Orient Securities released a research report stating that Google's v8t architecture can scale a single cluster to over one million chips. The combination of Scale-up and DCN architectures continues to drive demand for Optical Circuit Switches (OCS). The adoption of the Boardfly architecture in the v8i chip demonstrates that OCS technology is also applicable to the vast market for Mixture-of-Experts (MoE) inference scenarios, thereby raising the potential demand ceiling. Furthermore, the incremental demand for liquid cooling driven by ASIC chips is significant. The current v8 series already features liquid cooling as a standard component across all models. As chip supply accelerates, the liquid cooling industry chain is expected to maintain a high growth trajectory.

Key viewpoints from Orient Securities are as follows:

The v8t introduces memory pooling, enabling a single training cluster to scale to over one million cards. On April 23, Google launched its new generation of TPUs, including the v8t and v8i chips. The v8t, designed for training, aims to reduce model development cycles from months to weeks. A single Pod can scale to 9,600 chips, delivering 121 EFlops of FP4 compute power, nearly triple the previous generation. The Scale-up ICI rate per chip has doubled to 19.2 Tb/s, while the Scale-out bandwidth per chip has increased to 400 Gb/s. The introduction of the Virgo network architecture utilizes a flat, two-layer non-blocking network expansion, capable of connecting over 134,000 chips within a single network structure and scaling a single Cluster to more than one million chips. The v8t's Scale-up network continues to use 3D Torus with OCS interconnects, and the DCN network also employs OCS. As cluster sizes reach the million-card level, model training demands are expected to persistently drive the need for OCS interconnects. Additionally, the v8t incorporates TPU Direct RDMA and TPU Direct Storage technologies, enabling direct data transfer between TPU memory and network interface cards, as well as direct access between TPUs and high-speed managed storage.

The v8i adopts the Boardfly architecture, expanding OCS applications into the inference domain. As a post-training and inference chip, the v8i features SRAM capacity three times greater than the previous generation, accommodating larger KV caches and significantly reducing idle time during long-context decoding. The Boardfly architecture used in the v8i notably reduces the network diameter to 7 hops compared to the 3D Torus architecture, making it more suitable for MoE inference. In the Boardfly architecture, 4 chips within a tray are directly connected via ICI. Every 8 trays form a group, achieving full interconnectivity through copper cables. Every 36 groups are interconnected via OCS to form a pod of 1,152 chips. While some investors believe OCS applications are relatively confined to training and DCN, the Boardfly architecture demonstrates that OCS can also be deployed in the broad market for MoE inference scenarios. With growing demand from applications like OpenClaw and coding, OCS is expected to see increasing adoption alongside TPUs in the networking requirements for post-training and inference scenarios at model developers such as Anthropic and Meta.

Liquid cooling is progressively becoming standard for the new generation of ASICs, with its adoption becoming unstoppable. Air cooling is no longer sufficient for both the TPU v8t and v8i, leading to the adoption of liquid cooling technology paired with fourth-generation CDUs. On April 22, Vertiv announced its Q1 2026 financial results, reporting net sales of $2.65 billion, a 30% year-over-year increase, primarily driven by revenue growth in the Americas and strong data center demand. Adjusted operating profit rose 64% year-over-year, with an adjusted operating profit margin of 20.8%, up 4.3 percentage points. The company anticipates continued strong data center market demand in 2026, planning to accelerate capacity expansion and strategic investments. It forecasts adjusted operating profit to increase 53% year-over-year to a midpoint of $3.2 billion, approximately $160 million higher than previous guidance, and an adjusted operating profit margin to rise 2.9 percentage points to a midpoint of 23.3%, up 0.8 percentage points from prior estimates. Beyond the GPU liquid cooling market, the incremental demand for liquid cooling from ASIC chips is considerable. The current v8 series already标配 liquid cooling, which is necessary for both training and inference applications. As chip supply accelerates, the liquid cooling industry chain is poised to sustain high growth.

Investment recommendations and targets include OCS-related companies such as Tengjing Technology, Juguang Technology, Dekeli, Guangku Technology, Haitai New Light, Changxin Bochuang, Fujing Technology, and Shuijing Optoelectronics. For the full liquid cooling chain, consider Yingweike; other liquid cooling-related companies include Lingyi Zhibao, Shenling Environment, Yidong Electronics, and Dingtong Technology.

Risks include AI development falling short of expectations, intensified industry competition, and geopolitical risks.