[Yao Yao Jiang] Has Chinese AI Caught Up with the US? The Computing Bill Behind Kimi K3, and Which Side Is Jensen Huang On?

This video provides an in-depth analysis of the rise of the Chinese AI model Kimi K3 and its significance in the US-China AI race. The video points out that while Kimi K3 performs excellently in some tests, its success does not mean that Chinese AI has comprehensively surpassed the United States. The article dissects Kimi K3's Mixture-of-Experts (MoE) architecture, long-context processing capabilities, and agent performance, while exploring the limitations of leaderboards, the true meaning of open-source models, and the massive computational costs required to run large models. At the same time, the video analyzes the interests and narrative strategies of various parties involved in the Kimi K3 event, including Moonshot AI, Beijing, US media, Washington policymakers, and NVIDIA CEO Jensen Huang, emphasizing that computing power, advanced chip manufacturing capabilities, and sustained exploration capabilities remain the keys to the AI gap between the US and China.

Video link:

Full text:

Hello everyone, I am Yao Yao Jiang. Over the past two weeks, an AI model from Beijing called Kimi K3 has suddenly shaken Silicon Valley. US media outlets claim that China has erased the US lead in AI, and White House science advisors warn that the US might lose the AI race. Jensen Huang, meanwhile, says that China's open models are excellent and that US companies should use them boldly. At the same time, Chinese media have woven Kimi and DeepSeek into a single narrative: the chip blockade has failed, and Chinese AI is about to comprehensively overtake the US. It sounds like Washington, Silicon Valley, and Beijing have reached the same conclusion, but are they really describing the same thing? Or are different interest groups, for their own various purposes, collectively amplifying the same anxiety? Today, we will not rely on foreign slogans or counter-propaganda. We will deeply analyze Kimi K3's leaderboard, the so-called myth of low-cost open source, hardware bills, and training compute, tracing the interests behind every claim to answer these three questions: Where exactly is Kimi strong? What is China truly constrained by? Does one model catching up to the US mean that Chinese AI is about to comprehensively surpass the US?

The protagonist of this round of anxiety on both sides of the US-China divide is called Kimi K3, developed by Beijing-based Moonshot AI. On July 16, Kimi K3 was officially released, and the figures announced by Moonshot AI were quite astonishing: the total model size reaches 2.8 trillion parameters, with approximately 104 billion parameters activated for each response. It can process both text and images simultaneously, accommodating up to 1 million tokens, meaning it can read very long documents and large codebases at once. The company subsequently released the model weights and technical reports, allowing external developers to download and deploy it. What truly ignited public opinion was the leaderboard; in the Arena's front-end programming tests, Kimi K3 surged to the top. In several agent, software engineering, and comprehensive capability tests, it also entered the global top tier. Its API price is also lower than the strongest commercial models in the US. Once the news broke, US tech media quickly used a highly impactful headline: China has erased the US lead in AI. White House science advisor David Sachs warned the US that if it cannot continue to expand infrastructure, energy, and technology exports, it risks losing the AI race.

Nvidia CEO Jensen Huang said Wall Street misjudged once when DeepSeek appeared, and has misjudged again facing Kimi. He believes that China's open models are exceptional, and US companies should certainly use them. Chinese media are even less likely to let this opportunity pass; the story of DeepSeek had just ebbed when Kimi took over. The US blocks advanced chips, yet China uses fewer resources to build world-class models. US models are expensive and closed-source, while Chinese models are cheap and open-source. If development continues at this speed, comprehensive surpassing seems only a matter of time. There is a fact that must be clarified first: Kimi K3 is neither a scam nor a paper tiger built on hype. A 2.8 trillion-parameter mixture-of-experts model requires considerable algorithmic and systems engineering capabilities to function properly in training communication, long-context, and agent tasks. Kimi taking the lead in some tests also demonstrates that Chinese AI teams are already able to compete in the world's top tier. However, acknowledging that Kimi is strong is entirely different from accepting that China is about to comprehensively surpass the US. A Chinese athlete defeating US competitors in a few matches can prove that this athlete has reached a world-class level, but it does not automatically prove that China has comprehensively surpassed the US in sports, scientific research, training systems, venues, commercial leagues, and all events.

