DeepSeek Turns Quality Into Margin War

DeepSeek matters most because it turned Chinese foundation models from a quality race into a margin war. Once one vendor showed it would sell inference near cost, Moonshot could no longer rely on premium pricing for long context alone. That pushed competition toward coding, reasoning, multimodal features, and distribution, where users pick the model that feels best inside the product they already use.

• DeepSeek made the price shock concrete. Its API docs advertised a promotional 0.1 yuan per million input tokens for cache hits through February 8, 2025, then later moved to higher standard pricing. That kind of anchor resets what developers think model calls should cost.
  2 deepseek 1 deepseek
• Moonshot is especially exposed because its early edge was long context and a consumer app, Kimi. When rivals match context length and go cheaper, the defense has to come from better coding, reasoning, multimodal workflows, and tighter product packaging, not from raw token economics.
  3 sacra
• MiniMax shows the other path. It spreads across text, video, audio, and music, so a customer can buy one stack for many generation jobs. DeepSeek is the purer model competitor, because it pressures the core API business directly and trains buyers to compare on cents and benchmark scores.
  4 sacra 3 sacra

The next phase is likely a two front fight. DeepSeek will keep squeezing the base model layer on price and benchmark performance, while Moonshot tries to move up into products where users pay for finished workflows instead of tokens. In that market, the winners will be the companies that make the model disappear inside a useful daily tool.