The competitive battleground in AI data centers is shifting away from the number of GPUs and toward the speed at which data moves between chips and memory. In a new optics industry report, Jefferies raised its outlook for 1.6T optical transceivers but said meaningful mass production of scale-up co-packaged optics, or CPO, may be pushed back to 2029. That view drew pushback from semiconductor industry commentator Sami, who argued Nvidia needs CPO to strengthen NVLink and support a lower HBM configuration in the Rubin generation.
Jefferies lifts AI transceiver shipment forecast
Jefferies raised its forecast for AI optical transceiver shipments, projecting nearly 180 million units in 2027. That figure includes about 90 million 800G units and more than 80 million 1.6T units. For 2028, the bank expects shipments to reach 245 million units.
Based on feedback from recent supply-chain visits last week, Jefferies said Lumentum’s transceiver gross margin could exceed 30%, above its earlier estimate in the mid-20% range.
The report also flagged valuation and deployment questions. Jefferies said the ratio of AI optical transceivers to XPUs would rise to 7.0x in 2028, while the attach rate for TPUs would be only 1.5, which it described as elevated. It added that meaningful scale-up CPO production may be delayed until 2029. Customers remain more optimistic about NPO, or near-packaged optics, though planning is still highly fluid.
Sami says CPO is central to Nvidia’s HBM cost strategy
Sami said Jefferies was too conservative. In his view, getting CPO into NVLink by the end of 2027 is Nvidia’s top priority if it wants to defend the economics of expensive HBM. He argued Nvidia is effectively telling customers that a lower HBM allocation in Rubin is acceptable because faster NVLink would make memory pooling more efficient.
Under that framework, if CPO materially improves NVLink bandwidth and transmission distance, multiple GPUs could share memory resources more effectively. That would reduce the amount of HBM needed per GPU and make total system cost easier for customers to accept.
Nvidia has not publicly disclosed such a plan. Still, the article said market talk in August that Rubin Ultra would reduce HBM capacity in a single rack and instead connect multiple racks through optical components had been confirmed by research firms including TrendForce.
Rack cost mix is changing
The article said Jefferies’ rack bill-of-materials analysis, published a day earlier, showed GPU share of total rack cost falling from 62.5% in the GB300 NVL72 generation to 46% in Rubin-era NVL576 Pod systems. Over the same span, the share for networking and fiber interconnect rose from 8.6% to 22.4%.
A former senior Microsoft engineer was also cited as identifying high-speed networking, photonics, and memory pooling as some of the most underappreciated themes in AI infrastructure.
Timing remains uncertain, and NPO still has support
Even with CPO drawing more attention, the rollout timeline remains unsettled. Jeff Pu of GF Securities cut his shipment forecast for NVL576 optical engines, citing a delay of T4-configured systems to the first quarter of 2028. He also said NPO is still likely to remain the dominant approach in this cycle.
Sami’s long and short positioning
Sami laid out a trade framework built on the idea that lower HBM content per GPU would pressure demand for related equipment, while increasing the value captured by optical interconnect.
- Long: Lumentum (LITE), Coherent (COHR), Soitec (SOI.PA), Sumitomo Electric (5802.T)
- Short: BESI, SUSS MicroTec SE (SMHN), Hanmi Semiconductor (042700.KS), ASMPT
The article added that several points still need to be verified through future supply-chain updates, including the CPO-versus-NPO path, Nvidia’s actual deployment schedule, and whether HBM demand will in fact decline. It said investors should keep watching specification updates for Nvidia’s next-generation platform and order visibility at major optical suppliers.
The original article stated that it does not constitute investment advice.