Key Takeaways
JPMorgan Chase & Co. finds that DeepSeek's new V4 large language model holds a structural cost advantage of up to 40 times over rivals, a gap that resets the competitive field for China’s artificial intelligence leaders KNOWLEDGE ATLAS and MINIMAX-W. The performance gap stems from DeepSeek operating its model on its own infrastructure, a first-party advantage rivals using third-party cloud channels cannot match.
"The promotional pricing of DeepSeek V4-Pro now defines the low-cost frontier, while GLM-5.1 anchors the preference end," a JPMorgan report said. "Model weights can be freely distributed, but the cost curve cannot."
The bank's analysis of incremental data since V4's launch shows a 40x gap in cache-hit input performance between DeepSeek’s official API and third-party cloud services. This efficiency in prefix cache reuse, traffic density, and compute allocation establishes a new low-cost tier in the market. According to OpenRouter data, usage of V4 is rising, while incumbent models from KNOWLEDGE ATLAS (GLM), MiniMax, and even older DeepSeek versions have not seen broad-based declines, suggesting the market is expanding rather than consolidating.
The report frames a new strategic challenge for the two main publicly traded AI players. For KNOWLEDGE ATLAS, its monetization now depends on extending its model leadership to justify its higher price. For MINIMAX-W, its historical edge in infrastructure performance is now directly challenged. JPMorgan assigned an Overweight rating to both companies, with a price target of HKD950 for KNOWLEDGE ATLAS and HKD1,100 for MINIMAX-W, signaling confidence in their ability to adapt but highlighting the increased pressure.
KNOWLEDGE ATLAS must now widen its performance lead to defend its pricing. Its GLM-5.1 model currently ranks ahead of DeepSeek's V4 in evaluations, anchoring it as the preferred choice for quality-sensitive users. To sustain this, JPMorgan said the next GLM cycle needs to broaden its advantage in complex, workflow-related tasks like agent-based coding and long-context reasoning, where quality and reliability are more important than token costs. Failure to extend this lead could force it to cede the price-sensitive market to DeepSeek, retaining a smaller, higher-margin business.
MINIMAX-W, which historically competed on throughput and latency, faces more direct pressure. DeepSeek’s new low-cost, one-million-context API on a highly efficient first-party stack neutralizes MiniMax's traditional infrastructure-led value proposition. According to JPMorgan, the successor to MiniMax's M2.7 model will need to demonstrate measurably better task-completion economics—fewer cycles and retries leading to lower overall costs—to maintain its differentiation against the successor to V4.
This article is for informational purposes only and does not constitute investment advice.
