Silicon photonics.
Light moves inside.

The bottleneck in AI isn't the chip anymore. It's the wire.

A GPU does math faster than copper can feed it. At today's speeds, a copper link carries a clean signal about a meter before it turns to mush, so the most expensive hardware ever built spends part of its day waiting on the cheapest part of the rack.

Light fixes that. Silicon photonics builds the parts that turn electrical signals into light (lasers, modulators, and detectors) using the same wafers and fabs as ordinary chips. Fiber then carries the signal farther and faster on a fraction of the power, and power is what data centers actually run out of.

Investors figured this out a while ago. For two years, the money came from pluggable transceivers, the optical modules that slot into the front of a switch and ship by the million, and Bank of America called photonics the best-performing theme in the world this year. Now everyone owns silicon photonics stocks. Lumentum trades above 100 times trailing earnings, and in one July session the leaders dropped about 10% on no news at all. That's a crowded trade.

It's still worth your attention, because the next leg is a different business. Co-packaged optics (CPO) takes the optics off the faceplate and puts them inside the chip package, right next to the switch silicon. Broadcom has shipped CPO switches for a year, and Nvidia's first CPO Ethernet switch ships with Vera Rubin this fall. 2027 is the year CPO shows up as revenue instead of slides.

That shift changes which silicon photonics stocks win. Here's what each company sells into it:

• Lasers: the one part that can't move inside the package.
• Switch owners: the chip companies deciding where the optics go.
• Factories: the fabs, assembly lines, and glass the whole thing runs on.
• Pluggable holdouts: the bet that today's architecture lasts longer than the CPO crowd thinks.
• Moonshot and Private: small and early, priced on what might happen.

CPO moves nearly everything into the package except the light source. Lasers run hot and fail first, and nobody wants to scrap a switch because one died. So they sit outside in swappable modules, and they have to be far more powerful than anything in a pluggable.

Lumentum NASDAQ: LITE

Lumentum makes those lasers. It's a laser supplier named in Nvidia's CPO supply chain, and it already holds orders for ultra-high-power lasers that deliver in the first half of 2027.

Fewer lasers per switch, each one harder to make, from a short list of fabs that can make them reliably. That's pricing power, and it shows: fiscal Q4 revenue doubled to $1.01 billion and adjusted gross margin crossed 50%. Lumentum also runs a fast-growing optical switching business, but the laser is why it leads this list.

Coherent NYSE: COHR

Coherent is betting hyperscalers don't want to buy lasers from one vendor, photonic chips from another and assembly from a third. They want one supplier who owns the whole engine and takes the call when it breaks.

That's PhotonLink, launched in September: lasers, silicon photonics, optics, fiber and assembly sold as one integrated platform for CPO. Coherent says it adds more than $30 billion to the market it can sell into by 2030, with revenue starting to ramp this quarter. Most companies would have to buy half those parts. Coherent already makes them, which is the hard part. If the bundle wins, Coherent stops being a component vendor and becomes the default.

These companies make the silicon the optics attach to, so they decide when and how light moves in. You get the theme inside a much bigger business.

Nvidia NASDAQ: NVDA

Nobody buys Nvidia for photonics. It's here because it sets the schedule. When Jensen Huang said in March that Nvidia would keep using copper alongside optics, the group sold off. Now Spectrum-X Photonics is in volume production for Vera Rubin, and getting named on that supplier list is the best thing that can happen to an optics company. Watch Nvidia for the timeline, not the exposure.

Broadcom NASDAQ: AVGO

Nvidia gets the headlines, but Broadcom got there first. Its Tomahawk 6 "Davisson" switch has shipped with co-packaged optics since October 2025, and it's the company's third generation of the technology.

Google, Meta and the other companies designing their own AI chips build them with Broadcom. When those clusters go optical, the optics come from the same vendor. AI chip revenue hit $16.7 billion last quarter, up 221%. CPO is a small slice of that today, so you get the most mature CPO franchise inside a stock you'd own for other reasons anyway.

Marvell NASDAQ: MRVL

Marvell closed its $3.25 billion purchase of Celestial AI in February, with up to $2.25 billion more in stock if revenue targets hit. Celestial's Photonic Fabric runs optical links underneath the chip instead of at its edge, which is where GPU-to-GPU networking is headed. Management's own plan puts it at a $500 million run rate by the end of fiscal 2028. This is a 2028 product you buy in 2027.

Somebody has to print the photonic chips, assemble the modules and pull the fiber. Nobody tweets about these jobs, and the contracts are long.

Tower Semiconductor NASDAQ: TSEM

If optics is turning into a chip business, somebody has to run the fab. Most photonics designers don't own a silicon fab, and Tower rents them one. It's the go-to independent foundry for silicon photonics.

