Xiwang is seeking another 2 billion yuan in financing, with its valuation doubling to 20 billion yuan within four months | Exclusive

By HAI Ruojing

Over the past year, domestic GPUs have been in the midst of a capital boom.

Waves has exclusively learned that Sunrise recently secured a new round of 2 billion yuan in financing, with a post-money valuation reaching approximately 20 billion yuan. Back in April this year, Sunrise completed a financing of over 1 billion yuan led by Hangzhou Capital, pushing its valuation above 10 billion yuan. In less than half a year, its valuation has nearly doubled.

Roughly calculated, since its spin-off from SenseTime at the end of 2024, Sunrise has raised a total of nearly 6 billion yuan in financing.

In response to the above news, a representative of Sunrise stated: We are not in a position to make further comments for the time being.

Sunrise has not been an independent entity for long, but its chip team has existed for many years. At the end of 2024, SenseTime pushed forward the "1+X" organizational restructuring, spinning off its high-investment, long-cycle chip business into an independent entity. XU Bing, co-founder of SenseTime, serves as the chairman of the new company. The two co-CEOs, WANG Yong, who previously served as a core architect at AMD and Kunlunxin and joined SenseTime in 2020 to head the chip business, and WANG Zhan, who once served as Vice President of Baidu.

Prior to the spin-off, the team had already mass-produced two chips: the multimodal visual inference chip S1 and the large model inference GPU S2. In the investment circle that values "continuous success", this is a major plus.

According to insiders and verified with industrial and commercial change information, the list of industrial capitals participating in Sunrise's new round of financing in August 2026 continues to expand, including CP Group, Andon Health, Infore Environment, Tongcheng Travel, etc. At the same time, "national team" funds including PICC Capital and CCB Principal Asset Management have entered the investment; financial institutions such as Janchor Partners, CAS Star, and Cowin Capital also participated in the financing.

Why did the valuation of an inference chip company nearly double in just a few months?

Looking at the mid-2026 timeline, the change in the structure of AI computing power demand has become relatively clear. In the past few years, the period when AI computing power was consumed the most was during the large model training phase; now, Agents are being widely deployed in production scenarios, and a single task may trigger dozens of model calls. 24/7 uninterrupted inference and token generation have become the key driver of computing power consumption growth.

Especially this summer, with the successive release of open-source models including Kimi K3, MiniMax H3 and DeepSeek-V4, the demand for AI coding, video generation and office scenarios has increased significantly. These types of tasks often generate a large amount of long contexts, multiple rounds of calls and tool usage. The larger the call scale, the more sensitive users are to the incurred costs.

This is also the reason why "Token Economics" has been repeatedly discussed this year. When inference becomes a lucrative business, how many tokens a single chip can stably generate, how much power and cost each token consumes, and whether the supply can be guaranteed, are all practical factors that purchasers will inevitably consider.

The S3 inference chip released by Sunrise in January 2026 is designed to seize the initiative in aspects such as inference efficiency, cost and supply certainty in advance. In previous media interviews, the Sunrise team mentioned that in the actual delivery of the S1 and S2 chips, the vast majority of computing power was used for inference scenarios, which prompted the S3 to be fully designed with a dedicated inference-only architecture. Different from mainstream GPUs, the S3 does not pursue integrated training and inference capabilities, but concentrates its transistor and power consumption budget on maximizing inference efficiency.

In terms of memory selection, the S3 does not adopt the high-bandwidth memory (HBM) that is standard for integrated training and inference GPUs, but chooses the low-power memory (LPDDR) commonly used in mobile phones and laptops.

If we compare video memory to a "warehouse", HBM is like a three-dimensional warehouse built at an ultra-wide high-speed intersection, with the advantages of extremely high bandwidth and fast data access, but it is expensive, and the advanced packaging capacity is concentrated in a small number of manufacturers such as TSMC, leading to tight supply. In comparison, the bandwidth of a single LPDDR chip is lower than that of HBM, but its capacity can be made larger, the cost is lower, and the supply chain is more reliable.

As the tasks processed by Agents become more and more complex, apart from computing speed, whether the video memory can accommodate the model, long contexts and temporary memory during operation, is becoming the key to inference efficiency. After expanding the "warehouse" capacity with LPDDR, and then reducing the amount of data that needs to be transferred for each token through low-precision computing, the overall inference efficiency can be better.

Apart from technology, in terms of independent and controllable supply chain, some of CXMT's LPDDR5X products have entered mass production in 2025, and the R&D of the next-generation LPDDR6 is nearing completion, which means domestic storage manufacturers have had initial supply capacity. For domestic inference chip companies that also adopt this route, they do not have to fully rely on overseas supply chains in the storage segment.

Since the end of last year, Moore Threads, Mthreads, Biren Technology and other companies have been listed one after another, and the pipeline of to-be-listed companies also includes Enflame Technology, Kunlunxin and other players. Inference computing power is still a bottleneck in the AI industry chain. Entrepreneurship focused on dedicated inference architectures, processing-in-memory, and computing power scheduling remains very active, with capital and talents continuing to pour in. However, after the boom, capacity release and technical route differentiation will push the industry into a performance realization phase.

Sunrise is right at this turning point.

The mass production of two generations of chips has provided sufficient engineering experience, and the differentiated route bet on by S3 is also showing new value at a time when the industry is competing for "unit token cost". But at present, the innovation is only half completed. Whether this technical combination can perform stably in large-scale delivery, and whether the cost curve of millions of tokens can decline as expected, the remaining half of the answer will be left to the market.