Always Approaching, Never Arriving

On July 16, a Chinese company called Moonshot released a new artificial-intelligence model named Kimi K3. Two things about it rattled the market. It was very good, near the top of the field, close enough to the best American models that the difference no longer looked meaningful. And it was cheap, offered at a fraction of what the leading U.S. frontier LLMs charge, with the company promising to give the underlying model away for anyone to download and run themselves.  

Within days the market had drawn its conclusion and sold accordingly. The companies that supply the picks and shovels of the AI build-out, the chipmakers and infrastructure names, fell hard. The story told itself. If capable intelligence can be had this cheaply, and from China, then the enormous American investment in building it starts to look overpriced and ill-conceived, and the advantage everyone assumed the U.S. held was a mirage.  

The market inferred the rapid decline of American dominance, the lead once measured in years, now in months, gliding down toward zero as a cheaper Chinese rival closes in release after release.  

I see something else.  

A Caveat Before We Continue  

None of this is a knock on the model. Kimi K3 is a real achievement, and the team behind it is legit. Even the doubters concede as much. Pressed on whether K3 was simply copied from American systems, a senior figure at a leading U.S. lab said plainly that it was not; copying “plays a role” but “clearly it’s not primary.”¹If the only question were whether China can build a top-tier model, the answer arrived on July 16, and it is yes.  

Selling AI stocks was the lazy response to an already hyper-competitive LLM world. A lead you cannot consolidate and make durable is, as I have argued before, an academic exercise. The competition is real, but it settles nothing. Winning means capturing and defending, being able to exploit the advantage. Building a model that benchmarks well is simply the cost of entry for that contest.  

What Has to Be True  

A thesis is only as strong as its attack surface. I have developed a discipline of considering a long idea from the perspective of a potential short seller, hunting weak data points that can be leveraged. This “U.S. fades, China wins” story needs three things to be true at once.  

Trust is a high hurdle. Here it helps to separate two things the selloff conflated, where the work goes and what the leader can charge for it. Cheap models are winning enormous volume, the routine, low-stakes requests a company is eager to fulfill for a fraction of the price. That is real, and it does squeeze what the leaders can charge for the commodity tier. But the work that actually determines business outcomes, the most critical work, stays with verifiable providers.  

Adoption is fickle. Chinese models are already everywhere, tucked inside other products and tools. But everywhere-by-accident is not the same as chosen. Most of that spread happened quietly and without fanfare because compute cost is real and losing competitive edge today trumps future consequences. The moment governments and large enterprises start to care, through new regulation, disclosure rules, or privacy concerns, adoption may reverse, or at least be more carefully considered.  

Staying power is key. What threatens the leaders’ position is not a model reached over the internet, but a downloadable one a rival can take, keep, and build on. In knee-jerk fashion, the market repriced on a benchmark, when models routinely behave differently in the wild than on the test. Independent work documents gaps as wide as a third between lab scores and real performance.² The same model can look excellent on a well-worn test and merely middling on a fresh one built to foil memorization. Further complicating the issue, a model this size will run only where big data centers can host it, so walling it off in-house, the thing that would make it a durable threat, is something most buyers cannot do. And even after the weights are out, confirming actual real-world performance is a matter of weeks of independent testing, not a launch-day headline.  

Whose Frontier Are They Standing On?  

There is a deeper reason that staying power matters, and it has to do with how a fast follower reaches the front of the pack in the first place.  

Real, legitimate engineering is not the whole story. Running right alongside it is a practice called distillation, which means training a smaller, cheaper model by having it learn from the answers of a bigger, better one. Akin to viewing the answer book before the test. This past winter Anthropic went public with what it called industrial-scale efforts by three Chinese labs, including Moonshot, to pull knowledge out of Claude, more than sixteen million conversations run through roughly twenty-four thousand fake accounts.³ Moonshot’s share ran to millions of exchanges, aimed at the very capabilities K3 is now praised for. One of the labs was caught redirecting nearly half its effort at the newest Claude model within about a day of its release.  

This is the part the low-cost story leaves out. The fast follower distills the result at a fraction of the cost and sells the world a cheaper future, one built on the giant’s shoulders and the giant’s spend. The practice is being challenged but is not yet settled, so for now remains an effective way to catch up fast. Effective, but not leadership — simply approaching a frontier only the leader can extend.  

That is the flaw in the bear case. The cheap copy feeds on a frontier only a motivated leader keeps producing, so that ultimate logic leaves nothing left to copy. All of this holds only while the door for this behavior stays open, and it is beginning to close. The major U.S. labs, ordinarily fierce rivals, now share the tools to detect and block the high-volume extraction that distillation requires, and Washington treats the practice as a national-security matter.  

The Takeaway  

Reality is more nuanced than the trade. Frontier LLM providers’ lead is real, and it is shrinking, and both can be true without adding up to the conclusion the market reached. A shrinking lead is not a lost race. A capable model is not a winning one. And an ability borrowed by standing on someone else’s shoulders is a dependency, however impressive it looks on announcement day.  

And there is a simpler point the selloff missed entirely. The market did not just mark down the model makers; it marked down the companies that supply the computing power underneath them, the chips and the data centers. But those companies are not paid on who wins the model or on what a model costs. They are paid on how much the models get used, and a cheaper model, as I have written before, does not get used less. It gets used more — Jevons paradox at work. Whatever the right level of spending, demand for compute is set by usage, not by who trained the model or what it cost to copy. A cheaper model enlarges that demand rather than shrinking it. And the spend the market calls wasteful is exactly what the cheap copy was built from. You cannot call the frontier a mistake and its imitation a triumph in the same breath.  

So the pressure the announcement put on AI stocks, and above all on the compute beneath it, is overdone at both ends, with the model layer sturdier than the market assumed, and the compute layer never the thing at risk. The market mistook a descent for a destination, treating a shrinking gap as a finished one when it was contingent all along. Those who mistake a shrinking lead for a finished race tend to misjudge the ultimate distance.  

Footnotes 

¹ The Hill, “Chinese model Kimi 3 adds pressure on Trump administration’s AI policy,” July 22, 2026.
² DigitalApplied, “LLM Benchmark Methodology 2026: Reading Leaderboards,” May 27, 2026.
³  Anthropic, “Detecting and preventing distillation attacks,” February 23, 2026.