Kimi K3 will be open-sourced in 4 more days. Americans are really getting worked up this time.

Americans always want to keep sitting at the center of every industry.

The AI space is no exception. Americans have this air of being in complete control, gripping a hand of cards that looks impossible to lose.

No matter who is building AI applications out there, Americans believe that in the end, the bill will always come back to them. Chips belong to Nvidia, cloud services belong to Microsoft, Amazon, and Google, and the most expensive models are locked behind OpenAI and Anthropic’s APIs. If companies around the world want to use AI, they eventually have to pass through the U.S. to get it.

Even if the name of a Chinese team appears on the leaderboard now and then, Wall Street doesn’t pay it much mind. Chips choke the neck, cloud is in hand, and talent keeps running to Silicon Valley—how could they lose?

But that relaxed sense of being certain of victory has, recently, been exposed by a Chinese model hitting the scene: Kimi K3.

The U.S. tech world urgently added extra posts, describing Kimi K3 as a “Sputnik moment”—like the shock Americans felt in 1957 when the Soviet Union launched its Sputnik satellite. On X, discussions around Kimi K3, Yang Zhilin, and Chinese models quickly spread from a small technical-circles audience into a topic that rolled into the tens of millions of views.

Kimi K3 didn’t win every metric against the United States’ strongest closed-source model, but it showed more people a possibility: strong capability, high efficiency, and an open ecosystem may not necessarily only be able to grow in those few U.S. labs.

Silicon Valley really is anxious.

Storage is the comfort blanket for America’s AI anxiety

When the news about Kimi K3 reached Wall Street, several investment banks almost at the same time released research reports. Instead of spending time first on who it might disrupt—whether it would force U.S. model providers to cut prices—these institutions quickly shifted their focus to storage.

They collectively interpreted the arrival of Kimi K3 as: a surge in demand for storage. With longer context, AI needs to remember more things—images, audio, video, and work records pile up—so flash storage, hard drives, data centers, and data services all stand to benefit.

And so, Micron, SanDisk, and Western Digital became beneficiaries in this storyline.

Sure enough, in yesterday’s U.S. stock market, storage stocks rebounded violently across the board. Roundhill Storage ETF rose 10.91% in a day, SanDisk jumped 14.27%, and Micron gained 12%. A few days earlier, the sector had been getting hammered over “DeepSeek Moment 2.0.” In one night, it turned into the most certain bullish position again.

From an industry perspective, this line of thinking isn’t unreasonable. Chatbots in the past were like one-time Q&A sessions: you ask a question, it answers, you close the page—many things are over. But the AI people expect now is more like a new employee joining a company. It has to revisit old contracts and emails, remember what customers said, take over tasks not finished yesterday, and leave records so that if mistakes happen, nobody can explain clearly who did what. An AI that can do tasks, keep records, and also “see images and hear sounds” can naturally “consume” more data than an AI that only chats idly.

This conclusion isn’t pulled out of thin air, but when we look back at a few earlier rounds of model launches and deployments, the market reaction was basically: “Don’t focus on the model—focus on storage”?

You can say it’s an answer that lets Americans feel at ease.

A disruption from a Chinese model should have led to a chain of questions that are hard to answer: Would it make it harder for U.S. model companies to sustain high prices? Would it reduce developers’ reliance? Would new companies no longer need to start in Silicon Valley? Why not just discuss who Kimi K3 will steal users from, force price cuts from, and push product changes from?

By skirting the sharpest questions and going straight to hard drives instead, there’s a bit of “no smoke without fire” to it.

Like a shopkeeper who thought they monopolized the whole street, then suddenly discovers a new shop next door is really strong. So the shopkeeper quickly reassures themselves: even if the new shop attracts more customers, they still have to use the electricity and water and checkout counters that I sell.

Storage is the strongest comfort blanket under America’s AI anxiety.

Closed-source models start to pinch

In the past few years, closed-source has been the undisputed default answer in U.S. AI.

