Key insights
- The author argues that market reactions to AI advancements like DeepSeek and TurboQuant are often driven by superficial headline analysis, leading to incorrect investment decisions in the semiconductor sector. They contend that TurboQuant's impact on HBM demand is limited, particularly in the context of training spend and upcoming fab capacity increases. The author suggests that the market overreacts to news, creating cyclical volatility.
https://nanonets.com/blog/google-turboquant-ai-memory-crunch/
Last year Jan when DeepSeek dropped, entire semiconductor market bled and the investor thesis was wrong. Now last month when TurboQuant dropped, market bled again. I am sure the thesis is wrong this time too, because people don't even care to read beyond the headlines and what a new algorithmic development actually means (it's not even that difficult to understand tbh).
People went gaga over DeepSeek because it's an efficient AI and people assumed demand will fall because now AI can be made cheaply. But when inference got cheaper it expanded who could afford to deploy AI at all. So memory demand drastically rose. That's what happens when something gets cheaper, people use more of it, not less. The sector recovered.
With TurboQuant it's even simpler. The algorithm only compresses KV cache and has negligible impact on training memory where the actual majority of HBM demand comes from. The $180B hyperscalers are spending on memory this year is mostly training spend. Also it's just a research paper as of now, that too sitting since 2025, even Google hasn't deployed it widely. The memory crunch ends when new fab capacities come online in 2027-28. An algorithm doesn't matter much here.
It just frustrates me that we're clearly stuck in a loop. News drops, headlines go crazy, people panic sell without reading past the abstract, stocks bleed, thesis turns out to be wrong, stocks recover, and then three months later we do it all over again.
At some point you'd think the market learns. Apparently not.