Wall Street: Kimi K3 Reinforces Rather Than Reduces Compute Demand
0xBroomberg
Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-source model that tops every overseas closed-source model on coding benchmarks — triggering a semiconductor sell-off. But new reports from UBS, Nomura, BofA, and Citi reach the opposite verdict: K3 is a compute accelerator, not a killer.
Why did the market panic again?
Kimi K3 surpassed every overseas closed-source model on coding benchmarks; the market labelled it "DeepSeek Moment 2.0" and sold semiconductor stocks.
This means → investors reflexively replayed the DeepSeek R1 script: stronger Chinese model = less compute needed = sell chip stocks.
But all four banks reached the opposite conclusion: K3's signature is not "efficiency" — it is "scale," and scale increases compute consumption.
Why would K3 actually consume more compute?
Citi analyst Peter Lee framed his July 19 note around the Jevons Paradox — when a good model gets cheaper, developers deploy more apps, process more tokens, and total compute rises.
K3's inference-side memory demand matches frontier models: the KV cache — temporary memory that stores context during inference — balloons with K3's 1-million-token context window, directly lifting server DDR5 and enterprise SSD demand.
In plain terms = the stronger and longer-context the model, the more memory each inference call devours; large-scale deployment requires 64+ GPU "supernode" clusters, pushing the compute bill up, not down.
How will U.S. frontier labs respond?
BofA analyst Vivek Arya wrote on July 17 that the U.S. labs' answer is "not less compute, but more" — if Chinese open-source models keep closing the gap, OpenAI, Anthropic, and Google must train larger and run heavier inference to stay differentiated.
This reflects an arms-race dynamic: one side's efficiency breakthrough → the other side scales up → industry-wide compute demand spirals higher.
Arya also noted that Google's Gemini 3.5 Pro has reportedly been delayed by several months; frontier leadership is "becoming increasingly hard to defend."
How much can the memory and storage sector earn?
UBS analyst Timo Arcuri's team noted that open-source models are typically more memory-hungry than closed-source ones — 2.8 trillion parameters plus a 1-million-token window mean KV cache demand keeps growing in absolute terms, even after quantization — a technique that compresses model precision to save memory.
This means → the more widely open-source models are deployed, the more rigid the demand for HBM — high-bandwidth memory — and storage becomes; memory makers benefit most directly.
UBS estimates the memory and storage sector's cumulative free cash flow through 2028 will reach roughly 30% of market cap — with Micron alone at 47%.
How is K3 fundamentally different from DeepSeek R1?
All four banks stressed the two models represent different types of shock: DeepSeek R1 signalled "efficiency"; K3 highlights "scale."
K3's feature set — 2.8 trillion parameters, 1-million-token context, always-on reasoning, native multimodality, MoE architecture (Mixture of Experts, where only a subset of parameters activates per task) — pressures inference, memory, networking, and storage simultaneously.
K3's pricing confirms its positioning: input at $3 per million tokens, output at $15; per-task cost roughly $0.94, below Claude Fable 5's ~$2.75 but far above DeepSeek V4 Pro's $0.04. In plain terms = K3 is not chasing the lowest price — it aims to match frontier capability at a lower cost.
Who benefits, and where is the risk?
Nomura analyst Duan Bing reiterated buy ratings on TSMC, ASE, and MediaTek, and flagged Zhongji Innolight and Suzhou Innolight as beneficiaries of supernode-driven networking demand.
Chinese AI models are penetrating global usage fast: per OpenRouter, Chinese models' share of global developer token traffic rose from under 2% a year ago to over 45% today; BofA data shows roughly 55% of U.S. enterprises now subscribe to an AI model, platform, or tool.
BofA flagged one tail risk: if efficiency gains outpace workload growth, infrastructure buildout could see a pullback — this is the critical validation point for whether the current compute-demand thesis holds.
Content is for reference only, not financial advice.