Model leaderboards measure a specific version of a model's performance on a set of prescribed tasks; they do not measure the manufacturing capability of advanced chips, how much computing power the country possesses, or who has the ability to fund the next round of tens of thousands of experiments. Jumping from Kimi winning a few tests to China erasing the lead misses precisely the part we are investigating today. To see clearly how this jump happens, we must first ask: who needs this story? This news is not the result of a few people conspiring in a backroom; more commonly and powerfully, several forces each starting from their own interests ultimately converge into a bizarre tropical storm. The first force comes from Moonshot AI and Beijing. Moonshot AI is a commercial company that needs capital, talent, customers, and developers. Training a model like Kimi K3 requires massive capital investment, and the company must prove that it is not just another follower after DeepSeek, but a frontrunner capable of building a global technology ecosystem. The higher the leaderboard ranking, the more shocked the media, and the more international developers download it, the more competitive it becomes in financing and market competition.

For the Chinese Communist Party (CCP), desperate striving has another layer of value. Since 2022, the US has continuously restricted advanced AI chips, semiconductor equipment, and high-bandwidth memory from flowing to China. What Beijing needs to prove most is not just that Chinese companies have made good products, but that the US blockade has failed. As long as a Chinese model approaches the US frontier, Beijing can rewrite a company's technological achievements into a systemic victory. The more the West besieges, the stronger Chinese innovation becomes; US models are closed and expensive, while China gives technology away to the world for free. Therefore, Moonshot AI needs a story of commercial success, while Beijing needs a story of national rise. The two are not entirely identical, but they are highly consistent in amplifying the significance of Kimi K3. The second force comes from US media. Kimi K3 reached a leading level in some public tests, but different tests have methodological differences. This is an accurate headline, but few people would sensationalize it. Doesn't a headline like "China erases US AI lead" sound more like a historical turning point? The media likes to translate complex technological competition into horse racing: who is first, who is second, by how much time, and when will the leaderboard be overtaken?

It also happens to provide seemingly objective data. But once numbers are detached from testing conditions, they are easily abused. Model companies can choose the most favorable items, testing frameworks affect scores, and allowing varying amounts of inference time also changes results. More importantly, leaderboards compare publicly released products rather than the total capabilities of two countries. To verify that this is not our bias, I conducted a uniform-caliber sampling survey of the New York Times, The Economist, the BBC, and Bloomberg over the past year, searching China economy and technology reports month by month, randomly sampling a maximum of ten articles per media outlet per month, coding a total of 374 headlines. Judging solely by the headlines, I classified whether China was portrayed as rising in capability, worsening in problems, neutral, or a combination of positive and negative. The result is actually not that Western media are universally hyping China up. Among 213 economic headlines, 88 were negative and 28 were positive, with critical reports more than three times the number of positive reports. What is truly abnormal are the reports on AI and high technology: among 127 headlines, 40 were positive capability narratives, while only 25 were negative.

And the differences are highly concentrated: Bloomberg's ratio of positive to negative reports on US-China AI and high-tech is 18 to 9. What about The Economist? It is 12 to 4, the BBC is 7 to 8, and The New York Times is 3 to 4. In other words, this wave of the idea that Chinese technology is about to overtake the US was not manufactured uniformly by four media outlets, but rather amplified by business and financial media and The Economist. After being stripped of question marks and premises by social media, it turned into widespread anxiety across the entire internet. The third force, meanwhile, comes from Washington. Debates among policy factions in the US are extremely fierce, but the narrative that China is catching up is useful to almost all factions. Those advocating for deregulation can argue that American companies can no longer be tied down by cumbersome rules, while those advocating for the expansion of energy and data centers can argue that China will not wait for US approval processes. Those advocating for industrial subsidies can demand more government funding, while those advocating for stricter controls can demand further blockades on chips and models. The policies they propose are not necessarily wrong—China's challenge truly exists—but we must recognize the incentive structures behind them: the more urgent and critical the China threat is made to appear, the easier it is for these policy proposals to secure budgets and support.

