AIBRIEF
Approaching.AI (趋境科技) raised the round with a Henan provincial government linked anchor; per machine token efficiency (the unit AI models read and write during inference) up 3x and capacity up 30x, per the company's own release.
A Chinese AI infrastructure startup, Approaching.AI (趋境科技), closed a Series A that pushed its cumulative six-month funding past 1 billion yuan (about $140 million), the company said in a July 13 release re-reported by QbitAI and other Chinese financial outlets.
Lead investor 河南投资集团汇融基金, a Henan provincial government-linked fund, anchored the round, per 东方财富. Existing shareholders 真知资本, 尚势资本, 星连资本, 上海国方创新, 弘晖基金, 华控基金, and 杭州福成 over-subscribed, the company said.
Approaching.AI sells what it calls "Token-as-a-Service" (ATaaS), positioned against the more common "Model-as-a-Service" pitch. The company argues that "tokens," the chunks of text a model reads and writes during inference, are the production unit that matters, and that inference is won by holding first-token latency, sustained throughput, concurrency, quality, and unit cost simultaneously under real load.
Approaching.AI credits model splitting, domestic Prefill-Decode disaggregation, and KVCache conversion across heterogeneous domestic compute, run under a "few models, deep optimization" strategy, the company said. Since Spring Festival 2026, per-machine token-production efficiency is up 3x and total capacity up 30x, with one unnamed trillion-parameter model running at a trillion tokens per day, according to the release on Sohu.
All growth figures and the marquee client name come from Approaching.AI's own announcement; the release does not include independent benchmarks or third-party validation.