DeepSeek's New AI Model Reveals Homegrown Chip Strategy to Sidestep Nvidia
A leading Chinese artificial intelligence firm has unveiled a new data format alongside its latest model, a move that signals a significant strategic shift toward building a domestic AI hardware and software ecosystem independent of Nvidia Corp.'s dominant technology.
Beijing DeepSeek Artificial Intelligence Co., Ltd. announced the launch of its DeepSeek-V3.1 large language model on Thursday. However, the most consequential development was a detail highlighted by the company: the use of a custom UE8M0 FP8 data format specifically designed for the next generation of Chinese-made semiconductor chips.
This quiet announcement represents a pivotal moment for China's AI ambitions. It is a direct technical response to the challenge of developing powerful AI on domestic hardware that cannot fully replicate the proprietary optimizations built into Nvidia’s market-leading GPUs. By creating a software standard tailored to local chips, DeepSeek is forging a distinct, self-reliant path for AI development.
The move comes amid heightened scrutiny of Nvidia's products by Chinese regulators and a broader national push for technological self-sufficiency. For investors and the global tech industry, it marks a deliberate effort at decoupling in a critical sector, fostering a new, symbiotic alliance between China's AI developers and its nascent chipmaking industry.
The Bottleneck of Precision
The race to build ever-larger AI models has pushed the industry toward lower-precision data formats. For years, FP32 (32-bit floating point) was the standard for its accuracy. But for models with trillions of parameters, it is too memory-intensive and slow.
The industry has since shifted to formats like FP16 and, more recently, FP8 (8-bit floating point). Using a lower-precision format like FP8 significantly reduces the model's memory footprint and can more than double training speed, or throughput, without a substantial loss in performance. For companies training massive, GPT-level models, this efficiency gain is not just an advantage—it is a necessity.
Nvidia's Walled Garden
While industry players including Nvidia, Intel, and Arm have collaborated on FP8 standards, Nvidia has embedded deep, proprietary optimizations into its hardware and software stack. Features like dynamic scaling strategies and specialized instructions for its Tensor Cores are not part of the open standard but are crucial for achieving maximum performance and stability on its H100 and Blackwell GPUs.
This has effectively created a "walled garden." To unlock the full potential of FP8, developers are steered toward Nvidia’s integrated ecosystem. The company's latest Blackwell architecture further solidifies this with a new "Microscaling format," MXFP8, which has been shown to deliver performance nearly on par with higher-precision formats, reinforcing its technological moat.
A Pragmatic Workaround
Chinese GPU manufacturers cannot simply replicate Nvidia's hardware-specific optimizations. Attempting to run standard FP8 formats on domestic chips can lead to numerical instability and failed training sessions, a problem known as "gradient explosion."
DeepSeek’s UE8M0 FP8 format is a pragmatic solution to this hardware limitation. It employs a "range-first" strategy, prioritizing the ability to represent a wide range of values over fine-grained precision. It is a technical compromise, sacrificing some detail to ensure that models can train stably on the underlying domestic hardware. This software-level adaptation effectively works around the hardware gap, ensuring domestic chips can be viably used for large-scale AI training.
A Nascent Domestic Alliance
DeepSeek’s tailored format is designed to align with an emerging generation of Chinese AI accelerators. For instance, MetaX Integrated Circuits (Shanghai) Co., Ltd. officially launched its Xiyun C600 chip in 2025, which features native support for FP8 and is now in small-scale mass production. Its next-generation C700 is already in development.
Similarly, Shanghai Enflame Technology Co., Ltd. introduced its L600 chip this year after two and a half years of development. The chip integrates training and inference capabilities and, crucially, includes native FP8 support, making it a suitable hardware partner for models like DeepSeek's. The UE8M0 parameter signals that China's AI model companies and domestic chipmakers are now collaborating to build a co-dependent ecosystem, moving beyond simply chasing Nvidia's standards to creating their own.