Analysis
Huawei is pushing near-packaged optics as an alternative to co-packaged optics in AI networking gear, arguing the economics of CPO have not worked out as promised, The Register reported. The distinction sounds like packaging trivia and is actually a bet on how AI data centers get wired for the next decade.
Co-packaged optics puts the optical engine on the same substrate as the switch ASIC, cutting the electrical distance a signal travels and therefore the power burned driving it. That is the pitch Broadcom made with Tomahawk 5 Bailly and Nvidia made with its Spectrum-X and Quantum-X photonics switches. The problem is manufacturing: bonding optics to a switch die hurts yield, and when a laser fails the customer replaces the entire switch rather than a pluggable module. Near-packaged optics keeps the optical engine close to the ASIC but separately serviceable, giving up some power savings for better yield and field replaceability.
“That is the pitch Broadcom made with Tomahawk 5 Bailly and Nvidia made with its Spectrum-X and Quantum-X photonics switches.”
The stakes scale with cluster size. Interconnect is a rising share of both the cost and the power budget of large GPU deployments, and at 100,000-accelerator scale the networking layer is where a meaningful fraction of the capital goes. Broadcom, Marvell, Nvidia, Coherent and Lumentum all have positions here, and Celero Communications just raised $275 million at a $3 billion valuation building data transfer technology for exactly this problem.
Huawei's participation carries its own asymmetry. Cut off from advanced US accelerators by export controls, the company has concentrated on the parts of the stack it can still build at scale, and networking is one of them. Promoting an architecture that relies less on leading-edge packaging is a technical argument that happens to align precisely with the constraints it operates under -- which does not make the argument wrong, but does explain who is making it. Pulse has previously covered Huawei's export-control workarounds across its Ascend chip line, of which this optics push is the networking-layer sequel.
The resolution will be empirical and visible. Watch whether the next generation of hyperscaler switch deployments specifies CPO or a near-packaged design, because that choice determines which suppliers capture the interconnect budget in the 2027 buildout. Marvell and Coherent both have near-packaged roadmaps; a shift away from full CPO would reprice the optical component supply chain quickly.
