Moonshot AI's Kimi K3 could encourage a surge in enterprise workloads and provide a long-term tailwind to chip demand
Moonshot AI paused new sign-ups for Kimi K3 over the weekend in response to overwhelming demand and compute constraints.
A cheap yet competitive artificial-intelligence model out of China has reignited the debate over the massive amounts of money that companies are spending on memory components and compute power.
But the emergence of low-cost Chinese AI isn't necessarily a bad thing for U.S. semiconductor heavyweights that have seen their shares soar on strong demand for expensive AI hardware, according to some analysts. They reason that these chip companies could end up benefitting, even as AI costs go down.
Last Thursday, Chinese AI developer Moonshot AI released Kimi K3, an open-weight model that ranked competitively to leading U.S. models across certain benchmarks. With 2.8 trillion parameters, Kimi K3 will be the world's largest open-source model when Moonshot AI shares the model's weights later this month. Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol, which were edged out by Kimi K3 on Arena's Frontend Code leaderboard, are closed-source models.
That a Chinese lab could develop a model with near-frontier capabilities, while having fewer financial resources and facing export controls on Nvidia's (NVDA) latest and most powerful chips, sent shock waves through the market. The Kimi K3 release added momentum to the chip-sector selloff on Friday, helping push the PHLX Semiconductor Index SOX into bear-market territory.
However, skyrocketing demand for Kimi K3 suggests that Chinese AI models may be just as compute-hungry as their Western counterparts. Moonshot AI posted on X over the weekend that exploding demand for Kimi K3 had pushed capacity to its limits, leading the company to temporarily pause new sign-ups to prioritize compute access for current members.
The development is "a positive tailwind for the memory trade," according to Anni Sen, managing partner of BluBird Capital.
Within Kimi K3's 2.8 trillion parameters, the model's active parameters are only roughly 50 billion, according to the open-source community Hugging Face. Parameters are the variables that AI models use to learn and recognize patterns during training, and serve as a sort of memory. While this makes Kimi K3 efficient to run, the requirement to hold 2.8 trillion parameters in active memory means the model would be difficult for enterprises to run on their own servers, Sen told MarketWatch.
Kimi K3 currently charges $15 for every million tokens of output - compared to OpenAI's GPT-5.6 Sol, which charges $30, and Anthropic's Claude Fable 5, which charges $50. In an example of the Jevons paradox - which states that making a resource more efficient to use can increase total demand - Kimi K3's lower token costs could lead to an increase in AI workloads.
"Kimi will generate more use cases" for AI as developers utilize the model's lower-cost structure, Sen believes, increasing demand for inference as well. Companies will increasingly migrate "midtier" AI tasks to cheaper models, while reserving frontier-model usage for more complex multistep agentic processes, Sen said.
"Open-weight models are accelerationist for AI diffusion across the economy by having the entry price for intelligence at a certain level of performance be lower," Nathan Lambert, an AI researcher and founder of the Interconnects AI blog, wrote in a Monday post. Open models offer additional customization opportunities that can make them more useful to specific businesses, he added.
As AI models get larger, the need for memory will grow to support more parameters, Wedbush analyst Matt Bryson said. That could either mean the amount of memory per AI chip increases, or chip clusters would have to get larger for more weights, Bryson said in a Monday note.
Therefore, if Chinese AI models continue to gain traction, Bryson said it would be "arguably good for memory vendors," considering demand for high-performance memory would increase. If chip clusters get larger due to increasing model parameters, he said that could also offer a boost to networking providers.
Micron Technology $(MU)$, SK Hynix $(SKHY)$ and Samsung Electronics (KR:005930) are the three main suppliers of high-bandwidth memory, and shortages of memory components have given the companies a lot of pricing power.
D.A. Davidson managing director Gil Luria told MarketWatch that as the cost to run AI models goes down, demand for compute will likely increase, which would benefit chip makers - especially when it comes to memory chips, which are experiencing a supply crunch.
Joseph DeYonker, CEO of PurePlay ETFs, shared a similar view that as AI models become more efficient and accessible, the surge in user adoption "immediately tests physical infrastructure limits," as in the case with Kimi K3. In his view, that confirms that the long-term growth trajectory for chip makers is intact.
"Running advanced frontier models with expansive context windows places immense pressure on memory and manufacturing pipelines," DeYonker told MarketWatch in emailed comments. A model's context window is the amount of data an AI model processes at one time.
Therefore, he expects companies specializing in HBM and advanced chip packaging to maintain strong pricing power.
Meanwhile, DeYonker also expects designers of graphics processing units and custom AI chips, such as Nvidia and Broadcom $(AVGO)$, to "continue to see robust order backlogs as enterprises and hyperscalers rush to secure the hardware necessary to avoid these exact operational caps."
In his view, the limit on Kimi K3 subscriptions is "a structural validation for the semiconductor sector rather than a cyclical warning sign," and investors should take the ongoing memory and compute bottlenecks as a sign that the chip sector has sustainable earnings-growth potential in the long term.
Shay Boloor, chief market strategist at Futurum, noted that Kimi K3 was already trained before the issue emerged of how to serve the model at scale. That "suggests inference demand is now becoming the binding constraint," Boloor told MarketWatch in emailed comments.
-Britney Nguyen -Christine Ji
(END) Dow Jones Newswires
