Optical transceivers in AI racks: 800G and 1.6T modules
See what a pluggable optical module does in an AI rack, who makes the modules and the lasers and chips inside, and what NVIDIA’s supplier names establish.
Pluggable 800G and 1.6T modules that convert electrical signals to light and back for links that leave the rack.
What does a pluggable transceiver do?
It turns the network card’s electrical signal into light for the fiber and back again. An OSFP module carries eight electrical lanes: 800G at 8 × 100G or 1.6T at 8 × 200G. Inside sit a laser (an EML, or a CW laser with a silicon-photonics modulator), a DSP that retimes each lane, and drivers and TIAs. In GB300 NVL72 each GPU gets 800 Gb/s of scale-out bandwidth through ConnectX-8, and NVIDIA’s DGX guide lists four single-port OSFP network cards per compute tray. At 200 Gb/s per lane, IEEE’s objective for copper twinax is a reach of at least 1 m against 500 m to 2 km over single-mode fiber, which is why links that leave the rack are optical.
Who makes the modules, and who makes what is inside?
In March 2025 NVIDIA named Coherent, Eoptolink, Fabrinet and Innolight as pluggable-transceiver industry leaders. Innolight, Eoptolink and Accelink list 1.6T OSFP modules in their 2025 annual reports, Coherent and Lumentum describe 800G and 1.6T modules in their 10-Ks, AOI received a first volume order of more than $200 million for 1.6T modules from an unnamed hyperscaler in March 2026, and Cisco says its 1.6T OSFP optics ship with its Silicon One G300 systems in 2026. Fabrinet is a contract manufacturer that builds to customers’ designs. Inside the modules, Lumentum, Coherent, Mitsubishi Electric, Sumitomo Electric and Broadcom make lasers, Marvell and Credo make DSPs, MACOM and Semtech make drivers and TIAs, and TFC supplies optical engines and passive parts.
What do the names and figures not establish?
NVIDIA’s list shows whom it works with, not whose modules ship in a given rack, at what share or in what volume. In March 2026 NVIDIA invested $2 billion each in Coherent and Lumentum with multibillion-dollar purchase commitments, but neither value nor schedule is disclosed. Customers are concentrated and mostly unnamed: Lumentum’s two largest end customers were 26.6% and 15.0% of FY2026 revenue, and NVIDIA’s share of Fabrinet’s revenue fell from 27.6% in FY2025 to 16.3%. Accelink, citing Omdia, ranks itself fourth in datacom components at 5.9%, a figure not combined with any other firm’s. Architecture can also move content: NVIDIA’s March 2025 release says its co-packaged-optics switches use 4x fewer lasers and give 3.5x more power efficiency than traditional designs.
How to research an optical transceiver company
Identify the role before the company: module brand, contract manufacturer, laser or DSP supplier, or optical-engine and passive-part maker.
Record the generation and format (800G or 1.6T, DR8 or 2xFR4, EML or silicon photonics). A strong position in one generation needs new evidence in the next.
Read a leaders or ecosystem list as a partner statement, not a supply share, and look for the company’s own filing that describes the product.
Check how much revenue is data-center optics: CATV products were 53.8% of AOI’s 2025 revenue, and data center was 30.3% of MACOM’s FY2025 revenue.
Check the listing: Innolight, Eoptolink, Accelink and TFC trade in Shenzhen with no US line, Innolight added Hong Kong H shares on July 30, 2026, and Fabrinet trades only on the NYSE.
Watch architecture shifts, such as linear-drive modules and co-packaged optics, that move content between module, laser, DSP and switch makers.
Use the list as a starting point: it is curated, not exhaustive, and the model is illustrative, not a bill of materials.
Coherent, Eoptolink, Fabrinet and Innolight named as pluggable-transceiver leaders; co-packaged optics with 4x fewer lasers and 3.5x more power efficiency, by NVIDIA’s claim.
Partner-disclosed: NVIDIA’s Spectrum-X Photonics release names Innolight among pluggable-transceiver industry leaders, and Innolight’s 2025 annual report lists its 1.6T and 800G modules.
Disclosed: Eoptolink’s 2025 annual report describes its 1.6T and 800G modules for AI clusters; NVIDIA’s release separately names it among pluggable-transceiver leaders.
Disclosed: Coherent’s FY2026 10-K describes its AI data-center transceivers and the NVIDIA supply agreement; NVIDIA separately names Coherent among pluggable-transceiver leaders and CPO ecosystem partners.
Disclosed: Lumentum’s FY2026 10-K and product pages describe its AI transceivers and lasers; NVIDIA’s release announces the $2 billion investment and laser purchase commitment.
Disclosed: Cisco’s fiscal 2026 results give the hyperscaler AI orders and revenue; its newsroom describes Silicon One G300, the N9100 and UCS AI servers.
Partner-disclosed: NVIDIA’s Spectrum-X Photonics release lists TFC Communication in its silicon photonics ecosystem, and TFC’s 2025 annual report describes its data-center products.
Disclosed: Marvell’s FY2026 10-K describes its custom AI silicon, electro-optics DSPs, AEC DSPs for AI scale-up and scale-out, PCIe retimers and Teralynx switches.
Disclosed: Mitsubishi Electric’s May 29, 2026 Semiconductor & Device briefing gives the EML share estimate, the ¥40 billion investment and the capacity plan.
Disclosed: Sumitomo Electric’s own pages and briefing describe its data-center lasers, fiber and connectors; NVIDIA names it as a silicon photonics ecosystem partner.