Consequently, the narrative that China has a very strong model naturally gets upgraded to the idea that the US might lose the future. The fourth force comes from Jensen Huang and his NVIDIA. On the one hand, Jensen Huang constantly emphasizes that Chinese engineers are excellent and that China is only a short time behind the US, while on the other hand, he opposes excluding China from the US technology ecosystem. His reasoning is that if the US does not provide chips and platforms to the world, China will build its own alternative system, ultimately causing CUDA and US technology to lose global influence. This judgment makes strategic sense, but it also completely aligns with NVIDIA's commercial interests. NVIDIA does not sell chatbots; it sells the computers that run chatbots. The more the world believes that AI competition is fierce, the more countries need to build data centers, the more popular open models become, and the more enterprises will purchase their GPUs. The closer China appears to get to the US, the more urgent the lifting of chip restrictions against China seems. Therefore, the current wave is not simply Chinese propaganda deceiving the US; the reality is much more interesting: Beijing needs a victory of breaking through the blockade, Moonshot AI needs global momentum, the US media needs drama, Washington needs crisis mobilization, and NVIDIA needs continuously growing computing demand.

Different people are saying similar things while pursuing completely different goals. Once we understand this, we can return to the technology itself. So, what exactly did this Kimi win, and what did it not win? Let's first look at what truly makes Kimi worthy of attention. Traditional large models are like a company where, whenever a problem is handled, almost all employees must hold a meeting at the same time. The larger the model, the more expensive this approach becomes. Kimi K3 adopts a Mixture of Experts architecture. You can imagine it as a company with many specialized departments: legal issues are handed over to the legal group, programming issues to the engineering group, and image issues to another group of experts. Therefore, the entire model has 2.8 trillion parameters, but how many are activated each time? Only about 104.2 billion. This does not mean the remaining parameters are useless; on the contrary, the massive expert pool provides capacity, while selective allocation reduces the burden of each computation. Distributing so many experts across different chips and making them rapidly exchange information is a difficult systems engineering task. Kimi's second highlight is its context: 1 million tokens is not simply stuffing a book into a chat window, but rather requiring the model to find relevant information across multiple materials, maintain task objectives, and collaborate with external tools.

This is very attractive for analyzing codebases, legal documents, and research materials. The third highlight is agentic capability. The AI receiving the most attention today is no longer just answering questions, but accepting a goal, reading documents on its own, invoking tools, writing code, testing results, and continuing to modify based on errors. Kimi has achieved good results in frontend programming and some agentic tests, demonstrating that it has real competitiveness in the race to get things done. But the limitations of leaderboards also start here. First, leaderboards measure selected task anomalies. Frontend programming tests can tell us whether a model can create web pages according to requirements, but they cannot tell us whether it is equally leading in scientific reasoning, cybersecurity, biological research, robot control, and military planning. Just as a car can achieve the lowest fuel consumption on city roads, it does not prove it will perform best on a race track, in the snow, or in heavy-load transport. Second, the scores on leaderboards do not come solely from the model. Kimi K3 is paired with Kimi Code during testing, OpenAI is paired with Codex, and Anthropic's models are paired with Claude Code.

