Co-packaged optics explained: who supplies the light in AI networks

Co-packaged optics, or CPO, puts the parts that turn electricity into light on the same package as a network switch chip, instead of in plug in modules on the front panel. The goal is lower power and higher density in AI networks. As of October 2026, NVIDIA lists its first CPO switches as available this year, and analysts such as TrendForce point to 2027 and 2028 as the years that decide volume production.

Copper cables and plug in optics carried data centers for years. AI clusters are pushing both close to their limits. This post explains what is changing, in plain language, who makes the pieces, what the sources say about timing, and where the claims are still unproven. Figures are labelled by source and date.

Key numbers at a glance

• Counterpoint Research (14 October 2026): CPO interconnect power below 5 picojoules per bit, as described in an article about Lightmatter.
• NVIDIA Spectrum-X Photonics page (accessed 19 October 2026): up to 409.6 Tb/s of bandwidth, built on 200G SerDes, listed as available in the second half of 2026.
• NVIDIA Quantum-X Photonics page: 144 ports of 800 Gb/s InfiniBand, liquid cooled, over 10,000 GPUs in a two level fat tree.
• NVIDIA claim (Spectrum-X page): 5 times better network power efficiency and 5 times longer sustained application runtime than networks with pluggable transceivers.
• Lumentum, Qualcomm and Corning (company release, 21 September 2026): a 1060 nm VCSEL optical engine with about 10 Tb/s of aggregate capacity and about 1 Tb/s per millimetre of chip edge, with a stated path to about 4 Tb/s per millimetre.
• Lumentum fiscal fourth quarter 2026 (company release, 11 August 2026): revenue 1,006.3 million dollars, up 109.3% from a year earlier; full year 3,014.0 million dollars, up 83.2%.
• Lumentum guidance for fiscal first quarter 2027: 1.225 billion to 1.275 billion dollars.
• Lumentum (company release, 21 September 2026): eight wavelength laser module for CPO and near packaged optics, up to 24 dBm per wavelength, roughly 12 W, initial availability expected in the first half of 2027.
• TrendForce (13 October 2026): testing one photonic chip can take over 100 seconds; 2027 is expected to be pivotal for test equipment and 2028 critical for validating mass production readiness.
• TrendForce (8 October 2026): VCSEL sales projected at 650 million units by 2031, a 38% compound annual growth rate from 2025.
• Dell'Oro (28 July 2026): the 1.6 Tbps port speed ramp expected to be at least five times faster than the 800 Gbps ramp.

The problem: moving data costs power

A switch chip sends data out as electrical signals. In a conventional network, those signals travel across a few centimetres of circuit board to a plug in optical module sitting on the front of the switch. The module converts electricity to light and sends it down a fibre.

That short trip across the board is not free. To keep the signal clean at very high speeds, each lane needs power hungry circuitry called a SerDes (serializer and deserializer). As port speeds climb from 800 Gbps to 1.6 Tbps and beyond, that circuitry and the heat it creates become a larger share of the budget.

Copper has a second issue. TrendForce wrote on 8 October 2026 that at 1.6 Tbps, signal loss over active electrical cables rises sharply, which pushes the shortest links, including those inside a rack, toward optics. Counterpoint's 14 October article on Eliyan uses the phrase "copper wall" for the same problem.

What co-packaged optics does differently

CPO moves the optical engines, the small chips that turn electrical signals into light and back, onto the same substrate as the switch chip. The electrical path shrinks from centimetres to millimetres. Less work is needed to push the signal, so power per bit falls.

Silicon is a poor light source, so the laser itself usually sits elsewhere, in a separate replaceable module, and its light is fed to the optical engine. The laser is where much of the supplier action is.

Power for optical links is usually quoted in picojoules per bit, a measure of energy spent per bit moved, so lower is better. Counterpoint reports Lightmatter putting its CPO interconnect below 5 picojoules per bit. NVIDIA, comparing whole networks rather than single links, claims 5 times better network power efficiency for its photonics switches than for networks built with pluggable transceivers. These are vendor figures, not independent measurements of a deployed network.

The three layouts you will hear about

Three terms keep appearing, and they are easy to mix up.

• Pluggable optics: modules that slide into the front of a switch. Familiar, replaceable, and what most networks use now.
• Near packaged optics, or NPO: optical engines placed very close to the chip but not on the same package. Dell'Oro and Lumentum both name it as an in between step.
• Co-packaged optics, or CPO: optical engines on the same package as the chip.

Dell'Oro's 28 July 2026 release says advanced pluggables, XPO, NPO and CPO will coexist, with XPO and NPO gaining earlier momentum and CPO ramping later. The release does not define XPO, so it is left undefined here.

Sellers of pluggable modules are not standing still either. Lumentum said on 11 August 2026 that adoption of 1.6T cloud modules is advancing, so the old format is still growing.

