Axera Semiconductor Unveils AX8860 AI Chip at Nearly Half the Price of NVIDIA DGX Spark
At the 2026 China International Industry Fair on October 13, 2026, Axera Semiconductor officially launched the AX8860 series, a high-compute AI chip family positioned as a computing foundation for large language models at the edge and on-device. The company also introduced the Yuanxi module A256S and showcased Yuanxi inference cards and servers.
The AX8860 integrates a 16-core Arm Cortex-A55 CPU and a proprietary NPU, delivering 256 TOPS of mixed-precision compute (native FP4, no sparsity). It supports up to 128GB of LPDDR5X memory with 273GB/s bandwidth on a 256-bit interface, and is built on a 6nm process with PCIe 5.0 connectivity. The chip supports various form factors, including a 40W module, a single-slot PCIe card with Gen5x4 connector, and a single-slot four-chip solution. Axera also plans Mini PC and pocket PC products based on the chip. For multi-chip scaling, dual-chip configurations reach 512 TOPS and four-chip setups are claimed to deliver up to 1024 TOPS, with about 80% scaling efficiency.
On the software side, the AX8860 supports mainstream inference frameworks such as vLLM, SGLang, and PyTorch, with out-of-the-box support for ONNX and Hugging Face model formats, and compatibility with OpenAI API interfaces. It comes with a complete toolchain covering quantization, conversion, compilation, profiling, and debugging.
In benchmark comparisons on the Qwen 27B model, the AX8860 is up to 25% faster in prefill and 30% faster in decode than NVIDIA's DGX Spark. Sipeed says system pricing for the AX8860 is close to half of DGX Spark's price, which starts at $4,999 for the 64GB version and $6,950 for the 128GB version. According to Wccftech, Axera's roadmap also includes the 500 TOPS Axera 8950 and the next-generation 8970 chip with over 700 TOPS. Axera founder and chairman Qiu Xiaoxin said at the event, "We want to be a long-term stable computing foundation for customers, letting enterprises focus on business innovation without worrying about compute migration."
