Huawei Pangu’s Second-in-Command Joins Startup, Valuation Rises 10x in Three Months

If you're a 3D printing enthusiast, you've likely had this experience: wanting to design a layered storage box for your headphones, only to find no pre-made models in the library that fit your exact dimensions.

Existing solutions on the market each have significant drawbacks. Manual modeling is low in productivity and high in cost, failing to keep up with the rapidly growing demands from both industrial and maker communities. Generic AI-generated 3D models, on the other hand, are little more than “hollow shells”—lacking dimensional annotations and assembly logic; changing a single hole spacing requires redrawing everything entirely. In simulation environments, objects pass through each other, and when exported for printing, the parts fail to align properly.

3D printing is undoubtedly one of the most watched growth sectors globally, yet structured, editable physical 3D data remains extremely scarce. Into this industry vacuum emerges a young company—BitInf—specializing in AI-generated vector 3D models.

Not long ago during WAIC, CCTV featured this startup, founded just over a year ago, focusing on its all-90s team. A year ago, Li Shiqi, with years of overseas AI-to-consumer startup experience, keenly identified a gap in physical 3D data and partnered with Huawei Pangu Large Model Dr. Zhou Pingyi and other members of the frontier team founded BitInfinity.

Building a physical 3D data flywheel for the real world, Bit Infinite has carved out a unique path: using 3D printing and a global community of millions of Makers as its entry point, its proprietary Arko-T model has achieved state-of-the-art performance worldwide on text-to-physical-3D tasks, garnering over a million impressions on overseas platforms.

Investors learned that, following the official announcement of its seed round in April, BitInfinity has since successfully completed its angel and angel-plus funding rounds, attracting prominent institutions such as Zhuoyuan Asia, Rongyi Investment, Puhua Capital, and Yide Capital. “In less than three months, our valuation increased tenfold,” said Li Shiqi, CEO of BitInfinity, when met by investors at ModSpeed Space in Shanghai in July.

Whoever can quickly close the gap in physical 3D data will hold the key to entering the era of physical AI,” Li Shiqi assessed, noting that the GPT moment for generative physical 3D is just around the corner.

In June this year, "Godmother of AI" Li Fei-Fei published a lengthy article, breaking down world models into a three-layer structure: Renderer, Simulator, and Planner.

She believes that simulators are the most impactful component, producing a structured representation that faithfully captures the state of the world in terms of geometry, physics, and dynamics—serving as both a foundation for computation and an interface for interaction between humans and computer programs. Robotics training, autonomous vehicle testing, architectural visualization, engineering simulation, and even drug discovery all rely on simulators as this core platform.

Around the same time, Li Feng, founding partner of Fengru Capital, made a similar observation: the reason world models and physics models have become the biggest investment focus in embodied AI is because “there is no data”—humans have never systematically accumulated data containing physical quantities and the laws governing interactions with the physical world.

Earlier, Li Shiqi had already recognized the data vacuum in cutting-edge industries and felt an urgent need to build a self-sustaining data engine: “We need an engine capable of continuously generating data from ‘real-world intentions and interactive feedback.’”

Looking back, only a few data engines have truly achieved scalable flywheels: Google built its first-generation visual data engine through large-scale text annotation across the web, while Tesla created its second-generation FSD data loop by collecting road condition data from vehicles. Both share the commonality of using real-world interactions from vast numbers of ordinary users as their data source.

Li Shiqi turned his attention to another group—the 500 million makers and 100 million active 3D printing users worldwide. With BitInfinity’s Sparkoh.ai, users only need to describe their ideas in natural language or sketch a simple drawing, and the system automatically generates printable 3D physical models. Each modification and feedback from users flows back into the model, making each subsequent generation more accurate. In this way, BitInfinity gradually evolves from a tool into a living, dynamic physical 3D data engine.

BitInfinity Product Demo Video

A team of AI experts quickly came together. Co-founder Zhou Pingyi, 36, holds a Ph.D. in Engineering from Wuhan University and previously worked at Huawei. Pangu Large Model Original founding member and second-core researcher of the second-generation Pangu trillion-parameter MOE model. Globally, only about 3,000 people have led the complete training of trillion-parameter models; 70% are at overseas tech giants, 20% at research institutions, and full-time entrepreneurs are extremely rare—Zhou Pingyi is one of them.

