Approaching.AI Wakes Dormant Compute into Tokens

Compute has been purchased, yet tokens cannot be produced—this is one of the sharpest paradoxes in today’s AI infrastructure. Average GPU utilization at domestic AI computing centers has long been low, with large numbers of accelerators ‘sleeping’ in data centers; on the other side, daily token usage has climbed from hundreds of billions in early 2024 to around 100 trillion tokens. What enterprises truly need is no longer abstract compute, but the high-quality tokens that compute can stably produce. Procuring compute is a matter of addition; converting compute into mass-producible and priced token capacity is a multiplication problem. Approaching.AI (趋境科技), a high-quality AI token production service provider, has chosen precisely to tackle this multiplication problem.

Its approach does not pursue broader model coverage or simply pile on hardware performance; rather, it organizes ‘dormant’ heterogeneous compute into an industrially operable token production line. Zhiyuan Ai (艾智远), Founder and CEO of Approaching.AI, defines the token factory as ‘a production system that organizes models, hardware, inference technology, and operational services to continuously deliver measurable and guaranteed high-quality AI token capacity.’ This definition clarifies the key to awakening ‘dormant compute’: it is not simply adding hardware, but enabling models, inference, and chips to coordinate under real production loads, and evaluating optimization outcomes based on the effective token output per unit of resource and the effective production time.

图片2.png

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

Approaching.AI’s deployed ‘High-Quality AI Token Factory Solution for Domestic Heterogeneous Inference’ moves this approach into production-grade validation. Relying on the ‘fewer models, deeper optimization, higher quality’ approach, Approaching.AI focuses on a small number of high-productivity models and uses its industry-first full-system heterogeneous collaborative inference technology as a key underpinning to achieve precise coordination of heterogeneous resources: compute-intensive resources handle the high-compute-density Prefill phase, while resources with greater memory bandwidth handle the Decode phase; the solution brings the CPUs, GPUs, memory, and high-speed networks across the cluster into unified inference scheduling, safeguarded by accuracy preservation, cross-GPU coordination, and millisecond-level fault tolerance, stably producing higher-quality, more economical AI tokens in real production environments.

The feasibility of a technical path must ultimately be demonstrated through a replicable build-and-operate model. At present, Approaching.AI has formed a replicable high-quality AI token factory build-and-operate model: it has production projects with daily capacity of several trillion tokens, strategic partnerships with regional industry ecosystems and leading clients, and deployments under construction. Year to date, the company’s average AI token production efficiency per unit of compute has risen month by month; its high-quality AI token daily output has grown more than a hundredfold, to over one trillion tokens; and over the same period its revenue has approached ten times last year’s full-year total, with its overall business crossing the break-even point.

This model is also accelerating its spillover into project delivery. Within half a year, Approaching.AI has partnered with industry partners to advance more than ten AI token factory projects with daily output exceeding one trillion tokens, promoting the scaled application of domestic compute, and—together with local state-owned capital, industry leaders, and ecosystem partners—building a new type of AI token supply network that connects compute infrastructure, large models, and industry applications. As projects continue to roll out, the constraint imposed by limited high-end compute supply is gradually being broken; domestic AI infrastructure will gain an independent and controllable, scalable, and replicable path to inclusive access to compute—the ‘dormant compute’ in data centers truly becoming mass-producible high-quality AI tokens.