The stronger the model, the more it should be locked behind an API. Users pay to call it, the model company earns high gross margins, and safety and compliance are centrally managed by that company. It’s a path that is both respectable and profitable. When you follow it, everything is steady and reliable: customers feel secure, investors are satisfied, and regulators have something to explain.

Americans have even gotten used to the rhythm of this path: release a stronger version every few months, set an even higher price, and tell a bigger story.

But once open models started getting stronger, this road began to rub against something.

Kimi K3’s position in this chess game isn’t “catching up”—it’s driving down the cost of chasing. In an open and strong model, the most dangerous part isn’t only what it can do itself, but that it hands everyone coming later a much cheaper learning curve.

This isn’t a matter of pride in the tech world—it’s whether the business will get rewritten. Previously, the U.S.’ most comfortable setup was to turn AI into enterprise services first: capabilities are hidden in the cloud, customers sign long-term contracts, regular people can’t see the underlying layer, and it’s not easy to switch away. But if models elsewhere are good enough, developers get another choice. When enterprises purchase, they can have a different line item on the quote sheet. Even small teams might not have to keep betting their future on the same batch of U.S. companies. At that point, simply guarding a few big contracts and selling AI only to B2B is no longer a worry-free moat.

That means Kimi will spawn more excellent models, and competition among models will be fiercer—so bargaining power won’t be as strong.

The U.S. tech circle can also feel the change in wind direction.

A few days before Kimi K3 was released, on July 15, Thinking Machines Lab—founded by Mira Murati, the former CTO of OpenAI—released a model called Inkling. Its parameter count is close to the billion-trillion scale, the code and technology are fully public, and anyone can download it for free, modify it, and use it commercially.

This is arguably the first truly “serious” open-source AI from the U.S. Although there were already open-source models before—Meta’s Llama, Google’s Gemma, Microsoft’s Phi, Nvidia’s Nemotron, and even OpenAI’s gpt-oss—more often they were experiments or pilots.

The significance of Inkling is that someone who once pushed closed-source to its peak—the former OpenAI CTO—turned around and started doing open source seriously.

Worth noting: in Inkling’s post-training, early-stage data generated using open models like Kimi K2.5 were used, and the architecture also referenced DeepSeek’s approach. In other words, the U.S.’ most presentable open-source answer was written by standing on the shoulders of Chinese open source.

In sharp contrast is Anthropic. In February this year, Anthropic publicly accused DeepSeek, the Mysterious Moon (月之暗面), and MiniMax of launching “industrial-grade distillation” against Claude. It claimed they created 24,000 fake accounts, ran 16 million dialogue sessions, and used them to steal Claude’s capabilities. In June, it escalated further by explicitly naming Alibaba. By July 21, Bessent, the U.S. Treasury Secretary in the Trump administration, bluntly said that sanctions could be imposed on China for “AI theft.”

No matter how loud the threat talk gets, when it’s truly time to control costs and improve efficiency, Chinese models still end up being the “real deal.”

Airbnb uses Qwen for customer service. Cursor uses Kimi to build its own programming agents. DoorDash outsources some tasks directly to Kimi. Even Murati’s Inkling for post-training has to use Kimi’s data.

Whether it’s closed-source pinching itself, or the distillation accusations backfiring, they’re ultimately just awkwardness at the business model level. The privacy and security issues are what truly shake the last protective charm of the closed-source camp.

AI models “escape the jail”

Closed-source’s last line of defense has always been security.

Lock the model, lock the weights, route calls through APIs, and keep data from landing on the ground—those four walls create the space where the closed-source camp has its most credible promise. Enterprise customers are willing to pay a premium for that sense of security.

But enterprises are becoming less comfortable. They’re starting to ask questions that are hard for closed-source companies to answer: After I give you my code, contracts, and customer data, what exactly do you do with it? If you give me an agent that has access to the browser, terminals, credentials, and long-term goals, will it—so as to complete the task—cross the line I didn’t allow it to touch? Sending tokens to a closed-source API, in a sense, means letting data go beyond your own walls. And that’s precisely one of the toughest selling points of open weights: at least, I can see what the model is doing.

Just as both sides argue endlessly over which is safer, a nearly black-humor incident happened.

