> ## Content Index
> Fetch the complete content index at: https://www.edgewisely.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Top 7 Silicon Photonics Companies in 2026
- URL: https://www.edgewisely.com/top-7-silicon-photonics-companies-2026/
- Published: 2026-10-02T04:49:10.000Z
- Updated: 2026-10-02T04:50:50.000Z
- Description: Co-packaged optics is shipping and Marvell just bought Celestial AI for $3.25B. We compare the seven silicon photonics companies that matter in AI data centers on specs, funding and who is actually in production.
- Author: John Karpentar
- Tags: Roundups, Chips

**TL;DR**

- Silicon photonics moved from research to revenue in 2025–26\. **Broadcom** has shipped co-packaged optics switches since **March 2024**; **Nvidia's** Spectrum-X Ethernet Photonics reached production with CoreWeave, Lambda, Meta, Microsoft and Oracle as named first adopters.
- The biggest consolidation event: **Marvell acquired Celestial AI**, announced December 2025 and **closed February 2, 2026**, for roughly **$3.25 billion upfront** and up to $5.5 billion with earn-outs.
- Startup funding stayed enormous. **Ayar Labs** raised a **$500M Series E** at a **$3.75B valuation** (September 2025); **Lightmatter** raised **$400M** at **$4.4B** (October 2024).
- The open argument is co-packaged optics versus pluggable transceivers. Broadcom and Nvidia are betting on CPO; **Coherent** and **Lumentum** are shipping pluggables at volume today and booking the revenue.

Silicon photonics builds optical components — waveguides, modulators, detectors — directly onto silicon using standard semiconductor manufacturing, so data moves as light instead of electrical signals. In AI data centers it solves a specific problem: copper interconnects run out of reach and power efficiency past a few hundred gigabits per second. The companies leading commercially in 2026 are Broadcom, Nvidia, Coherent, Marvell, Lumentum, Ayar Labs and Lightmatter.

## How we picked these

Ranked by commercial traction in AI infrastructure — what is actually shipping, to whom, with what disclosed. Our criteria:

- **Shipping status** — in production, sampling, or early access.
- **Named adopters** — customers or partners the company has publicly identified, not implied.
- **Published technical specs** — bandwidth, port counts, power per bit. Marketing adjectives do not count.
- **Financial disclosure** — segment revenue in earnings, or disclosed funding and valuation.
- **Architectural significance** — does the approach change how AI clusters get built?

Companies without a live independent presence were excluded, which is why Celestial AI appears in context rather than as an entry.

## Quick comparison

| Company     | Best for                       | Technology                              | Status                                 |
| ----------- | ------------------------------ | --------------------------------------- | -------------------------------------- |
| Broadcom    | Hyperscale Ethernet switching  | CPO switch ASICs (Bailly, Davisson)     | Shipping since 2024                    |
| Nvidia      | GPU-cluster scale-out fabric   | CPO InfiniBand and Ethernet switches    | Quantum-X shipping; Spectrum-X 2H 2026 |
| Coherent    | Pluggable datacenter optics    | SiPh, InP and VCSEL transceivers        | Shipping 1.6T                          |
| Marvell     | Custom XPU optical integration | 3D silicon photonics engine             | Design-in stage                        |
| Lumentum    | Lasers and optical components  | InP lasers, external light sources, OCS | Shipping                               |
| Ayar Labs   | Optical I/O inside the package | TeraPHY UCIe chiplet                    | Scaling to production                  |
| Lightmatter | Photonic interposers           | Passage 3D photonic platform            | Early access                           |

## 1\. Broadcom

[Broadcom](https://www.broadcom.com/?ref=edgewisely.com) has the longest commercial track record in co-packaged optics. **Bailly**, delivered in **March 2024**, integrates eight 6.4 Tbps silicon photonics optical engines with a StrataXGS Tomahawk 5 switch for **51.2 Tbps** total, supporting 128 ports of 400G or 64 of 800G. Broadcom says the design uses roughly 70% less optical interconnect power than pluggable transceivers. The follow-on **Tomahawk 6 "Davisson"** doubles capacity to **102.4 Tbps** using sixteen 6.4 Tbps engines, and began shipping in 2025.

What separates Broadcom from most of this list is that it names partners. ByteDance, Juniper Networks, H3C and Micas Networks were all quoted in the Bailly launch.

**Best for:** Hyperscalers and OEMs building Ethernet fabrics at maximum radix.

**Pros**

- Two successive CPO generations shipping, with the first in customers' hands since early 2024.
- Named production partners including ByteDance and Juniper, rather than unnamed "evaluations".
- Published, specific power and density figures tied to named silicon.

**Cons**

- CPO versus pluggable adoption is still contested; trade coverage has characterized Davisson as Broadcom's third attempt to make CPO stick.
- The proprietary Broadcom Fiber Connector introduces vendor lock-in at the physical layer.
- Co-packaged optics are harder to service in the field than a swappable transceiver — a real operational cost at scale.

