Skip to content
ContentLora

    Tip: press / anywhere to search.

    Explainer

    How optical interconnects and silicon photonics work

    Optical interconnects carry data between chips as light instead of electrical signals, which cuts the energy spent moving bits. In AI data centers, pluggable optical transceivers were reported to use about 10% of GPU compute power, which is why co-packaged optics moved into production in 2026.[1][2]

    Editor reviewedUpdated Next-gen computing hardwareSemiconductorsComputing

    Why light

    Inside a data center, thousands of AI chips must constantly swap data. Traditionally, an electrical signal leaves the chip, runs across a circuit board to a plug-in module at the front of the switch, and only there is converted into light for the fibre. Each step costs energy. IEEE Spectrum reported that these pluggable optics consume about 10% of the total compute power of GPUs in AI data centers.[1]

    In AI data centers, pluggables were reported to consume about 10% of total GPU compute power, more than half of it in the lasers.[1] NVIDIA’s Ian Buck said that in a 400,000-GPU factory this would be about 40 megawatts.[3] A pluggable transceiver bundles lasers, optical circuits, digital signal processors and other electronics, and converts between electrical bits at the switch and light in the fibre.[4] Shortening the electrical path is the main lever for cutting energy per bit.[5]

    Silicon photonics in brief

    Silicon photonics makes optical engines with semiconductor manufacturing processes, so they can be built and packaged like chips: Lightmatter’s Passage M1000 is built on GlobalFoundries’ Fotonix silicon photonics platform, and Ayar Labs’ TeraPHY combines silicon photonics with CMOS manufacturing in a chiplet.[6][7] The lasers are made from non-silicon materials, and in co-packaged switches they stay outside the package; NVIDIA’s design shares one laser across eight data links.[5][8]

    A modulator is the part that turns electrical ones and zeros into flashes of light. Different companies use different kinds: some are larger and robust, others are tiny but need to be kept at a steady temperature.[9]

    Broadcom uses Mach-Zehnder modulators, which split light into two arms, shift the phase of one and recombine them; this is the basis of pluggable optics. NVIDIA uses microring resonators, which are far more compact but temperature-sensitive, so each needs a built-in heater that consumes power.[9]

    Three levels of integration

    • Pluggable modules. Today’s common setup: optics sit in removable modules at the front of the box.[1]
    • Co-packaged optics (CPO). Optical engines move into the same package as the switch chip, so electrical signals travel millimetres, not centimetres.[5]
    • Optical I/O on processors. Optical chiplets or photonic interposers sit right next to the compute chips themselves.[7][6]
    • Pluggables: front-panel modules, field-replaceable, but with long host traces.[1]
    • CPO switches: optical engines around the switch ASIC on a shared substrate. Broadcom’s Tomahawk 6-Davisson uses 16 optical engines built with TSMC’s COUPE process for 102.4 Tb/s.[10][11]
    • Optical I/O for XPUs: Ayar Labs’ 8 Tbps TeraPHY chiplet uses a UCIe electrical interface and a 16-wavelength light source; Lightmatter’s M1000 is a photonic interposer of more than 4,000 mm² with 114 Tbps of bandwidth.[7][6] Celestial AI, bought by Marvell in 2026, claims up to 25x more off-package bandwidth.[12][13]

    Where things stand in 2026

    NVIDIA announced Quantum-X and Spectrum-X Photonics switches in March 2025 and said in May 2026 that Spectrum-X Ethernet Photonics was in production.[14][2] Broadcom began sampling its third-generation CPO switch in October 2025.[10] NVIDIA’s Spectrum-6 switch offers 102.4 Tb/s with 200 Gb/s co-packaged optics.[15]

    The next step is moving optics into the scale-up network that links GPUs. NVIDIA said in March 2026 that Vera Rubin Ultra NVL576 will join eight racks into one NVLink domain using copper and direct optical connections.[16] A July 2026 report citing SemiAnalysis said that configuration might slip or ship in low volume until the optics mature; NVIDIA said its roadmap was intact.[17] Optical engines are also becoming a foundry product: Broadcom’s Tomahawk 6-Davisson uses engines built with TSMC‘s COUPE technology.[11]

    The power figures quoted so far come from vendors, and independent measurements at scale are still scarce; see the co-packaged-optics page for details.[18][19]

    Questions readers ask

    Why not just keep using copper wires?

    Moving data electrically over distance costs a lot of power; vendors such as Ayar Labs argue optical links are needed to solve power-density limits in AI systems.[20][1]

    What is co-packaged optics?

    Placing silicon optical engines in the same package as a switch or processor, with lasers kept outside, instead of using pluggable modules at the front panel.[5]

    How much power can co-packaged optics save?

    Broadcom claims about 70% lower optical interconnect power than pluggables for Tomahawk 6-Davisson; NVIDIA claims 5x better power efficiency for Spectrum-X Ethernet Photonics.[19][18]

    Sources

    Each numbered claim is a statement we checked against the sources listed with it. Status shows how well established it is.

