TECH

Chinese GPU-Maker Challenges Nvidia in Three-Year Development Plan

Published: Jan. 27, 2026 7:59 p.m. GMT+8

00:00

00:00/00:00

Listen to this article 1x

Photo: Iluvatar CoreX

Shanghai Iluvatar CoreX Semiconductor Co. Ltd., a Hong Kong-listed GPU-maker, has announced an aggressive development roadmap for its artificial intelligence (AI) chip architecture, which the company claimed will allow it to rival industry leader Nvidia Corp. in the near future.

Under the roadmap announced on Monday, Iluvatar CoreX said that its next-generation AI chip architectures are expected to surpass Nvidia’s Blackwell architecture by 2026, with the ultimate goal of outperforming the U.S. company’s most advanced Rubin architecture by 2027.

You've accessed an article available only to subscribers

Subscribe today for just $.99.

VIEW OPTIONS

Unlock exclusive discounts with a Caixin group subscription — ideal for teams and organizations.

Subscribe to both Caixin Global and The Wall Street Journal — for the price of one.

Share this article

Open WeChat and scan the QR code

DIGEST HUB

Digest Hub

Back

Explore the story in 30 seconds

Shanghai Iluvatar CoreX aims to surpass Nvidia’s Blackwell AI GPU architecture by 2026 and Rubin by 2027.
The company claims its chip is about 20% more efficient than Hopper for DeepSeek V3 model reasoning and supports over 400 models.
Major Chinese firms like Moore Threads, Cambricon, and Huawei are also advancing GPU development, but entering big internet firms' supply chains is challenging for startups.

AI generated, for reference only

Who’s Who

Shanghai Iluvatar CoreX Semiconductor Co. Ltd.: Shanghai Iluvatar CoreX Semiconductor Co. Ltd. is a Hong Kong-listed GPU-maker aiming to rival Nvidia in AI chip architecture. They plan for their next-gen AI chips to surpass Nvidia's Blackwell architecture by 2026 and their Rubin architecture by 2027. They've optimized hardware and software for usability and claim their chip architecture is 20% more efficient than Hopper in specific reasoning tasks.

Nvidia Corp.: Nvidia Corp. is an industry leader in AI chip architecture, with its Blackwell and Rubin architectures being key benchmarks. Its Blackwell-based GPUs, utilizing TSMC's 4nm technology, are used by major U.S. developers for model training. The Rubin-based chip, on 3nm technology, began mass production in January 2026. Nvidia's Hopper-based chips were also a popular choice for cloud vendors in 2023.

TSMC: Nvidia's Blackwell-based GPUs, utilizing TSMC's 4-nanometer technology, are currently used by major U.S. model developers. TSMC also produces the 3-nanometer technology used in Nvidia's Rubin-based chips, which entered mass production in January 2026.

Moore Threads Technology Co. Ltd.: Moore Threads Technology Co. Ltd. (摩尔线程智能科技（北京）有限责任公司) is a Chinese GPU-maker mentioned in the article. In December 2025, the company launched its next-generation chip architecture, claiming its performance to be between Nvidia's Hopper and Blackwell architectures.

Cambricon Technologies Corp. Ltd.: Cambricon Technologies Corp. Ltd. is currently the only domestic Chinese chip designer that has secured substantial orders from a top-tier internet firm. They supply chips to ByteDance Ltd. for both model training and inference.

ByteDance Ltd.: ByteDance Ltd. is a top-tier internet firm in China. They currently use chips from domestic designer Cambricon Technologies Corp. Ltd. for model training and inference. This highlights ByteDance's role in supporting Chinese semiconductor firms as they race to develop advanced AI chips.

Huawei Technologies Co. Ltd.: Huawei Technologies Co. Ltd. is a Chinese technology company that produces Ascend chips. These chips are widely adopted by telecom operators and state-owned enterprises within China.

AI generated, for reference only

What Happened When

2023：: Hopper-based chips were the first choice for global cloud vendors when ChatGPT took the world by storm.

December 2025：: Moore Threads Technology Co. Ltd. launched its next-generation chip architecture, claiming its performance is between Hopper and Blackwell.

January 2026：: Rubin-based chips, utilizing 3-nanometer technology, entered mass production.

By 2026：: Iluvatar CoreX expects its next-generation AI chip architectures to surpass Nvidia’s Blackwell architecture.

January 27, 2026：: Iluvatar CoreX announced its development roadmap and stated it had optimized hardware/software designs; also announced chip architecture was approximately 20% more efficient than Hopper in DeepSeek V3 reasoning tasks.

As of 2026：: Over 400 models are powered by Iluvatar CoreX chips.

2026：: Blackwell-based GPUs are currently used by major U.S. model developers for model training.

2026：: Cambricon Technologies is the only domestic chip designer to secure significant orders from a top-tier internet firm, supplying chips to ByteDance Ltd.