The global recognition of Chinese AI star, DeepSeek, is expected to serve as a catalyst for small AI chip companies to massively scale up around the world.

The Hangzhou-based startup rattled Silicon Valley last month with the launch of its low-cost, but powerful AI model, which wiped out around half a trillion US dollars in chip leader Nvidia's market cap because it was built without using cutting edge chips, due to American export restrictions.

Speaking on RTHK's Trending China programme, Esther Wong, founder of 3C AGI Partners, noted that DeepSeek's breakthrough came as the global AI cycle transitioned from "training" to "inferencing".

Inference refers to the act of applying AI to make predictions or decisions based on the processing of new information and data in real time, instead of training the AI model to learn patterns, relationships or features.

"It's equivalent to Generative AI, more focused on AI's real applications where AI can serve users differently in a tailored way," Wong explained.

"[So if] inferencing is going to be used by some seven billion people in the world. You can imagine the explosion of demand is going to be so much that in the words of Jensen Huang, the inferencing demand is literally one billion times bigger than the training demand."

"[But] If we do not change how AI is being manufactured now, the world will run out of power very, very soon," she said.

Wong, who previously worked as managing director at Chinese AI leader SenseTime, noted the new opportunities following DeepSeek's rise as the firm's open-source model would help other small AI chip firms to challenge the current market landscape that is dominated by big players such as Nvidia, with investors eyeing alternatives to their chips.

"The moment DeepSeek emerged, it can actually help the proliferation of all those basic chips, i.e. chips apart from Nvidia, that are also very good for inferencing purposes. And that applies for both domestic chips in China or chips produced by startups in the West," she said.

The veteran banker and investor also stressed that despite the narratives that AI breakthroughs could be achieved with smaller computing power and lower investments, demand for such power as well as funds would only increase.

"The inferencing demand is one billion times bigger than the training demand, and by making inferencing so cheap and accessible [due to DeepSeek's model being open-sourced], we actually need a lot more computing powers so that more people can afford inferencing," she said.

"We are seeing that the Large Language Models [LLMs] are quickly being commoditised, so it's difficult to make money out of it because the marginal cost of producing intelligence is close to zero. But that is very good for the application layers where a lot of usage can be proliferated as they can be commercialised."