The emergence of China’s artificial intelligence (AI) start-up DeepSeek marks an inflection point in new paradigm of model training and inference. The company achieved a monumental innovation in software architecture that allows it to deliver on par performance with industry leader OpenAI’s existing models, but at a much lower cost.
AI industry shift?
There used to be a belief that companies in the AI industry have no option but to spend their way into building dominance and deep moats. Under the scaling law in AI, this means whoever could buy the most advanced chips in large quantity and provide strong raw training power had an edge and this would, usually be just a handful of tech giants in the US. But that dynamic now appears to be changing.
The release of DeepSeek‘s models raised questions among investors that whether the massive spending on semiconductors and technology hardware could be more cyclical in the short term, regardless of structural tailwinds.
This in turn, could result in deflation in the cost of AI-related build-out and thus, less need for astronomical capital expenditures at the outset. In fact, investors are starting to closely scrutinize and justify the expectations built in existing AI capital expenditure (capex). As a result, many AI infrastructure-related stocks, particularly those hardware names tasked with intense computing and data processing, appear to have stretched valuations and there is a greater chance of de-rating with the lofty expectations and fading momentum. At the same time, companies are increasingly turning to application-driven engines, like Internet of Things (IoT) and software, in anticipation of potential reducing computing cost and lowering entry barrier. This means greater emphasis in the ecosystem is likely to shift away from AI infrastructure to AI application, and from AI enablers to AI adopters.
The implications for China
We think this industry-wide shift is poised to benefit Chinese AI companies. As access to the most advanced AI training chips becomes less critical, Chinese firms have a chance now to narrow the gap with their US competitors, since inference chips are more accessible to Chinese companies, than advanced training chips amid tight US export-control.
More broadly, unlike their training-focused counterparts, inference chips are more energy efficient, less complex and do not require the same cutting-edge manufacturing, which China is fully capable of producing domestically. This transition to utilize more inference chips should also allow Chinese AI firms to reduce their reliance on the US chip designers like Nvidia.
Tina Tian, Portfolio Manager in Chinese equities:
“The breakthrough in DeepSeek illustrates that China has structural advantages in delivering fast paced innovations. In particular, we see how the strengths in data, research and development, and talent in China have propelled the success of DeepSeek, despite the constraints in high-end computing hardware due to export control. The success of DeepSeek also introduced some nuanced dynamics to the AI infrastructure demand, since large language model (LLM) performance is not solely driven by the amount of computing power. However, the lower cost of LLM means we will likely see an acceleration in AI adoption which could drive replacement of devices as more AI features can be afforded and could also drive the AI application in software, where companies are able to enhance product offering and improve operational efficiency.”
Sherry Qin, Greater China internet and software analyst:
“While it’s natural for savvy investors to be cautious about sharp share price movement, I see real efficiency improvement and new consumer value creation in long-term. This implies AI application is likely to take off in China and China may even move faster than US on applications/ IOTs, given its lower inference cost (around 5% of that in the US) and strong domestic manufacturing supply chain. This could underpin an acceleration in software development. A good example is Kingdee, a domestic ERP (Enterprise Resource Planning) software leader which has been gaining shares from SAP and Oracle. They have already developed new products with more AI features to enhance efficiency and user experience. Elsewhere, the recent rally in Alibaba boosted by its cloud business also signalled strong prospects for cloud platforms in China. The overlooked segment has been struggling with slow growth, fierce competition and subdued margin for some time. Now a surge in cloud services demand, combined with accelerated cost reduction could unlock potentials in cloud business.”
Broadly speaking, AI involves fast-moving progress, it’s still too early to say which participants will capture future revenue streams, but it is clear that the race for AI leadership is no longer just about who owns the best chips, it’s about who can put them into best use. This provides good conditions for bottom-up stock pickers to distinguish winners from the rest, among sectors and companies.