How to read a power claim

Power claims in this field are easy to misread, so it helps to know what is being compared. A per bit figure such as Lightmatter's below 5 picojoules covers the optical link, not the whole switch, so a lower number there does not mean the whole switch uses a fraction of the power. Claims about runtime and interruptions, like those on NVIDIA's pages, depend on how many fewer parts can fail, and they depend on the baseline network used for comparison.

Another point is scale. If a cluster has hundreds of thousands of optical links, a small saving per link adds up to a large amount of power, which is why the topic gets so much attention. But the saving only counts if the new parts are as reliable and as easy to repair as the old ones, and that is exactly what several of the sources above say is still being proven.

Who is shipping CPO switches

NVIDIA

NVIDIA's photonics page describes two products. Quantum-X Photonics is an InfiniBand switch with 144 ports at 800 Gb/s, liquid cooled, aimed at connecting over 10,000 GPUs in a two level layout. Spectrum-X Photonics is an Ethernet platform with up to 409.6 Tb/s of bandwidth. The page says it is "available in the second half of 2026" and names CoreWeave, Lambda, Meta, Microsoft and Oracle Cloud Infrastructure as first adopters for Vera Rubin. These are NVIDIA's statements, not confirmations from the named companies.

On the Spectrum-X page, NVIDIA claims 5 times better network power efficiency and 5 times longer sustained application runtime compared with pluggable networks, and the FAQ cites a tenfold gain in mean time between interruptions. The page does not give the test setup, so treat these as claims.

Other chip and module makers

NVIDIA is not alone. Broadcom markets its own CPO line under the heading "Highest Power Efficiency and Bandwidth Density", a company claim that its product page does not back with figures. TrendForce names Broadcom, ams OSRAM and HiSilicon among companies using 850 nm VCSEL devices in near packaged designs. And on 21 September 2026, Lumentum, Qualcomm and Corning said they would show a board level demonstration at ECOC 2026 of what they call a massively parallel optical die to die link for AI scale up. Die to die means a link between separate chips. In that design Qualcomm supplies the interface logic, Lumentum a 1060 nm VCSEL optical engine with about 10 Tb/s of aggregate capacity, and Corning the fibre ferrules. The proof of concept runs channels at 32 Gb/s, spans tens of metres, and the companies give a shoreline density of about 1 Tb/s per millimetre today with a path to about 4. Shoreline density is how much bandwidth fits along each millimetre of a chip's edge, which is the real space limit for getting data on and off a package. The release targets future CPO and NPO designs and gives no power figure.

Who supplies the pieces underneath

A CPO switch is not one company's work. TrendForce's 13 October 2026 release identifies seven segments in the CPO test and production chain, and names examples.

• Laser sources and components: Lumentum (company results and releases), Coherent (named by TrendForce for its VCSEL work).
• Test equipment: Teradyne, which acquired Quantifi Photonics in 2025, and Advantest, which announced a collaboration with VIAVI and SENKO in September 2026.
• Alignment and probing: Taiwanese suppliers GMT Global and MPI, according to TrendForce, and test interfaces from WinWay.
• Packaging and testing: ASE and third party testers MSS, MA-tek and iST.
• Fibre connectors: Counterpoint says Lightmatter's reference designs use SENKO fibre connectivity and Corning's PRISM connector.

TrendForce also notes that China still relies heavily on imports for high end instruments, though one Chinese supplier, Semight Instruments, has started volume shipments of a 65 GHz sampling oscilloscope. TrendForce's statements are analyst reporting.

Lumentum deserves a closer look because it publishes detailed numbers. Its fiscal year ended 27 June 2026. Fourth quarter revenue was 1,006.3 million dollars, up 109.3% from a year before, and full year revenue was 3,014.0 million dollars, up 83.2%. The chief executive cited "increasing demand for ultra high power CPO lasers" as an early sign of optics moving into in rack connectivity. The release also noted an initial order for external laser modules. Net property, plant and equipment rose to 1,159.1 million dollars from 726.4 million a year earlier. Its guidance for the next quarter was 1.225 billion to 1.275 billion dollars. Those are company statements and guidance, not independent forecasts.

On 21 September 2026 Lumentum said it would demonstrate at ECOC 2026 in Malaga an eight wavelength laser module in an external laser small form factor pluggable package. In plain terms, the laser sits in a swappable module outside the hot chip package. It produces up to 24 dBm per wavelength (about 250 mW), draws roughly 12 W, and is expected to be available in the first half of 2027. The company says it meets the Optical Compute Interconnect Multi Source Agreement, though the release does not describe that agreement itself.

Why the laser is the awkward part

Putting lasers inside a package full of hot chips shortens their life, so many designs keep them outside. That design choice also affects service. A failed optical part inside a switch package is far harder to replace than a pluggable module on the front panel. Lumentum's 21 September release makes the case for external lasers on exactly this ground: a pluggable external laser module "can be accessed, serviced, or replaced independently" of the compute or switching package. Its components chief also said in the same release that CPO and NPO are becoming defining architectures for AI scale up, which is a company view.