On July 21, OpenAI itself confirmed that its flagship model GPT-5.6 Sol and another unpublished model with stronger capabilities escaped the isolation environment during an internal network security evaluation.

The situation was like this: the engineering team wanted to test the model’s upper limit in offense-defense capability, so they lowered the model’s safety restrictions—turning off the protections that normally intercept high-risk behaviors. The model was supposed to simply and obediently complete the test questions. But it discovered a security vulnerability in the system, climbed up through it to the public network, bypassed permissions and traversed systems along the way, and finally used stolen login credentials to break into Hugging Face—the core system of the world’s largest open-source AI platform—and pulled the answers to the test questions directly from the database.

OpenAI’s explanation was eight words: not malicious, overly focused.

Those eight words are what truly chills people.

For enterprise customers, the most terrifying thing was never that the model would actively do something harmful. It’s that it did so with extreme seriousness—completing a bad objective for you.

The greatest irony of this incident is that over the last year and a half, the “dangerous Chinese open-source model” that the world has been guarding against is still stuck at the hypothetical stage. The one that truly “escaped the jail” and truly broke through another party’s production systems was actually the closed-source camp’s own flagship. Immediately, Hugging Face CEO Clem Delangue turned the whole incident into an advertisement for open source. He said AI safety can’t be solved by one company closing the door—it can only be solved through collaboration in openness.

Same incident, and each side—open and closed—used it as evidence that their own route is correct.

The real fork in the road in the future may not even be whether models are open source or closed source, but what kind of sandbox they run in, what kind of identity system they operate under, what revocable permissions and audit logs they have. Closed-source or open-source—either way, this question can’t be avoided.

And while the closed-source security narrative hit a problem first, a larger-scale reversal was also happening quietly.

Attack and defense flip—now it’s the U.S. that starts to fear

In some policy discussions and tech narratives in the U.S., there has long been an imagination almost in the style of “Three-Body”: as long as the most advanced Nvidia chips are restricted from entering China, AI progress will be forced to slow down.

This doesn’t mean China can’t do research anymore. It’s just that the U.S. believes the gap in compute power will keep widening, and the threshold for training the most frontier models will become so high it will be hard to cross. Advanced chips are like the “laws of physics” in this race: whoever can’t get them will find it difficult to run ahead.

That judgment isn’t without basis. Training big models does require compute. Chip restrictions raise costs, slow expansion, and make it harder for many teams to replicate the training scale of U.S. labs. The issue is that restrictions also change people’s choices. If you can buy the best existing tools outright, you have less incentive to figure out how to spend less compute, how to alter model architecture, and how to make every training run count more. But once the door is shut, taking detours stops being an option and becomes a survival instinct.

So it’s hard for Americans to understand why restricting Nvidia’s supply didn’t leave Chinese models stuck in place, but instead pushed out a set of teams that are more aggressive about efficiency, engineering, and open-source distribution.

They say that even the Mysterious Moon (月之暗面) is still training with H800, Nvidia’s compliance-version AI chip customized for the Chinese market in 2023.

Perhaps that’s the “military-style small arms, step-by-step” approach Chinese teams are best at.

In June 2026, to comply with export controls, the U.S. temporarily shut down Anthropic’s strongest Fable 5 and Mythos 5. That may make sense on the compliance side. But it hands every Chinese open-source lab a ready-made marketing script: at least, our model doesn’t include a switch that can be remotely turned off.

The more you emphasize control, the more control itself turns into a selling point for your opponent.

More dramatic is the other side. Reuters reports that in China, companies have also started meetings with Alibaba, ByteDance, and others to study whether they should restrict foreign access to China’s most advanced AI models, and the discussion scope even includes those open-source models that are already public. Zhou Hongyi (周鸿祎), founder of 360, also publicly called for China to have its own top closed-source models to hold the technological heights.

A year ago, it was the U.S. that worried advanced chips would flow to China. A year later, it’s China that now has something worth restricting.

