional devices. But "functional" wasn't the specification-"within spec at -40°C over 20 years" was the specification, and that 2nm shift suggested something had drifted in the growth process.

You don't build that culture in five years.

 

The Component Shell Game

 

Follow the actual parts through a "made in China" optical transceiver and you'll find a geography lesson.

Laser diodes: Japan, with increasing competition from Chinese domestic suppliers like Accelink for lower-speed parts. The performance gap at 25G PAM4 and above remains significant.

Driver and TIA integrated circuits: Taiwan dominates through TSMC's photonics-adjacent processes and fabless design houses in Hsinchu. Semtech, Macom, and Broadcom all tape out there. China has been trying to build domestic alternatives for a decade with limited success-the EDA tools alone create dependency chains.

Optical isolators and circulators: Mostly Chinese now. Casix, Agiltron's Chinese operations, and a dozen smaller players in Fuzhou have essentially commoditized these components. Quality is fine for datacom.

Fiber array units and ferrules: Japanese precision manufacturing still leads for tight-tolerance MT ferrules. Chinese suppliers handle volume on standard LC/SC connectors.

Ceramic packages and submounts: Split between Japanese specialists like Kyocera and Chinese volume manufacturers. The thermal management requirements for high-power transmitters still favor Japanese sources.

PCBs and flex circuits: Taiwan and mainland China, with some specialty high-frequency boards from Japanese suppliers.

 

The Tariff Scramble

 

The 2018 tariffs reshaped the industry's manufacturing geography in ways that are still playing out.

Before Section 301, the calculation was simple: manufacture in China for cost, ship to the world. Hyperscalers and enterprises bought through distribution or direct, paid the China price, and didn't think much about country of origin.

Then 25% tariffs landed. Suddenly every purchasing manager needed to understand the difference between HTS codes, substantial transformation rules, and what exactly constitutes "manufacturing" versus "assembly" in the eyes of CBP.

The initial response was chaos. I know companies that literally air-freighted partially assembled modules to Mexico for final kitting, stuck "Assembled in Mexico" labels on them, and hoped the tariff exclusion held. Some got away with it. Some didn't. The enforcement was-and remains-inconsistent.

The more sophisticated players established actual manufacturing presence in tariff-friendly locations. Vietnam attracted significant investment. Malaysia picked up some overflow. Thailand got a few operations. Innolight opened a facility in Tijuana. FS built out Mexican capacity.

But here's the thing nobody talks about: the component supply chains didn't move. The laser diodes still come from Japan. The driver ICs still come from Taiwan. The passive optics still come from China. Performing final assembly in Vietnam doesn't eliminate China exposure-it just adds a logistics hop and a compliance headache.

The hyperscalers learned to live with it. They have purchasing teams who do nothing but manage tariff exposure. Smaller buyers got squeezed.

 

What Actually Happens Inside the Clean Room

 

Most optical transceiver content you'll read glosses over the actual manufacturing process. The glossy factory tour videos show robots and gleaming equipment. The reality is more complicated.

Die bonding: The laser or VCSEL chip gets attached to a submount using eutectic solder-typically gold-tin at 280°C-or conductive epoxy. Placement accuracy matters. For edge-emitting lasers, you need lateral alignment within a few microns to hit the fiber core. For VCSELs talking to multimode fiber, tolerances are looser but you're placing multiple emitters in an array.

I've watched experienced operators hit first-pass yields over 95% on die bonding. I've also watched less experienced operators destroy expensive laser bars by overheating them during reflow. The difference is touch, pattern recognition, knowing when the solder joint looks right versus when something went wrong.

Wire bonding: Gold or aluminum wires connect the die to the circuit traces. For RF-critical connections-the high-speed signal paths from driver to modulator-you use ribbon bonding to reduce inductance. The wire bonds look trivial but they're a major yield limiter. One contaminated bond pad, one misplaced loop that shorts to the next trace, and the module fails final test.

Fiber attachment: Getting light from the laser into the fiber requires alignment precision that borders on the absurd. Active alignment systems drive the fiber to the position that maximizes coupled power, then UV-cure epoxy freezes everything in place. The epoxy shrinks slightly during cure. Good process engineers compensate. Bad process engineers wonder why their yields dropped.

Hermetic sealing: High-reliability modules get sealed under dry nitrogen in a windowed metal can. The seal has to hold for 20 years in a telecom environment. Seam welding or resistance welding does the job. Laser welding is faster but introduces thermal stress.

Every one of these steps can be automated, semi-automated, or performed manually. The equipment costs scale accordingly. A fully automated 400G production line runs into eight figures. A manual line can be set up for under $500K. Both produce working modules. The difference shows up in consistency, throughput, and long-term reliability.

