PARIS, October 6 — The economics of artificial intelligence are undergoing a significant transformation, driven by the rapid emergence of Chinese open-weight AI models that are approaching Western capabilities while offering substantially lower prices, according to a new research report from Natixis Corporate & Investment Banking. The report, authored by Eric Benoist, Tech & Data Research Specialist, and Rita Boutros, Tech & Data Research, explains that the AI market is naturally dividing into tiers, with different models suited to different workloads.
The report notes that Chinese companies such as DeepSeek, Alibaba, Z.ai, and Moonshot have developed models that challenge the dominance of Western AI systems, prompting businesses to reconsider their approach to artificial intelligence investment and implementation.
How Open-Weight Models Work
Open-weight models, defined as AI systems whose core components are publicly released, allow users to download, run, study, and modify them on their own computers. This fundamental difference from proprietary systems creates new possibilities for implementation and customization. Chinese developers have advanced efficiency through several technical innovations including ‘efficient attention,’ ‘mixture-of-experts’ architectures, ‘multi-token prediction,’ and ‘quantization.’ These techniques enable powerful AI systems to operate on standard, less expensive hardware, dramatically reducing the infrastructure requirements that have traditionally made advanced AI accessible only to well-funded organizations.
While these innovations are not exclusively Chinese, they have been effectively combined and deployed at scale by Chinese firms. The report notes that domestic competition and government support, including investment funds and computing subsidies, have accelerated development and deployment.
The gap between the release of leading Western models and competitive Chinese alternatives has narrowed dramatically, from months to weeks, sometimes days. This compressed timeline has made it increasingly difficult for Western companies to maintain technological superiority in the AI space.
Understanding True AI Costs
Comparing AI costs solely by token price is insufficient, according to the report. Open-weight models may be free to download but can require substantial infrastructure investments for operation, including GPUs, networking, electricity, and engineering resources. For near-frontier models, self-hosting infrastructure costs can reach millions of dollars.
While application programming interfaces reduce upfront investment, they introduce usage-based costs that can accumulate over time. Licensing terms also affect the economics, particularly for large commercial deployments where proprietary restrictions may apply.
The ultimate value of any AI model depends on the specific task it’s designed to perform. For straightforward work like document summarization or data extraction, lower token prices yield direct savings. For complex multi-step AI agents, however, reliability at each step compounds, making the cost of successfully completing an entire task a more relevant measure than token cost alone.
As the AI landscape continues to evolve, businesses will need to carefully consider their specific requirements when selecting between different model types and implementation strategies.
