Skip to content
ContentLora

    Tip: press / anywhere to search.

    technology

    Co-packaged optics (CPO)

    Also known as CPO, optical I/O, silicon photonics switch, photonic interconnect

    Co-packaged optics (CPO) places silicon optical engines in the same package as a network switch or processor, replacing pluggable transceivers and shortening electrical paths.[1] NVIDIA said its Spectrum-X Ethernet Photonics CPO switches were in production by May 2026, and Broadcom began sampling a 102.4 Tb/s CPO switch in October 2025.[2][3]

    Editor reviewedUpdated Next-gen computing hardwareSemiconductorsComputing
    Key facts

    Co-packaged optics (CPO) is a way of building network switches and, increasingly, processors so that the conversion between electrical signals and light happens inside the chip package rather than in plug-in modules at the front of the equipment.[1] It moved from demonstrations to production in 2025-2026, driven by the energy cost of connecting very large AI clusters.[2][4] The chiplets page covers the packaging it relies on.

    How it works

    In a CPO switch, silicon optical-transceiver chiplets surround the switch chip in one package, while the lasers, made from non-silicon materials, stay outside. Signals travel millimetres instead of across a circuit board, which removes components and reduces power.[1] Broadcom’s Tomahawk 6-Davisson, for example, combines optical engines built with TSMC‘s COUPE technology with substrate-level multi-chip packaging.[5] Because the lasers are external and shared, NVIDIA’s design needs only one laser for every eight data links.[6]

    The two main vendors chose different modulators, the devices that encode bits onto light. Broadcom uses Mach-Zehnder modulators, the same type used in pluggable optics, while NVIDIA chose much smaller microring modulators, which are sensitive to temperature and each need a controlled heater.[7]

    Products and milestones

    NVIDIA announced Quantum-X Photonics (InfiniBand) and Spectrum-X Photonics (Ethernet) switches in March 2025, with Quantum-X expected that year and Spectrum-X in 2026.[8] It said in May 2026 that Spectrum-X Ethernet Photonics, with 200Gb/s SerDes, was in production.[2] Its Spectrum-6 switch offers 102.4 Tb/s across 512 lanes with 200 Gb/s co-packaged optics.[9]

    The next step is NVLink, the link that ties GPUs into a single scale-up system. In March 2026 NVIDIA said Vera Rubin Ultra NVL576 will join eight 72-GPU racks into one NVLink domain using copper and direct optical connections, and that its Kyber racks will later scale to NVL1152 with similar optical rack-to-rack links.[10] In July 2026 Tom’s Hardware, citing SemiAnalysis, reported that Kyber had slipped to 2028, that the CPO-linked NVL576 might slip or ship in low volume, and that a production-ready CPO NVSwitch was not expected before the Feynman generation; NVIDIA said only that its roadmap was intact.[11]

    Broadcom began sampling Tomahawk 6-Davisson, its third-generation CPO switch with 102.4 Tb/s of capacity, in October 2025, and says a fourth generation targeting 400 Gbps per channel is in development.[3][12]

    Micas Networks announced volume production of a CPO switch based on Broadcom technology in March 2025, promising 40% power savings over standard pluggables.[13]

    Startups and acquisitions. Lightmatter unveiled the Passage M1000 photonic interposer, offering 114 Tbps of optical bandwidth, in March 2025.[14] Ayar Labs unveiled an 8 Tbps TeraPHY optical chiplet with a UCIe interface the same day.[15] Marvell completed its roughly $3.25 billion purchase of Celestial AI in February 2026; Celestial claims up to 25x more off-package bandwidth.[16][17]

    Claimed benefits

    Vendors say CPO uses far less power and fails less often than plug-in optics. Broadcom claims about 70% less power for the optical links; NVIDIA claims five times better power efficiency and five times longer uptime.[18][19] These are company figures, not independent measurements.

