EVAS Raised RMB 2 Billion Five Months After Its Series B. Nobody Outside China Has Benchmarked the Chip.

EVAS Intelligence, the Beijing-based designer of RISC-V cloud AI accelerators known in Chinese as 奕行智能, has closed a new funding round of nearly RMB 2 billion (roughly $295 million) at a post-money valuation estimated at close to RMB 15 billion (about $2.2 billion), according to reports that surfaced on September 18 and were confirmed across Chinese financial media the next day. More than 20 institutions joined the round, which lands five months after the company's RMB 1.5 billion Series B and arrives with its second-generation cloud chip already taped out.

The investor list reads like a map of China's state-aligned semiconductor capital: Huatai Innovation, Eastern Bell Capital, SMIC Capital (the venture arm tied to foundry SMIC), Andon Health, Renai Capital, packaging house Tongfu Microelectronics, the Saiyi Industrial Fund, Xinxin Leasing (a sub-fund of the first phase of China's National Integrated Circuit Industry Investment Fund, the so-called Big Fund), Hengxu Capital, Shangqi Capital, JAC Capital's Zhanlu fund and Rongyi Capital. Chinese outlet Zhidongxi called it the largest financing ever for a domestic RISC-V company.

The pace is striking even by the standards of China's overheated AI chip market. On April 22, EVAS announced a RMB 1.5 billion Series B co-led by four Beijing municipal funds. On May 8 it added a strategic investment of several hundred million yuan from China Mobile's Chain Leader Fund.

Founded in early 2022, EVAS is led by founder and CEO Liu Hui, a 20-year veteran of STMicroelectronics, Fujitsu Semiconductor, Cadence and Socionext. Rather than cloning Nvidia's GPU model or licensing Arm, the company pairs the open RISC-V instruction set and its vector extension with a TPU-style matrix engine of its own design, EVAMIND, and a virtual instruction set, VISA, meant to insulate software from hardware changes between generations.

In an interview with Jiemian News published in May, Liu argued that “the compute war has moved from a contest of single-point performance to a group battle spanning full-stack infrastructure and the whole ecosystem.” He was dismissive of the strategy, common among Chinese rivals, of chasing CUDA compatibility, describing Nvidia's software layer as a programming paradigm co-designed with its hardware, “like a key made specifically to open Nvidia's hardware lock,” and warning that a follower on that path “will always be five years behind Nvidia.” He said EVAS booked meaningful revenue in the fourth quarter of 2025, that sales were growing around 200 percent a year, and that the company was actively planning an IPO.

The first-generation Epoch chip, which the company describes as China's first mass-produced RISC-V cloud AI accelerator, natively supports block-quantized FP8 and is shipping in volume. At the World Artificial Intelligence Conference in Shanghai in July, EVAS unveiled what it billed as the industry's first RISC-V AI supernode, a liquid-cooled rack with a backplane-free interconnect fabric called ELink that the company says delivers 3.2 terabits per second of per-chip bandwidth across 64 to 128 fully meshed accelerators. Within roughly two months, a 64-node cluster built on its E200-L liquid-cooled OAM modules went live in a telecom operator's data center in Shijiazhuang, Hebei, with the company claiming 100-nanosecond-class latency across all 64 chips.

The second-generation chip, now taped out, is said to deliver several-fold gains in compute and interconnect performance and to add four-bit formats such as EXFP4 and MXFP4. Paired with a second-generation supernode designed for 128 cards per rack and up to 16,384 cards per pod, EVAS claims inference throughput in prefill-decode disaggregated deployments can rise more than tenfold. The company has committed to a one-generation-per-year cadence and a 400-fold system-level performance gain within three years. Shi Xianshuai, its chief AI officer, has framed AI chip competition as shifting from a race on raw compute to a race on R&D efficiency.

None of these figures has been independently verified. Epoch has not appeared in MLPerf or any comparable third-party benchmark, and EVAS has not disclosed its foundry or process node, a material omission given SMIC's role as an investor and its limited access to advanced lithography.

Why It Matters

The round says as much about Beijing's priorities as about one startup. Days before the announcement, the Ministry of Industry and Information Technology and the National Development and Reform Commission published the electronics manufacturing plan for the 15th Five-Year period, which explicitly calls for accelerating RISC-V industrialization and supporting RISC-V chips in AI. RISC-V International relocated from the United States to Switzerland in 2019 precisely to reduce exposure to American trade rules, and Washington has floated but never enacted export licensing for RISC-V work with China. For investors, an open ISA with a domestic full-stack software layer is a structurally safer bet than architectures dependent on foreign licenses.

It is also a demand signal. As this newsletter noted earlier this week, Chinese AI chipmakers have been raising prices amid the global HBM squeeze, a sign that domestic accelerators are being bought at volume rather than merely subsidized. EVAS's round was the largest of a cluster of Chinese compute-infrastructure deals on the same day; Tech Times noted concurrent raises of roughly RMB 100 million each by Arclight Quantum, photonic-interconnect startup Xili Optoelectronics and embodied-AI data firm Shutu Technology. State and industrial capital is clearly willing to underwrite the validation phase for a homegrown alternative to both Nvidia and Huawei's Ascend line. The harder question is software: data-center AI is decided by compiler maturity and operator libraries, and EVAS's stack has no production track record outside China.

What to Watch

The near-term signposts are concrete: whether second-generation Epoch silicon comes back from the fab on schedule and whether EVAS publishes any third-party benchmark data for it; how quickly the Shijiazhuang cluster is followed by additional carrier, internet and financial-sector deployments; and whether the company files for a domestic listing, which Liu has said is in active planning. Watch, too, for whether Beijing's RISC-V push draws Alibaba's XuanTie unit and other domestic players deeper into cloud-scale AI silicon, turning EVAS's lonely bet into a crowded race.