Why DeepSeek’s Permanent Price Cut Aims at a $10 Trillion AI Market

The industry consensus is that DeepSeek wields great influence but does not earn profit because it offers only API charging and no subscriptions; on May 22 it announced a permanent 75% price cut effective June 1.

Girish Dilip Patil, AWS generative‑AI lead for the Singapore region, wrote an article titled “DeepSeek’s $10 Trillion War Strategy,” dissecting the underlying logic from a technical to an industry‑wide perspective.

The low‑price strategy is not merely a traffic grab; it is an “unbinding” that reduces reliance on expensive HBM memory. From V3 to V4, DeepSeek focuses on saving memory and inference cost. Using a 1 M‑token context with 8‑bit KV precision and 16‑bit indexing, DeepSeek V4’s KV cache occupies only 5.48 GB HBM, compared with GLM‑5’s 60 GB and Qwen‑3‑235B‑A22B’s 89 GB.

DeepSeek V4’s MoE, MLA, DSA, CSA, and HCA designs compress long‑context data so it can be stored on cheaper media such as SSD, NAND flash, or LPDDR. This reduces the need for HBM and Nvidia GPUs, which are currently scarce and have lead times of many months.

Lower API prices dramatically cut the cost of experimentation, allowing developers and hardware vendors to adapt with minimal investment. This creates a positive feedback loop: cheaper APIs → higher usage → larger total inference load → greater demand for GPUs, storage, and power. The OpenAI‑AMD strategic agreement, which sets gigawatt‑scale compute deployment milestones, illustrates that AI infrastructure competition is becoming an energy competition.

Consequently, data‑center power capacity, cooling efficiency, and energy cost become core competitive factors. Investments such as CATL’s stake in DeepSeek, VNET’s data‑center operations, and Zhongheng’s HVDC power systems highlight the growing importance of the power side of AI infrastructure.

The financing round, involving the national AI fund, Tencent, and CATL, is portrayed as building an AI‑infrastructure alliance rather than a simple startup raise.

Compared with peers—Zhipu, MiniMax, and Kimi—DeepSeek’s strategy differs: while Zhipu focuses on enterprise on‑premise deployments and MiniMax on multimodal products, DeepSeek aims to become the pricing and technical standard that the entire AI supply chain aligns to, influencing hardware manufacturers, chip makers, and developers alike.