Reporter |
Editor | Sophia Wen
On May 8, domestic RISC-V AI computing chip company, YiXing Intelligence, announced that it had secured strategic investment from China Mobile’s Chain Long Fund. Just half a month ago, on April 22, the company had completed a 1.5 billion yuan B-round financing.
The intense capital injection is driven by the industry’s urgent need for new computing architectures and the company’s accelerated commercialization process.
According to reporters, YiXing Intelligence’s core product, Epoch, the first domestic RISC-V AI computing chip, has achieved mass production. The latest round of funding will focus on the mass production and delivery of Epoch series products, as well as the development of next-generation products.
Founded in 2022, YiXing Intelligence’s founder and CEO, Liu Hang, previously stated in an interview that the current computing power war has evolved from a single-point performance competition to a comprehensive battle involving the entire stack and ecosystem.
In fact, globally, the computing infrastructure boom is reshaping the semiconductor industry landscape.
According to Gartner’s latest report, global semiconductor revenue will reach $793 billion by 2025, with a 21% year-over-year growth. AI semiconductors will account for nearly one-third of global sales. Gartner predicts that global AI infrastructure spending will exceed $130 billion by 2026, further solidifying the dominance of AI chips in the industry.
Under this macro trend, Liu Hang pointed out that the industry focus has shifted from simply competing on computing power to a comprehensive consideration of scale, efficiency, and ecosystem.
He believes that the final product delivered by large model service companies to customers is essentially a token, with each token representing actual computing power and energy consumption.
This means that the cost of computing power has become a core competitive point for large model commercialization.
Currently, domestic leading manufacturers have been able to compress the unit cost of tokens to an extremely low level through algorithm innovation and software and hardware optimization. Therefore, in Liu Hang’s view, AI computing is essentially specialized computing.
Facing the market barriers constructed by NVIDIA, the conventional approach is to try to be compatible with the CUDA ecosystem. Liu Hang believes that CUDA is not simply a software library, but a programming paradigm designed in conjunction with NVIDIA’s hardware architecture, like a key specifically designed to unlock NVIDIA hardware.
He pointed out that following the compatibility route means passively adapting to the opponent’s old version, and the overall progress will always be 5 years behind NVIDIA. Moreover, under the constraints of process technology, the gap will be further widened.
Under the macro requirement of autonomy and controllability, finding alternative paths has become a key to breaking through. YiXing Intelligence’s answer is to adopt the RISC-V architecture combined with self-developed virtual instruction sets.
According to Global Market Insights, the global RISC-V market is expected to grow from $2.3 billion in 2025 to $8.57 billion in 2030. Its open architecture has attracted heavy investment from global tech giants and has been listed as a strategic priority by several countries, including China.
Liu Hang introduced that YiXing Intelligence’s core product is the Epoch series AI computing chip and related computing platform products. As the industry’s first general-purpose AI computing chip that combines RISC-V and self-developed VISA virtual instruction sets, the product was taped out and returned in 2025 and completed the Bring Up lighting test in 3 hours. It has now achieved mass production.
Liu Hang told reporters that RISC-V has Turing completeness, ensuring general computing capabilities while allowing manufacturers to freely expand specialized AI computing instructions for customization and acceleration. RVV (vector extension) natively supports complex vector calculations and supports mixed-precision calculations for multiple data types. At the same time, its open-source attribute highly aligns with the country’s autonomous and controllable security strategy.
YiXing Intelligence’s other core technology is its self-developed VISA virtual instruction architecture.
Liu Hang stated that the “one-year-one-generation” iteration speed of hardware often brings huge challenges to software adaptation. VISA can establish a middle abstraction layer between software and hardware, allowing upper-layer operators and AI compiler software to be built on top of VISA, thereby isolating the impact of hardware changes on upper-layer software and effectively solving software compatibility and adaptability challenges.
In terms of data precision evolution, Liu Hang believes that the AI reasoning market is leaning towards low-bit-width and high-precision directions. Therefore, YiXing Intelligence’s products natively support multiple data types, and through the innovation of shared computing units at the bottom layer, the chip natively supports these formats, greatly reducing memory bandwidth and capacity occupation and doubling the computing density.
Liu Hang revealed that with the support of Alpha customers in early funding, algorithm deployment, and other aspects, the company achieved significant sales revenue in the fourth quarter of 2025, with overall sales revenue growing at a rate of around 200% per year.
In the second half of the industry competition, capitalization capabilities are also a part of the hard power. As the competition in the AI chip field enters the deep water zone, Liu Hang frankly stated that going public early can help obtain strategic resources, and the company is currently planning and taking action to advance the IPO process.
