The technology that was supposed to solve AI’s bandwidth problem has run into a manufacturing one. Co-packaged optics, the approach that puts lasers and switch silicon in a single package, are producing usable parts at a yield of roughly 19%, according to a report from SemiAnalysis, far below the level needed for volume production.
The result is a delay that ripples through every AI data-center roadmap: mass production is now expected in 2028 or 2029, later than the industry had hoped, and the companies that bet their next-generation switches on the technology are scrambling for alternatives.
Co-packaged optics matter because of a simple physics problem. AI models train across tens of thousands of chips, and moving data between them is becoming as costly as computing it. Copper wires degrade over distance and consume power at high speeds; optical links are faster but expensive at the conversion points. CPO puts the optics next to the switch, cutting power and latency, in theory.
The theory met the factory floor. SemiAnalysis, a research firm that tracks semiconductor supply chains, found that yields on the most advanced co-packaged modules are stuck near one in five, a level at which unit costs are multiples of the target and no customer will design a product around the part.
The numbers behind the stall are concrete. Nvidia’s Spectrum 6 switch, a showcase CPO product, shows insertion loss above 3.5 decibels, meaning too much signal is leaking at the connection points between silicon and glass. And TSMC’s advanced packaging process, the same technology that assembles AI accelerators, has become the bottleneck, with capacity already consumed by the chip packages that pay better.
The financial fallout arrived quickly. Morgan Stanley has sharply cut its 2027 forecast for optical engine shipments, a move that rippled through suppliers and pushed investors to ask which companies in the optical supply chain still have a story. The answer, for now, is the ones that make the old technology.
The old technology is pluggable optics, the transceivers that plug into switch faces and have powered data centers for a decade. The delay extends their run: cloud builders planning 2027 data centers will buy pluggables, not co-packaged modules, and suppliers of those components are revising their own outlooks upward as the CPO timeline slips.
There is a middle path that gained attention as CPO stumbled. Linear-drive optics, which simplify the electronics between the switch and the optical module, promise much of the power savings without the yield problem, and several switch makers have begun planning around it. It is not the full CPO dream, but it works with existing packaging.
The delay reshuffles the competitive picture. Nvidia has pushed CPO hardest, betting that co-packaging gives it the bandwidth edge its next accelerators need; a 2028 or 2029 timeline means its next two switch generations will carry pluggable optics after all. Rivals betting on the same physics gain nothing from the delay, but they lose nothing either.
For the supply chain, the question is who absorbs the pain. Optical module makers that invested in CPO production lines face idle capacity; packaging houses that reserved space for co-packaged modules will fill it with other work; and the laser companies whose components go into the modules face a longer wait for volume orders.
Analysts said the episode follows a familiar pattern: every generation of AI infrastructure hits a manufacturing wall, and the wall moves rather than falls. Memory was the bottleneck, then power, then packaging, and now the optics between everything. Each time, the industry solves it a year later than planned and a generation sooner than feared.
The practical read for the next two years:
The delay hits the companies that bet early. Switch makers that designed around co-packaged modules must re-spin products, cloud builders that reserved space for the new optics will fill it with older technology, and startups that raised money on CPO roadmaps face a longer wait for revenue. The startups expected to survive are the ones with pluggable and linear-drive products to sell in the meantime.
Yield problems compound. A part that fails half the time doubles every upstream cost, and at 19% the economics fail exponentially rather than linearly. Suppliers said the practical threshold for volume production is roughly 80% to 90%, and closing the gap requires process work measured in years, not months, because every fix in the assembly line tends to break something else. AI data centers will get faster switches, more of them, and a lot more glass in the form of pluggable optics and traditional fiber. The co-packaged future is not cancelled; it is delayed, and the companies with the patience and the balance sheets to wait are the ones positioned to win the second half of the decade.
For now, the 19% yield figure is the number the industry is quoting, and it explains why every roadmap quietly moved. The fix, it turns out, is a manufacturing problem wearing a technology disguise, and it isn’t due until 2028.