    1. [1]

      IEEE Spectrum reported that pluggable optical transceivers consume about 10% of total GPU compute power in AI data centers, more than half of it to drive lasers. reportedas of 2025-03-25

    2. [2]

      By May 31, 2026, NVIDIA said its Spectrum-X Ethernet Photonics co-packaged-optics switches with 200Gb/s SerDes were in production. confirmedas of 2026-05-31

    3. [3]

      NVIDIA's Ian Buck said that in a 400,000-GPU AI factory pluggable optics would draw about 40 megawatts, more than half of it for lasers. reportedas of 2025-03-25

    4. [4]

      Pluggable optical transceivers combine lasers, optical circuits, digital signal processors and other electronics, and translate between electronic bits at the switch and light in the fibre. confirmedas of 2025-03-25

    5. [5]

      In co-packaged optics, silicon optical-transceiver chiplets sit beside the switch chip inside one package, while the lasers, made from non-silicon materials, stay outside, shortening electrical paths and reducing components. confirmedas of 2025-03-25

    6. [6]

      Lightmatter unveiled the Passage M1000 on March 31, 2025, a photonic interposer of more than 4,000 square millimetres offering 114 Tbps of optical bandwidth, built on GlobalFoundries' Fotonix platform. confirmedas of 2025-03-31

    7. [7]

      Ayar Labs unveiled on March 31, 2025 an 8 Tbps TeraPHY optical I/O chiplet with a UCIe electrical interface, powered by a 16-wavelength light source. confirmedas of 2025-03-31

    8. [8]

      IEEE Spectrum reported that NVIDIA's co-packaged optics need only one laser for every eight data links. reportedas of 2025-03-25

    9. [9]

      Broadcom's co-packaged optics use Mach-Zehnder modulators, the basis of pluggable optics, while NVIDIA chose more compact microring modulators, which are temperature-sensitive and each need a controlled heating circuit. confirmedas of 2025-03-25

    10. [10]

      In October 2025 Broadcom began sampling Tomahawk 6-Davisson, its third-generation co-packaged-optics Ethernet switch with 102.4 Tb/s of capacity, to early-access customers. reportedas of 2025-10-08

    11. [11]

      Broadcom's Tomahawk 6-Davisson uses optical engines built with TSMC's Compact Universal Photonic Engine (COUPE) technology. reportedas of 2025-10-08

    12. [12]

      Marvell completed its roughly $3.25 billion acquisition of optical-interconnect startup Celestial AI on February 2, 2026. reportedas of 2026-02-02

    13. [13]

      Celestial AI claims its Photonic Fabric can raise off-package bandwidth between data center chips by up to 25 times while substantially reducing power. reportedas of 2026-02-02

    14. [14]

      NVIDIA announced Quantum-X Photonics InfiniBand switches for availability in 2025 and Spectrum-X Photonics Ethernet switches for 2026. confirmedas of 2025-03-18

    15. [15]

      NVIDIA's Spectrum-6 switch offers 102.4 Tb/s with 512 lanes and 200 Gb/s co-packaged optics, replacing pluggable transceivers. confirmedas of 2026-03-16

    16. [16]

      In March 2026 NVIDIA said Vera Rubin Ultra NVL576 will join eight 72-GPU racks into one 576-GPU NVLink domain using copper and direct optical connections, and that its later Kyber racks will scale to NVL1152 with similar direct optical rack-to-rack links. confirmedas of 2026-03-16· forecast

    17. [17]

      In July 2026 Tom's Hardware, citing SemiAnalysis, reported that NVIDIA's Kyber rack had slipped to 2028, that the eight-rack NVL576 linked by co-packaged optics was likely to slip or ship in low volume, and that a production-ready co-packaged-optics NVSwitch was not expected before the Feynman generation; NVIDIA responded only that its roadmap was intact. reportedas of 2026-07-06

    18. [18]

      NVIDIA claims Spectrum-X Ethernet Photonics delivers 5x better power efficiency and 5x longer AI uptime than networks using traditional transceivers. confirmedas of 2026-05-31

    19. [19]

      Broadcom says Tomahawk 6-Davisson cuts optical interconnect power by about 70% compared with pluggable transceivers. reportedas of 2025-10-08

    20. [20]

      Ayar Labs' chief executive argues optical interconnects are needed to solve power-density limits of copper in large AI systems. confirmedas of 2025-03-31

    Revision history (2)
    1. Page created.
    2. Added how pluggables and modulators work, laser sharing, NVLink optical scale-up plans and the July 2026 delay report.

    Created Oct 10, 2026. Last reviewed by an editor on Oct 10, 2026. Next scheduled review: Jan 10, 2027.

    Cite this page

    "How optical interconnects and silicon photonics work." ContentLora, updated Oct 10, 2026. https://contentlora.com/explain/how-optical-interconnects-work

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.

    Keep exploring

    ☰ Contents