Cambricon Technologies, a leading developer of AI processors from China, achieved its first-ever quarterly profit in late 2024. This financial upturn occurred after years of being in the red, reports the South China Morning Post. The company's products have become significantly more popular in China in recent quarters after the U.S. government banned sales of Nvidia's advanced AI GPUs to the People's Republic.
Cambricon's revenue surged nearly 70% in 2024, reaching approximately ¥1.2 billion ($163.7 million), which is a fraction of the $90 billion earned by Nvidia globally, but represents significant growth. Cambricon's quarterly profit for Q4 2024 ranged between ¥240 million and ¥328 million ($32.74 million to $44.74 million), marking a turnaround after losing ¥724 million ($98.76 million) in the first three quarters. This brought its full-year loss for 2024 down to ¥396 - ¥484 million ($54 million - $66 million), a substantial improvement compared to previous years.
As the adoption of Cambricon's hardware expanded in China over the past several quarters (including adoption by Huawei), the company has also seen its stock on the Shanghai Stock Exchange’s Star Market skyrocket by over 470% in the past year, amid the shortage of Nvidia's GPUs in China. The company's stock climbed from ¥120.80 to ¥695.96 over the past year.
Cambricon was founded in 2016 to develop AI processors for a wide range of applications from edge devices to cloud data centers. Nowadays, the company offers a family of AI accelerators based on the MLU instruction set architecture that can be used for both training and inference.
The company's most powerful product is its MLU290-M5 accelerator which boasts 512 INT8 TOPS, 256 INT16 TOPS, and 64 CINT32 TOPS performance and comes with 32GB of HBM2 memory featuring a 1,228 GB/s bandwidth. The company has a more sophisticated but less powerful MLU370-X8 48GB LPDDR5 accelerator that supports FP32, FP16, BF16, INT16, INT8, and INT4 formats, but which performance is limited to 256 INT8 TOPS and 24 FP32 TFLOPS.
To put the numbers into context, Nvidia's A100 from 2020 offers INT8 performance of 624 TOPS/1248 TOPS (with sparsity) and FP32 performance of 19.5 TFLOPS, so Cambricon is about 4-5 years behind Nvidia. Additionally, Nvidia has a major edge over virtually all of its rivals due to its software ecosystem and developer support. However, for Cambricon's customers who cannot get enough Nvidia GPUs, this advantage of the Nvidia platform is hardly a concern.
Cambricon's latest AI processors are made on a 7nm-class process technology, though it is unclear where the company produces them since it is on the U.S. Department of Commerce's Entity List, which largely prohibits it from using TSMC's services. Changjiang Securities estimates that China’s AI semiconductor market will grow to ¥178 billion ($24.28 billion) this year. Cambricon is a key player in China's AI market, and with its rapid growth and improving financials, the company is poised to capitalize on the country's exploding demand for AI hardware.