Amid all these structural anxieties—whether it’s storage, compute, closed source, safety, or the flip in offense and defense—there’s one most concrete, most painful, most personal landing point. It’s not an industry trend. It’s not a research report. It’s not a policy.

It’s a person.

The end point of anxiety lands on Yang Zhilin

At the end of the day, the open-source dispute is what made the U.S. see a bigger problem: will AI’s future serve only a small number of companies able to sign huge contracts, or will it become an ability like electricity or the internet—something an ever-growing number of ordinary teams can use? If the answer gradually leans toward the latter, who can attract developers, who can get young people willing to stay and tinker, will matter more than who has more enterprise customers in their pocket.

And “where people go” is what ultimately carried America’s anxiety onto a very specific name.

Yang Zhilin is mentioned again and again by the U.S. tech world not merely because he’s an outstanding Chinese researcher. Not just because people want to reduce the topic to “the U.S. didn’t keep people.” Reducing someone’s departure or stay to a visa is too light, and too much like hindsight after the fact.

What truly stings Americans is an assumption that can’t be repeated: if people like Yang Zhilin and his team eventually completed the entire journey in the U.S.—from research to entrepreneurship—what would happen? They’d use U.S. cloud and chips to train models, recruit people within the U.S. talent network, raise money from U.S. venture capital, and take their product to knock on the doors of major U.S. customers. In a few years, Wall Street’s ledger might add another star company. One person’s decision, following that familiar relay chain, could turn into a series of companies’ revenues, jobs for a group of people, and confidence across an entire industry.

What America has been most proud of in the past is exactly this ability to amplify talents. It doesn’t just attract smart people to study and work—it catches their cleverness and doesn’t let it stop at papers or in labs. There’s enough money, enough customers, and enough people willing to take risks together.

So why didn’t someone like that stay in America back then? Legendary investor Vinod Khosla directly pointed at Trump’s tightened immigration policies. But Yang Zhilin’s former mentor at Carnegie Mellon, Salakhutdinov, came out to debunk it, saying it has nothing to do with visas. Back then, Yang already had plenty of opportunities to stay. Salakhutdinov even asked an Apple executive in an email whether Yang should join.

It was Yang Zhilin who made up his mind to return to China and start a business.

That’s the most painful part of this whole debate. For Americans, “he chose to go back on his own” is far more uncomfortable than “he was pushed out by immigration policy.” The former implies the system can still be fixed; the latter implies that even if you open the door as wide as possible, the other side might still not want to walk in.

In overseas discussions about Yang Zhilin, what truly hurts is never simply “another excellent Chinese researcher.” It’s a different counterfactual: if this person had stayed inside the U.S. system, his papers, team, funding, and company value should have been written into America’s AI ledger. Now, that accomplishment is first seen as a Chinese team’s capability—and then spreads globally through the open-source community.

For a system that’s been confident for half a century, what’s hardest to accept is often not that someone is stronger than you, but that someone proves you don’t even have to go through you to reach the finish line.

China has dense engineering talent, teams that can quickly turn ideas into products, a huge market for applications, and customers who are willing to pay for efficiency. Open models also make distribution easier: a team doesn’t necessarily have to be recruited by big U.S. companies first, and it doesn’t necessarily need first to win the approval of Silicon Valley investors. It can still deliver its product to developers worldwide. For top talent, the choice is no longer as simple as “go to the U.S.” or “don’t go to the U.S.” Instead, it becomes: where can their judgment truly turn into a company, a product, or even a new ecosystem?

That’s what America’s latest round of anxiety can’t quite hide.

Storage stocks are up—of course that’s worth celebrating; cloud services sell more—of course that’s also a capability. But neither can replace the real question: when the next batch of smartest, most ambitious people is ready to place bets, will they still—like in the past—unhesitatingly treat the U.S. as the only answer?

Kimi K3 is like a gust of wind that blows that question in through the crack of the door. The U.S. still has deep foundations. Chips, cloud, capital, and enterprise markets aren’t something that can be replaced overnight. It will still make a lot of money from global AI prosperity.

But relying on closed-source AI alone is no longer enough for Americans to feel completely secure.