China's chip self-sufficiency push runs into cloud rules and memory prices

Nvidia narrows the cloud route

The guidance is not mandatory. It applies to future contracts, not existing deals. Nvidia also asks cloud customers to disclose who the end users of the chips are. Sales reps and regional teams communicated the instruction verbally, according to the Financial Times. The target is a gap in US export controls. The rules restrict chip shipments but generally say nothing about remote access through cloud services.

That gap matters beyond one vendor. Chinese AI labs can still rent compute abroad. But the hardware they buy at home is getting dearer. One HBM3E memory unit now sells for about $4,000 in China, against several hundred dollars a year ago.

HBM smuggling fills the hardware gap

A smuggler owns a chip design firm. He pairs cheap, low-performing processors with banned HBM from Samsung and SK hynix. Those processors fall under US and South Korean export controls. The rules look at the performance of the server processor, not at the other components attached to it. So the memory slips through. Tom's Hardware reported the details, citing The Information.

An AI chip needs six to eight high-end memory chips. That puts $24,000 to $32,000 of memory into every GPU the broker supplies. Domestic supply is not filling the gap yet. CXMT has begun risk production of HBM3E. It is expected to mass produce by the end of this year, but low yield makes its output expensive.

Empyrean buys a domestic EDA piece

Empyrean said in a Shenzhen Stock Exchange filing on Sunday it would buy a 50.38 per cent stake in Easy-Logic Technology. The price is 218.8 million yuan, or US$33 million. Easy-Logic was founded in Shenzhen in 2021. It builds logic optimisation tools that cut iteration time and resources for chip designers. Empyrean says the start-up's algorithms have won international EDA competitions. Nearly 100 digital chip design firms use its products. The South China Morning Post reported the deal is meant to address a weak link and support Huawei's Tau Scaling Law.

EDA is the software layer where Chinese firms still lean on Western vendors. The deal targets the design stage rather than the fab.

Yields and prices show the bottleneck

Investors have already marked the sector down. China's STAR50 index of AI and semiconductor companies collapsed by 40 percent over four months. That wiped out $480 billion in market value, Reuters reported, as relayed by CleanTechnica. Unitree's shares have halved since its August IPO. CXMT has lost more than a quarter of its value over the past month.

Huawei has pushed the problem downstream. It told customers to find the HBM their servers need themselves. Huawei cannot buy from Samsung or SK hynix, and Chinese memory makers lack capacity.

That instruction is an invitation to the grey market. Tom's Hardware reported that the resulting demand is likely to lift prices further. It is also likely to make smuggling more lucrative.

Regulators eye the next loophole

US authorities had been looking into Chinese AI companies' access to chips through foreign data centers. This followed allegations involving Moonshot AI, the developer behind the Kimi chatbot. Taiwan and Malaysia have been ramping up anti-smuggling activities. The smuggler told The Information he fears South Korean and US authorities will clamp down soon.

The controls have not stopped the trade. They have raised its price. That price now sits inside what Chinese AI companies pay for compute.