Zhou Pingyi, Co-founder of BitInfinity

CEO Li Shiqi brings years of overseas AI-to-consumer startup experience, having successfully launched two AI tools each with over a million monthly active users abroad, and possesses deep expertise in operating creator communities in Europe and the U.S. CTO Chen Litaoo has six years of full-stack experience in Huawei’s simulation and post-training technologies, with a strong grasp of both physical constraint implementation and large model deployment.

In its early stages, BitInfinity operated largely "under the radar." Dr. Zhou Pingyi led his team in developing a global AI-generated vector 3D specialized model, while quietly completing three rounds of financing. In April this year, the seed round came to light, led by Zhuoyuan Asia, a Tsinghua-affiliated investor with a track record of backing prominent AI projects such as Shengshu Technology and Jiliu Technology. This was followed by the angel and angel-plus rounds, bringing total funding to RMB 45 million, with investors including Rongyi Capital, Puhua Capital, and Yide Capital.

A noteworthy signal is that Rongyi Investment previously invested in embodied robotics companies such as Unitree Robotics and Galaxy General; Puhua Capital also invested in Zhi Square, an embodied AI unicorn. Now, both have independently turned their attention to BitInfinity—positioning themselves at the critical juncture of physical 3D data.

The logic isn't complicated. The humanoid robotics sector has already produced leading core companies like Unitree and Galaxy General, while rapidly expanding markets such as 3D printing, industrial design, and general world models are growing at high speed. However, the standardized physical 3D data infrastructure that supports industry-wide iteration remains largely absent. Whoever can secure a position in this area has the opportunity to become a super player in the next generation of AI infrastructure.

This might be a new FOMO. For investors, physical 3D data is the crucial missing piece in reaching the next paradigm—the strategic positioning is self-evident. Just as BitInfinity emerged from underwater, its valuation surged tenfold within three months. It has been revealed that the company is currently raising a new round of funding, with a noticeable increase in investors seeking to connect.

Bit Infinite has taken a somewhat different path.

Most 3D generation on the market simply adds AI features on top of traditional modeling software, still centered around software commands and manual operations. BitInfinity has reimagined the entire design process from the ground up for AI-native interaction—enabling AI to understand spatial geometry, structural relationships, and motion principles, directly transforming descriptions into editable, manufacturable physical 3D models.

The underlying logic of this system is a parameterized code generation model that inherently carries structural information and editable attributes, making it directly usable for 3D printing, robotics simulation, and Physical AI training.

In the context of 3D printing, users describe their needs in natural language, and the system automatically decomposes the task, generates code, performs execution verification, and iterates through multiple rounds of refinement. This precisely aligns with makers' most authentic needs—they don’t want “beautiful models,” but “functional models,” such as drawers with slides that open smoothly, where every component’s dimensions and coefficients of friction withstand physical validation.

Underpinning all of this is Arko-T, BitInfinity’s proprietary, specialized physical 3D large model.

Most text-to-3D models generate shapes, while Arko-T, with 4 billion parameters, generates editable designs. On the Text2CAD-Bench benchmark, Arko-T was comprehensively evaluated against seven leading general-purpose large models across 12 metrics, ranking first in eight metrics and second in three others, with inference costs仅为ChatGPT, DeepSeek One-tenth or even less of general-purpose large models.

Data is another moat. Zhou Pingyi explained that BitInfinity’s training data comes from two sources: first, its own proprietary end-to-end data engineering involving tens of millions of data points; second, the free modeling tool sparkoh.ai, which is used by hundreds of millions of DIY enthusiasts worldwide. Every modification and adjustment made by users flows back into the model, automatically enabling continuous training and optimization.

The barrier to entry for this closed-loop replication system is high. In just one year, BitInfinity has successfully established a scalable data production pipeline. More importantly, user growth is accelerating. As Sparkoh.ai’s reputation spreads across global Makers communities, an increasing number of users are generating printable models using natural language—each use further fueling the data flywheel.

Even though big companies may not lack computing power, building a mature overseas creator operations team and accumulating a large base of active users is no easy task. BitInfinity, however, already has a first-mover advantage and is continuously providing genuine manufacturing feedback data.

