Large quote: Jensen Huang simply spent every week watching his firm’s valuation get hit by the specter of a Chinese language open-weight mannequin. His response was to go on the document defending it. Talking to Axios in Fort Price, Texas, on the opening of a brand new part of the Wistron plant that builds Nvidia’s AI infrastructure, the Nvidia CEO stated American firms ought to “completely” be free to run Chinese language fashions. “These Chinese language fashions are wonderful,” he stated. “Open-source fashions which might be wonderful needs to be used.” Requested whether or not China may displace American labs, he was blunt: “Zero chance.”
The timing is what makes it fascinating. Moonshot AI launched Kimi K3 on July 16. Impartial evaluators put it third on Synthetic Evaluation’ Intelligence Index, solely behind Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol, and first on LMArena’s blind Frontend Code Area board.
The market took it about in addition to it took DeepSeek. The Philadelphia Semiconductor Index fell 12.5% in every week, its worst stretch in 15 months. Taiwan’s benchmark dropped greater than 6%, Japan closed down 4%, and Chinese language AI shares really obtained hit tougher than American ones, with Zhipu down as a lot as 30% in Hong Kong and MiniMax off 16%.
“The market misunderstood the influence of DeepSeek the primary time,” Huang instructed Axios, including that Wall Road has “misunderstood the influence of Kimi once more this time.”
His pitch is one Nvidia has made earlier than, delivered with out a lot subtlety this time: “Free AI needs to be nice for {hardware}. Free AI needs to be nice for chips. Free AI needs to be nice for knowledge facilities.” The logic is that as low cost, succesful fashions get used extra, not much less, extra utilization means extra chips operating someplace. He additionally argued open fashions do not cannibalize OpenAI and Anthropic, however they provide individuals a free first style, and loads of these individuals find yourself paying for one thing quicker and extra dependable.
Bloomberg’s evaluation of K3 famous that the place DeepSeek’s story was about cheaper coaching, Moonshot’s is a couple of a lot bigger mannequin that leans tougher on reminiscence infrastructure.
The safety argument, flipped
Huang waved off the concept that a downloaded Chinese language mannequin is a few sort of backdoor to Beijing, declaring you may examine the weights, tweak them, and run the entire thing sealed off from the web if you would like. His argument is that openness really makes issues safer, as a result of extra persons are on the lookout for issues. Lock all the things into one closed system, he stated, and “if all the things simply turns into one single mannequin, one single level of assault, one single supply of failure, I feel the world is far, rather more susceptible.”
– Kimi.ai (@Kimi_Moonshot) July 16, 2026
Then he turned that very same logic on an American firm. Huang stated Anthropic ought to open up Claude Mythos, its cybersecurity mannequin that is at the moment restricted to a vetted group of companions, calling it one thing that “needs to be obtainable as a service” and arguing “holding Anthropic again is just not in the advantage of the USA.” His framing: “Simply because Mythos is just not obtainable, open fashions can be found anyhow.”
It’s price noting that Nvidia is among the companions with entry to Mythos already, by way of Anthropic’s Venture Glasswing program. Glasswing has grown from round 50 companions to roughly 200 throughout about 15 nations, and Anthropic has stated fashions this succesful will possible be broadly obtainable inside 6 to 12 months no matter what it decides.
In the meantime, in Washington…
– Treasury Secretary Scott Bessent (@SecScottBessent) July 22, 2026
Huang’s feedback landed hours after Treasury Secretary Scott Bessent instructed Fox Enterprise the administration is wanting into whether or not Chinese language AI fashions have been constructed on stolen US mental property.
“If we see … that abroad fashions are stealing from our nice firms, we’ve the flexibility to sanction them due to this theft,” Bessent stated, pointing to what he known as “watermarks” of US fashions displaying up inside Chinese language ones, with motion potential inside days or perhaps weeks. He additionally floated whether or not US firms ought to must confide in prospects once they’re operating Chinese language fashions.
– Director Michael Kratsios (@mkratsios47) July 22, 2026
Behind the scenes, Axios says this combat has been simmering for some time. US Commerce has been contemplating blacklisting Chinese language AI labs since final yr, and the White Home had a draft government order that may’ve made US firms answerable for operating Chinese language fashions. All of it stalled, however Kimi’s launch appears to have introduced it again to life.
Jensen Huang break up the distinction on the “they stole our work” query: “Distillation, studying from AI, studying from different sources of data, is prime to intelligence.”
Learn the incentives
Granted, all this commentary is just not coming from a impartial celebration. Nvidia’s China income is mainly zero proper now, down from a enterprise Huang as soon as stated could possibly be price $50 billion a yr. Huang has spent two years arguing towards export controls, so after all he’ll say demand for AI is elastic and that restrictions can backfire. That stated, he has additionally endorsed conserving Blackwell and Rubin out of China’s fingers.
The excellent news is that his core declare may quickly be verified. Each quantity on Kimi K3 thus far comes from Moonshot itself or from early API testing. The total weights go public on July 27, and at that time the benchmarks both maintain up below impartial testing or they do not.
Anthropic has accused Moonshot of coaching on 3.4 million Claude conversations earlier this yr, although analyst Nathan Lambert’s take is that even when a few of that occurred, it isn’t sufficient to clarify how good K3 really is. Individually, Epoch AI estimates the hole between the very best open and closed fashions is now all the way down to round three months.
If that hole retains shrinking, the actual query stops being whether or not US firms needs to be allowed to make use of Chinese language fashions, and begins being whether or not banning them would even do something as soon as the weights are already sitting on servers all over the place.
