The recent spotlight on KV cache (KV Cache) compression does not constitute an issue that would undermine the medium-to-long-term investment logic for the memory sector
KV Cache image. (Image from datamagiclab.com)
Nvidia’s KVTC (KV Cache Transform Coding) and Google’s Turbo Quant have drawn attention and been cited as a backdrop to the sharp decline in memory semiconductor stocks; however, these technologies already sit within existing research trends, suggesting that market concerns are somewhat excessive.
Son In-jun, a researcher at Heungkuk Securities, explained in a report on March 27 that the core purpose of KV cache compression is to alleviate memory bottlenecks that arise during large-scale inference and the resulting slowdown in AI workload processing.
According to the report, Nvidia’s KVTC can achieve approximately 20-fold compression by selectively retaining only important information, while Google’s Turbo Quant can deliver approximately six-fold compression through random rotation and quantization.
However, Son pointed out that these technologies are not brand-new concepts that suddenly emerged. KVTC was made public in November last year, and Turbo Quant in April last year, and both papers are being revisited as they have been accepted to the AI conference ICLR 2026 being held next month.
In particular, given that similar KV cache compression studies—including KVQuant and KIVI—were published earlier in significant numbers, interpreting this issue as a signal of industry deterioration is a stretch.
Son viewed that if KVTC or Turbo Quant had been overwhelmingly innovative approaches compared with existing techniques, they likely would have garnered far greater attention across the AI industry since last year. Therefore, he assessed that the likelihood of these technologies being widely deployed in the short term is limited, and that improvements in AI performance ultimately continue to occur along the established axes of computing power, data, and algorithms.
He added that it is also worth noting that hyperscalers, including Google, are preparing long-term supply contracts on a 3~5-year basis to address memory supply shortages.
Son identified dispersed market attention, rather than technological change itself, as the main backdrop for the recent correction in the memory sector. As non-memory topics came to the fore in quick succession—such as Nvidia’s Groq3 LPX unveiling, Tesla’s Terafab plan, and Arm’s server CPU strategy—investor attention shifted temporarily. Groq3 LPX, an SRAM-based accelerator, heightened concerns about encroachment on the HBM market, while the Terafab and Arm-related developments were analyzed to have stoked expectations for the leading-edge foundry and server CPU markets, respectively.
Even so, Son assessed that within AI hardware, memory currently exhibits both the highest profitability and the lowest valuation. Coupled with the expansion of long-term supply contracts of up to 5-year duration, which reconfirms memory’s role as a core bottleneck in AI infrastructure, he diagnosed the recent plunge as a temporary adjustment driven by supply–demand dynamics and a shift in attention rather than any damage to fundamentals. He projected that, underpinned by a prolonged upcycle, upward revisions to earnings estimates, and valuation re-rating, the memory sector’s stock prices will maintain an upward trajectory.
Kim Eun-jin
댓글
0