单条 2026-10-24 1 条帖 8 段 1095 字在 X 打开 ↗

DeepSeek 的 80% 毛利率从哪来：组织形态与「退一步进两步」

DeepSeek 的高毛利率来自独特的组织形态：没有 KPI 壁垒，允许为长期架构做短期退步。

DeepSeek组织推理服务

核心主张

DeepSeek 的高毛利率来自独特的组织形态：没有 KPI 壁垒，允许为长期架构做短期退步。

zartbot 的文章指出：其他实验室各团队 KPI 独立，infra 优化（如 KV cache 压缩）会被算法团队否决；DeepSeek 能妥协、能试验，V4 Pro 的暂时落后是退一步进两步，最终成就 4.1 Flash 与 CSA2。

要点

组织差异

无 KPI 壁垒，infra 与算法团队可以互相妥协。

代价

激进架构试验可能产出暂时落后的模型（V4 Pro）。

回报

退一步进两步：4.1 Flash + CSA2 的进化基础。

结果

80% 毛利率。

金句

DeepSeek’s unique organizational setup is what enabled it to achieve an 80% gross margin
DeepSeek 独特的组织形态，才是它能做到 80% 毛利率的原因。

everyone started saying DeepSeek was past its prime
所有人都说 DeepSeek 过了巅峰。

this was one step back to take two steps forward
这是退一步进两步。

后续观察

CSA2 架构在后续模型上的延续
开源「印钞机」后的生态影响

01A new post from @zartbotF.@zartbotF 的新文章。

02It’s an essay on DeepSeek open-sourcing its much-discussed “money-printing machine.” Zartbot argues that DeepSeek’s unique organizational setup is what enabled it to achieve an 80% gross margin.这篇随笔讨论 DeepSeek 开源其广受讨论的「印钞机」。Zartbot 认为，DeepSeek 独特的组织形态，才是它能做到 80% 毛利率的原因。

03At other labs, for example, each team has its own KPIs, leaving little room for compromise.比如在其他实验室，每个团队都有自己的 KPI，几乎没有妥协余地。

04Even if an infrastructure engineer proposes compressing the KV cache further, they’ll inevitably face pushback from the algorithm team if it hurts model performance.即使基础设施工程师提出进一步压缩 KV cache，只要影响模型性能，就必然遭到算法团队反对。

05But the most interesting point in the piece goes a step further: even if the algorithm team agrees to experiment with a new architecture that enables greater KV cache compression, the experiment may fail to produce a leading model.但文章更有意思的一点更进一步：即使算法团队同意试验一套能实现更大 KV cache 压缩的新架构，这个实验也可能产不出领先的模型。

06We saw this with V4. When DeepSeek V4 Pro came out and lagged behind GLM 5.3 and Kimi K3, everyone started saying DeepSeek was past its prime.我们在 V4 上见过这一幕。DeepSeek V4 Pro 发布并落后于 GLM 5.3 和 Kimi K3 时，所有人都说 DeepSeek 过了巅峰。

07But Zartbot argues that this was one step back to take two steps forward, and that the work played a major role in the subsequent evolution of 4.1 Flash with CSA2.但 Zartbot 认为这是退一步进两步：这些工作为后来 4.1 Flash 与 CSA2 的进化起了重大作用。

08A short but fascinating read. I’d encourage everyone to check it out.一篇短小但精彩的随笔，推荐大家去读。

原帖在 X 打开 ↗

A new post from @zartbotF.

It’s an essay on DeepSeek open-sourcing its much-discussed “money-printing machine.” Zartbot argues that DeepSeek’s unique organizational setup is what enabled it to achieve an 80% gross margin.

At other labs, for example, each team has its own KPIs, leaving little room for compromise.

Even if an infrastructure engineer proposes compressing the KV cache further, they’ll inevitably face pushback from the algorithm team if it hurts model performance.

But the most interesting point in the piece goes a step further: even if the algorithm team agrees to experiment with a new architecture that enables greater KV cache compression, the experiment may fail to produce a leading model.

We saw this with V4. When DeepSeek V4 Pro came out and lagged behind GLM 5.3 and Kimi K3, everyone started saying DeepSeek was past its prime.

But Zartbot argues that this was one step back to take two steps forward, and that the work played a major role in the subsequent evolution of 4.1 Flash with CSA2.

A short but fascinating read. I’d encourage everyone to check it out.