On April 17, Xu Bing, co-founder of SenseTime, hosted a private dinner in Beijing. He invited Wang Zhan, a member of Baidu’s founding team, and brought two intermediaries along. Over the meal, Xu asked Wang directly: “Are you interested in joining a chip startup?”
Wang, known for keeping a low profile, has little written about him online. But in China’s internet industry, his name carries weight. After graduating from Peking University’s physics department, he joined Baidu in 2000 when the company had fewer than 20 employees. He became a product manager at Baidu and later served as the legal representative of the overarching group, a role he held until his departure in 2016.
Over 16 years, Wang managed teams of up to 8,000 and oversaw both search and commercialization. Many of his former subordinates now hold senior roles across China’s largest internet firms.
When Xu extended the offer, Wang quickly agreed and took on the role of co-CEO of Sunrise.
Five years ago, to address both the need for technological self-reliance and the rising cost of computing power, SenseTime began investing heavily in chip development. That initiative spun out at the end of 2024 under SenseTime’s “1+X” strategy, which separated high-investment, long-cycle businesses with strategic value. The new unit was named Sunrise and tasked with building inference chips for large models.
“SenseTime has nurtured us like a mother for ten months, but now it’s time for the chip team to grow independently,” Wang said. Explaining the name “Sunrise,” he added:
“It represents the breaking dawn of China’s artificial intelligence chip era, where we act as guardians of the morning light.”
The day after their first dinner, Wang and Xu met again in Beijing, talking until half past midnight. Within 24 hours of adding each other on WeChat, they had outlined the company’s direction, division of responsibilities, and leadership structure.
Nearly a decade after leaving Baidu, and already financially independent, Wang decided within a single day to return to the industry. As he put it, building AI chips is a “difficult but deeply passionate and honorable endeavor.”
At a time when Chinese AI chipmakers are gaining visibility, Wang spoke with 36Kr about SenseTime’s five-year chip strategy and Sunrise’s roadmap as an independent company.
SenseTime’s chip bet
The year 2025 is shaping up to be pivotal for Chinese chipmaking. Nvidia’s market capitalization has surpassed USD 4 trillion, while geopolitical tensions have fueled speculation over who could be “China’s Nvidia.”
Chinese chipmakers are lining up their answers. Cambricon, already publicly listed, saw its market capitalization surge past RMB 600 billion (USD 84 billion), with its share price briefly overtaking liquor giant Kweichow Moutai. Meanwhile, a wave of Chinese GPU companies, including Moore Threads, MetaX, Biren Technology, Enflame, and Vastai Technologies, have entered the pre-IPO advisory stage.
This listing momentum marks a shift from capital-intensive technology breakthroughs to a phase of commercialization. Wang told 36Kr that although Sunrise only spun out late last year, it intends to build on SenseTime’s five years of groundwork for a fast start.
Sunrise has already completed its leadership team. Its other co-CEO is Wang Yong, a chip veteran with 20 years of experience at AMD and Baidu’s Kunlunxin. He led the design of Kunlunxin’s second-generation architecture and oversaw two chip generations at SenseTime.
Wang Yong now leads product and technology development, while Wang Zhan and Xu Bing steer operations, commercialization, and fundraising. In 2025, Sunrise’s team expanded by 50% to nearly 200 employees, many from AMD, Intel, Alibaba, and SenseTime.
Alongside the spinoff, Sunrise has been raising funds and pushing forward on products. According to 36Kr, the company secured more than RMB 1.5 billion (USD 210 million) in just six months of 2025. Wang Zhan added that SenseTime invested over RMB 1.1 billion (USD 154 million) into chip development over the past five years.
On the product side, Sunrise has already mass produced two chips:
- S1 chip, a visual inference chip for cloud and edge computing. First taped out in 2019, it powered SenseTime’s computer vision products. Its NPU IP was licensed to Sony and Xiaomi, with more than 10,000 units shipped.
- S2 chip, a general-purpose GPU for large model inference, mass produced in 2024 and compatible with Nvidia’s CUDA architecture.
The next-generation S3 chip will continue targeting inference workloads. Its highlight feature, Wang said, is cutting inference costs by a factor of ten through architectural innovations. It is slated for release in 2026.
The technical team explained that S3 will use a higher proportion of low-precision compute units with larger, cost-efficient memory. It is designed to handle both compute-intensive prefill and memory-intensive decode phases of large model inference, echoing elements of Nvidia’s newly launched Rubin chip.
Sunrise’s cost advantage stems from pursuing fully proprietary IP, from instruction set design and GPU architecture to compiler toolchains.
“In inference chips, cost-performance is the lifeline,” Wang said. “Reducing inference costs is the driving force for mass adoption of large models.” He added that Sunrise’s R&D efficiency is above the industry average despite its team being only one-third the size of typical competitors. It has also shortened the cycle from tapeout to mass production to 14 months, compared with the industry norm of 22 months.
The latecomer’s advantage
For years, the biggest challenge facing AI chip startups has been adoption. Sunrise is addressing this by working closely with industrial investors.
In its recent funding rounds, 70% of Sunrise’s backers were strategic investors, including Sany Group’s Huaxu Fund, 4Paradigm, Yoozoo Games, and Midea. Wang said these partnerships ensure Sunrise designs chips with commercialization in mind, building around specific use cases.
Standing on SenseTime’s shoulders is another advantage. Sunrise chips will interface with SenseTime’s Ririxin large models and AI cloud infrastructure, tapping Ririxin for validation and SenseTime’s computing capacity for scaling.
China’s GPU startup wave began in 2020, driven by urgency around tech self-sufficiency and returning talent from Nvidia and AMD. Nearly 20 GPU startups emerged within a short time. But Wang argues that 2020 was not the real starting point. The industry’s real window, he said, is opening now.
In earlier years, local players could only capture small fragments of Nvidia’s market, with shares below 10%. Today, surging inference demand and Nvidia’s constraints have created a genuine opportunity.
Financial results reflect the shift. Cambricon’s first-half revenue grew 4,347% year-on-year, turning profitable for the first time. Kunlunxin won a major China Mobile order. Alibaba’s AI chip strategy announcement added RMB 260 billion (USD 36.4 billion) to its market cap overnight. Moore Threads’ first-half 2025 revenue exceeded its past three years combined. MetaX shipped more than 25,000 chips, while Hygon crossed 100,000 units for the first time.
So is Sunrise too late to the party?
Wang doesn’t think so. He argues that chip startups benefit from a latecomer’s advantage: entering too late risks missing the window, but entering too early means wasting resources on wrong bets.
Over the past five years, Chinese AI chipmakers have wrestled with two questions: whether to build CUDA-compatible ecosystems or their own, and whether to focus on general-purpose GPUs or specialized scenarios. The market has shown that CUDA compatibility drives adoption, and specialization is more pragmatic. Those lessons form the basis of Sunrise’s strategy today.
Looking forward, Wang sees three decisive factors for AI chipmakers:
- Software-hardware integration: GPUs require strong capabilities across the stack.
- Market foresight: Companies must anticipate trends and position early.
- Timing: With Chinese AI chips on the verge of commercialization, the current vacuum is a one-shot chance to capture share.
Wang believes China’s GPU market could support three to five leading players. But 2025 will be decisive:
“Any company that hasn’t passed tapeout and shipped a working product may never get another chance.”
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Qiu Xiaofen for 36Kr.