    NVIDIA’s 2025 claims were 4x fewer lasers, 3.5x better power efficiency and 10x better resiliency; its 2026 claims for Spectrum-X Ethernet Photonics are 5x power efficiency and 5x longer uptime versus transceiver-based networks.[20][19] Broadcom cites roughly 70% lower optical interconnect power.[18] Resiliency claims address link flaps, transient failures of pluggable connections; a Micas executive told IEEE Spectrum these lengthen AI training runs and that CPO should have fewer because the signal path has fewer components.[21] NVIDIA’s Ian Buck put the stakes at about 40 megawatts of pluggable optics in a 400,000-GPU AI factory, more than half of it for lasers.[22]

    Open questions

    How serviceable sealed-in optics are compared with swappable pluggable modules is not yet documented in public sources, and the power and uptime figures above all come from vendors.[19] A UC Santa Barbara CPO researcher told IEEE Spectrum that NVIDIA would only adopt CPO once data centers could not afford the power of pluggables, suggesting economics, not technology, had held it back.[23] The next frontier is optical links directly between GPUs rather than only in switches, and the reported uncertainty over NVL576 shows that timing is not settled.[10][11] Microring modulators also carry a power cost of their own, because each needs a heater to hold its temperature.[7]

    Questions readers ask

    What problem does co-packaged optics solve?

    Pluggable optical transceivers were reported to consume about 10% of total GPU compute power in AI data centers; CPO shortens the electrical path and cuts components.[4][1]

    Who ships CPO switches?

    NVIDIA said Spectrum-X Ethernet Photonics was in production by May 2026, and Broadcom began sampling Tomahawk 6-Davisson in October 2025.[2][3]

    How big are the claimed savings?

    Broadcom claims about 70% lower optical interconnect power than pluggables; NVIDIA claims 5x better power efficiency. Both are vendor figures.[18][19]

    Is CPO used between GPUs, not just in switches?

    Not yet in volume. NVIDIA plans optical NVLink links between racks for Rubin Ultra NVL576, though a July 2026 report said that system may slip; startups such as Ayar Labs and Lightmatter make optical I/O for processors.[10][11][15][14]

    Sources

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

    1. [1]

      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

    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]

      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

    4. [4]

      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

    5. [5]

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

    6. [6]

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

    7. [7]

      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

    8. [8]

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

    9. [9]

      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

    10. [10]

      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

    11. [11]

      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

    12. [12]

      Broadcom says a fourth-generation co-packaged-optics platform targeting 400 Gbps per channel is in development. reportedas of 2025-10-08· forecast

    13. [13]

      In March 2025 Micas Networks announced volume production of a co-packaged-optics switch based on Broadcom technology, promising 40% power savings over systems with standard pluggable transceivers. reportedas of 2025-03-25

    14. [14]

      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

    15. [15]

      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

    16. [16]

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

    17. [17]

      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

    18. [18]

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

    19. [19]

      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

    20. [20]

      NVIDIA said in March 2025 that its silicon photonics switches use 4x fewer lasers and deliver 3.5x more power efficiency and 10x better resiliency than traditional methods. confirmedas of 2025-03-18

    21. [22]

      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

    22. [23]

      A UC Santa Barbara co-packaged-optics researcher told IEEE Spectrum that NVIDIA would only adopt CPO once GPU-heavy data centers could no longer afford the power of pluggables. reportedas of 2025-03-25

    Revision history (2)
    1. Page created.
    2. Added NVIDIA NVLink optical scale-up plans and the July 2026 delay report, Spectrum-6, laser sharing, modulator choices and Micas; corrected the link-flap attribution.

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

    Cite this page

    "Co-packaged optics (CPO)." ContentLora, updated Oct 10, 2026. https://contentlora.com/wiki/co-packaged-optics

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

    Keep exploring

    Connections
    Chiplets and advanc… Integrates viaNVIDIA Developed byTSMC Optical engines made by
    ☰ Contents