Silicon photonics revenue grew more than 270% year over year in Q2, to a $680 million annual run rate, and Tower expects to pass a $1 billion run rate in the fourth quarter. About $1.3 billion of 2027 silicon photonics revenue is already under customer contract. Silicon photonics is already more than a third of Tower's revenue and by far its fastest-growing piece.

Fabrinet NYSE: FN

Fabrinet assembles a big share of the world's high-speed optical modules in Thailand for Cisco, Nvidia, Amazon and others. Orders hit its factory floor before they hit anyone's earnings, which makes it the group's best demand gauge.

In August it posted record revenue of $1.32 billion, guided higher, and still dropped nearly 20% in a day. The business didn't break. The market showed you how little slack it gives a contract manufacturer at these prices. When Fabrinet guides up again, the rest of the group usually follows.

Corning NYSE: GLW

Every photon in this story travels through glass, and Corning makes the glass. Meta signed a multiyear deal worth up to $6 billion for Corning fiber and connectivity, and two more hyperscalers have signed agreements of similar size. It also supplies the fiber connectivity inside Broadcom's CPO switches.

That's demand booked years out, from a 175-year-old company. Even Corning dropped 14% in a day in September when it set up a $2 billion stock sale, so nothing in this trade is sleepy. But if you want the theme with contracts underneath it, start here.

CPO is coming, but pluggables and copper will still carry most of the traffic through 2027. These two companies get paid while the switch takes its time, which makes them the silicon photonics stocks that don't need CPO to hurry.

Applied Optoelectronics NASDAQ: AAOI

AOI is the purest bet that pluggables have a longer life than the CPO crowd expects. It makes its laser chips and some of its transceivers in Texas, which matters if Washington follows through on banning new Chinese-made modules, the source of most of the world's supply. Capacity goes from about 200,000 units a month to 650,000 by year-end, and management says demand outruns supply into mid-2027.

You're paying for that build in shares. AOI has set up three at-the-market stock programs this year, worth up to $1.7 billion, most recently $600 million in August. Growth funded by dilution, in the most volatile stock on the list.

Credo NASDAQ: CRDO

Credo is the hedge. Its active electrical cables are the reason copper still owns the short runs inside a rack, and they're selling fast: revenue was $479 million last quarter, up 115%, at a 68% adjusted gross margin. It also sells the DSP chips that go inside optical transceivers.

If light takes over faster, Credo sells more DSPs. If copper hangs on, it sells more cables.

POET Technologies NASDAQ: POET

POET has a clever idea and almost no revenue. Its interposer builds optical engines at the wafer level without the slow, expensive alignment step everyone else uses. In May it landed a $50 million first order from Lumilens, under a deal that could reach $500 million over five years.

Q2 revenue was $569,925. Not million. Either that order ships and POET becomes a real company, or it doesn't. Size it like a lottery ticket.

The most ambitious CPO designs are still private. These two set the price tag on the technology for everyone else.

Ayar Labs

Ayar Labs makes optical I/O chiplets that sit beside the GPU so chips can talk to each other over light. Nvidia, AMD and Intel are all investors, and in June it joined Nvidia's NVLink Fusion program so its parts plug straight into Nvidia systems. It raised $500 million in March at a $3.75 billion valuation and another $150 million in September. If it files to go public, the S-1 will be the first honest look at what CPO economics actually are.

Lightmatter

Lightmatter's Passage is a 3D photonic interposer that turns the whole surface of a chip into optical I/O, rated at 114 terabits per second from a single package. Its last priced round, in 2024, valued it at $4.4 billion. It competes head-on with the technology Marvell just paid $3.25 billion for, so every chipmaker without an optical story knows what the going rate is. It's a takeover story before it's an IPO story.

Co-packaged optics could do what a tariff can't: pull the optics business away from China.

Most of the world's pluggable transceivers come from Chinese module makers, which is why a U.S. ban on new Chinese-made modules would matter so much. A pluggable is a box that slides into the front of a switch, and a hyperscaler can buy it from whoever is cheapest that quarter. Cheap, fast assembly at enormous scale is exactly the game Chinese factories win.

CPO takes the box away. The optical engine moves inside the switch package, and the switch maker, not the hyperscaler's purchasing team, decides whose lasers, photonic chips and assembly go into it. Broadcom and Nvidia make that call once per generation, from a short list of suppliers they've qualified.

Pluggables will still carry most of the traffic for years, so the Chinese module makers aren't going anywhere soon. But every rack that goes co-packaged is a rack where the optics money flows to suppliers picked by two American chip companies. In 2027, that short list starts showing up as revenue, and silicon photonics stocks will trade on who made it.