Why AI's Next Networking Upgrade Is Stuck at the Test Bench

AI data centers are running out of bandwidth. The pluggable optical modules that front every switch and router were built for a slower era, and the GPU clusters now driving generative AI need roughly 10 times the per-link bandwidth that scale-out networks used to deliver. The industry's answer is co-packaged optics, or CPO: the optical interface moves off the front panel and onto the same package as the switch chip, cutting the electrical path and the power that goes with it. Near-package optics, or NPO, is the intermediate step, with the optical engine mounted on the board next to the chip rather than on it. Both are now in the supply chain, according to Semiconductor Engineering: Nvidia and Broadcom have already introduced CPO and NPO parts, pulling hyperscalers in behind them.

The technology is no longer the bottleneck. The test bench is.

Every CPO design today needs its own custom test setup, because there are no shared standards for the fiber connectors that survive high insertion counts, no agreed product test specifications, and no common format for the data that comes off the test floor. Custom-only testing is the reason time to market is longer, per-design cost is higher, and the same automated test equipment cannot be reused from one CPO program to the next. "The lack of standards is what's gating the ecosystem right now," says Vishal Chandrasekar, director of product management at Ayar Labs, a CPO vendor. "Without standard test specifications and data formats, profitability across the CPO ecosystem is delayed."

The clock is short. Ayar Labs projects near-package optics will reach volume production in 12 to 18 months, and on-substrate CPO in 18 to 24 months for the two jobs AI data centers care about: scale-up, the high-bandwidth fabric that links the GPUs in one server or rack, and scale-out, the network that connects racks and rows of servers across the building. The Ayar Labs forecast is a vendor projection, not shipped product, and it lines up with what the rest of the optical supply chain is now quoting publicly. If the standards are not in place before those windows close, the rollout lands custom-test-heavy, which means longer lead times and higher cost per port for every hyperscaler that wants to deploy it.

The standards gap is not abstract. A fiber connector rated for a few mating cycles is fine for a pluggable module that gets swapped once a year. A CPO connector has to survive thousands of insertions across a board-level test flow, and right now each CPO vendor qualifies a different connector against a different test spec, so the test cells cannot be shared. The data coming off those cells is also non-portable: yield learning that one supplier collects on a die cannot be compared directly to the next supplier's run, because the test formats and traceability chains are not aligned. That is what Chandrasekar means when he says the bottleneck is not the chip.

Three things have to standardize before volume production lands, and they map to three different organizations. The first is a smaller set of board-mountable fiber connectors qualified for high insertion counts, which the Optical Internetworking Forum and COBO have begun to scope but have not yet published. The second is shared product test specifications for CPO and NPO, the kind of document an automated test equipment vendor (Advantest, Teradyne, FormFactor) could build a turnkey cell around. The third is end-to-end traceability across the CPO supply chain, so a die tested at a foundry can be correlated with its package-level optical data and with the system-level yield the hyperscaler eventually reports. None of the three is in production today.

Ayar Labs' own roadmap is tied to the same clock. The company's 12-to-18-month NPO projection is the near-friction number to watch: if NPO is in volume on schedule, the industry has had time to land at least the connector and test-spec standards; if it slips, the slip will trace back to the test floor rather than to the foundry. The standardization question is whether the optical networking layer that the AI buildout is counting on arrives as a shared platform or as a stack of one-off designs.