July 13, 2026. Qujing Technology closed its Series A funding round, led by Henan Investment Group's Huirong Fund. Total raised across all rounds in just six months: over 1 billion yuan.
That speed is unusual, even for a sector used to large fundraises. If you've been tracking the China AI startup scene, this one is worth slowing down on. The Qujing Technology AI Token factory Series A funding Tsinghua University story isn't just a milestone - it's a signal about where institutional capital is moving, and why inference infrastructure has become the most interesting part of the AI stack.
The Tsinghua University Team Behind Qujing Technology's Series A
Qujing Technology launched formally in December 2023. Its founding team came directly from Tsinghua University's Institute of High Performance Computer Science.
CEO Ai Zhiyuan holds a PhD from that institute. CTO Chen Xianglin shares the same background. Academician Zheng Weimin - a member of the Chinese Academy of Engineering and a leading figure in computer system architecture - serves as chief scientific advisor. Professor Wu Yongwei from Tsinghua's Computer Science department is chief scientist. Associate Professor Zhang Mingxing co-initiated the company and has stayed hands-on with technology strategy throughout.
This Tsinghua-linked AI funding model - research team spins out, completes formal university technology transfer, builds a commercial company - is the same path that produced Zhipu AI. Both are now held up as flagship examples of Tsinghua's commercialization pipeline.
In March 2026, Dr. Wu Wenjie joined as President. PhD in Finance from the University of Hong Kong. Deep experience in corporate finance and global strategy. That hire signals a deliberate shift from "research team with a business attached" to "business backed by serious research." It matters for how investors read the company.
The Contrarian Bet Nobody Was Making in 2023
Here's where the story gets interesting.
When Qujing's founders set the company's direction, training dominated the AI infrastructure conversation. Building foundation models - creating larger, smarter "brains" - was where the prestige and the capital were. Almost every AI infrastructure startup in China was chasing training workloads.
Qujing went the other direction. They chose inference.
The distinction sounds simple. Training creates a large model. Inference runs it - constantly, at scale, for real users and real enterprise applications. CEO Ai Zhiyuan described the logic directly: "Training is the cost item, reasoning is the profit item." The judgment was that inference generates sustained economic value every single second - every query, every API call, every enterprise workflow.
The team saw the AI computing infrastructure opportunity in enterprise inference as real and largely unpursued. So they built for it before the market caught up.
Three years later, it did.
Inside Qujing Technology's AI Token Factory Infrastructure
Running a large language model at enterprise scale is harder than it sounds. You need low initial token latency - the time before that first output character appears. High concurrency, because enterprise deployments mean thousands of simultaneous requests. Stable output quality over time, not just in a controlled demo. Structured results, function call support, and compatibility with trillion-parameter models. All while keeping unit costs low enough that businesses can actually profit.
Most infrastructure companies get one or two of these right. The genuinely hard problem - and this is hard - is achieving all of them simultaneously, under real production loads, sustained over months of operation. Qujing's analysis shows that different capability combinations can produce production efficiency differences of 10 times or more. That gap directly affects whether enterprise AI deployments are profitable or not.
Qujing's engineering answer combines three proprietary techniques: "system-wide heterogeneous collaboration" (coordinating different hardware types across the entire compute environment), "storage-based computing" (restructuring how storage and compute interact to eliminate throughput bottlenecks), and "virtual-physical isomorphism" (a software architecture enabling efficient scaling across physically diverse cluster configurations).
Since early 2026, that engineering stack has driven a 30-fold increase in high-quality AI token output from equivalent compute. That's a production figure, not a benchmark.
For context on how this fits the broader AI agent infrastructure investment landscape, the scale of capital moving into inference infrastructure now matches - and in some cases exceeds - what training attracted two years ago.
Token as a Service: What Enterprise Customers Actually Buy
Qujing packaged their capability into a model called Token as a Service, or TaaS. Their platform, ATaaS, converts raw compute hardware investment into standardized, high-quality token output.
Think of it like a utility. You don't build a power plant to run your business. You connect to the grid. TaaS applies that same logic: enterprises don't need to hire inference engineers, manage cluster failures, or tune model parameters. They connect to ATaaS and get stable, production-ready token output.
The domestic AI chip supply chain complexity, heterogeneous hardware management, and optimization overhead all sit inside Qujing's platform. Customers get the tokens. Qujing absorbs everything else.
The product strategy is deliberately narrow. Ai Zhiyuan's framing: rather than building a supermarket stocked with every available model, Qujing wants to be a boutique specialty store. A small number of leading models, optimized extremely deeply. The market data backs it up: "less than 10% of the top models in China occupy the vast majority of the AI Token market." Broad compatibility isn't the advantage here - depth on the right models is.
