Why Citi and Bank of America Say Kimi K3 Is Actually Bullish for Memory Chips

On July 16, 2026, Moonshot AI unveiled Kimi K3, a 2.8-trillion-parameter open-weight model that matched or exceeded frontier US models at roughly half the cost. The market reaction was immediate and brutal: the PHLX Semiconductor Index fell 12.5% in a week, Nvidia dropped 2.5% in a single day, and investors replayed the same fear that gripped markets during DeepSeek-R1 in January 2025 — that cheaper, more efficient AI would gut demand for the chips that power it .

Two Wall Street banks took the opposite view. Citi and Bank of America Securities both argued that the selloff was precisely the wrong bet, and they invoked a 160-year-old economic theory to explain why: the Jevons Paradox.

The Jevons Paradox argument in a nutshell

Citi and Bank of America Securities both argue that Kimi K3's efficiency breakthrough will increase total memory-chip demand rather than reduce it, because cutting the per-token cost of inference makes AI dramatically more accessible and widely deployed — a textbook Jevons Paradox outcome. The selloff on July 16 was, in their view, the wrong directional bet .

Citi's position: Peter Lee explicitly invokes the paradox

Citi semiconductor analyst Peter Lee explicitly invokes the Jevons Paradox: improved technological efficiency lowers the per-unit cost of usage, which historically leads to an increase — not a decrease — in total resource consumption . In a Citi Research podcast recorded months before K3's launch, Lee stated directly: "I think this kind of software efficiency isn't a threat to memory. I think it's positive for demand because it makes AI cheaper and more useful, which drives a purchasing cycle of even higher demand for the advanced chips. So I think it's like Jevons paradox" .

His reasoning for why Kimi K3's architecture specifically drives more memory demand:

• 1M-token context window — Each long-context inference call consumes a massive KV cache that must be stored in high-bandwidth memory (HBM) and server DRAM. Cheaper per-token pricing means users will fill that million-token context more often, not less .
• MoE sparsity is a mirage for memory — Although Kimi K3 activates only 16 of 896 experts per token (~50B active parameters), the full 2.8T weights must still be loaded into system memory to support dynamic expert routing. Self-hosting the open-weight release (July 27) will require clusters with terabytes of server DDR5 and enterprise SSDs to hold and page the entire model .
• Agent-based workflows — Agentic, multi-step reasoning loops (which K3 is specifically optimized for) generate many sequential inference calls per task, each consuming memory bandwidth. Cheaper inference makes longer agent chains economically viable, compounding per-task memory consumption .
• Open-weight self-hosting — When full weights go public on July 27, enterprises and researchers will run their own instances rather than relying solely on API access. Each self-hosted deployment requires its own dedicated memory pool (HBM on the GPUs, DDR5 on the host, and enterprise SSDs for model storage/swapping), multiplying the total memory installed base .

Citi has also separately raised memory price forecasts sharply, projecting DRAM prices could surge up to 200% in 2026 and HBM pricing to rise 30% QoQ in Q4 2026, indicating they already expected an AI-driven memory supercycle before Kimi K3 . Citi had also warned as early as September 2025 that both DRAM and NAND flash memory would enter supply shortage by 2026 .

Bank of America Securities' position: Vivek Arya applies the same logic

Bank of America has not explicitly used the phrase "Jevons Paradox" in published commentary about Kimi K3, but their pre-existing and post-selloff calls align with the same logic:

• "Summer reset, autumn rebound" — BofA characterized the semiconductor pullback (the SOX index fell 11% in Q3) as a seasonal mood swing and a healthy reset, not a structural demand reversal. They forecast a rebound in autumn as AI infrastructure buildout resumes .
• Memory is "undervalued and poised for expansion" — In a July 7 report, BofA noted memory chips trade at just 10x earnings, significantly below the broader semiconductor sector, and recommended them as a buying opportunity .
• Long-term AI demand thesis intact — BofA's 2026 semiconductor outlook projected global semiconductor sales surpassing $1 trillion for the first time (up 29% YoY), with memory (HBM, DDR5, NAND) as the primary growth driver. They have consistently argued that more capable, cheaper AI models broaden the addressable market rather than shrink it . Bank of America later raised its 2026 semiconductor TAM forecast even further, to $1.3 trillion .
• Post-Kimi K3, BofA told investors not to panic — On July 17, Vivek Arya attributed the K3-triggered selloff to "seasonal mood swings" and cost-inflation fears in memory, not a demand collapse, advising that the fundamentals for memory chipmakers remain intact .

Why the architecture itself reinforces the bull case

Architectural feature	Why it drives memory demand
2.8T total parameters (MoE)	Even with sparse activation, 2.8T weights must reside in system memory. Open-weight self-hosting requires servers with large DDR5 pools and HBM-equipped GPUs .
1M-token context window	The KV cache scales linearly with context length. Cheaper inference encourages longer prompts, increasing per-query memory consumption .
Agentic / multi-step workflows	Each reasoning step is a separate inference call. More steps per task = more memory bandwidth used per user session .
Open-weight release (July 27)	Enables widespread self-hosting. Every self-hosted deployment adds a fixed memory footprint that a shared API endpoint would have consolidated .
Competitive pricing (~$3/M tokens input)	Historically, every major AI price drop multiplied total token volume far outstripping the efficiency gain .

The historical precedent both banks lean on

The DeepSeek-R1 selloff in January 2025 followed the same pattern: a cheaper, efficient Chinese model caused a semiconductor panic, and then chip demand actually accelerated as deployment broadened . Citi and BofA see Kimi K3 as "DeepSeek 2.0" — the same fear, and likely the same eventual outcome for memory makers like Samsung, SK Hynix, and Micron .

Industry data supports the thesis. TrendForce projected HBM demand would increase 70% year-over-year in 2026 alone, with HBM consuming 23% of total DRAM wafer output . Analysts estimate AI data centers could consume ~70% of high-end DRAM in 2026 . Memory manufacturers themselves — including Micron CEO Sanjay Mehrotra — have acknowledged shortages will persist beyond 2026 .

Even the Korean market recognized the argument: on July 20, the Chosun Ilbo reported that a securities industry memo citing Jevons Paradox caused Samsung and SK Hynix shares to briefly turn positive before the broader selloff resumed .

What the banks are not saying

It is important to note the limits of this thesis. Neither bank claims the selloff was irrational — only that it was directionally wrong. BofA explicitly acknowledged that legitimate "cost inflation fears in memory" were partly driving the decline . Citi's memory price forecasts, while bullish, also imply pricing pressure that could squeeze downstream AI companies. And Bank of America has not explicitly used the phrase "Jevons Paradox" in published Kimi K3 commentary; the alignment of their logic is inferred from their pre-existing and post-selloff calls .

Bottom line

Both banks acknowledge the near-term market fear, but their consistent position is that Kimi K3's lower cost and open-weight release will broaden AI adoption, increase total inference volume, and raise memory demand — not strand it. The Jevons Paradox, first observed in 19th-century coal consumption, may prove to be the most reliable framework for understanding AI's semiconductor demand in 2026.