Cheap AI Models Drive Infrastructure Demand, Jefferies Favors Memory Makers

The arrival of cheaper artificial intelligence models is not cooling AI spending. It is likely to push computing infrastructure demand higher, according to a Jefferies report issued this week.

The report points to Z.ai's launch of the GLM-5.2 model, a Chinese AI system that delivers performance near leading enterprise models at a fraction of the operating cost. Jefferies called it "another DeepSeek moment," referencing the price war started by Chinese developers earlier this year. Falling token costs are also prompting companies to shift AI workloads off public clouds and onto local corporate servers, the brokerage said. "GLM-5.2 proves enterprises no longer have to sacrifice intelligence for privacy. We are seeing a massive acceleration in companies pulling their AI workloads out of the public cloud and back onto local corporate servers," the report said.

The mechanism behind the bullish view is the Jevons Paradox, an economic pattern where greater efficiency leads to higher overall consumption. Lower inference costs will eventually increase total demand for computing power, Jefferies believes, and that should strengthen pricing and volume for dynamic random access memory (DRAM) chips. Memory makers are the clearest beneficiaries, the report said.

Jefferies has already acted on that conviction. It added South Korean memory maker SK Hynix and Japanese flash memory company Kioxia to its model portfolios. The brokerage also increased its weighting in Samsung Electronics while trimming exposure to internet companies such as Alphabet and Alibaba. The shift reflects a bet that hardware suppliers will capture more value as AI workloads move on-premise and scale up.

The report also noted strong AI-driven investment momentum in Taiwan. The country's economy, exports and semiconductor capital expenditure continue to benefit from the global AI infrastructure build-out led by companies like TSMC, Jefferies said.

A confirming factor for the thesis would be a sustained rise in DRAM contract prices or memory chip orders from enterprise customers. A weakening factor would be a noticeable slowdown in cloud-provider capital expenditure or disappointing return-on-investment numbers from early AI deployments. Jefferies acknowledged that the biggest long-term risk is whether companies will generate sufficient returns on their large AI investments. The brokerage added that those concerns remain theoretical for now. Investment momentum shows "zero sign of AI capex slowing," the report said.