A data-center operator in Beijing is no longer selling raw compute by the card. It is selling finished answers — millions of them, every second, at a price customers can actually budget.
That shift, from renting GPUs to delivering guaranteed-quality AI tokens, is where one of China's fastest-funded AI-infrastructure startups just put its newest cheque.
Approaching.AI (趋境科技), a Beijing company spun out of Tsinghua University's high-performance computing group, closed a Series B of more than RMB 1.3 billion (≈ US$183M / HK$1.43B) announced on 26 October 2026. Together with earlier rounds this year, the raise lifts its 2026 fundraising to roughly RMB 2.4 billion (≈ US$338M / HK$2.64B) — an unusually heavy pace for a firm founded only in March 2024.
Key takeaways
- Approaching.AI (趋境科技) closed a Series B above RMB 1.3 billion (≈ US$183M / HK$1.43B) announced 26 October 2026, its fourth priced round in 2026 alone.
- The round brings 2026 total fundraising to about RMB 2.4 billion (≈ US$338M / HK$2.64B), per Chinese financial media and the company.
- Lead investors are almost entirely local state and industrial capital: Zhejiang Finance Innovation, Xiaoshan Capital, ICBC Investment, Zhongguancun Science City, Xinglian Capital and Qianjiang Century City, with SenseTime (商汤科技) and Wangsu Technology (网宿科技) among follow-on backers.
- The company sells high-quality AI tokens — production-grade model outputs — via an ATaaS (Approaching.AI Token as a Service) platform, betting that inference (推理), not training, is where Chinese enterprises will spend next.
- Approaching.AI claims its daily output of high-quality tokens has crossed one trillion, and says it is building a 10,000-GPU-scale "token factory" in Hangzhou — figures the company states rather than independently audited.
From renting GPUs to shipping tokens
The thesis is simple to state and hard to execute. Most enterprises do not want a cluster; they want a reliable stream of model outputs they can plug into a product. Approaching.AI packages that stream as a service, defining a "high-quality AI token" as one anchored to a real production scenario — high intelligence, low latency, stable throughput, controllable cost.
Technically the firm leans on heterogeneous coordination: running prefill and decode on different chips, converting KV cache between architectures, and fusing domestic and foreign accelerators so that Chinese-made silicon can serve trillion-parameter large models (大模型) in production. The company says it has run leading trillion-parameter models on domestic chips at a cost-performance it claims beats the most advanced international compute. Those are engineering claims worth treating as directional until benchmarked independently.
The round at a glance
The Series B was led jointly by Zhejiang Finance Innovation (浙财创新), Xiaoshan Capital (萧山资本), ICBC Investment (工银投资), Zhongguancun Science City (中关村科学城), Xinglian Capital (星连资本) and Qianjiang Century City (钱江世纪城). Follow-on investors included Yunchuang Capital, Hunan High-Tech Zongheng, Jian Tou Huake, Shenwan Innovation Investment, CITIC Juxin, Zhuhai Technology Industry Group, SenseTime (商汤科技), Wangsu Technology (网宿科技), Gansu Shuzhi, Changjin Control, Yuanzhi Future, Qingzheng Xingzhu and Xiaoka Capital, with existing shareholder Shangshi Capital (尚势资本) increasing its stake.
What is striking is the geography of the cap table. The lead investors read like a map of Yangtze Delta and Beijing state capital: Zhejiang, Xiaoshan, Hangzhou's Qianjiang Century City, Beijing's Zhongguancun. The message is that local governments now treat inference capacity as regional infrastructure, the way they once treated ports and highways.
Why inference, why now
Training a model is a one-time project; running it is a perpetual operating cost. As agents and assistants make many model calls to finish a single task, inference has become the line item enterprises actually feel. Approaching.AI argues the early market chased raw token volume; the next phase prices quality — latency, stability, context integrity — and that is where it wants to sit.
Its route is "fewer models, deeper optimization, higher quality": squeeze more useful tokens from a given batch of compute, and keep that compute effectively utilized. The firm also designs, builds and operates "token factories" for customers who already own or plan to own hardware, then shares in the economics — a relatively asset-light model it has been deploying in the second half of the year.
Where money meets the model
The domestic-compute angle matters because advanced foreign chips remain constrained in China. Approaching.AI proposes "domestic prefill-decode (PD) heterogeneous collaboration," "high-performance heterogeneous KV cache conversion" and "heterogeneous compute fusion" — jargon for making different chips cooperate so that a model trained for one architecture still runs well on another. If the approach scales, it lowers the barrier for enterprises locked out of the newest accelerators.
The state-capital signal
A Tsinghua pedigree helps, but the real story is who is writing the cheques. When industrial banks' investment arms and municipal capital lead an AI round, the bet is on deployed capacity, not a lab demo. The investor base also signals a broader 2026 pattern: Chinese hard-tech funding is consolidating around operators that can show real throughput rather than slideware.
Honest limitations
Every headline number here rests on company disclosure or Chinese financial media, not an independent audit. The RMB 1.3 billion Series B and RMB 2.4 billion 2026 total are reported by outlets including Zhitongshi, EE Focus and TechStartups, citing the company; treat them as company-anchored. The "more than tenfold" growth in operating metrics, the "one trillion daily tokens" claim, and the planned 10,000-GPU Hangzhou facility are company-stated and unverified externally. Valuation was not disclosed. The domestic-chip cost-performance claim is the firm's own framing and awaits independent benchmarking. Inference is a crowded, rapidly commoditizing field, and several well-funded peers — Infinigence (无问芯穹), SiliconFlow (硅基流动) and others — are chasing the same customers, so capital intensity alone does not guarantee a durable edge.
What readers can do now
- If you buy or build AI infrastructure in China, pilot Approaching.AI's token service against your own workload and price per finished answer, not per GPU-hour; that is the metric the company is actually competing on.
- For investors, separate the round's size from durability: track whether the 10,000-GPU factory materializes and whether unit economics hold once subsidies and related-party demand fade.
- If you follow China's AI-capital cycle, watch whether more municipal and bank-led funds target inference — it tells you where Beijing expects recurring spend to land.