Kyle Hulett Co-Head: Investments at Sygnia

A new AI model from Chinese company Moonshot sent a jolt through global tech markets in July, dragging down shares of chipmakers, memory producers and data-centre builders. It has deepened a concern that has been building for a while: Will the hyperscalers (tech giants like Meta, Google, Microsoft, Amazon and Oracle) pouring hundreds of billions of dollars into AI actually see a return on that spending?

Moonshot’s new model, Kimi K3, ranked ahead of Anthropic’s Opus 4.8 according to independent testing group Artificial Analysis, making it the first Chinese open-weight model to surpass a leading US model at this level (see chart). China is already known for producing cheaper AI models; Kimi K3 suggests it can now compete on quality too. Alibaba shares rose on similar news after the company previewed its flagship Qwen3.8 Max model, which it described as second only to Anthropic’s Claude Fable 5, though Kimi K3 reportedly delivers comparable performance at roughly one-third of Fable 5’s cost per task. These launches have intensified investor concern that cheap, cutting-edge Chinese AI could push US companies to rely more on these models instead of building their own processing capacity, reducing demand for chips and data centres. Apple, for instance, has already said it will use Alibaba’s Qwen AI and Baidu’s search engine to serve the Chinese market. The selloff in semiconductor stocks deepened as signs of China’s progress in advanced memory and chipmaking weighed on global rivals (see China section for more details).

AI-related capital expenditure is at record highs and has been one of the biggest drivers of stock market gains this year. Bloomberg forecasts AI capex will reach $963 billion in 2027, an acceleration pushing hyperscalers’ prospective free cash flow into negative territory. Google’s share price fell 7% on its latest quarterly results despite 82% growth in AI-driven cloud revenue, marking the com pany’s first quarter of negative free cash flow since going public. From here, the way forward could take several routes:

• Commoditisation. Chinese models could erode the pricing power of US AI models. Prices for AI tokens have been falling, and the Silicon Data LLM Token Expenditure Index has dropped nearly 30% from its May peak, coinciding with the recent underperformance in AI hyperscaler stocks.

• Efficiency gains. Meta’s stock rose after an internal memo suggested the company can build AI computing capacity more cheaply than expected, a dynamic that could trigger a rotation out of semiconductor names and back into hyperscaler stocks.

• Demand expansion. The more optimistic view is that cheaper models will simply expand the addressable market and that the drop in token prices reflects additional usage rather than a price war. Token consumption rose from 11.4 billion tokens per week in 2024 to 15.6 trillion tokens per week in July 2026, an increase of roughly 1 370 times.

The path forward will probably be a combination of the three. US security concerns are likely to limit the scope for Chinese model adoption domestically, while underlying compute demand shows little sign of weakening. For now, the global economy is experiencing a period of exceptional growth driven by the AI capex boom, and the upcoming round of hyperscaler earnings and capex forecasts will be closely watched for signs of which path is dominating the outlook.

ENDS