The 2026 Sustainable Global Leaders Conference was held from October 20 to 22 at the Huangpu Expo Park in Shanghai.
Liao Ruoxue, Partner and Senior Vice President of Technology at Deep Potential Technology, attended and delivered a speech.
In her remarks, Liao Ruoxue, building on the sharing by Mr. Wei, pointed out that in scientific application scenarios, the quality of large model tokens is the core pain point, and the industry faces many practical difficulties.
If enterprises purchase GPUs on their own, build large models from scratch, and complete deployment and inference, the entire process is complex and cumbersome.
During project implementation, Deep Potential Technology has repeatedly found that tokens provided by some suppliers are watered down or excessively quantized, directly causing a decline in model output quality.
She introduced that AI applications in the scientific field have two distinct characteristics.
First, token consumption is enormous, reaching the trillions, while the demand for token quality is extremely high.
As Mr. Wei mentioned regarding CTFB-related requirements, scientific computing hopes to use original models and avoid excessive quantization, aggressive optimization, and other means that damage the model's native capabilities — this is also the most important point in scientific scenarios.
Second, scientific scenarios do not require high inference real-time performance, and there are large numbers of batch and offline computing needs.
Such business characteristics can effectively improve the usage efficiency of tokens and GPU resources, helping token operation vendors raise GPU utilization to 80% to 90%.
Liao Ruoxue stated that although tokens are called in the same way, scientific scenarios differ significantly from general internet applications in terms of quality standards, latency indicators, and TTFT (time to first token), and the supply logic of general large models cannot be directly applied.
Comments