AI’s copper problem is turning into a full-on sector trade, and the market is doing what it always does when physics gets inconvenient: it starts pricing the workaround. Silicon photonics is moving from lab-speak to budget line, and the names tied to that shift are suddenly getting much more interesting than your average “AI adjacent” cosplay.
Here’s the clean version: AI clusters are getting too big for copper to handle efficiently, so the market is chasing optical interconnects, co-packaged optics, lasers, and the plumbing that lets chips gossip at light speed without melting the power bill.
The core issue is simple. As AI clusters scale to tens and even hundreds of thousands of processors, copper wiring starts to hit a physical wall. At the speeds these systems need, a passive copper cable can carry a clean signal for less than a meter before the signal degrades. Push further and you burn more power just to keep data intact, which means more heat, more cost, and more excuses from executives with glossy slide decks.
Silicon photonics is the fix because it uses light instead of electrons. That lets data travel farther, carry more bandwidth, and waste less power. The market is already leaning in: the broader optical transceiver market is projected to jump about 60% in a single year to roughly $26 billion in 2026, while the silicon photonics market is estimated at $3.6 billion in 2026 and projected to reach $15.7 billion by 2033. If you enjoy following money with no patience for poetry, this is your aisle.
Marvell announced a definitive agreement on December 2, 2025, to acquire Celestial AI for a total potential value of over $5.5 billion, expected to close in Q1 2026. That matters because Celestial AI’s Photonic Fabric technology replaces copper-based electrical signals with light-based optical interconnects and, per Wedbush, cuts interconnect power consumption by over 80%. Translation: Marvell is buying a direct ticket into the part of AI infrastructure that exists because copper keeps embarrassing itself.
Trading profile: this is not some sleepy utility stock. It’s a strategic acquisition story sitting inside a fast-growing hardware transition, which usually means investors will treat every update like a referendum on the future of data center architecture. The reward is obvious: Marvell gets a stronger silicon photonics hand. The risk is also obvious: execution, integration, and customer adoption can turn a “smart move” into a very expensive science project.
Investor takeaway: Marvell is one of the cleanest names if you want exposure to the optical shift without betting on a tiny pure-play that could get whiplash from one customer order.
Broadcom began mass-shipping its Tomahawk 6 “Davidson” switch, described as the industry’s first 102.4 Tbps CPO platform. Wedbush says it reduces interconnect power from 30 pJ/bit to under 5 pJ/bit, which is the sort of efficiency jump that makes old copper setups look like they were engineered by a committee during a lunch break. This is the co-packaged optics version of telling the network to eat a salad and stop guzzling power.
Trading profile: Broadcom is the heavyweight here, which means the market tends to view it less like a speculative moonshot and more like a toll booth on the data center highway. That doesn’t make it boring; it makes it dangerous in a different way. When a giant starts getting more embedded in AI infrastructure, expectations get sticky and disappointments get punished with corporate-grade efficiency.
Investor takeaway: Broadcom gives investors a large-cap way into the same theme, especially if they want exposure to switch silicon and co-packaged optics instead of betting only on component makers.
Credo entered a definitive agreement on April 13 to acquire DustPhotonics for about $750 million in cash and stock, bringing silicon photonics PIC technology in-house. That is a very specific kind of move: not just riding the wave, but buying a bigger surfboard before the wave changes shape. The message is pretty blunt. Credo wants tighter control over the optical stack as AI interconnect demand keeps moving up the food chain.
Trading profile: acquisition-driven stories are nice until lawyers and regulators start stretching the calendar. The current expectation is that the deal closes in Q2 of calendar 2026, subject to regulatory approvals. That means investors get a thesis with a timeline, which is adorable until time starts charging rent. Still, bringing photonics capability inside the company is the kind of move that can matter when customers want speed, power savings, and fewer handoffs between vendors.
Investor takeaway: Credo is worth watching as a smaller, more direct beneficiary of optical networking demand, but the deal and integration timeline add the usual dose of corporate friction.
GlobalFoundries management stated that silicon photonics revenue doubled in 2025 and is expected to nearly double again in 2026, with a clear line of sight to a $1B+ run rate by end of 2028. It also plans to launch a silicon photonics and advanced packaging CPO facility in Malta, New York in 2025. That’s not just commentary; that’s industrial real estate trying to keep up with the AI arms race.
Trading profile: this is the picks-and-shovels version of the trade. GlobalFoundries does not need to be the face on the conference badge to make money from the build-out. It sits where manufacturing capacity, packaging, and photonics intersect, which gives it leverage on a trend that is still early enough to be messy but large enough to be unavoidable.
Investor takeaway: GlobalFoundries offers a less flashy, more infrastructure-heavy way to play silicon photonics, and the revenue trajectory suggests the company is seeing real demand, not just buzzword inflation.
Coherent is described as a global leader in the optical technology feeding AI data centers, and demand for its datacenter transceivers has surged as cloud giants build out. The company has also deepened a partnership with Nvidia to pioneer next-generation silicon photonics, which puts it near the center of the transition from copper to light. In other words, Coherent is not trying to invent the hype cycle. It’s trying to sell the parts that make the hype physically possible.
Trading profile: optical component makers are cyclical and lumpy, and their fortunes can swing hard from quarter to quarter based on a handful of huge customers. That is a polite way of saying the revenue line can behave like it has a caffeine problem. But when demand is tied to a structural shift in data center architecture, the volatility can be the price of admission rather than a red flag.
Investor takeaway: Coherent remains one of the clearest pure exposure names to the shift, but it comes with the same warning label as the rest of the optical crowd: high upside, high drama, no mercy for sloppy positioning.
This trade is not about who tweets AI the hardest. It is about which companies make the wires, lasers, switches, and manufacturing tools that let giant AI clusters keep scaling after copper taps out. Marvell, Broadcom, Credo, GlobalFoundries, and Coherent each sit somewhere on that path, but none of them are sleepy, and none of them deserve a position sized like a church donation.