Bit Infinite's proprietary Harness Agent scheduling framework bridges the final gap from concept to physical product. Traditional industrial modeling software features outdated interfaces and fragmented workflows; Harness establishes a fully automated closed loop: decomposing user requirements, generating structured models, automatically verifying dimensions and assembly issues, and iteratively improving不合格 designs—all without any human intervention.

On the commercial side, BitInfinity employs a dual-track strategy: for C-end users targeting overseas Makers, the Web Agent uses a token-based subscription billing model; for B-end clients such as industrial design, robotics simulation, and Physical AI development companies, API access is provided, enabling customers to batch-generate simulation assets compliant with the OpenUSD standard.

A bidirectional, parallel, and mutually reinforcing commercial cycle is becoming clear—C-end interaction data accumulates and continuously refines the model; the improved model generates higher-quality physical 3D data, which in turn supports B-end robotics companies.

Just one year after its founding, BitInfinity already holds several key "tickets."

Not long ago, the team joined NVIDIA’s Inception startup acceleration program, and BitInfinity is the youngest member on the Zhipu official ecosystem partner list. In addition, the company has moved into Shanghai Mousu Space and is eligible for specialized AI computing power subsidies.

Within China’s startup ecosystem, it is rare to find companies that simultaneously hold NVIDIA Inception membership, official ecosystem partnership status with Zhipu, and access to computing power support through Mosu Space. The combined resources from these three ecosystems provide tangible support in reducing compute costs, adapting large models, and accelerating commercial deployment. For an early-stage team founded just one year ago, these advantages translate into real cost savings and efficiency gains.

But in Li Shiqi’s view, the deeper significance goes beyond this. These resources are positioning BitInfinity at a more critical juncture, bringing them one step closer to their “GPT moment” in generative physical 3D.

Looking back at the internet era, the vast amount of textual data gave rise to large language models. The emergence of ChatGPT has leveled the creative barrier—ordinary users no longer need to learn complex prompt engineering or understand Transformer architectures; they can freely create various types of text using just natural language.

We are now entering the era of physical AI, where humanity needs vast amounts of three-dimensional data that accurately reflect real-world geometry and mechanics to enable the practical deployment of robotics and intelligent manufacturing. However, the industry is currently stuck in a frustrating gap: demand side craves physically realistic interaction data, while the supply side—AI—can only produce “visual shells.”

The other side of the challenge is opportunity. Recognizing the long-term demand for physical 3D data engines, Li Shiqi is confident that the generative physical 3D sector is on the verge of its own “GPT moment.” When that happens, the data vacuum faced by 3D printing makers, as well as industrial and robotics sectors, will be profoundly transformed.

The most direct example is the 3D printing industry. Manufacturing barriers have been completely lowered—any idea you conceive can be designed, printed, and used on the same day. Of course, this will further propel the 3D printing market toward genuine explosive growth.

Industrial product development will be the area with the most noticeable efficiency gains. Engineers will no longer need to spend extensive time drafting basic components; AI will automatically generate fully assembled designs with complete fit and dimensional constraints, significantly reducing development cycles and labor costs.

The efficiency of robot training is currently being bottlenecked by the "modeling" phase. Whoever breaks through first will secure their ticket to the era of embodied intelligence. As Zhou Pingyi noted, the fields of embodied intelligence and world models are likely the biggest beneficiaries. Once the supply bottleneck for mid-level simulation assets is resolved, virtual training data for robotic arms and humanoid robots will no longer be scarce. The pace at which general-purpose robots become practical may be faster than most people anticipate.

Within BitInfinity, the timeline to reach this milestone has been clearly defined—a window of approximately 18 months from the current stage to the GPT moment for generative physical 3D.

The clock has turned. This startup, which began by addressing pain points in 3D printing data, is transforming everyone’s designs, trials, and iterations into fuel for AI evolution, striving to rebuild the foundational supply system for industrial manufacturing and embodied intelligence, step by step constructing the bridge between digital code and the physical world.

"Think it, and you get it"—this may sound distant now, but who could have imagined just two years after ChatGPT’s debut at the end of 2022 that the large model landscape would evolve into what it is today? Right now, we stand on the eve of a turning point.

This article is from the WeChat public account "Investment Daily" (ID: pedaily2012), authored by Zhou Jiali.