## 2\. Nvidia

![Nvidia Quantum-X InfiniBand Photonics co-packaged optics switch](https://storage.ghost.io/c/54/5a/545a66b3-60ef-480c-80ae-765bac52f6ec/content/images/2026/10/top-7-silicon-photonics-companies-2026-2.jpg)

Image: [Nvidia](https://www.nvidia.com/en-us/networking/products/silicon-photonics/?ref=edgewisely.com)

[Nvidia](https://www.nvidia.com/en-us/networking/products/silicon-photonics/?ref=edgewisely.com) entered co-packaged optics in 2025 and brought its installed base with it. The **Quantum-X InfiniBand Photonics** switch (Q3450-LD) delivers **144 ports of 800 Gb/s** on a liquid-cooled design to handle the thermal load of onboard photonics, and can connect over 10,000 GPUs in a two-level non-blocking fat tree. **Spectrum-X Ethernet Photonics** scales to **409.6 Tb/s** and is slated for the second half of 2026\. Both are built on 200G SerDes.

Nvidia claims 5x better power efficiency than pluggable transceivers and 5x sustained AI application runtime from improved resiliency. Because CPO removes DSP retimers, it also cuts latency. Manufacturing partners include TSMC, Coherent, Corning, Fabrinet, Foxconn, Lumentum, Senko and Sumitomo Electric — several of whom appear elsewhere on this list.

**Best for:** Operators standing up very large [GPU clusters](https://www.edgewisely.com/top-7-gpu-cloud-providers-for-ai-training-and-inference-2026/) on Nvidia fabric.

**Pros**

- Named first adopters for Spectrum-X Photonics: CoreWeave, Lambda, Meta, Microsoft and Oracle Cloud Infrastructure.
- Eliminating DSP retimers removes a latency and power stage that pluggable designs cannot avoid.
- A deep manufacturing partner bench across TSMC, Foxconn, Fabrinet and major optics suppliers.

**Cons**

- Spectrum-X Ethernet Photonics is not generally available until the second half of 2026.
- The 5x efficiency and resiliency figures are Nvidia's own; no independent benchmark has been published.
- Liquid cooling is a hard requirement for Quantum-X, which rules out air-cooled facilities without retrofit.

## 3\. Coherent

![Coherent 1.6T-DR8 silicon photonics optical transceiver module](https://storage.ghost.io/c/54/5a/545a66b3-60ef-480c-80ae-765bac52f6ec/content/images/2026/10/top-7-silicon-photonics-companies-2026-3.jpg)

Image: [Coherent](https://www.coherent.com/news/press-releases/silicon-photonics-based-transceiver-modules?ref=edgewisely.com)

[Coherent](https://www.coherent.com/?ref=edgewisely.com) — formed when II-VI acquired the original Coherent in 2022 and took its name — is the volume optics supplier with the clearest financial disclosure. Its **1.6T-DR8** transceiver pairs silicon photonics with Marvell's 3nm Ara optical DSP across 200 Gbps electrical and optical lanes, targeting over 20% lower power dissipation. Unusually for this field, Coherent spans silicon photonics, indium phosphide EMLs and VCSELs, which hedges against any one technology winning.

The company reported **800G revenue up roughly 80% sequentially to about $200 million** in a quarter, and began 1.6T revenue shipments in Q4 FY2025.

**Best for:** Buyers who want merchant pluggable optics at volume, independent of switch vendor.

**Pros**

- Discloses actual segment revenue and growth in earnings — the only pure-play optics supplier here that does.
- Shipping 1.6T transceivers commercially, not sampling.
- Three parallel technology lines (SiPh, InP, VCSEL) reduce exposure to a single architectural bet.

**Cons**

- Its pluggable-first model is structurally exposed if CPO adoption accelerates at Broadcom and Nvidia.
- Two divestitures in the same window — aerospace and defense to Advent for $400M, materials processing tools to Bystronic — suggest portfolio-wide margin pressure.
- Transceivers are a commodity market with persistent pricing dynamics, regardless of volume growth.

## 4\. Marvell

[Marvell](https://www.marvell.com/?ref=edgewisely.com) sells merchant silicon and custom XPU design services, and photonics is now central to both. Its **3D Silicon Photonics Engine**, announced January 2025, delivers 6.4 Tbps across 32 channels of 200G and is positioned as a CPO building block for customer-designed accelerators. Marvell claims twice the bandwidth and I/O density and about 30% lower power per bit versus comparable 100G-interface devices, with XPU-to-XPU reach far beyond electrical cabling. Its COLORZ optical modules have shipped for over eight years with more than 10 billion device field hours by the company's count.

