The AI industry is now entering a new “inference-driven” cycle. As multimodal and AI agent technologies evolve, demand for inference compute is growing exponentially. The industry has long been caught in a familiar arithmetic trap: advanced compute is expensive and in tight supply, while standard compute is constrained by memory bandwidth and engineering coordination bottlenecks, making it difficult to align compute investment with effective token output. Relying solely on hardware upgrades yields diminishing marginal returns, and token economics has become an unavoidable barrier to deploying large models at scale.
For Approaching.AI, a provider of high-quality AI token production services, the underlying value innovation lies precisely in converting compute into effective tokens. Rather than pursuing mere hardware upgrades, the company reshapes the input–output logic of compute through system-level heterogeneous collaborative engineering. This approach has been translated into a deliverable solution: at the 2026 “AI+” Conference, Approaching.AI, together with Zhipu AI and Moore Threads, released the “High-Quality AI Token Factory Solution for Domestic Heterogeneous Inference.”
Drawing on Approaching.AI’s industry-first full-system heterogeneous collaborative inference technology, the solution achieves precise orchestration of heterogeneous resources: compute-intensive resources handle the high-compute-density Prefill phase, while resources with greater memory bandwidth handle the Decode phase. Approaching.AI brings CPUs, GPUs, memory, and high-speed networks across the cluster into a unified inference scheduling framework. Underpinned by end-to-end preservation of model accuracy, fine-grained cross-GPU collaborative networking, and a millisecond-level fault-tolerance mechanism with zero data loss, it overcomes single-chip performance bottlenecks and the engineering barriers to heterogeneous coordination, and leverages system-level engineering to substantially reshape token economics.
The value of this system-level engineering capability extends well beyond isolated efficiency gains. Using the AI token factory as its implementation vehicle, the solution is the first to close the industrial loop across models, inference, and chips. It transforms standard compute into AI capacity with production-grade stability, enabling standard compute to be deployed in enterprises’ core production workflows—more than doubling effective token output for the same hardware investment while substantially reducing overall costs. This is underpinned by a “fewer models, deeper optimization, higher quality” technical approach: anchored by three globally pioneering foundational technologies—heterogeneous Prefill–Decode (PD) coordination, high-performance heterogeneous KV cache conversion, and heterogeneous computing fusion—it conducts end-to-end deep optimization across model adaptation, inference execution, KV cache management, resource orchestration, and service operations and maintenance (O&M); uses effective token output per unit of resource as the key metric; and builds an intelligent, low-latency, and highly stable delivery foundation.
The viability of this technical approach is being validated by operating metrics. According to the company’s official disclosures, Approaching.AI has raised a cumulative RMB 2.4 billion in funding this year—reflecting market recognition of this approach on the capital side. Its high-quality AI token production capacity has grown more than tenfold in half a year; revenue for the year to date has reached nearly ten times last year’s full-year total; and its overall business has crossed the break-even point—validating its capacity for delivery at scale from the operational side. This demonstrates that system-level optimization at the inference layer is the key lever for unlocking the value of domestic compute, and it establishes a reusable economic model for deploying large models at scale.
From the perspective of industrial collaboration, within the collaborative framework of “Beijing Models, Beijing Inference, and Beijing Chips,” this approach uses inference acceleration to address the critical gap in “Beijing Inference efficiency,” enabling diverse compute resources to enter real production scenarios more broadly and strengthening the engineering foundation for a closed-loop domestic full-stack AI ecosystem.
