Global Industry Home Beyond Compute Stacking: Approaching.AI Builds a Domestic AI Token Factory

Beyond Compute Stacking: Approaching.AI Builds a Domestic AI Token Factory

As multimodal and AI agent technologies advance rapidly, demand for inference compute is growing quickly, and compute cost and reliable supply have become key constraints on scaling AI across the industry. Against this backdrop, Zhiyuan Ai (艾智远), Founder and CEO of Approaching.AI (趋境科技), drew on first-hand engineering practice in building a high-quality AI token factory to argue at the CITICS Capital Market Strategy Conference 2027 that the industry’s benchmark is shifting from the amount of compute resources deployed to the number of high-quality AI tokens each unit of compute can produce. What enterprises really need is not the peak performance of a single device, but the ability of underlying compute resources to be steadily converted into stable, usable, cost-controlled high-quality AI tokens.

Zhiyuan Ai (艾智远), Founder and CEO of Approaching.AI

Zhiyuan Ai noted that in real business environments, enterprises must handle response latency, output speed, concurrency capacity, service stability, output quality on specific tasks, and unit cost at the same time — and these factors carry different weights for different businesses. The same pool of resources can yield different levels of token output efficiency and different delivery outcomes depending on model adaptation, the inference system, and operating practices. With hardware costs hard to change quickly, simply adding hardware cannot solve the twin problems of output speed and cost under high concurrency. Inference system optimization has therefore become an important path to overcoming the cost bottleneck and improving the efficiency of high-quality AI token production.

With the path defined, how is this optimization put into practice? Approaching.AI dynamically decomposes model computation tasks and achieves heterogeneous Prefill–Decode (PD) coordination: resources with greater compute capability carry the high-compute-density Prefill phase, resources with larger memory bandwidth handle the Decode phase, and CPUs, GPUs, memory, storage, and high-speed networking are brought into a unified inference system. Task routing, cross-device collaboration, and continuous operations ease the constraints imposed by single-chip performance. At the same time, the company tiers intermediate states such as the KV Cache and reuses them on demand through a “memory-for-compute” (以存换算) approach, reducing repeated computation. In parallel, it uses “virtual–physical isomorphism” (虚实同构) to simulate runtime behavior across different model, device, and configuration combinations before deployment, and then continuously calibrates those simulations with real-world data after launch. As a result, token production no longer depends solely on the performance of a single device; it becomes a system-level engineering effort in which compute, memory, storage, and networking work together.

Building on this practice, Approaching.AI released the High-Quality AI Token Factory Solution for Domestic Heterogeneous Inference with its ecosystem partners, closing the industrial loop across models, inference, and chips. The solution lets different types of compute resources play to their strengths and has achieved a several-fold increase in effective token output for the same hardware investment in real production environments. It is understood that its domestic heterogeneous PD solutions have been fully deployed in production environments, where they stably support trillion-parameter ultra-large models and continuously handle high-concurrency real-world traffic. This means system-level collaboration is not confined to laboratory optimization; under real workloads, it can be steadily converted into usable token production capacity.

This article is sourced from the internet. http://news.ystp.com.cn/home/251.html

发表回复

您的邮箱地址不会被公开。 必填项已用 * 标注

联系我们

联系我们

15521078730

在线咨询: QQ交谈

邮箱: 1039585111@qq.com

Working hours: Monday to Sunday, 9:00-19:30
关注微信
微信扫一扫关注我们

微信扫一扫关注我们

关注微博
Back to top