2026-03-25 · 个股论点

讨论 Google TurboQuant 对 DRAM/NAND 的影响，认为更像效率提升而非需求坍塌。

中文翻译

谷歌的 TurboQuant…… 以及它对 $SNDK、$MU、海力士和其他公司的影响：它做的事情是： -> KV cache 内存占用减少 6 倍 -> H100 GPU 速度提升 8 倍它本质上是一个压缩算法。那现在问题来了……它会把存储压下去吗？ -> 大概率不会。不过这也许对 $ARM 和其他公司是利好，因为你可以本地跑 AI，而不是依赖 DRAM-heavy 的数据中心。但话说回来： -> 这基本上就是 DeepSeek round 3。你能让算法更高效，但这并不会替代存储或者 GPU。 -> 它可能会在结构上略微降低 DRAM 需求。 -> 而且到目前为止，好像也只在 Gemma、Mistral 和 Llama-3.1 这些小模型上测试过（而且那篇论文已经发了一年了）另外，市场还把 DRAM 和 NAND 混在一起看……这个算法压的是 KV cache（DRAM），并不会对 NAND 存储有什么作用？不管怎样：算法总会变得更高效。大家老说杰文斯悖论，这没错，因为这只是把用途规模继续放大。真正该看的还是超大规模云厂商的 CapEx 预期，而不是让事情更高效的 Google 算法。我更觉得这是个叙事层面的逆风，而不是对盈利有实质影响。

英文原文

Google's TurboQuant... And it's effect on $SNDK, $MU, SK Hynix, and others: What it does: -> 6x reduction in KV cache memory footprint -> 8x Speedup on H100 GPUs It's a compression algorithm. Now... Will it beat down memory? -> Prob not. Implications might be bullish for $ARM and others though where you can run AI locally, rather than DRAM heavy DCs. However: ->This is basically DeepSeek round 3. You can make algorithms more efficient. But that doesn't replace either memory or GPUs. -> It could structurally (and slightly) reduce DRAM demand. -> think it's only been tested on small models so far like Gemma, Mistral, and Llama-3.1 (and paper's been out for a year) Also, markets conflated DRAM with NAND... this algo compresses the KV cache (DRAM). Doesn't do anything to NAND storage? Regardless: Algorithms will always get more efficient. People keep saying Jevons Paradox, which is true since this just scales use cases. Main thing to look out for is hyperscaper capex projections, not Google Algorithms that made things more efficient. Feels more like a narrative headwind than anything material to earnings.

在 X 上查看原推 ↗