What Is Silicon Photonics? Why the Laser Is the Stuck Step

EXPLAINED · Mechanism · Last reviewed: 25 Oct 2026 · Next review: after Tower Semiconductor reports third-quarter 2026 results and Lumentum and Coherent report their September 2026 quarters

Silicon photonics is a way of making the optical parts of a data link, the circuits that guide, switch and detect light, on silicon wafers in ordinary chip factories.

Why an investor cares: every AI cluster that grows past a single rack moves its data as light, and silicon photonics is how the industry is trying to make those optical links cheaper, cooler and more numerous. Our optics in the AI cluster explainer covers why the network binds GPU scale. This page goes one level down, to the chip that does the optical work, and to the one part of the link that silicon cannot make.

What is silicon photonics?

Silicon photonics is optics built the way chips are built. Instead of assembling a transceiver from many separate lenses, crystals and tiny components, a designer draws the optical circuit and a foundry etches it onto a silicon wafer, many copies at a time, on the same kind of tools used for ordinary chips.

The optical circuit has a few basic parts. A waveguide is a narrow strip of silicon that traps light and carries it across the chip, the way a copper trace carries current. A modulator switches that light on and off, or shifts its phase, billions of times a second to write the data. A photodetector at the far end turns the light back into an electrical signal.

Simplified. Real links add packaging, fibre attach and test steps. Company examples come from the releases and calls listed in References.

How does silicon photonics work in an AI data centre?

It sits at each end of the optical links that join GPUs, servers and switches. Inside a rack, copper still does much of the work. Across racks, the electrical signal is turned into light, sent down fibre, and turned back into electricity at the other end.

Today most of that conversion happens in pluggable transceivers, small modules that slot into the front of a switch, many with a silicon photonics chip inside. The next step is co-packaged optics, where the optical engine moves into the same package as the switch chip.

The reason is power. In an August 2025 technical post NVIDIA said the electrical path from a switch chip out to a pluggable module can lose up to 22 dB of signal at 200 Gb/s per lane, and that a pluggable interface often draws about 30 watts. With the optics co-packaged, it said the loss falls to about 4 dB and the power to as low as 9 watts per interface. Our rack-scale AI explainer covers why power per rack is now the design limit.

Vendor claims, each against its own pluggable baseline. Not independently tested by Second Order.

Our arithmetic, as an illustration only: at those figures, 1,000 interfaces would draw about 30 kilowatts as pluggables and about 9 kilowatts co-packaged, a saving of about 21 kilowatts. Scale that to 100,000 interfaces and the gap is about 2.1 megawatts of power that could go to GPUs instead. The interface counts are ours, not a real cluster.

Why can’t silicon make its own light?

Because silicon is a poor light emitter. In physics terms it has an indirect bandgap: when an electron drops to a lower energy level in silicon, it usually gives off heat rather than a photon. A University of California, Santa Barbara textbook chapter on silicon lasers calls the lack of a laser source on silicon the key hurdle for silicon photonic circuits.

So the light comes from somewhere else. Data-link lasers are made from indium phosphide, a compound semiconductor that emits light efficiently. It is grown on its own wafers in its own fabs: Coherent said on its August 2026 call that it is moving from 3-inch to 6-inch indium phosphide wafers, while the silicon photonics foundries run 200 mm and 300 mm lines. Our read: that makes laser capacity the slower side to expand.

The laser is either bonded beside the silicon chip or kept in a separate external module that feeds light in over fibre. NVIDIA’s and Broadcom’s co-packaged switches both keep lasers in replaceable front-panel modules, so the part most likely to fail can be swapped.

Our read: this is the defining fact of the business. Silicon photonics moves most of the optical link onto cheap, scalable silicon, but it cannot move the laser. Every link still needs indium phosphide, and co-packaged designs need more powerful lasers than before, because one laser now feeds several channels. The Lumentum sold-out Wire is what that dependence looks like when demand runs ahead.

What is co-packaged optics, and how is it different?

Silicon photonics is how the optical chip is made; co-packaged optics is where it is placed. A co-packaged switch puts several silicon photonics engines on the same package as the switch chip, so the electrical signal only has to travel millimetres before it becomes light.

NVIDIA announced its Spectrum-X Photonics and Quantum-X Photonics switches in March 2025, at 1.6 Tb/s per port. It said they use four times fewer lasers and deliver 3.5 times better power efficiency than pluggables, built with TSMC, Coherent, Corning, Foxconn, Lumentum and SENKO, with the Ethernet version due in 2026.

Broadcom announced Tomahawk 6-Davisson in October 2025, a 102.4 Tb/s switch with 16 optical engines of 6.4 Tb/s each. It said the engines are built on TSMC’s COUPE silicon photonics platform and cut optical interconnect power by 70%. In April 2026 TSMC said COUPE on substrate, a true co-packaged version, begins production in 2026, with twice the power efficiency and a tenth of the latency of a pluggable on the board. CoWoS and hybrid bonding are the packaging steps that put chips this close together.

There is also a halfway house, near-packaged optics, which puts the optical engine on the board beside the chip. Lumentum and Coherent both reported a jump in customer work on it in August 2026, and Coherent said its content is very similar either way.

Who makes silicon photonics chips?

Mostly specialist foundries, making chips to other companies’ designs. Three names matter most for AI optics: Tower Semiconductor, GlobalFoundries and TSMC.

Tower Semiconductor is the clearest window, because it reports the number. In its second-quarter 2026 results on 4 August, Tower said silicon photonics revenue reached an annual run rate of $680 million, up from $180 million a year earlier, and that it plans to pass a $1 billion run rate in the fourth quarter of 2026. Our arithmetic: $680 million a year is about $170 million a quarter, roughly 37% of Tower’s $460 million of second-quarter revenue, and the run rate grew about 3.8 times in a year.

