en.Wedoany.com Reported - On July 13, China's AI Token production service provider Qujing Technology announced a new round of infrastructure construction plans, aiming to increase AI Token production capacity reserves, upgrade its self-developed ATaaS high-efficiency AI Token production service platform, and build AI Token factories targeting leading large models, internet platforms, and regional industrial ecosystems. The project will also expand the access scale of domestic heterogeneous computing power, enabling computing chips of different architectures to enter core production scenarios such as large model inference, agent operation, and enterprise-level generation services.
An AI Token factory is not merely about increasing the number of servers; it involves forming a continuous production chain comprising chips, memory, storage, networks, inference engines, and scheduling systems. When a user submits a model task, the system must complete request classification, model loading, computing resource allocation, prompt preprocessing, first Token generation, subsequent Token continuous decoding, and result return. Resource waiting, cache invalidation, or communication blocking at any stage directly impacts the first Token latency, Tokens per second output, and concurrent processing capacity. The focus of Qujing Technology's expansion is to ensure that new computing power can be stably converted into deliverable Token production capacity, rather than forming a computing resource pool with only hardware scale but lacking inference efficiency.
The ATaaS platform will continue to handle resource scheduling across different computing devices. The platform can decompose large model inference tasks into hardware components such as GPUs, CPUs, memory, and SSDs, preventing all computations from being concentrated on a single type of accelerator card. For clusters with both domestic and non-domestic chips, it must also identify the computing power, video memory, bandwidth, and software adaptation capabilities of different devices before deciding which device should run model layers, operators, or requests.
The deployment of domestic heterogeneous computing power also involves unified management of underlying inference software. Different chips typically use different computing architectures, operator libraries, drivers, and communication components, meaning the same model cannot simply be copied to run on all devices. ATaaS needs to complete adaptation for quantization formats, operator compatibility, video memory allocation methods, and node communication strategies before the model enters the production environment, and distribute tasks across different hardware based on actual workloads. Short-text Q&A, long-context processing, code generation, structured data output, and agent tool invocation have different requirements for computing and storage resources. The system must dynamically select devices based on business load, rather than using a fixed configuration long-term.
Large model inference is typically divided into two stages: prefill and decode. The prefill stage requires concentrated processing of user input context, with high computational density. The decode stage generates Tokens one by one, demanding higher memory access speed, cache capacity, and sustained scheduling capability. Qujing Technology employs a compute-storage collaborative approach to manage KV Cache, storing intermediate states of computed context in memory or storage systems. When multiple requests contain identical or similar system prompts, knowledge base content, and historical dialogues, the system can directly read the cache, reducing repetitive model computations. The technical system previously disclosed by the platform also includes cross-cluster cache sharing, inference pipeline isolation, elastic scaling, and quality monitoring, used to handle cache reuse, resource contention, and service fluctuations in high-concurrency tasks.
The new AI Token factories will target different types of production loads. Leading models require long-term operation of large-scale inference clusters, demanding high concurrency, fast first Token return speed, and output stability. Internet platforms experience significant peak-valley changes in requests, requiring the system to quickly add or release computing nodes. Regional industrial ecosystems may simultaneously handle tasks from government, manufacturing, finance, healthcare, and enterprise office sectors, with different businesses requiring independent models, data permissions, and service quality parameters. ATaaS converts these requirements into underlying computing power configurations and controls resource usage through task queuing, model instance scheduling, and node scaling.
Qujing Technology's previously disclosed ATaaS platform has achieved a daily processing capacity of nearly one trillion Tokens, capable of supporting tens of thousands of artificial intelligence inference demands. The platform adopts a "fewer models, deeper optimization" operational approach, concentrating resources on a small number of production models and conducting specialized tuning around first Token latency, Tokens per second output, structured output, and function call stability. Currently, its inference services are used in businesses related to models such as China's Zhipu GLM and China's Moonshot Kimi.
The funds required for this construction come from Qujing Technology's newly completed Series A financing round. Within six months, the company's cumulative financing amount has exceeded 1 billion yuan. This round was led by Huirong Fund, part of China's Henan Investment Group, with continued participation from existing shareholders including China's Zhenzhi Capital, China's Shangshi Capital, China's Xinglian Capital, China's Shanghai Guofang Innovation, China's Honghui Fund, China's Huakong Fund, and China's Hangzhou Fucheng. The financing amount, investment proportions of each institution, and the company's latest valuation have not been disclosed, but the use of funds has been clearly focused on computing power reserves, ATaaS platform upgrades, and AI Token factory construction.
During the expansion, the quantity of computing power is not the sole indicator of production capacity. AI Token production also requires continuous monitoring of first Token latency, decoding speed, concurrent request volume, hardware utilization, cache hit rate, task failure rate, and structured output success rate, with scheduling strategies adjusted based on model versions and business loads. If a computing cluster only adds accelerator cards without simultaneously expanding high-speed interconnects, memory, storage, and inference software, the new hardware may become idle due to communication waiting or data movement. Qujing Technology's construction places domestic heterogeneous computing power and the ATaaS platform within the same production system, aiming to have multiple chips jointly undertake inference tasks and convert dispersed computing resources into measurable and schedulable Token production capacity.
Currently, Qujing Technology has not disclosed the scale of new computing power, chip models, number of AI Token factories, project locations, or initial production timelines. The confirmed construction content includes expanding capacity reserves, upgrading the underlying inference system, increasing domestic heterogeneous computing power production loads, and building dedicated Token production capabilities for leading models, internet platforms, and regional projects.