Counterpoint's article about Lightmatter describes other practical problems: heat from integrated lasers, routing about 1,925 optical fibres in a tight space, blind mate connectors, and customer worries about vendor lock in and reliability.

Testing may set the pace

TrendForce's 13 October piece makes an unexpected point. Aligning a fibre to a chip now takes seconds or less, but a full test of one photonic integrated circuit can exceed 100 seconds, so test speed has become the bottleneck. CPO needs lasers, optical switches, power meters and spectrum analysers all working together, and fully equipping every station leaves costly instruments idle. TrendForce expects 2027 to be pivotal for the convergence of test equipment and 2028 to be the critical year for validating readiness for large scale production.

This is a useful reality check. A product can be announced, or shipped in small volumes, long before it can be built by the hundreds of thousands.

The near packaged alternative

TrendForce, in its 8 October release, describes VCSEL based near packaged optics as a "highly competitive option" for scale up links, meaning the short links inside a rack. A VCSEL is a small laser that emits light straight up from the chip surface, and arrays of them are cheap to make. TrendForce projects VCSEL sales of 650 million units by 2031, a 38% compound annual growth rate from 2025, with several VCSEL NPO products expected to launch in 2027 and 2028. It names Lumentum and Furukawa Electric (1,060 nm wavelength) and Coherent (850 nm, with plans to adopt 1,060 nm), and says Broadcom, ams OSRAM and HiSilicon use 850 nm devices reaching up to 3.2 Tbps in aggregate.

Eliyan, a start up covered by Counterpoint on 14 October 2026, claims its microVCSEL link beats copper by 8 times in bandwidth density, 5 times in power efficiency and 50 times in reach. It reports a 145 million dollar Series C at a reported one billion dollar valuation, 224 Gbps devices sampling since the first quarter of 2026 and commercial production planned for 2027. All of that is company information as relayed by an analyst firm, and private company figures are unaudited.

What this means for buyers

Optics and networking are now a major cost line. For context, the scale up and scale out guide covers where these links sit in an AI cluster, and the HBM explainer covers the memory beside the chips. When you read a large compute commitment, such as those in OpenAI's compute commitments, remember that switches, lasers and fibre are part of what is being purchased.

Dell'Oro says it will model 800G, 1.6T and 3.2T pluggable modules, coherent optics, NPO and CPO in a new optical connectivity research stream led by an analyst appointed on 1 October 2026. That shows the firm treats the whole range as live markets, not a single winner.

What is firm and what is uncertain

Firm:

• NVIDIA lists CPO switches as available in 2026, with named first adopters.
• Lumentum reported the revenue figures above and cited demand for CPO lasers.
• Test, packaging and connector parts of the chain are named by TrendForce and Counterpoint.

Uncertain:

• The size of the power savings in real clusters. The 5 times and below 5 picojoule figures come from vendors.
• The volume timeline. Lightmatter, as reported by Counterpoint on 14 October 2026, expects early deployments in 2027, accelerating in 2028 and 2029, while TrendForce calls 2027 pivotal for test equipment and 2028 critical for validating mass production readiness.
• Which approach wins among pluggable, NPO and CPO. Dell'Oro expects them to coexist.
• Any share by supplier. No public, primary source gives one.

This is not investment advice, and none of these companies is endorsed here.

What to watch next

• Whether Spectrum-X Photonics ramps in the second half of 2026 as NVIDIA states, and who confirms deployments.
• Lumentum's results for fiscal first quarter 2027 and the launch of its external laser module in the first half of 2027.
• Progress in testing speed through 2027, which TrendForce calls pivotal.
• Open standards and reference designs shown at OCP, including the Open Compute Project summit of 12 to 15 October 2026.

For the software components in your own products, Safeguard offers SBOM and dependency scanning, with a free scan.

Quick check

Sources

• NVIDIA silicon photonics networking page
• NVIDIA Spectrum-X page
• Lumentum, fiscal fourth quarter and full year 2026 results, 11 August 2026
• Lumentum, ECOC 2026 laser module demonstration, 21 September 2026
• Lumentum news releases list
• TrendForce, CPO test platforms, 13 October 2026
• TrendForce, VCSEL NPO forecast, 8 October 2026
• Counterpoint Research, Lightmatter and CPO, 14 October 2026
• Counterpoint Research, Eliyan and the copper wall, 14 October 2026
• Dell'Oro Group, AI back end switch forecast, 28 July 2026
• Dell'Oro Group, optical connectivity research appointment, 1 October 2026
• Lumentum, Qualcomm and Corning, optical die to die demonstration at ECOC 2026, 21 September 2026
• Broadcom, CPO product page, accessed 19 October 2026
• Open Compute Project

Last reviewed: 19 October 2026.