DeepSeek published the recipe and the West quietly cooked it. That's the core of this post: MLA architecture, Compressed Sparse Attention, and the follow-ons in DeepSeek-V4.1-Flash (CSA2, cross-layer cache reuse, causal encoder-decoder, FP4 caching) brought the KV cache to 890 bytes per token — a claimed 437x reduction versus V1, or 389.12 GB down to 890 MB per million tokens. Since KV cache VRAM is a dominant serving cost for long-context workloads like coding, that lands directly on the invoice. The author's evidence for "adoption, not stealing": the pricing. Opus 5.5 cut cache-read pricing 60% ($0.50 to $0.20 per million tokens); GPT-6.1 Sol cut it 80% ($0.50 to $0.10), alongside input and cache-write drops. Both released quietly. The claim is plausible but circumstantial — the source offers no architecture confirmation, just price behavior consistent with cache-efficiency gains. Worth noting the author's framing: export controls forced Chinese labs to prioritize optimization, and they published it anyway. Jensen will explain that the more you compress, the more you save, which is somehow a discount this time. Open question: if the inference-margin story is real, why is nobody at Anthropic or OpenAI saying the word "MLA" out loud?