China's Moore Threads unveils new AI chips after IPO

Moore Threads Technology, a Beijing-based chipmaker founded by ex-Nvidia executive Zhang Jianzhong, has unveiled a new line of AI chips after its Shanghai IPO.

The company said its new Huagang architecture will raise computational density by 50% and improve energy efficiency by 10x.

Production is expected to begin in 2026, with the technology enabling data centers to connect more than 100,000 chips for AI training.

Previously focused on gaming and visual rendering chips, Moore Threads is now targeting the AI accelerator market.

It also introduced updated Lushan-series GPUs, an integrated Changjiang SoC chip, and upgrades to its MUSA computing platform.

Blacklisted by the US in 2023, the firm is positioning its products as alternatives to Nvidia, as its IPO shares surged fivefold.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

Huagang’s 2026 production timeline faces manufacturing hurdles under export controls

Teams need the foundry (the contract chip manufacturer) and the process node (the manufacturing generation e.g. nanometer class). They also need advanced packaging capacity (the high-density assembly that links chips to memory) for the claimed 100,000+ chip clusters under US export limits on advanced semiconductor gear 1.
ChangXin Memory Technologies (CXMT), a Chinese DRAM maker, is mass-producing HBM2 in 2024, with HBM3 or HBM3E not expected until 2027 2. If Moore Threads’ AI GPUs need 8 HBM sites as claimed 3, sourcing enough memory by 2026 looks hard.
The 100,000+ chips claim also hinges on interconnect. Readers need the network fabric (the high-speed links among chips), the topology, and end-to-end bandwidth. The announcement lacks these details, so the scale could be a plan rather than deployed gear.

Software migration firms can win work as enterprises adopt MUSA-based domestic accelerators

Third-party software companies and enterprise IT system integrators can offer CUDA-to-MUSA porting for Chinese enterprises that must use domestic accelerators. Moore Threads’ Musify toolkit (its CUDA translation tool) translates CUDA code 1. Teams still need tuning and kernel work, plus benchmarking.
The torch_musa framework (Moore Threads’ PyTorch backend) supports over 1,000 operators and offers tools for building MUSA extensions 4. Production AI workloads still need model optimization, debugging, and validation. This creates work for consultancies and vendors as firms shift off Nvidia gear.
Demand is front-loaded. Firms that move early can build expertise and win reference customers before the market crowds.