Summary: A sharp decline in memory-related equities after Google introduced TurboQuant has prompted Bank of America to call the move a buying opportunity. The firm contends that efficiency advances like TurboQuant do not erase the need for high-bandwidth memory (HBM) because AI capital expenditure - rather than single efficiency measures - ultimately determines hardware demand.
Shares of Micron Technology fell 7% on Thursday amid the selloff, and the Philadelphia Semiconductor Index (SOX) dropped nearly 5% as investors reacted to the TurboQuant announcement. In a Friday note, BofA analyst Vivek Arya argued that concerns about a structural reduction in AI memory demand are overstated.
BofA highlighted that Google described TurboQuant as a technique that "effectively reduces KV cache memory footprint…by up to 6x without sacrificing accuracy." Still, the bank emphasized that such compression techniques are not new, noting Google has published similar approaches over the past 18 months.
The firm also pointed to concurrent company-level behavior and hardware trends to underline continued memory demand. Google raised its 2026 capital-expenditure outlook by 100% to about $180 billion, and BofA noted that Google's next-generation TPU v7 will feature "6x more HBM at 192GB vs. its prior TPUv6…at just 32GB." The bank further observed that forthcoming chips from other major vendors including Nvidia, AMD and Meta incorporate higher memory across generations.
BofA's central view is captured in the line: "AI capex remains the ultimate proof point of AI spend/demand, not efficiency measures." The firm added, "We continue to believe demand for AI memory remains strong," and suggested that model-efficiency gains are more likely to enable larger model sizes rather than reduce overall memory requirements.
As context for why short-term fears can prove temporary, BofA recalled the DeepSeek episode in early 2025, when similar efficiency concerns faded and Chinas AI capital expenditure still rose 66% in 2025. Based on this assessment, the bank identified preferred areas of exposure as AI compute, semiconductor capital-equipment (semicap) suppliers, networking and memory vendors.
Specific companies flagged by BofA as relevant to these exposure themes include Nvidia, Broadcom, Advanced Micro Devices, Lam Research, Applied Materials, Marvell Technology and Credo Technology.
Key points
- Market reaction: Micron shares fell 7% and the SOX declined nearly 5% following headlines around TurboQuant.
- BofA view: Efficiency measures like TurboQuant are unlikely to replace demand signaled by AI capital expenditure, which the bank regards as the definitive indicator of memory needs.
- Sector focus: Bank of America favors AI compute, semicap equipment, networking and memory names, citing rising memory on next-generation chips and elevated cloud capex guidance.
Risks and uncertainties
- Efficiency headlines - Compression techniques such as TurboQuant can trigger investor concern and near-term valuation pressure in memory and semiconductor equities.
- Demand sensitivity - If AI capital-expenditure trends were to weaken materially, the long-term demand outlook for HBM and related hardware would be less certain; the article references capital expenditure as the ultimate proof point.
- Timing and persistence - Past episodes like the DeepSeek story in early 2025 show efficiency worries can be temporary, but the duration and market impact of any similar episode are uncertain.
The analysis presented is based on BofAs note and the market moves cited therein. It reflects the bank's reading of cloud capex guidance, vendor roadmaps showing increased HBM per chip generation, and historical observations about how efficiency narratives have interacted with capital spending. The commentary identifies specific companies and sectors that the bank views as positioned to benefit from continued AI memory demand.