The AI Race Just Got Awkward：西方模型悄然采用 DeepSeek 的 KV cache 优化

If you read the news headlines these days, you would be forgiven for thinking that the Western labs are getting spawn-camped by Chinese labs en masse.

The Distillation Drama

Not a week goes by when Anthropic doesn’t release another article on how the Chinese are distilling their models, becoming a danger to humanity itself, etc.

It’s beneficial for them to say that because it sets the ground for these models to be restrained legally and regulatorily later on.

But it’s clear that the days of mindless distilling are over.

Not just over. The new game in town is adopting Chinese labs’ advances. Note how I call this adoption instead of the more vitriol-infused “stealing” that Anthropic tends to use.

That’s because, unlike the Western companies, the Chinese are pretty much giving away their recipes.

A Different Game

The latest one shamelessly copied without acknowledgement is the breakthrough in KV cache optimizations that DeepSeek has generously shared with the world.

It is a mind-blowing optimization that basically dropped the KV cache footprint for certain use cases that use a long session context, like coding, by a factor of roughly 437x compared with DeepSeek-V1.

They were the first ones to release the MLA architecture, which compressed the cache by roughly 15x, and then followed it up with ‘Compressed Sparse Attention’ and ‘Heavily Compressed Attention.’ The latest DeepSeek-V4.1-Flash pushes it even further with CSA2, cross-layer cache reuse, a causal encoder-decoder architecture, and FP4 caching, bringing the global KV cache down to 890 bytes per token.

Why do these things matter? Because for serving long-context models, one of the largest costs is the VRAM needed to hold this cache in GPU memory.

437 equal-area squares shrink to one: 389.12 GB of KV cache becomes 890 MB per million tokens. Approximately 99.77% less memory.

Below is the graph showing just how crazy this whole thing is:

KV cache memory per million tokens: DeepSeek-V1, 389.12 GB; V3.2, 48.07 GB; V4-Flash, 3.51 GB; V4.1-Flash, 890 MB. The latest cache is 437.2 times smaller than V1.

Follow the Cache Money

Compare that with what the same tier cost roughly two months ago. All prices below are per 1 million tokens:

Pricing per million tokens, before and after. Anthropic Opus 5 to 5.5: input $5 to $4, cache write $6.25 to $5, cache read $0.50 to $0.20, 60% cheaper. OpenAI GPT-5.6 Sol to GPT-6.1 Sol: input $5 to $2, cache write $6.25 to $2.50, cached input $0.50 to $0.10, 80% cheaper.

A Very Quiet Thank-You

All this must mean the Western AI companies are now extremely inference-margin positive.

The constraints on access to advanced GPUs forced Chinese labs to make performance optimization a number one goal, and it shows in the results.

Now I don’t know why they would freely give away such a breakthrough, but they just did, and for once both Anthropic and OpenAI released models that are basically top-tier and are using these optimizations.

They do seem to be a little embarrassed by the copying. Hence the silent releases without much pre-announcement for both Claude Opus 5.5 and GPT-6.1 Sol.

The user reviews have been stellar w.r.t. usage, and the quality doesn’t seem to be that far off compared to their flagship models (Claude Fable 5.1 and GPT-6 Astra).

The cache read costs are the proof of the adoption. They dropped sharply: Opus 5.5 cut cache-read pricing by 60% versus Opus 5, while GPT-6.1 Sol cut it by 80% versus GPT-5.6 Sol’s late-July pricing.

So the Chinese labs have thrown a lifeline to the Western loss-making labs, and I just have no clue as to why.