China’s AI chipmakers report first profits, but will the trend continue?

China’s pure-play artificial intelligence chipmakers crossed an important milestone in the third quarter as several firms swung into profit, but severe supply chain hurdles and soaring memory costs cloud their long-term growth prospects.

Domestic graphics processing unit (GPU) developers including Moore Threads, MetaX and Iluvatar CoreX are racing to fill the market void left by Nvidia, whose high-end processors face strict US export controls.

These home-grown firms are competing fiercely to win over domestic enterprise clients and penetrate major cloud service providers (CSPs).

Beijing-based Moore Threads, one of the “four little dragons” in China’s GPU space, led the earnings turnaround, posting an adjusted net profit of 49.26 million yuan (US$7.4 million) for the quarter, driven by a 1,774 per cent year-on-year surge in revenue. It was the chipmaker’s first profitable quarter.

Shanghai-based MetaX recorded a net profit of 26.5 million yuan, marking its second profitable quarter of the year.

The momentum faces headwinds, however. The domestic sector is grappling with an acute shortage of high-bandwidth memory (HBM), an indispensable component in the systems for training and running large AI models.

As US export controls reduced supply, factory prices for 36GB HBM3e modules skyrocketed to more than US$4,000 on grey markets from around US$500 before the memory squeeze, according to recent supply-chain checks by Morgan Stanley.

The cost surge threatens the margins that enabled the chipmakers’ early profits. Moore Threads’ gross margin contracted by more than 8 percentage points year on year in the first three quarters as it stockpiled expensive components.

For MetaX, another little dragon, limited scale among tier-one CSPs makes absorbing these spikes even harder, despite raising prices for its flagship C600 processor, according to Morgan Stanley, which has lowered its 2027 and 2028 gross margin forecasts for the chipmaker.

The industry’s challenges extend beyond memory pricing to production and adoption delays.

Although MetaX secured an order for around 20,000 units of its C600A variant from e-commerce giant JD.com, Morgan Stanley said that this provided “insufficient evidence of broader hyperscaler adoption”, leaving the company reliant on government and regional computing projects.

MetaX’s Xiyun C600, which entered mass production earlier this year, pairs HBM3e memory with a precision architecture suited to model training.

Volume shipments for MetaX’s next-generation C700 processor could also be delayed until late 2027, risking a critical window to win over enterprise accounts, the bank said.

Meanwhile, hardware rivalry is intensifying. Last month, Shanghai-based Iluvatar CoreX held a live demonstration of its new Tiangai 300 GPU, claiming it matched or even outperformed Nvidia’s H200 in real-world AI workloads.

Chipmakers focused on application-specific integrated circuits (ASICs) rather than GPUs also face operational strains.

Cambricon Technologies, widely seen as one of China’s leading alternatives to foreign AI hardware suppliers, is grappling with the financial weight of securing its supply chain.

The company’s inventory surged by 3.42 billion yuan quarter-on-quarter to 11.6 billion yuan at the end of September, leading to an inventory impairment loss of 502 million yuan over the first nine months of the year.

Even so, Cambricon has quickly penetrated mainstream cloud infrastructure. Having delivered its MLU692 chip to a major CSP for testing in the third quarter, the company was projected by Morgan Stanley to ship nearly 350,000 units to CSP clients next year alongside the roll-out of its next-generation MLU580.

In a recent research note, Goldman Sachs analysts said long-term survival for Chinese AI chipmakers would hinge on sustaining profitability while improving local fabrication capacity and production yields.

Despite immediate supply challenges, Nomura analysts argued in a report last month that financial markets continued to underestimate “both the breadth of AI demand and the earnings that such strong demand can generate across the value chain”.

They added that sustained cuts to computing commitments or significant delays to funded projects would challenge that outlook.