The market's instinct on Moonshot AI's Kimi K3 was to reach for the DeepSeek playbook and sell anything connected to AI hardware.
That instinct may be aimed at the wrong target.
The Beijing-based startup unveiled Kimi K3 on Friday, claiming performance close to Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol at a fraction of the cost.
The reaction was brutal: Bloomberg's Asian semiconductor index slid more than 6%, Nasdaq 100 futures dropped 2%, and Chinese rival Z.ai lost as much as 30% in Hong Kong, its worst session since listing in January.
But the comparison with early 2025, when DeepSeek's cheap reasoning model wiped hundreds of billions of dollars from Nvidia's valuation, obscures a crucial difference.
DeepSeek's breakthrough was about doing more with less compute.
Kimi K3 does the opposite on at least one dimension: at 2.8 trillion parameters, it is the largest open-weight model ever released, nearly triple the size of its trillion-parameter predecessor K2.
Larger parameter counts mean more memory is needed simply to hold the model's weights.
That leaves two options for anyone deploying it: fit more high-bandwidth memory onto each accelerator, or build bigger chip clusters with more accelerators to spread the load.
Either outcome is good news for memory vendors, because both routes lift demand for high-performance memory.
The second route arguably goes further, boosting accelerator counts and networking requirements as well, which would flatter Nvidia and the advanced packaging lines at Taiwan Semiconductor Manufacturing Co rather than threaten them.
The question is whether Chinese models actually gain the traction to move the needle, and there are still genuine doubts on that front.
Kimi K3 sits below Fable 5 and GPT-5.6 Sol in overall rankings, even if it beat Claude Opus 4.8 and GPT-5.5 across coding and agent evaluations.
Its full weights are not published until 27 July, so independent verification of Moonshot's benchmark claims remains pending.
Polymarket still puts Moonshot at single-digit odds of holding the leading model position by year-end.
If the traction does come, though, the memory complex looks well placed to absorb it.
SK Hynix controls roughly 58% of the high-bandwidth memory market and its chief executive has said HBM demand will persist beyond 2030.
Micron, whose market capitalisation recently passed $1 trillion, forecasts the HBM addressable market growing from about $35 billion in 2025 to around $100 billion by 2028.
Memory stocks were caught in last week's downdraught alongside everything else AI-adjacent.
If the logic of bigger models holds, that sell-off looks less like a verdict and more like an entry point being mispriced in real time.