How Qujing Technology Raised Over 1 Billion Yuan in Six Months
The funding timeline moved fast.
February 2026: Angel++ round closes. Parallel Technology invested.
May 2026: Pre-A closes, co-led by Xingling Capital and Huakong Fund. Honghui Fund, Tianhao Energy, Shangshi Capital, Tianjin Renai Hongsheng, and Hangzhou Fucheng follow on. Hillhouse Capital (GL Ventures) continued adding to its existing position.
July 13, 2026: Series A closes, led by Henan Investment Group's Huirong Fund. All major existing shareholders - Zhenzhi Capital, Shangshi Capital, Xinglian Capital, Shanghai Guofang Innovation, Honghui Fund, Huakong Fund, Hangzhou Fucheng - chose to oversubscribe. Total across all rounds: over 1 billion yuan. The next round is reportedly already underway.
That last detail is the most meaningful one. Existing shareholders oversubscribing means the people with the clearest view into Qujing's actual operations chose to put in more at higher valuations. They're not buying the pitch. They're buying the data.
In the context of high-tech startup funding dynamics across China's tech sector, insider conviction at this scale, compressed into six months, is genuinely uncommon.
The Operating Numbers Behind the Round
China's National Bureau of Data Science reports daily AI token usage hit 140 trillion tokens by March 2026. A 1,000-fold increase in two years. The China AI sector’s explosive growth driving that demand is structural - enterprise AI deployment across industries, not just consumer apps.
Qujing's internal metrics since Spring Festival 2026 tell an equally striking story:
- Token production efficiency per unit of compute: up more than 3 times
- Total high-quality AI token output: up more than 30 times
- June 2026 revenue alone: exceeded all of 2025's total revenue combined
Their flagship deployment - a trillion-parameter model - is now producing trillions of high-quality AI tokens per day. Not in a lab. In production.
Why Capital Is Shifting From Training to Inference Right Now
There's a larger reallocation happening in AI infrastructure investment. Training captured most of the capital from 2022 through 2024. Now, with foundation models increasingly mature and enterprise adoption accelerating, value is migrating toward whoever can run those models profitably at scale.
AI-driven industry transformation across verticals is compressing procurement timelines. Companies that would have evaluated AI deployments over years are now moving in quarters. That creates immediate, sustained demand for production-grade inference infrastructure.
The China AI industry growth forecast consistently identifies inference as the fastest-growing AI infrastructure spending category. And recent changes to China tech startup listing rules increasingly favor companies with recurring enterprise revenue over pure model developers still searching for monetization - a shift that directly benefits TaaS-model platforms like Qujing.
This pattern is familiar from earlier tech infrastructure cycles. Once the foundational technology stabilizes, commercial value migrates toward efficient deployment. In AI, that migration is happening right now.
Qujing Technology vs. Zhipu AI, SiliconFlow, and the Broader Field
The comparison to Zhipu AI comes up often, and it's partially useful.
Both completed technology transfer from Tsinghua University. Both built founding teams from the Institute of High Performance Computer Science. Both are now considered flagship examples of Chinese AI companies global rise through Tsinghua's commercial spinout pipeline.
But the business models split sharply after that. Zhipu AI builds and commercializes its own large models. SiliconFlow operates as a broad inference platform with wide model support. Qujing is narrower and deeper - it focuses exclusively on inference infrastructure, deliberately limits model coverage to high-value targets, and doesn't train or sell models at all.
That positioning matters for enterprise relationships. You can't have a clean partnership conversation with a vendor that's also building the models you'd otherwise use. Qujing sidesteps that entirely.
How this plays out in long-term valuation metrics and AI unicorn rankings will become clearer as the inference infrastructure segment matures. But the revenue trajectory is already making the argument.
What Qujing Technology's Series A Round Actually Signals
The Qujing Technology AI Token factory Series A funding and Tsinghua University founding team together represent something more than a single company's fundraising milestone.
Training built the foundation models. Inference has to run them - at scale, profitably, for enterprises building real products. That transition is the defining moment in AI infrastructure right now, and it's where serious institutional capital is reallocating. Henan Investment Group and Hillhouse Capital don't oversubscribe rounds because of pitch decks. They do it because operating metrics are real.
Thirty times more token output from the same compute base. Monthly revenue exceeding a full prior year in a single month. Informed existing shareholders increasing their bets rather than managing their positions.
Qujing Technology's AI Token factory built for a market that wasn't obvious in 2023. Now the market has arrived. The Tsinghua University-rooted team has 1 billion yuan and a Series A to build further into it - and the next round is apparently already forming.
Whether this becomes the dominant inference infrastructure layer in China, or faces structural pressure from hardware-native approaches down the road, those are real questions worth watching. But right now, the operating data points in one direction. And so does the money.