For a decade the optics industry told itself that the pluggable transceiver was a dead end and that co-packaged optics would replace it the moment the technology matured. The technology has matured. Broadcom has been shipping its 102.4T Tomahawk 6 Davisson CPO switch since October 2025, and Nvidia put its 409.6T Spectrum-X Ethernet Photonics switch into mass production in August 2026 with CoreWeave, Lambda and Oracle as launch customers. Yet the architecture winning hyperscaler purchase orders in 2026 is not CPO. It is the humble OSFP module with the digital signal processor ripped out.
The DSP was the problem all along
A retimed 800G pluggable burns 14 to 16 watts, and roughly half of that is the DSP that cleans up the electrical signal between the switch ASIC and the optics. Linear pluggable optics delete that chip, leaving only a driver, a transimpedance amplifier and a continuous-time linear equalizer, and lean on the switch SerDes to do the equalization it was already doing anyway. Arista, which pioneered the category, lists its LPO-800G-DR8 at about 9 watts. A slide deck from Arista's Robert Friskney at NetUK in July 2026 put the comparison bluntly: 16 watts versus 8 watts, 110 nanoseconds of latency versus under 3, and a calculated per-optic MTBF of 143 years against 97 for the retimed part. His summary line was that if you can use LPO, you should, and Arista's 7060X-E7 platform, its first with 1.6T ports and liquid-cooled options, now advertises roughly 60 percent lower optical interconnect power with LPO.
At 1.6T the arithmetic gets sharper. A fully retimed 200G-per-lane module lands in the 23 to 25 watt range. Semtech, whose FiberEdge linear TIAs and drivers are now deployed at several hyperscalers in the United States and China, targets about 10 watts for the LPO equivalent. On a 512-port Tomahawk 6 system that is roughly 500 watts saved at the module level and close to a kilowatt once cooling overhead is counted. Semtech booked its first LPO revenue in the quarter ending January 2026 and told analysts in May that linear solutions, counting LPO, LRO and active copper cables, will represent more than 25 percent of the transceiver mix within two years.
LRO is the compromise that makes 1.6T real
Full LPO has a real limitation: with no retimer on either end, the link budget depends on the host SerDes on both sides, which makes multi-vendor interoperability at 200G per lane genuinely hard. Linear receive optics keep a DSP on the transmit path, where it does the most good for signal quality, and go linear on the receive path, where most of the power was being wasted. The result is a module that behaves like a retimed part from the far end of the fiber but saves perhaps a third of the power. An IEEE Electronics Packaging Society overview published in March 2026 argued that above 30 watts the thermal budget alone makes LRO the more viable near-term option at 1.6T, and Cignal AI noted that every vendor demonstrating 1.6T LPO at OFC 2026 also showed an LRO version of the same product. Eoptolink now sells its 1.6T line in three flavors, fully retimed, LRO and LPO, and lets the customer pick per port. The OIF ran a live interop at OFC 2026 across 40 member companies covering retimed, half-retimed and linear modules.
What CPO actually buys, and what it costs
Broadcom's own numbers are the fairest benchmark because they come from the company that sells both. In its Davisson briefing Broadcom cites a 65 percent optics power saving for CPO over retimed pluggables at 100G per lane, and 35 percent over LPO. A CPO port at 800G runs in the 4 to 5 watt range by most vendor accounts, against 8 to 9 watts for LPO and 16 for retimed. That remaining gap is real, and at a 400,000-GPU campus it adds up to tens of megawatts. CPO also cuts laser count: Nvidia says Spectrum-X Photonics uses four times fewer lasers than an equivalent pluggable fabric and claims five times better power efficiency and five times longer AI uptime, the last of which is the number hyperscalers actually care about.
But CPO changes the failure model. When an optical engine on a Tomahawk 5 Bailly package dies, the fix is a chassis swap, not a module swap. Meta has been testing exactly this at scale. At the Optica Executive Forum in March 2026 Meta optical engineer Drew Alduino reported more than 50 million device-hours on second-phase Bailly systems, with first-phase data showing about 1.47 million hours MTBF against 0.71 million for the 2x400G FR4 pluggables used as a control, roughly a twofold improvement, rising past 8 million hours once a surface-mount defect in the external laser modules was excluded. Alduino's framing, as reported by Converge Digest, was that the industry has answered whether it can build these systems and now has to decide whether it should, because every watt saved carries a cost in reliability, availability and serviceability at hyperscale. Meta co-founded the OCI-MSA in March 2026 with seven other companies to standardize optical scale-up precisely because it does not want to be locked to one vendor's engine.
Who is betting on which side
The vendor map is clearer than it was a year ago. Nvidia is all-in on CPO for its own fabric, with Quantum-X for InfiniBand and Spectrum-X for Ethernet feeding the Vera Rubin generation, and Gilad Shainer told LightCounting's July 2026 conference that CPO and near-packaged optics will coexist, which is as close as Nvidia comes to conceding that pluggables are not going away. Broadcom straddles every option, selling Davisson CPO, Sian DSPs for retimed modules, and ASICs explicitly spec'd to interoperate with LPO at 200G per lane. Marvell, the dominant PAM4 DSP house, now ships a 1.6T LPO TIA and driver chipset alongside its Ara DSPs, and LightCounting's own forecast says plainly that volume linear-drive deployments will dampen DSP chipset growth from 2027 onward. Arista is the loudest pluggable advocate, has put a 12.8T liquid-cooled XPO module through an MSA with 45 manufacturers, and expects copper and pluggables to dominate through 2027 with open CPO trials that year and broader deployment in 2028 and 2029.
Delay or complement? The honest answer is sequencing
LPO is not killing CPO. It is deciding when CPO gets deployed and where. For scale-out fabrics at 800G and the first 1.6T wave in 2027, linear pluggables and LRO deliver 50 to 60 percent of the available power saving with zero change to operations, sparing, or multi-sourcing. That is enough to win the next two refresh cycles at most hyperscalers, and it pushes CPO toward the places where pluggables physically cannot go: scale-up domains beyond 256 accelerators, 400G-per-lane switches, and liquid-cooled racks where faceplate density is the binding constraint. The panel consensus at OFC 2026 was that CPO reaches roughly 30 percent of scale-up and scale-out ports by 2028, which implies linear and retimed pluggables still carry the other 70. One distributor estimate puts CPO at about half a percent of AI data center optical modules in 2026, climbing toward a third by 2030.
The market is large enough to carry both. LightCounting now projects 120 percent growth in Ethernet transceiver sales in 2026 and more than $2 billion in 1.6T chipset revenue this year. The DSP vendors will keep growing in dollars while losing share, and CPO will own the top of the pyramid. Expect the real inflection around 2028 with Broadcom's fourth-generation 400G-per-lane CPO and whatever Nvidia pairs with the post-Rubin platform. Until then the most consequential optical decision in an AI data center is not pluggable versus co-packaged. It is whether you can live without the DSP, and the answer at 800G is increasingly yes.