SC25 Power is becoming a major headache for datacenter operators as they grapple with how to support ever larger deployments of GPU servers - so much so that the AI boom is now driving the adoption of a technology once thought too immature and failure-prone to merit the risk.

We’re talking, of course, about co-packaged optical (CPO) switching.

At the SC25 supercomputing conference in St. Louis this week, Nvidia revealed that GPU-bit barn operators Lambda and CoreWeave would be adopting its Quantum-X Photonics CPO switches alongside deployments at the Texas Advanced Computing Center (TACC).

Nvidia has some competition: Broadcom showed off its own Tomahawk 5- and 6-based CPO switches. But while CPO switching is poised to take off in 2026, getting to this point has been a journey - and it’s far from over.

So what’s driving the CPO transition? In the words of former Top Gear host Jeremy Clarkson, “speed and power.”

AI networks require exceptionally fast port speeds, up to 800 Gbps, with Nvidia already charting a course toward 1.6 Tbps ports with its next-gen ConnectX-9 NICs.

Unfortunately, at these speeds, direct attach copper cables can only reach a meter or two and often require expensive retimers. Stitching together a few tens or hundreds of thousands of GPUs means potentially hundreds of thousands of power-hungry pluggable transceivers.

Networking vendors like Broadcom have been toying with CPO tech for years now. As the name suggests, it involves moving the optical components traditionally found in pluggable transceivers into the appliance itself using a series of photonic chiplets packaged alongside the switch ASICs. Instead of QSFP cages and pluggables, the fiber pairs are now attached directly to the switch’s front panel.

While each transceiver doesn’t consume that much power – somewhere between 9 and 15 watts depending on port speed – that adds up pretty quickly when you’re talking about the kinds of large non-blocking fat-tree networks used in AI backend networks.

A compute cluster with 128,000 GPUs could cut the number of pluggable transceivers from nearly half a million to around 128,000 simply by moving to CPO switches.

Nvidia estimates its Photonics switches are as much as 3.5x more energy efficient, while Broadcom’s data suggests the tech could lower optics power consumption by 65 percent.

What took so long?

One of the biggest barriers to CPO adoption has been reliability and the blast radius when they fail.

In a traditional switch, if an optical pluggable fails or degrades, you might lose a port, but you don’t lose the whole switch. With CPO, if one of the photonic chiplets fails, you don’t just lose one port – you might lose 8, 16, 32, or more.

This is one of the reasons most CPO vendors, including the two big ones — Broadcom and Nvidia — have opted for external laser modules.

Lasers are one of the more failure-prone components of an optical transceiver, so by keeping them in a larger pluggable form factor, they’re not only user serviceable, but can also compensate for failures by boosting the output of the others in the event of a failure.

But as it turns out, many of these concerns appear to be unfounded. In fact, early testing by Broadcom and Meta shows that the technology not only offers better latency by eliminating the number of electrical interfaces between the optics and switch ASIC, but is also significantly more reliable.

Last month, Meta revealed that it had not only deployed Broadcom’s 51.2 Tbps co-packaged optical switches, codenamed Bailly, in its datacenters, but had recorded one million cumulative device hours of flap-free operation at 400 Gbps equivalent port speeds.

If you’re not familiar, link flaps occur when switch ports go up and down in quick succession, introducing network instability and disrupting the flow of information.

Nvidia, meanwhile, claims its photonic networking platform is up to 10x more resilient, allowing for applications, like training workloads, to run 5x longer without interruptions.

The current state of CPO switching

As we mentioned earlier, Broadcom and Nvidia are among the first to embrace CPO for packet switching.

At GTC this spring, you may recall that Nvidia teased its first set of CPO switches in both InfiniBand and Ethernet flavors with its Spectrum-X and Quantum-X.

Nvidia’s Quantum-X Photonics platform is fully liquid-cooled and sports 144 ports of 800Gbps InfiniBand using 200 Gbps serializer-deserializers, putting its total bandwidth at 115.2 Tbps.

Here’s a closer look at Nvidia’s 144 Tbps Quantum-X Photonics switch, which boasts 144 800 Gbps optical InfiniBand ports - Click to enlarge

These are the switches that TACC, Lambda, and CoreWeave have announced plans at SC25 this week to deploy across their compute infrastructure going forward.

For those who would prefer to stick with Ethernet, the options are more diverse. Nvidia will offer multiple versions of its Spectrum-X Photonics switches depending on your preferred configuration. For those wanting maximum radix, aka lots of ports, Nvidia will offer switches with either 512 or 2,048 200 Gbps interfaces.

Meanwhile, those looking for maximum performance will have the option of either 128 or 512 800 Gbps ports.

Here’s a closer look at the front panel of Broadcom’s 102.4 Tbps Davisson co-packaged optical switch - Click to enlarge

Nvidia’s Photonic Ethernet kit won’t arrive until next year and is already facing competition from the likes of Broadcom. Micas Networks has begun shipping a 51.2 Tbps CPO switch based on Broadcom’s older Tomahawk 5 ASICs and Bailly CPO tech.

Alongside the switch, Broadcom also showed off its latest generation Davisson CPO platform, featuring a 102.4 Tbps Tomahawk 6 switch ASIC that can break out in up to 512 200 Gbps interfaces.

• SC25 gets heavy with mega power and cooling solutions
• Scientific computing is about to get a massive injection of AI
• Europe joins US as exascale superpower after Jupiter clinches Top500 run
• Nvidia’s green500 dominance continues as France’s Kairos super takes efficiency title

What comes next?

So far, Nvidia has focused most of its optics attention on CPO switches, preferring to stick with QSFP cages and pluggable transceivers on the NIC side of the equation - at least through its newly announced ConnectX-9 family of SuperNICs.

However, Broadcom and others are looking to bring co-packaged optics to the accelerators themselves before long. You may recall that back at Hot Chips 2024, Broadcom detailed a 6.4 Tbps optic engine aimed at large scale-up compute domains.

Several other companies are also looking to bring optical I/O to accelerators, including Celestial AI, Ayar Labs, and Lightmatter, to name a few.

At SC25, Ayar Labs showed what an XPU using its CPO chiplets (the eight dies located at the ends of the package) might look like. - Click to enlarge

Both Ayar and Lightmatter showed off live demos of their latest CPO and optical interposer tech at SC25. In the case of Ayar, the startup showed off reference design developed in collaboration with Alchip to integrate eight of its TeraPHY chiplets into a single package using a combination of UCIe-S and UCIe-A interconnects, which will eventually provide up to 200 Tbps of bidirectional bandwidth to chip to chip connectivity.

Lightmatter, meanwhile, is approaching the optical I/O from two directions. The first is a CPO chiplet that the company claims will deliver up to 32 Tbps of bandwidth using 56 Gbps NRZ, or 64 Tbps using 112 Gbps PAM4.

Here Lightmatter demonstrates data moving across its M1000 silicon photonic interposers. - Click to enlarge

Additionally, Lightmatter has developed a silicon photonic interposer called the Passage M1000 that’s designed to stitch multiple chiplets together using photonic interconnects for both chip-to-chip and package-to-package communications.

Eventually, these technologies may eliminate the need for pluggable optics entirely, and even open the door to more efficient scale-up compute domains containing thousands of accelerators acting as one. ®