Marvell's **acquisition of Celestial AI closed February 2, 2026** — roughly $1.0 billion cash plus about 27.2 million shares, with earn-outs taking the total to as much as $5.5 billion. Celestial's Photonic Fabric targets memory disaggregation, a different problem from switch interconnect. Marvell has guided to revenue contribution starting in the second half of FY2028.

**Best for:** Hyperscalers designing [custom accelerators](https://www.edgewisely.com/top-7-ai-chip-makers-training-inference-2026/) that need optical I/O integrated.

**Pros**

- Eight years and over 10 billion field hours of shipping silicon photonics underneath the new CPO work.
- The Celestial AI acquisition adds memory-disaggregation IP that complements rather than duplicates the switch roadmap.
- Supplies the optical DSP inside competing vendors' transceivers, giving it visibility across the market.

**Cons**

- CPO for custom XPUs is pre-production. Marvell will only say multiple customers are evaluating, with none named.
- Celestial AI generated negligible disclosed revenue before acquisition, and Marvell's own targets are roughly two years out.
- Absorbing an acquired photonics roadmap and team carries integration risk that has not yet been tested.

## 5\. Lumentum

![Lumentum AI infrastructure optical components solution graphic](https://storage.ghost.io/c/54/5a/545a66b3-60ef-480c-80ae-765bac52f6ec/content/images/2026/10/top-7-silicon-photonics-companies-2026-5.jpg)

Image: [Lumentum](https://www.lumentum.com/en/solutions/ai-infrastructure?ref=edgewisely.com)

[Lumentum](https://www.lumentum.com/?ref=edgewisely.com) makes the light sources and optical components that other people's photonics depend on. Its AI infrastructure portfolio covers indium phosphide lasers, **external laser source (ELS) modules** — which matter specifically because co-packaged optics designs move the laser off the switch package — and optical circuit switching. Its **$705 million acquisition of Cloud Light in November 2023** bought manufacturing scale in data center optical modules.

Lumentum is a named technology partner on Nvidia's silicon photonics program, which is a more concrete adoption signal than anything on its own product pages.

**Best for:** Supplying lasers and components into CPO and pluggable designs alike.

**Pros**

- Positioned to supply both camps: external laser sources for CPO and complete modules for pluggables.
- Named partner on Nvidia's co-packaged optics program alongside TSMC and Corning.
- Cloud Light acquisition added datacenter module manufacturing capacity at scale.

**Cons**

- Publishes the least product-level technical detail of the seven — no bandwidth or efficiency figures on its AI infrastructure pages.
- Legacy telecom and industrial laser segments dilute the AI-optics story in its financials.
- Component supplier margins are structurally thinner than switch silicon, and it faces the same CPO transition risk as Coherent.

## 6\. Ayar Labs

![Ayar Labs TeraPHY UCIe optical I/O chiplet delivering 8 Tbps](https://storage.ghost.io/c/54/5a/545a66b3-60ef-480c-80ae-765bac52f6ec/content/images/2026/10/top-7-silicon-photonics-companies-2026-6.jpg)

Image: [Ayar Labs](https://ayarlabs.com/teraphy/?ref=edgewisely.com)

[Ayar Labs](https://ayarlabs.com/?ref=edgewisely.com) attacks a different layer: optical I/O inside the chip package. Its **TeraPHY** chiplet and **SuperNova** multi-wavelength light source move data optically straight off the die rather than across the board to a transceiver. Published specs are specific — **8 Tbps bidirectional per chiplet**, **10 nanosecond latency**, bit error rate below 1e-12, reach from millimeters to 2 kilometers, and 16 WDM channels per port at 512 Gbps. It uses NRZ modulation, avoiding forward error correction and the latency that comes with it.

In **April 2025** Ayar announced what it called the industry's first **UCIe-compliant** optical interconnect chiplet, which matters because UCIe is the standard the [chiplet design ecosystem](https://www.edgewisely.com/top-7-semiconductor-eda-software-platforms-2026/) is converging on. Its **$500M Series E in September 2025**, led by Neuberger Berman at a **$3.75 billion valuation**, included Nvidia, AMD, MediaTek and Alchip as investors. Total raised is about $870 million.

**Best for:** Accelerator designers who need optical bandwidth inside the package.

**Pros**

- Published per-chiplet specs including latency and BER, not just peak bandwidth.
- UCIe compliance means it plugs into the standard chiplet ecosystem rather than a proprietary one.
- Both Nvidia and AMD are investors — rare for a company selling into both.

**Cons**

- Investors are not customers. No production design win has been publicly named.
- Full specifications remain under NDA, which limits independent evaluation.
- Its headline bandwidth-density and power claims are company marketing, with no third-party benchmark published.

