Author｜Ren Tianqin

Editor｜Chen Xiaoran

The AI industry’s competitive landscape has shifted significantly, moving beyond traditional comparisons of large model parameters. The new battleground now lies in token production capacity during the inference stage.

On July 13th, Beijing Qujing Technology Co., Ltd. (Qujing Technology), a Tsinghua University-affiliated company, completed its Series A financing round, raising over 1 billion yuan in cumulative funds within just six months. This achievement was made possible through collective investments from state-owned capital and seasoned venture capital firms.

As a service provider focused on building a “High-Quality AI Token Factory,” Qujing Technology is leveraging its technological and commercial data to demonstrate the arrival of a pivotal moment in the sector.

Intensive Capital Influx

This financing round was led by the Huirong Fund of Henan Investment Group, with strong continued support from existing shareholders such as Zhenzhi Capital, Shangshi Capital, Starlink Capital, Shanghai Guofang Innovation, HighLight Capital, Huakong Fund, and Hangzhou Fucheng, all of whom made oversubscribed follow-on investments.

Completed merely two months after the previous Pre-A round, this latest funding achievement was driven by proactive investor outreach, resulting in oversubscription.

Analyzing the investor composition reveals distinct strategic alignments: frontline financial venture capital firms prioritize sector growth potential, local state-owned enterprises aim to meet regional smart computing center transformation needs, and industrial capital focuses on technology implementation for domestic computing power adaptation.

Unlike conventional financial investment approaches, lead investor Henan Investment Group has gone beyond mere capital contribution. It has partnered with Qujing Technology to establish a large-scale AI Token factory capable of daily production in the trillions, marking a shift towards industrial co-construction.

The confidence underpinning this capital influx is rooted in the company’s impressive operational data.

Since the 2026 Spring Festival, Qujing Technology has more than tripled AI token production efficiency per unit of computing power, with total capacity surging over 30-fold. Notably, daily production of high-quality AI tokens for a leading trillion-parameter large model has consistently exceeded one trillion.

Commercially, revenue generated in June 2026 alone surpassed the entire year’s earnings of 2025. Some mature business segments have already crossed the cost break-even point, addressing a common challenge faced by AI infrastructure companies—revenue growth without profitability.

Delving into the team’s background, Qujing Technology originated from the High-Performance Computing Institute at Tsinghua University’s Department of Computer Science. Chinese Academy of Engineering academician Zheng Weimin serves as Chief Scientific Advisor, while Professor Wu Yongwei holds the position of Chief Scientist.

CEO Ai Zhiyuan earned his PhD in High-Performance Computing from Tsinghua, and the core management team combines expertise in scientific research, capital operations, and industrial implementation.

Qujing Technology has also successfully commercialized its technology through equity transfers from Tsinghua. Its research-oriented technical foundation is a key factor behind the bold capital commitments it has received.

Technological Route Breakthrough

While the industry predominantly focuses on MaaS model marketplaces and accumulating computing power and model quantities, Qujing Technology has forged a differentiated path since its inception in 2023. It targets the inference sector with a “Fewer Models, Deeper Optimization” strategy, building a token industrial production system centered on the ATaaS platform.

First, it is essential to clarify core concepts. Qujing Technology’s definition of “high-quality AI tokens” extends beyond simple model output metrics. These tokens represent the minimal production units that integrate computing power, models, and business applications. They must simultaneously meet stringent requirements for low first-token latency, high concurrency, stable output, structured invocation, and cost control, all while maintaining performance under sustained high-load production. Efficiency gaps between different industry solutions can vary by tens of times.

To overcome these challenges, Qujing Technology has developed proprietary technologies, including “Full-System Heterogeneous Collaboration,” “Compute-for-Storage,” and “Virtual-Physical Isomorphism.”

The ATaaS platform deeply integrates underlying computing power with model inference systems at the lower level, while seamlessly connecting with real enterprise business scenarios at the upper level. It orchestrates full processes, including model partitioning, memory management, cross-cluster scheduling, and self-healing failure recovery.

The “Fewer Models” approach concentrates on refining top large models that dominate market demand, abandoning the superficial appeal of compatibility with hundreds of models. The “Deeper Optimization” strategy focuses on two core metrics—token output efficiency per unit of computing power and effective computing power utilization rates—to maximize asset efficiency.

This logic reflects a clear commercial judgment: enterprise customers pay for business outcomes, not the number of compatible models.

Simultaneously, Qujing Technology actively engages with open-source ecosystems. It leads the KTransformers open-source project in collaboration with Tsinghua’s team, partners with Tsinghua University, Kimi (Yuezhian'anmian), Alibaba Cloud, Ant Group, and others on the Mooncake project, and participates in global mainstream AI inference communities like vLLM and SGLang.

This open-source approach not only refines technology but also builds industry trust, forming a closed loop of “technology R&D - open-source validation - enterprise implementation.”

Market Share Competition

From an industry perspective, the industrialization of AI tokens has become an irreversible trend.

According to the National Data Bureau, China’s daily token invocation volume exceeded 140 trillion by March 2026, marking over a 1,000-fold increase from 100 billion in early 2024.

Goldman Sachs projects that China’s daily AI model token consumption will surge from 350 trillion in 2026 to 4,600 trillion by 2030, with API and subscription revenue climbing from 35 billion yuan in 2026 to 879 billion yuan in 2030.

IDC forecasts that China’s AI server market will reach 350 billion yuan by 2026.

Inference computing demand is comprehensively surpassing training needs. Chinese Academy of Engineering academician Wu Hequan notes that global inference computing will account for 70-80% of total AI computing workload by 2026.

IDC expects inference workloads to reach 73% by 2028. Industry veterans estimate that AI inference computing demand reached 4-5 times training demand by 2026, with inference computing rental prices surging nearly 40% in six months.

Facing this massive market, Qujing Technology has developed two primary monetization paths:

Under the direct sales model, Qujing Technology leases computing power to produce and sell AI tokens to leading model companies, internet platforms, and large enterprises. The co-construction model provides overall design, construction, and joint operations for AI token factories to customers with existing computing resources.

The latter precisely meets the transformation needs of regional smart computing centers shifting from “selling computing power” to “selling high-quality AI tokens.”

On the technical front, Qujing Technology has invested heavily in R&D, proposing solutions like “Domestic Prefill-Decode Heterogeneous Collaboration” and “High-Performance Heterogeneous KVCache Conversion” to enable formal production deployment of domestic computing power in high-standard scenarios.

However, this booming sector faces multiple hidden challenges:

The industry landscape shows rapidly intensifying homogenized competition. In June 2026, Silicon Flow, another token factory contender, secured over 2 billion yuan in Series B financing, while internet giants’ self-developed inference optimization systems continue open-sourcing, further dividing market share.

Cost pressures remain significant, with storage hardware prices doubling in Q1 2026, directly raising comprehensive expenses for computing equipment procurement and daily operations.

At the implementation level, profitability logic remains unproven at scale. Successful profitability in single projects does not guarantee nationwide replication, as management complexity in cluster operations and quality control escalates dramatically with expansion.

Moreover, the sector grapples with capital-driven phased bubbles, as massive capital inflows risk triggering blind capacity expansion unrelated to actual demand, creating long-term supply-demand mismatches.

Thus, while Qujing Technology stands at the industry’s forefront through early technological accumulation and commercial achievements, its long-term market position hinges on whether its token factory model can achieve nationwide deployment while continuously deepening technological innovation to raise computing power-to-token conversion efficiency.