Shares of Nvidia (NVDA +2.10%) and most of the AI-related semiconductor sector sold off last week after Moonshot, a China-based AI start-up, released its Kimi 3 model.
Kimi made waves across the industry, as the open-weights model displayed impressive performance against even the latest frontier models by Anthropic and OpenAI.
But the knee-jerk reactions to Kimi 3 seem like an echo of the DeepSeek and TurboQuant sell-offs of early 2025 and 2026, respectively. In both cases, innovations that made AI much more efficient didn’t derail the AI build-out; in fact, one could argue they accelerated it by lowering adoption costs.
While these past cases aren’t perfect mirrors of Kimi 3, here’s why Nvidia investors shouldn’t panic over this new model.
Today’s Change
(2.10%) $4.26
Current Price
$207.54
Key Data Points
Market Cap
Day’s Range
$204.01 – $208.63
52wk Range
$164.07 – $236.54
Volume
37.1K
Avg Vol
156.2M
Gross Margin
74.15%
Dividend Yield
0.14%
Why Kimi sent a shudder through U.S. AI stocks
Although Moonshot and other Chinese AI labs may have smuggled in some Nvidia chips illegally, Moonshot likely doesn’t have access to nearly as many Nvidia chips for model training as the leading U.S. labs. There is also some uncertainty about whether Moonshot merely “distilled” a leading LLM from either Anthropic or OpenAI, essentially copying the weights from the U.S. labs.
Either way, Kimi 3 appears to have been trained at a small fraction of the cost of leading U.S. models, leading to panic over whether the U.S. giants should and will keep spending on high-end, very expensive Nvidia GPUs.
Another reason why Kimi may have spurred a sell-off in Nvidia and AI memory stocks is that it displayed a novel innovation called Kimi Delta Attention (KDA). This architecture enables the model to selectively read prior tokens to process new ones, rather than reading all prior tokens. The result is a 75% decline in KV cache, essentially an AI’s short-term memory required to run the model, and a sixfold increase in speed. That means the model requires less memory and processing power, all things being equal.
Kimi doesn’t lower inference requirements as much as feared
Regardless of how Kimi was trained, if consumers and enterprises want to use it, the model has to run. And while KDA certainly makes more efficient use of KV cache, other architectural features make it somewhat compute-intensive, requiring high-end hardware such as the latest Nvidia racks.
First, Kimi 3 is a massive 2.8 trillion-parameter model that requires 1.5 terabytes of high-bandwidth memory. Second, Kimi 3 uses 896 experts in a “mixture of experts” architecture. A mixture of experts means a query can go to a specific, specialized “subnetwork” of the entire model, so each query doesn’t have to run the entire model.
While that theoretically frees up space and lowers speed and cost, Kimi 3’s experts aren’t loaded entirely onto a GPU but rather are split across 16 experts per GPU, requiring at least 56 chips to hold and inference the model. Spreading the experts over more chips is a technique called WideEP.
According to chip research firm SemiAnalysis, this means that to run the model efficiently, one will need high-end chip systems with the required number of chips and associated networking, such as the Nvidia GB300 NVL72 reference architecture. Moreover, SemiAnalysis says that the lower KV cache per chip requires a subsequent massive scale-up in bandwidth to coordinate the dozens of chips required. That means a greater focus on rack-level networking and, therefore, Nvidia’s NVLink technology.
Image source: Nvidia.
Don’t forget U.S. regulations or the Jevons paradox
Finally, even if Kimi does deliver certain efficiencies, many workloads likely won’t be able to run Chinese models, especially if they have been distilled — a fancy word for “pirated” — from leading U.S. labs. Regulations will likely still spur many U.S. enterprises to adopt U.S.-based models, or at least take security precautions that will also increase costs.
Meanwhile, even if Kimi 3 still provides much more efficient frontier-level AI usage, the Jevons paradox, an economic concept that states as technology makes resource use more efficient, overall resource consumption increases rather than decreases, indicates this will only unlock greater adoption and usage, offsetting any efficiencies regarding Nvidia chips or memory.
Just as the DeepSeek and TurboQuant scares of 2025 and early 2026 proved to be buying opportunities in AI names, it appears as though the Kimi 3-inspired sell-off looks to be another such opportunity for long-term investors.
