DeepSeek-Led Open-Weight AI Surge Intensifies the Global Battle Over AI Economics

The most striking evidence comes from Vercel’s AI Gateway, a platform that allows developers to access different AI models through a common interface. On August 25, open-weight models accounted for 54 per cent of token volume on the platform, while their share had reached a record 62 per cent several days earlier. On June 24, open models had represented only 28 per cent. DeepSeek-V4-Flash emerged as the leading model by token consumption, while models from Chinese developers StepFun and Zhipu AI, also known internationally as Z.ai, ranked among the most heavily used systems. The speed of this reversal matters because token volume provides an indication of how intensively models are being used inside applications rather than simply how much attention they receive from consumers.

Cost is becoming one of the decisive factors behind that migration. Generative AI applications can consume enormous quantities of tokens, particularly newer autonomous agents that repeatedly reason through problems, generate and revise code, search databases and interact with other software. Small differences in per-token pricing can therefore translate into substantial operating-cost differences at enterprise scale. DeepSeek’s official pricing illustrates the strategy: after its August pricing adjustment, V4-Flash was priced during off-peak hours at US$0.22 per million uncached input tokens and US$0.66 per million output tokens, with higher rates during peak periods. More broadly, industry research cited by Jefferies showed average enterprise AI inference prices falling to roughly US$1.16–US$1.18 per million tokens in early August, compared with US$2.04 at the end of May. Chinese open-weight models have become one of the forces accelerating this global price competition.

The implications extend beyond DeepSeek. Reuters reported in August that Chinese open-weight systems had become increasingly popular in Silicon Valley because they combine low cost, customisability and performance sufficient for many business tasks. Corporate use of platforms offering open-weight and Chinese-developed models has also been rising, encouraging US technology companies to respond. Meta has renewed its commitment to releasing open models, while Nvidia is expanding its own model portfolio. This suggests that Chinese companies are influencing the competitive structure of the global AI industry even without dominating the most powerful computing hardware. Instead of trying to match American companies solely through enormous training budgets, Chinese developers are competing through model efficiency, software optimisation and aggressive pricing.

The distinction between “open-weight” and genuinely “open-source” AI remains important, however. Open-weight developers generally make the trained model parameters available for download, allowing companies to run and modify models on their own infrastructure, while training datasets and significant portions of the underlying development process can remain private. US and European companies also continue to have concerns about data security, regulatory exposure and geopolitical risk when using Chinese technology. One solution has been to run Chinese open-weight models through American cloud infrastructure, keeping sensitive prompts and corporate data away from servers operated by the original Chinese developers. This reduces some security concerns while still allowing businesses to capture the models’ cost advantages.

For investors, the larger story is a potential redistribution of value across the AI industry. If capable models become cheaper and increasingly interchangeable, maintaining very high margins on proprietary model access may become more difficult. The beneficiaries could instead include cloud infrastructure providers, AI-agent platforms, inference services and businesses that can turn inexpensive intelligence into commercially useful applications. Chinese developers such as DeepSeek are accelerating this transition by demonstrating that competitive AI does not necessarily require premium pricing. The development does not mean US frontier models are becoming irrelevant; OpenAI and Anthropic may continue to hold advantages in the most demanding reasoning and coding tasks. But as AI moves from experimentation into mass deployment, businesses increasingly have to ask whether marginal performance improvements justify substantially higher operating costs. That shift from a race for the “best model” toward a race for the best performance-to-cost ratio could become one of the most consequential changes in the global AI market—and one in which China is increasingly setting the pace.