Customers are paying ahead to secure that supply. On its call Tower said it received $290 million of prepayments from silicon photonics customers in the first quarter of 2026, mostly for 2027 capacity, and that it has contracts for about $1.3 billion of silicon photonics revenue in 2027. It expects fourth-quarter 2026 wafer starts above three times its second-quarter silicon photonics shipments, and new 300 mm capacity in Japan by late 2027. Its 2028 target is $3.6 billion of revenue, which it called fully spoken for.

GlobalFoundries bought Singapore’s Advanced Micro Foundry in November 2025. GF said the deal makes it the largest pure-play silicon photonics foundry by revenue, adds AMF’s 200 mm silicon photonics platform in Singapore with plans to move to 300 mm, and sets up a research centre with A*STAR, Singapore’s public research agency, focused on 400 Gb/s data transfer. Our Singapore markets door tracks the city’s data-centre side of the same build-out.

TSMC brings silicon photonics into its advanced packaging line through COUPE, which is why the same company that makes the GPUs is also building the optical engines that sit beside the switch chips. Our TSMC Stack page covers that packaging business, and who makes advanced packaging lists the other players.

Who gets paid from silicon photonics?

Four groups. Foundries are paid per wafer for the photonic chips. NVIDIA and Broadcom are paid for switches that carry co-packaged optics. Module makers are paid for pluggables, which still carry most links. And laser makers are paid for the light source every one of those products needs.

The growth numbers are large across the chain. Lumentum reported revenue of $1,006.3 million for the quarter to 27 June 2026, up 109% from a year earlier, with its components business, mainly lasers and optical chips, at $649.4 million. Coherent reported datacenter and communications revenue of $1.62 billion for the quarter to 30 June 2026, up 59% and 79% of its total. Coherent said customer orders now extend into calendar 2028 and long-term agreements run to the end of the decade.

Margins say more about who holds the scarce part. Lumentum’s non-GAAP gross margin rose to 50.4% in its June quarter from 37.8% a year earlier.

Whole-company margins on different bases (non-GAAP for Lumentum and Coherent, reported for Tower), so this is a rough signal, not a like-for-like comparison. Our read is in the text.

Our read: the further a company sits toward the laser, the more pricing power it has right now. Foundry wafers can be added by converting and expanding silicon lines, as Tower and GlobalFoundries are doing. Indium phosphide laser capacity takes longer, which is the scenario our AI optics sold-out analysis works through.

Where is the stuck step?

The laser. Both of the big laser makers said so on their August 2026 calls, in plain terms.

Coherent’s chief executive, Jim Anderson, said indium phosphide capacity continues to be the company’s primary constraint. He said Coherent is not constrained in transceiver assembly and test, and is really just constrained by the ramp of indium phosphide production. Coherent said it is on track to double its internal indium phosphide output year on year by the end of the September quarter, and to more than double it again by the end of 2027.

Lumentum’s chief executive, Michael Hurlston, said that on high-power lasers for co-packaged optics the company is very much further behind its ability to supply, because demand accelerated faster than it could add output. He also cited a significant supply–demand imbalance for continuous-wave lasers, the steady light sources silicon photonics transceivers use. Lumentum expects ultra-high-power laser revenue of about $50 million a quarter by the end of 2026, and a bigger ramp in the second half of 2027.

The fixes are physical and slow. Lumentum is expanding its two indium phosphide fabs in Japan and converting a former Qorvo fab in Greensboro, North Carolina into its fifth. It also signed a supply and capacity agreement with AXT for indium phosphide substrates running to 31 December 2031, starting with a $43.5 million deposit. Our read: when a laser maker pays deposits to lock in the raw wafers a decade ahead, the bind has reached the material itself.

This mirrors a pattern across the semiconductors layer. As with HBM and IC substrates, the part with the fewest qualified suppliers and the longest build time sets the pace for everything around it.

What would change the view?

• Laser supply catches up. If Coherent’s doubled indium phosphide output and Lumentum’s new fabs close the gap, pricing power would drift back toward the foundries and switch makers.
• Co-packaged optics slips. If large deployments move beyond 2028, demand for high-power lasers would ramp later than the laser makers now expect; near-packaged designs would carry the interim.
• Foundry capacity overshoots. Tower expects fourth-quarter wafer-start capacity above three times its second-quarter silicon photonics shipments. If laser supply cannot keep up, photonic chips could wait for light, and foundry utilisation would show it first.

What to check next

Watch Tower Semiconductor’s third-quarter 2026 results for whether silicon photonics stays on track for the $1 billion run rate, and Lumentum’s and Coherent’s September-quarter results for laser shipments, indium phosphide capacity and the first co-packaged optics revenue. The Taiwan markets door covers TSMC’s monthly sales, where COUPE sits inside advanced packaging.

Common misconceptions

• People assume silicon photonics means silicon makes the light. Actually silicon guides, switches and detects light; the laser is still made from indium phosphide.
• People assume co-packaged optics and silicon photonics are the same thing. Actually silicon photonics is the chip technology; co-packaged optics is one way of placing it, inside the switch package.
• People assume the foundries hold the bottleneck. Actually the foundries are adding capacity quickly, and Tower expects to triple wafer starts; the laser makers say they are the ones behind demand.

FAQ

Related: optics in the AI cluster; foundry and advanced packaging; chiplets; interposers; GPU clusters; who makes AI chips.

Stack doors

Layer: Semiconductors. Also: Hyperscalers and AI Data Centers. Markets: Taiwan and Singapore, where silicon photonics wafers are made.

Get the Wire — short daily notes when chips, lasers and AI capacity move.

Not investment advice. Do your own research.

References