 

 

Testing Is Where Corners Get Cut

 

A properly tested optical transceiver goes through a gauntlet:

Parametric testing at room temperature: optical power, extinction ratio, receiver sensitivity, eye diagram mask compliance. This is table stakes. Everyone does this.

Temperature cycling: run the same tests at -40°C, +85°C, and several points between. This is where marginal units fail. Laser threshold current shifts with temperature. Thermal expansion stresses die bonds. Receiver sensitivity degrades as photodiode dark current increases.

Burn-in: run modules at elevated temperature under continuous traffic for 24, 48, or 168 hours depending on customer requirements. Infant mortality failures happen in the first few hundred hours. Catching them before shipment is cheaper than catching them in the data center.

Here's the problem: testing costs money and takes time. Temperature chambers aren't free. Burn-in consumes floor space. Every hour a module sits in test is an hour it's not shipping.

The pressure to reduce test time is constant. A module that needs 168-hour burn-in but gets 24 hours instead will probably work fine. Probably. The failure distribution shifts-instead of failing in your factory, marginally-bad units fail in the customer's rack three months later.

I've seen test coverage vary by an order of magnitude between suppliers claiming equivalent quality levels. The customer spec says "burn-in required." It doesn't say for how long, or at what temperature, or under what traffic pattern.

Ask your supplier what their test coverage actually looks like. Most won't answer honestly. The ones who will are usually the ones you want to buy from.

 

The Compatible Module Market

 

Major switch vendors-Cisco, Arista, Juniper-charge substantial premiums for their branded optical modules. A Cisco-branded 100G-LR4 might list at $3,500. A "compatible" equivalent from a Shenzhen assembler runs $300-600.

The modules use identical or near-identical components. The laser diodes come from the same Japanese suppliers. The driver ICs are the same Semtech parts. The difference is: the OEM module went through the vendor's qualification program and the compatible module didn't.

Some compatible modules work flawlessly for years. Some fail in ways that are annoying but manageable. Some fail in ways that take down a production network at 2 AM and create finger-pointing between your switch vendor's TAC and your module supplier's sales rep.

The failure mode I've seen most often: thermal management under sustained high-bandwidth operation. The module works fine in the lab. It works fine for the first month in production. Then summer hits, the data center runs a few degrees warmer, the module runs 3°C hotter than design margin, and the laser ages prematurely.

You can avoid most of this by buying from reputable third-party suppliers and doing your own qualification testing. The cost savings compound over thousands of modules. But you have to actually do the testing, not just assume the spec sheet tells the truth.

 

 

Who Makes What Where: A Rough Guide

 

What 800G and CPO Mean for Geography

 

The transition to 800G changes things. Power density increases. Thermal challenges multiply. The DSP complexity ratchets up.

At 800G, the signal integrity margins are razor-thin. PAM4 at 100+ gigabaud per lane demands DSP implementations that only a few companies can execute. Marvell. Broadcom. Credo. Maybe one or two Chinese players in the next generation.

The implication: 800G consolidates design capability into fewer hands even as assembly remains distributed. You can still build 800G modules in Shenzhen, but you're buying the silicon from a small oligopoly of suppliers.

Co-packaged optics pushes further. When the transceiver function integrates into the switch ASIC package, the manufacturing shifts from module factories to semiconductor fabs. TSMC and Intel become the bottleneck, not the Longhua industrial park.

This is years away from mainstream deployment. But it's coming. And it will restructure who makes what where in ways that aren't obvious yet.

 

Practical Advice for People Who Buy This Stuff

 

If you're deploying thousands of modules per year:

Qualify multiple suppliers. The tariff situation can change. Suppliers can exit the market. Single-source risk is real and it will bite you eventually.

Actually test incoming quality. Don't just trust the vendor's certificate of conformance. Run a sample through your own test setup. Track failure rates by supplier, by date code, by lot.

Understand where the components come from. Your "US-manufactured" module might have critical components from China. Your "China-free" module might not be as China-free as the sales rep claims. The supply chains are opaque and the incentives to obscure origin are significant.

If you're running a smaller operation:

Buy from established third-party suppliers with real engineering teams, not trading companies that rebrand whatever's cheapest this week. The price difference between good and sketchy is maybe 20%. The reliability difference is infinite when you're troubleshooting link flaps at 3 AM.

For everyone:

The manufacturing geography matters less than the manufacturing quality. A well-run factory in Shenzhen produces better modules than a poorly-run factory in America. Judge suppliers by their processes, their test coverage, their field failure rates-not by the flag on their marketing materials.

The technology works. Billions of modules ship every year and the vast majority perform exactly as specified. The challenge is finding the suppliers whose "exactly as specified" includes the corner cases your network will eventually encounter.