A Morgan Stanley rebuttal of SemiAnalysis's Nvidia 800V delay claim has split the AI infrastructure trade, with CPO downgrades confirmed but power architecture timelines contested.
Back
Morgan Stanley backs SemiAnalysis CPO delay, disputes 800V timeline
Morgan Stanley's Greater China semiconductor and hardware teams on June 10 issued a split verdict on SemiAnalysis's June 9 report that triggered a rout in US optical communication stocks: they agreed that co-packaged optics shipments will fall far short of market expectations through 2027, but disputed the claim that Nvidia's 800V DC power architecture has been pushed to post-2028.
"CPO near-term volume will be meaningfully below consensus, but the long-term thesis from 2028 onward remains intact," the Morgan Stanley team said in a note seen by Edgen. The bank estimated 2027 global light engine shipments at 6 million to 7 million units, versus market expectations of 20 million to 30 million — a gap of as much as 24 million units that it said will keep near-term sentiment under pressure.
Manufacturing bottlenecks are the primary constraint. TSMC plans to expand photonic integrated circuit capacity to 10,000 wafers per month by the first quarter of 2027, but system-on-integrated-chip yields remain at 50 percent to 60 percent, while downstream assembly yields are as low as 20 percent to 50 percent, the bank said. The 2026-to-2028 period will see a coexistence of pluggable optics, CPO and copper interconnects, with 1.6T and 3.2T products still dominating the market.
The stakes are high for the AI infrastructure supply chain. Optical component suppliers including Coherent and Lumentum face continued headwinds as the CPO ramp disappoints, while power equipment makers such as Delta Electronics stand to benefit if the 800V timeline holds. Morgan Stanley maintained overweight ratings on core CPO names including TSMC, arguing the real inflection begins in 2028.
CPO: A 20-million-unit gap
Morgan Stanley's CPO forecast represents a roughly 75 percent discount to the sell-side consensus of 20 million to 30 million light engines in 2027. The bank attributed the gap to two compounding yield constraints at TSMC's advanced packaging lines.
TSMC's SoIC technology, which integrates the photonic and electronic dies, is running at 50 percent to 60 percent yield, the bank said. The downstream module assembly step — where individual components are packaged into final optical engines — is even more constrained at 20 percent to 50 percent yield. Together, these bottlenecks cap total addressable output well below what the market has been pricing in.
The bank said the transition period from 2026 through 2028 will see three interconnect technologies competing: traditional pluggable optics (1.6T and 3.2T), CPO and near-packaged optics, and copper-based direct-attach cables. CPO's share of total optical shipments will remain small until 2028, when yields improve and hyperscaler adoption reaches critical mass.
800V: The ±400V divergence
On the power architecture front, Morgan Stanley's hardware team took direct issue with SemiAnalysis's timeline. Nvidia stated at its Taipei GTC event that 800V DC power cabinet development is on track for mass production readiness in the third quarter of 2026, according to the bank's supply chain checks.
Delta Electronics is positioned to be the first manufacturer to deliver standalone 800V DC power cabinets, with initial shipments to a North American hyperscaler expected in the fourth quarter of 2026, Morgan Stanley said. Early volumes will be limited, as 800V DC protection devices, UL certification and industry safety standards are still being finalized.
The more consequential disagreement concerns the fate of the ±400V DC architecture. SemiAnalysis argued that ±400V will continue in parallel with 800V, serving hyperscalers' own ASIC deployments. Morgan Stanley's supply chain work suggests the opposite: major cloud providers have shifted research and development resources from ±400V to 800V. Both views cannot be correct for the same set of customers and the same deployment window, creating binary risk for the power supply chain.
Three events will resolve the dispute: Delta's fourth-quarter 2026 cabinet deliveries and their recipient; year-end ±400V sidecar orders and which supplier wins them; and Nvidia's next public comments on its Rubin Ultra and Kyber product roadmaps. Until then, the AI power supply chain trades on competing narratives rather than confirmed orders.
This article is for informational purposes only and does not constitute investment advice.
