Responding to SemiAnalysis? NVIDIA Posts Video of First CPO Switch Deployments

Just days after the bearish report landed, NVIDIA responded with a hardware showcase.

NVIDIA recently posted a video on X, publicly showcasing its Quantum-X InfiniBand Photonics Q3450-LD switch equipped with its silicon photonics CPO technology. The device has been deployed in Lambda's production-grade GPU clusters, improving power efficiency by 3.5 times and network resilience by 10 times. Additionally, CoreWeave, Meta, Microsoft, and Oracle Cloud Infrastructure have joined the initial deployment cohort.

However, just before this progress, the market was thrown into turmoil by a bearish report on NVIDIA's CPO prospects. On June 9, SemiAnalysis downgraded its shipment expectations for NVIDIA's CPO switches and stated that mass shipments of its native 800V DC power solution would be delayed until after 2028, causing a sharp decline in the U.S. optical communication sector. Currently, the divergence in market expectations regarding the two technical routes led by NVIDIA—CPO and 800V DC power—remains a core variable influencing the trend of the AI infrastructure sector.

CPO Hardware Implementation: From PowerPoint to Production Deployment

NVIDIA's CPO switch technology has moved from the conceptual stage to actual mass production deployment. The Q3450-LD switch directly encapsulates optical components internally, replacing traditional pluggable transceivers, and supports full-duplex bandwidth of up to 115.2 Tbps.

Power consumption is the core selling point of this technology. Standard switches consume approximately 7 kW, while the CPO version requires only about 3.95 kW, saving roughly 3.05 kW per unit. Lambda summarized this value proposition as "more tokens per watt." In large-scale AI training scenarios, the tenfold increase in network resilience is equally significant—a single network interruption can force tasks spanning thousands of GPUs to restart from scratch, incurring extremely high costs.

Lambda had already announced the completion of its integration with NVIDIA's silicon photonics networking solution in November 2025, and the public debut of the Q3450-LD marks the fulfillment of this early commitment. NVIDIA positions the CPO switch as essential infrastructure for running "agent AI workloads" in its "AI factories."

SemiAnalysis and Morgan Stanley: Consensus Bearishness on CPO, Divergent Views on 800V

There is a significant expectation gap in the market regarding the two technical routes of CPO and 800V.

Regarding CPO, Morgan Stanley and SemiAnalysis hold largely consistent views. Morgan Stanley expects global optical engine shipments to reach only 6 million to 7 million units in 2027, far below the market expectation of 20 million to 30 million units. The core bottleneck lies in manufacturing: TSMC's expansion of photonic integrated circuit capacity is slow, with system integration chip yields at only 50% to 60%, and downstream assembly yields even lower at 20% to 50%. The period from 2026 to 2028 will be a transition phase where pluggable optical modules, CPO, and copper interconnect solutions coexist.

Regarding the 800V power architecture, Morgan Stanley's stance is diametrically opposed to that of SemiAnalysis. According to Morgan Stanley, NVIDIA stated that 800V power cabinets will be ready for mass production in Q3 2026. This is not entirely contradictory to SemiAnalysis's claim that "mass shipments will be delayed until after 2028," as "readiness for mass production" and "large-scale volume ramp-up" are distinct milestones. Delta Electronics is expected to become the first manufacturer to achieve mass production, with initial deliveries to North American hyperscale cloud service providers scheduled for Q4 2026, albeit on a limited scale initially.