Jefferies, in a research note reported by ANI on Jun 26, 2026, argues that the emergence of lower-cost AI models is unlikely to reduce AI investment and could increase demand for computing infrastructure. The brokerage highlights Chinese model GLM-5.2 from Z.ai as delivering performance close to leading enterprise models at lower operating cost and quotes the report calling the moment another "DeepSeek moment." Jefferies' report also states that falling token costs are encouraging more firms to run inference on local servers rather than public cloud, and that OpenRouter usage shows rapid uptake of lower-cost Chinese models. The note invokes Jevons Paradox to explain why lower per-unit inference cost could raise overall compute consumption, and it explicitly flags memory chip makers, especially DRAM suppliers, as beneficiaries. Jefferies is quoted as saying there is "zero sign of AI capex slowing."
Paying for LLM inference by the kilowatt-hour instead of per token