Orient: Chip and model help boost demand for domestic computing power XPO drives demand for liquid cooling.

date
15:12 07/04/2026
avatar
GMT Eight
The continuous increase in the speed of optical modules drives the growth of power consumption, and the demand for liquid cooling of optical modules is expected to gradually be released.
Orient released a research report stating that the daily call volume of Byte Beans has exceeded 120 trillion, and the domestic large model Tokens call volume is "moving from virtual to real," while tools such as "lobster" based on large models are quickly spreading, further amplifying the accelerating growth of domestic Tokens demand. The commercialization path of large models is becoming increasingly clear, and the demand for domestic computing power industry chain is expected to accelerate. Recently, Arista jointly released a new pluggable optical module XPO standard with more than 45 industry companies. XPO's liquid-cooled cold plate can support module power consumption exceeding 400W. As the speed of optical modules continues to increase, driving power consumption growth, the demand for liquid-cooled optical modules is expected to continue to grow. Key points from Orient: The daily call volume of Byte Beans exceeded 120 trillion, and the domestic large model Tokens call volume is "moving from virtual to real". On April 2nd, the Volcano Engine held the 2026 Force Link AI Innovation Roadshow - Wuhan Station, revealing that by the end of March this year, the daily call volume of Byte Beans' large models exceeded 120 trillion Tokens, nearly doubling from the end of 2025. Compared to the call volume at the end of May 2024, it has achieved a high-speed growth of 1,000 times. The company has become the third player to consume over 100 trillion Tokens daily following OpenAI and Google. Behind the significant growth of Tokens is mainly due to the rapid popularization of Seedance representing AI video creation and AI agents represented by OpenClaw "lobsters". For example, Seedance 2.0 consumes over a million Tokens to generate a 1-minute 720P video; in a complex Agent task, the consumption of Tokens can easily reach tens of thousands or hundreds of thousands. In addition, the number of C-end users of Byte Beans continues to grow, with the Byte Beans App having nearly 78 million daily active users and total downloads exceeding 85 million. Orient believes that the current call volume of large models Tokens is "moving from virtual to real." From the early measures such as advertising and red envelopes to attract traffic, after several iterations of domestic large models, they have gradually reached a leading industry level. From the outstanding performance of domestic Seedance 2.0 to the shutdown of overseas Sora App, it proves the transition of domestic models from "available" in terms of cost and performance to truly "usable". This also drives the upward demand for underlying computing power from single text to high-consumption video demand. In addition, tools like "lobsters" based on large models are spreading rapidly, further amplifying the accelerating growth of domestic Tokens demand, and the commercialization path of large models is becoming increasingly clear. The demand for domestic computing power industry chain is expected to accelerate. The competitiveness of domestic chips continues to improve, and by 2025, the market share of domestic chips will reach 41%. According to Cloud Headlines, IDC announced that the sales volume of AI accelerator cards in the Chinese market in 2025 will reach about 4 million, a 48% increase compared to over 2.7 million cards sold in 2024. In terms of structure, NVIDIA's shipments are about 2.2 million cards, accounting for a decreased share of 55%, while the share of domestic chip manufacturers has increased to 41%, accelerating the process of localization substitution. With the accelerated landing of applications such as AI agents, it is expected to continuously improve the ecosystem of domestic computing power chain. The continuous increase in optical module data rates is driving power consumption growth, gradually entering the era of liquid cooling. On March 11th, Arista jointly released a new pluggable optical module standard XPO (ultra-high-density pluggable optics) with over 45 industry companies. XPO has 64 200Gbps PAM4 high-speed electrical channels, with a single module bandwidth of up to 12.8Tbps. It adopts a Belly-to-Belly architecture, with two independent 32-channel PCB boards in the middle cooled by a liquid-cooled cold plate. The high-power heating elements on the PCB, such as transmission lines, laser drivers, DSP digital signal processors, are installed on the inner side facing the cold plate, and the cold plate provides heat dissipation for the top and bottom PCBs through direct contact. XPO's liquid-cooled cold plate can support module power consumption exceeding 400W. With the continuous increase in optical module data rates driving power consumption growth, the demand for liquid-cooled optical modules is expected to continue to grow. Targets: Domestic computing power related companies such as Cambricon (688256.SH), Hygon Information Technology (688041.SH), Verisilicon Microelectronics (Shanghai) Co., Ltd. (688521.SH), Brite Semiconductor (688691.SH), Semiconductor Manufacturing International Corporation (688981.SH), Hua Hong Semiconductor (688347.SH), Huaqin Co., Ltd. (603296.SH), Suzhou Centec Communications (688702.SH), etc. OCS-related companies such as Optowide Technologies (688195.SH), Focuslight Technologies Inc. (688167.SH), Wuxi Taclink Optoelectronics Technology (688205.SH), Advanced Fiber Resources (300620.SZ), Qingdao Novelbeam Technology (688677.SH), EverProX Technologies (300548.SZ), Castech Inc. (002222.SZ), Zhejiang Crystal-optech (002273.SZ), etc. Liquid cooling related companies such as Shenzhen Envicool Technology (002837.SZ), Lingyi Itech (002600.SZ), Guangdong Shenling Environmental Systems (301018.SZ), YD Electronic Technology (301123.SZ), Dongguan Dingtong Precision Metal Co., Ltd. (688668.SH), etc. Risk warning: AI development falls short of expectations, intensifying industry competition risks, geopolitical risks.