AI PCs have been on the market for a couple of years now, with applications and use cases progressively finding ways to put local AI models to work. China’s version of that story looks different from the rest of the world’s. Since earlier this year, interest in OpenClaw-like AI agent harnesses has driven a wave of ultrasmall desktop PCs there, and unlike the high-end, DGX Spark-class boxes getting attention elsewhere, most of these are inexpensive, modestly specced machines built to host an agent rather than run a large model locally.
Where the lobster talk comes from
All of this traces back to a single piece of open-source software called OpenClaw, an autonomous AI agent that can be given a task in plain language and left to carry it out on its own, browsing the web, editing files, and calling other tools as needed. Its logo features a lobster-like claw, and Chinese users picked up on that and started calling the software 小龙虾, or crayfish, a nickname that stuck to everything that followed, including the hardware people bought to run it. Naming for that hardware hasn’t really settled, but some call it 龙虾盒子, so we’re calling them lobster boxes here for simplicity.
OpenClaw’s rise from around February through April 2026 drove a hardware surge, and it was global rather than a China-only phenomenon. Mac mini sold out in multiple markets because the software simply worked better on macOS, with tighter access to the operating system’s own apps and a quiet, low-power, always-on form factor well suited to running an agent around the clock. What was distinctive about China is how quickly it moved past that Mac dependency. Chinese startups and retailers began shipping their own inexpensive, purpose-built boxes within weeks, and those boxes worked despite modest specifications because they were calling cheap, cloud-hosted AI services rather than running a model locally, so most of them never needed the expensive memory that on-device inference requires.
Tencent and Baidu both ran free installation clinics at their headquarters in Shenzhen and Beijing, drawing crowds in the hundreds as company engineers set up the software machine by machine for anyone who showed up.
Tokens that are cheap
Part of why this took off in China specifically comes down to what it costs to keep an agent thinking. Western AI subscriptions have generally settled around $20 a month for services like ChatGPT Plus or Claude Pro. Chinese equivalents often run under $10 a month, and a large share of casual use runs through free consumer apps entirely, with paid tiers unlocking higher concurrency rather than gating basic access. There are some indications that pricing could increase soon, but this new class of hardware emerged without needing much memory in the first place, precisely because cloud-based models were affordable enough to lean on.
The rise of the OPC
Chinese researchers and businesses have started using a formal term, OPC, short for one-person company, to describe a business built around a single person making the decisions while a cluster of AI agents handles the execution and the day-to-day output. It is a notable enough trend that Lenovo’s own marketing for its lower-cost Baiying-branded AI PC, a roughly $560 Windows box, describes the product as built for OPC operators, sometimes called super-individuals, with use cases like AI content creation, e-commerce operations, and market analysis.
There is also a cultural aspect to consider. Much of the anxiety Chinese workers have felt about AI replacing their jobs pushed people toward learning to run the tool themselves, which helps explain why installation clinics and DIY tutorials found such a ready audience. That desire to learn the tool fits naturally with the OPC idea, since China already has a way to turn it into a business.
Where the growth is headed
Most of these agent-hosting devices are what IDC classifies as ultrasmall desktops, which we expect to top one million units in China in 2026 and grow to roughly 1.8 million by 2027. That is a small slice of China’s overall PC market, which IDC expects to total around 51 million units in 2026, so this remains a niche category. Lenovo has been the most visible vendor building for this category specifically, with its lower-cost Baiying and Tianxi-branded AI hosts, and Xiaomi joined in August with its own prototype, the AI Cube, built around three of its own chips. Many of the units behind that growth today are still the cheap, low-end boxes that lean on cloud AI rather than local processing, but interest in local token generation is picking up, and that points to some shift toward pricier, more capable hardware.
Likely to stay a China story
These low-end boxes look likely to stay a China phenomenon, since the conditions behind them are not really present in the same combination elsewhere. The rest of the world is increasingly exploring NVIDIA’s DGX/RTX Spark, AMD’s Ryzen AI Halo-based systems, and Microsoft’s Project Zenith, announced together with AMD at IFA earlier this month, which are all aimed at running large models locally. These are on the other end of the price spectrum but are based on the premise that generating tokens locally will be more economically effective than paying for cloud-based tokens.