Cheap Chinese AI may be a chip-stock spark, analysts say
Moonshot AI’s Kimi K3 rattled chip investors, but analysts say surging demand for the low-cost model could lift memory and compute demand.
By Frankie Delgado · News Reporter
3 min read
Moonshot AI had to pause new sign-ups for Kimi K3 over the weekend after demand for the Chinese AI model ran into compute limits, the company said on X.
That capacity crunch is now feeding a counterargument on Wall Street: cheaper AI may not crush chip demand. It may create more of it.
Kimi K3, released last Thursday by Chinese developer Moonshot AI, is an open-weight model with 2.8 trillion parameters. MarketWatch reported that it ranked competitively with leading U.S. models on some benchmarks, including Arena’s Frontend Code leaderboard, where it edged Anthropic’s Fable 5 and OpenAI’s GPT-5.6 Sol.
The launch unnerved investors because Moonshot AI built a near-frontier model despite tighter access to Nvidia’s newest and most powerful chips under export controls, according to MarketWatch. The release helped add pressure to chip stocks Friday, when the PHLX Semiconductor Index moved into bear-market territory.
But the weekend scramble for Kimi K3 access told a different story to several analysts and investors: even efficient AI still needs a lot of hardware when users pile in.
Cheap tokens, bigger workloads
Kimi K3 charges $15 per million output tokens, according to MarketWatch. That compares with $30 for OpenAI’s GPT-5.6 Sol and $50 for Anthropic’s Claude Fable 5.
Anni Sen, managing partner of BluBird Capital, told MarketWatch that the model could be a positive force for memory-related chip trades. She said lower costs may push developers to create more AI uses and move midtier tasks onto cheaper models, while saving more advanced systems for harder, multistep agentic work.
Sen pointed to the Jevons paradox, the idea that making a resource more efficient can increase total consumption. In AI terms, lower token prices could mean more queries, more applications and more inference demand.
Hugging Face said Kimi K3 has about 50 billion active parameters inside its 2.8 trillion total parameters. Sen told MarketWatch that structure may make the model efficient to operate, but enterprises would still struggle to run it on their own servers because the full parameter set must be held in active memory.
Nathan Lambert, an AI researcher and founder of the Interconnects AI blog, wrote Monday that open-weight models speed AI adoption across the economy by lowering the cost of capable intelligence. He also said open models can be customized in ways that make them more useful for specific businesses.
Memory makers get the spotlight
Wedbush analyst Matt Bryson said in a Monday note cited by MarketWatch that larger AI models require more memory to support more parameters. That could mean more memory on each AI chip or bigger chip clusters to hold more model weights, he said.
Bryson said broader adoption of Chinese AI models would be arguably positive for memory suppliers if it lifts demand for high-performance memory. He added that larger clusters could also help networking companies.
Micron Technology, SK Hynix and Samsung Electronics are the three major suppliers of high-bandwidth memory, according to MarketWatch. Shortages in memory components have given those companies pricing power.
D.A. Davidson managing director Gil Luria told MarketWatch that falling AI operating costs should lift compute demand, which would help chip makers, especially memory-chip companies facing tight supply.
Joseph DeYonker, chief executive of PurePlay ETFs, told MarketWatch in emailed comments that Kimi K3’s sign-up limit shows how quickly broader AI access can strain physical infrastructure. He said advanced models with large context windows put pressure on memory and manufacturing pipelines.
DeYonker said companies focused on high-bandwidth memory and advanced chip packaging should keep strong pricing power. He also said GPU and custom AI-chip designers such as Nvidia and Broadcom should continue to see strong order backlogs as enterprises and hyperscalers try to avoid compute caps.
Shay Boloor, chief market strategist at Futurum, told MarketWatch that Kimi K3 had already been trained before the scale-up problem appeared. In his view, that suggests inference demand is becoming the main constraint.
This story draws on original reporting from MarketWatch.