Buying a GPU cluster is easy. Getting it to deliver a stable, cheap stream of model output all day is harder. A Beijing startup is raising money on the gap between those two things, and its investors are betting that the gap is worth more than the hardware.
Key takeaways
- Approaching.AI (趋境科技), formally Beijing Approaching Technology Co., Ltd., announced a B round above ¥1.3 billion (≈ US$182M / HK$1.42B) in late October 2026, according to Shanghai Securities News and China Securities Journal.
- The company says its cumulative 2026 financing has reached ¥2.4 billion (≈ US$336M / HK$2.62B), following an A round in July 2026.
- The round was co-led by Zhejiang Cai Chuangxin, Xiaoshan Capital, ICBC Investment, ZGC Science City, Xinglian Capital and Qianjiang Century City.
- Company-disclosed operating claims include daily high-quality token (词元) output growing more than 100-fold to a trillion-token scale and 2026 revenue approaching ten times the 2025 full year.
The round
The financing was reported by several mainland financial wires on the same company announcement. The lead group combines provincial and district capital, a bank investment arm, a Zhongguancun platform and a market-oriented fund. A long list of followers includes SenseTime (商汤科技), Wangsu Science & Technology (网宿科技), CITIC-affiliated funds and several regional industrial funds. Existing shareholder Shangshi Capital added to its position.
The company says the proceeds go to four uses: expanding high-quality token capacity, upgrading its ATaaS (Approaching.AI Token as a Service) platform, research on inference infrastructure and new products, and building a commercial delivery system.
The investor list is itself a signal of how this category is being financed. Strategic investors from chips, content distribution and cloud infrastructure sit alongside local state capital. For a company selling infrastructure, those shareholders can also be customers or suppliers.
What a "token factory" sells
The company does not describe itself as a model vendor or a compute rental service. Its stated position is a token production service: turning compute resources into a measurable, commercially stable output stream.
The unit is the token, the countable piece of text a model generates. The company's founder and chief executive, Ai Zhiyuan (艾智远), describes quality as a bundle of metrics rather than speed alone: first-token latency, output throughput, concurrency stability, service-level guarantees, model-effect consistency and unit cost.
The engineering approach is reported as "few models, deep optimisation" (少模型、深优化) — supporting a small number of base models intensively rather than many superficially. Named techniques include Prefill-Decode (PD) heterogeneous collaboration, heterogeneous KV-cache conversion and separating storage from compute, all aimed at extracting more inference throughput from domestic chips.
That framing has a commercial logic. If a customer needs a fixed volume of usable output at a predictable cost, the vendor that controls scheduling and software can sell a service contract rather than bill by the hour.
The claims that need time
The operating figures attached to the round are company disclosures, not audited results:
- Daily high-quality token output rose more than 100-fold over six months to a scale above one trillion tokens.
- 2026 revenue reached close to ten times the 2025 full-year figure.
- The business crossed its cost line and now has two repeatable models: self-operated token services and jointly operated token factories.
- More than ten projects for factories producing a trillion tokens per day are under development with partners.
- A token line built with SenseTime is described as having daily trillion-token production and stable consumption capacity.
- A 10,000-card-scale token factory is planned in Hangzhou's Qianjiang Century City, with regional projects in Henan and Hainan.
Partners named in reporting include Parallel Technologies, Hongxin Electronics, Wangsu, Moore Threads, MetaX and Iluvatar CoreX — a mix of domestic GPU vendors and infrastructure providers.
A 100-fold increase from a small base is easier than a ten-fold increase from a large one, and the company has not published absolute output or revenue figures. "Trillion tokens per day" describes capacity, not necessarily sold volume. These are the numbers to watch in a later disclosure rather than to accept as settled.
Why the category is attracting capital
Three market conditions sit behind the round.
First, model training spending has slowed relative to inference demand in the accounts of many Chinese AI companies. Inference is the part that scales with usage, and usage is what enterprise agents generate.
Second, domestic chips are available in volume but software stacks vary in efficiency. A vendor that can make mixed hardware deliver consistent output addresses a real integration cost.
Third, agents (智能体) consume far more tokens per task than chat interfaces. A single coding or analysis task can use hundreds of thousands of tokens, which turns unit cost into a budget line rather than a technical footnote.
Honest limitations
All funding amounts, investor names and operating metrics come from reports by Shanghai Securities News, China Securities Journal and related financial wires, which cite the company's own announcement. Approaching.AI is private and publishes no audited accounts, so revenue and output claims cannot be independently checked here. No valuation was disclosed in the reports reviewed.
What readers can do now
- Distinguish installed capacity from sold volume when you see a "trillion tokens per day" figure; ask which one the number describes.
- Watch whether the next funding announcement repeats the same growth claims with absolute figures attached.
- If you buy inference services, benchmark first-token latency and consistency under sustained load, not a short demo run.
