The new Chinese model scares Wall Street?
The market might be spooked, but the real story is elsewhere
Headlines about Kimi K3 created the impression that the new Chinese model was responsible for the drop in technology stocks.
But those who only look at the headlines miss the bigger picture.
News about Kimi K3 was published on Friday, and precisely on that day, we saw for the first time buyers entering and relative strength in technology stocks that had been hit during the week.
Therefore, it's hard to argue that the model itself was the trigger for the wave of declines.
But what's much more interesting is that according to SemiAnalysis's analysis, the impact of Kimi K3 might even be positive for AI infrastructure.
The first reason is the model's size.
With over 2.8 trillion parameters, Kimi K3 is one of the largest AI models built to date.
Such a model does not run on regular hardware, but requires enormous computing systems, precisely the area where Nvidia's GB200 and GB300 NVL72 systems provide a significant advantage.
The model's architecture also plays in favor of infrastructure providers.
The Linear Attention technique reduces the need for KV Cache, but the WideEP mechanism distributes hundreds of "experts" among a large number of GPU processors.
This means an enormous demand for bandwidth between processors, and therefore systems with extremely fast communication become critical.
The memory market also isn't left out. The model weights alone occupy more than 1.5 terabytes of HBM memory, leaving almost no room for KV Cache, the result is a shift of some data to DDR5 and NVMe drives, which increases demand for memory and fast storage.
Perhaps the most impressive data point comes from the company that developed Kimi K3, which clarified that to fully utilize the model's capabilities, a rack containing at least 64 chips is required, this is not a system that reduces hardware demand, it's a system built on enormous amounts of hardware.
And above all looms Jevons' paradox:
The more efficient and cheaper AI becomes, the more companies adopt it.
In the long run, efficiency does not reduce demand for infrastructure, it actually increases it. More usage means more GPUs, more HBM and DRAM memory, more communication equipment, and more storage.
From a technical perspective, it's impossible to ignore that the market broke significant levels, including the 50-day moving average. On the other hand, key support levels still hold, and on Friday, initial signs already appeared that buyers are returning to affected stocks.
The coming week is expected to be one of the most volatile of recent times, with a combination of geopolitical tensions and the earnings season for tech giants.
The market is looking for a reason to fear, but the headline isn't always what tells the real story.
The market loves drama, but in the end, it prices infrastructure, not headlines.