AI Agents Reshape AI Computing: CPU Demand Surges, Unisound Sees Agent Revenue Jump
AI agents are lifting CPU demand and agent revenue, as Nvidia shows CPU racks and Unisound reports growth.
According to Leiphone, the shift reflects a change in AI workloads. Agents need to understand a goal, break it into tasks, orchestrate steps and call external tools repeatedly. A complex agent task can run 20 to 50 steps, with much of the load falling on CPUs. A senior server vendor told Leiphone that in one observed workflow, CPU load exceeded 80%. The CPU-to-GPU ratio in AI data centers may move from about 1:8 to 1:4 or even closer to 1:1.
Demand is already visible in pricing. Global server CPU supply was unbalanced in Q1 2026, and average selling prices rose nearly 30% from Q4 2025 to Q2 2026. Intel and AMD signaled price increases of over 10%. Intel's Chen Baoli said a leading domestic model maker's CPU demand grew fivefold year-on-year, but he cautioned that the final ratio is not settled. Xu Qingwei of Lanxin Computing sees a long-term price rise, while Feng Jiayu of Yizhidian argues that intense domestic competition limits the room for increases.
Investors have taken notice. Leiphone reported that chip investor Li Hua, who had previously passed on an Arm CPU startup, contacted the company again after Nvidia's demo. Li Yang of SenseTime Guochuang Capital said his firm began systematic research on CPU opportunities in late 2025, with agent-based reasoning adding a new layer to their thesis. Many experts see the next five years as a critical window for domestic CPU companies.
The first beneficiaries are companies with CPUs already in mass production. Huawei's Kunpeng and Hygon's C86 remain leading forces, and Arm-route startups such as Yizhidian and Hongjun Micro are expanding into Internet customers; Yizhidian has completed validation with Tencent in cloud gaming and cloud phone scenarios. For RISC-V companies, the challenge is the ecosystem. Xu Qingwei said his company has mass-produced the industry's first RISC-V+AI fusion server CPU, but the high-performance ecosystem is immature. Ze Yuan, a veteran with nearly 20 years of CPU experience, said Internet companies with strong software capabilities may be the first RISC-V adopters.
CPU design for agents is far from settled. Chen Baoli said CPUs take two to four years to develop, making many 'agent-designed' claims marketing. Chen Zhengsheng of Yuxian Micro emphasized multi-threading and I/O, noting Nvidia's Vera CPU uses dual-threading. There may be a split: fewer-core, high-single-core-performance CPUs for orchestration, and many-core CPUs for agent sandboxes and concurrent scheduling. This uncertainty keeps x86, Arm and RISC-V in the race, with system-level integration and ecosystem as the final differentiators.
On the application side, QbitAI reported that Unisound, called the first AGI stock in Hong Kong, posted first-half 2026 revenue of 562 million yuan, up 38.7% year-on-year. Agent business brought in 478 million yuan, up 35.7% and accounting for 85.1% of total revenue. Token business revenue was nearly 30 million yuan, with Q2 up more than 500% quarter-on-quarter and gross margin above 60%. Repeat purchases made up over 60% of revenue, contract value grew 65%, and orders in hand exceeded 1.5 billion yuan. Gross profit rose 42% to 186 million yuan, while losses narrowed 20.2% to 237 million yuan.
Unisound follows a heavy-delivery-then-reuse model similar to Palantir. It has served more than 470 medical institutions, completed profile governance for 524,400 Xiamen enterprises, and helped a new-energy vehicle maker boost orders by 10%. Its agent platform UniAgentOS has accumulated 1,773 instantiated agents, with one module reused up to 119 times; the project implementation cost ratio dropped from 21% to 17%. In a typical hospital case, the medical agent cut single-record review time to under 10 seconds, lifted quality-control coverage to 100%, and raised manual efficiency by more than 80%. The company invested 284 million yuan in R&D in the first half, accounting for 78.5% of its combined R&D, sales and management expenses, and its 327 R&D staff make up 67.7% of employees. Its flagship model U2, a native agent model with 260 billion total parameters activated sparsely at about 10 billion per inference, has reached the first tier in agent benchmarks.