## 7\. Lightmatter

![Lightmatter Passage L20 photonic interconnect module](https://storage.ghost.io/c/54/5a/545a66b3-60ef-480c-80ae-765bac52f6ec/content/images/2026/10/top-7-silicon-photonics-companies-2026-7.jpg)

Image: [Lightmatter](https://lightmatter.co/products/passage/?ref=edgewisely.com)

[Lightmatter](https://lightmatter.co/?ref=edgewisely.com) is building the most architecturally ambitious product here. **Passage** uses 3D photonic integration to put an optical layer underneath the compute die — what the company calls "edgeless I/O", where any point on the chip can send data optically rather than routing to the package edge. The **M1000** evaluation kit spans **114 Tbps across a 4,000mm² interposer**. The **L200** delivers 32–64 Tbps aggregate via co-packaged optics, and the L20 offers 12.8 Tbps at four times pluggable density.

Lightmatter publishes efficiency figures competitors mostly avoid: **1.9 pJ/bit at 3.2 Tbps per fiber** on the EVK100, and 2.6 pJ/bit at 800 Gbps on the EVK50\. Its **$400M Series D in October 2024**, led by T. Rowe Price with Fidelity and GV participating, valued it at **$4.4 billion**, up from $1.2 billion ten months earlier. Total raised: $850 million.

**Best for:** Research and advanced-development teams evaluating post-CPO interconnect.

**Pros**

- Publishes picojoule-per-bit efficiency at named bandwidths — the most falsifiable numbers in this roundup.
- A roadmap spanning pluggable-adjacent modules through full 3D interposers, so customers can stage adoption.
- Backing from T. Rowe Price, Fidelity and GV at a $4.4 billion valuation.

**Cons**

- Passage is early access only. No general availability date has been disclosed.
- No named production customer, and the company does not disclose revenue.
- 3D photonic interposers are unproven at manufacturing volume, with yield and thermal questions still open.

## How to choose

**If you are buying switches for an Ethernet AI fabric today:** Broadcom. Two shipping CPO generations and named partners beat architectural elegance.

**If your cluster is Nvidia end to end:** Nvidia's Quantum-X or Spectrum-X Photonics, provided you can supply liquid cooling and can wait until the second half of 2026 for the Ethernet variant.

**If you want optics decoupled from your switch vendor:** Coherent for transceivers, Lumentum for lasers and external light sources. Both ship at volume now.

**If you are designing a custom accelerator:** Marvell, for photonics integrated into the XPU design flow — but treat CPO-for-XPU as pre-production.

**If you need optical bandwidth inside the package:** Ayar Labs, with UCIe compliance as the practical argument.

Photonics also underpins several [quantum computing platforms](https://www.edgewisely.com/top-7-quantum-computing-cloud-platforms-2026/), but that is a separate market from data-center interconnect and the vendors barely overlap.

**If you are three to five years out:** Lightmatter. Passage is the most interesting architecture here and the furthest from general availability.

## Frequently Asked Questions

### What is silicon photonics?

Silicon photonics builds optical components — waveguides, modulators, photodetectors — on silicon wafers using standard CMOS manufacturing. That lets optical and electronic circuits share a chip and a production line, making optical interconnect cheap enough for data centers rather than only long-haul telecom.

### How does silicon photonics work?

An external laser supplies light that a modulator encodes with data by varying the refractive index of a silicon waveguide. The light travels through waveguides or fiber, and a photodetector converts it back to an electrical signal. Wavelength-division multiplexing runs many channels down one fiber simultaneously.

### What is the difference between co-packaged optics and pluggable transceivers?

Pluggable transceivers sit in a faceplate cage and can be swapped individually. Co-packaged optics put the optical engine on the same substrate as the switch ASIC, cutting the electrical path and saving power — Nvidia claims 5x — at the cost of field serviceability, since a failure means servicing the whole package.

### Can photonic chips replace silicon chips?

Not for general computation. Photonics excels at moving data and at specific linear-algebra operations, but has no practical equivalent of a transistor for logic or dense memory. The realistic outcome is hybrid: electronic compute with optical interconnect, which is what every company in this roundup is building.

### Is silicon photonics the future of AI data centers?

For interconnect, it is already the present — Broadcom has shipped CPO since 2024 and Nvidia named production adopters in 2026\. The unsettled question is whether co-packaged optics displaces pluggable transceivers or the two coexist, with CPO at the highest-radix switches and pluggables everywhere else.

**Editor's note — sources:** Company product pages, press releases and newsrooms observed September–October 2026: Nvidia Silicon Photonics, Broadcom Bailly and Tomahawk announcements, Coherent 1.6T-DR8 press release, Marvell 3D Silicon Photonics Engine and Celestial AI acquisition disclosures, Lumentum AI infrastructure solutions, Ayar Labs TeraPHY, Lightmatter Passage. Funding figures from company announcements. Specifications and availability dates are as published by each vendor as of October 2026 and are subject to change. Where a figure is a company claim rather than an independently verified benchmark, we have said so.