A new report from JPMorgan highlights a growing performance gap in China's AI sector, with DeepSeek's infrastructure efficiency creating a new low-cost frontier.
Back
DeepSeek V4's 40x Cost Advantage Pressures 2 Chinese AI Rivals, JPM Says
A JPMorgan report finds that DeepSeek’s V4 large language model has a structural cost advantage that is putting pressure on Chinese AI competitors KNOWLEDGE ATLAS (02513.HK) and MINIMAX-W (00100.HK). The analysis, released three weeks after the V4 model's launch, suggests that only DeepSeek's proprietary infrastructure can operate the model at peak economic efficiency, creating a new competitive dynamic in the rapidly growing sector.
The bank's report highlights a structural first-party advantage in how the model handles prefix cache reuse, traffic density, and compute allocation. "In cache-hit input performance, there is around a 40x gap between DeepSeek’s official API and third-party cloud channels," JPMorgan noted. The firm concluded that while model weights can be distributed, the underlying cost curve cannot, giving DeepSeek a significant edge.
According to data from OpenRouter, the launch of DeepSeek V4 has not led to a corresponding drop in usage for competitors like GLM and MiniMax, suggesting the market is experiencing supply-constrained growth rather than zero-sum substitution. The report frames the market with DeepSeek V4-Pro defining the low-cost frontier and KNOWLEDGE ATLAS's GLM-5.1 anchoring the high-preference end, leaving MINIMAX's M2.7 model caught in the middle.
For investors, JPMorgan assigned an "Overweight" rating to both KNOWLEDGE ATLAS and MINIMAX-W, with price targets of HKD950 and HKD1,100, respectively. However, the report stresses that both companies must strengthen their strategic positioning to compete effectively against DeepSeek's cost efficiencies.
Zhipu AI's Premium Under Threat
For KNOWLEDGE ATLAS, which operates the Zhipu AI model, JPMorgan believes its monetization now depends on extending its model leadership. While its GLM-5.1 currently ranks ahead of DeepSeek's V4 in evaluations, justifying a price premium, that lead must widen. To sustain its pricing power, the next GLM version needs to broaden its preference advantage in complex, workflow-related tasks like agent-based coding and long-context reasoning, where the cost of retries and quality is more important than raw token costs. Failure to do so could see it lose price-sensitive customers to DeepSeek.
China's AI Usage Exceeds US for 3 Weeks
The intense competition comes as China's AI usage continues to surge. According to the latest estimates from OpenRouter for the week of May 11-17, China's large model token usage was 1.81 times that of the United States, marking the third consecutive week it has held the top global spot. Chinese models recorded token usage of 7.693 trillion, while US models saw usage of 4.24 trillion. Two of the top three models by global token usage were Chinese, including TENCENT's (00700.HK) Hy3 preview, which ranked first with a 210% week-over-week spike in usage to 2.66 trillion tokens.
MINIMAX, meanwhile, faces incremental pressure on its infrastructure-led value proposition. Historically a competitor on throughput and latency, it must now contend with DeepSeek's low-cost, one-million-context API and a service stack that appears more efficient. JPMorgan suggests the successor to MiniMax's M2.7 model will need to prove it can deliver lower overall costs through fewer cycles and retries to maintain its differentiation.
This article is for informational purposes only and does not constitute investment advice.
