DeepSeek V4 may not have made as much of a splash as the R1 model launched last year, but the company's real strategy is to become the world's cheapest AI infrastructure.
DeepSeek chose to compete by offering low prices. Photo: Bloomberg.
In early 2025, the launch of DeepSeek R1 caused Nvidia to lose $600 billion in market capitalization in a single day. At that time, the global tech world was shaken by this fledgling model from China. A year later, DeepSeek V4 was released and seemingly failed to make an impact.
DeepSeek is no longer a mystery. The V4 model is built on the MoE architecture with 1.6 trillion parameters, but each inference only triggers 49 billion parameters. This "call on demand" mechanism drastically reduces inference costs without affecting performance. The context window reaches one million tokens, allowing users to put entire command lines or reports of hundreds of thousands of words into a single conversation.
DeepSeek acknowledges that the V4 is lagging behind closed-loop models by about 3-6 months. However, when the computing cluster running the Huawei Ascend 950 chip becomes operational later this year, the price of the V4 will drop sharply.
"This is a statement from a company that has already planned its next move, not from a company reacting under pressure," one industry analyst commented.
The issue with the V4 isn't about top performance. It's just good enough to be the industry minimum standard. Analysis from Union Bancaire Privee suggests that the performance gap between the top models will be almost insignificant for the vast majority of real-world users.
When everything is good enough, the battle shifts to price, ecosystem, and who becomes the default choice in the global developer workflow. DeepSeek understands this very well. Their open-source strategy is a way to attract tens of thousands of global developers to build applications on their platform without incurring marketing or sales team costs.
DeepSeek has not yet received external investment. Tencent and Alibaba, two companies that have developed their own AI models, appear to be competing to participate in DeepSeek's first round of investment. This is because DeepSeek's "maximum efficiency plus open source" strategy has the potential to shape the global AI infrastructure standard.