These tools determine which files a model can read, how it searches, when it executes, and how it handles code failures. The final score is the result of the underlying model, software framework, prompts, computation, budget, and safety rules working together. American models sometimes refuse to answer due to cybersecurity or dangerous task concerns; such refusals may cause them to lose points in certain capability tests, but this does not necessarily mean they lack the ability. Conversely, a model that almost never refuses might score higher, but it could also pose greater risks. If these conditions are not unified across different tests, rankings can hardly be interpreted as pure differences in intelligence. Third, the tests published by model companies are inherently selective. This is not a problem unique to Chinese companies; when American companies release new models, they also pick the metrics where they perform best. Truly reliable judgments must be based on multiple independent tests, real user experiences, and stable performance over a period of time, rather than just a single chart from a press conference. Fourth, public products are not equivalent to a nation's entire technological reserve. We do not know the exact level of models that have not yet been released by OpenAI, Google, and Moonshot AI, so we cannot rely on imagination to claim that the U.S. must be hiding something stronger.

However, one cannot use public leaderboards to declare that China has eliminated all of America's leads either. What we can compare are only the products placed on the counter on a given day. Therefore, Kimi's performance can support three conclusions: China possesses world-class AI talent, Chinese teams have mastered the ability to train ultra-large Mixture-of-Experts (MoE) models, and Chinese models can already compete head-to-head with public American products on several important tasks. But it cannot support three other conclusions: that China has mastered the world's most advanced chips, that China possesses total computing power equivalent to that of the U.S., and that China has achieved comprehensive dominance in next-generation AAA research. This is precisely the key sleight of hand in this public opinion manipulation: it packages proximity at the product level as a transcendence of the entire system. And this sleight of hand is easy to succeed because of an extremely attractive term: open source. Kimi and DeepSeek are often called open-source models; this term has become deeply ingrained, but it is not accurate in news headlines. The source code of ordinary software is like a complete recipe; you can see every step, know what ingredients were added and how they were processed, and you can modify it according to your own needs and then recreate it.

AI models are different. When training begins, there are hundreds of billions or even trillions of parameters in the system that have not yet formed capabilities. It reads massive amounts of data and adjusts these parameters through calculation after calculation; the final set of numbers obtained is the model weights. Weights are more like a cake that has already been baked. You can take the cake home, cut it, decorate it, or even create new products based on it. But just getting the cake does not mean you know where the flour came from, what equipment was used in the kitchen, how many times it failed, nor does it mean you can rebuild the same production line. Kimi K3 has indeed made very valuable things public: model weights, technical reports, architecture descriptions, and partial running tools. Developers can download the weights and modify, quantize, and deploy them within the scope of the license. This is certainly much more open than closed models that can only be called via a company's API. But Moonshot AI has not made public the training data, the entire data processing workflow, the records of every experiment, or the complete training pipeline sufficient to fully reconstruct Kimi. The outside world can study the model it created, but cannot replicate from scratch how Moonshot AI made this model.

DeepSeek is in a similar situation; it has made its weights, papers, and part of its inference code public, which is commendable in terms of engineering transparency, but it has not handed over all its training data and the complete production process to the world. Therefore, a more accurate name is actually an open-weights model. This is not a play on words. Open weights bring three truly important freedoms: First, data can remain local; hospitals, banks, governments, and enterprises do not have to send sensitive information to the servers of foreign companies. Second, developers can adjust the model; it can be fine-tuned into a legal assistant, a programming tool, or a local language model, and content restrictions can also be changed. Third, a country does not have to rely forever on the API of a single American company; even if commercial relationships or international politics change, as long as the weights have been obtained, it is theoretically still possible to continue running. But open weights do not provide three other things: they do not let the outside world know all the training materials, so it is difficult to thoroughly audit whether the model used copyrighted content, or whether it absorbed a large amount of output from other models, which is what everyone talks about every day as model distillation.

It did not disclose the complete training process, so third parties cannot determine which techniques are original and which are the result of repeated trial and error. Most importantly, it did not provide the chips, data centers, and funding required to train models of the same level to those who download it. Open weights allow China to rapidly spread the usage rights of its models globally without automatically granting other countries China's research and development capabilities. This is also why Beijing favors this strategy; it can build a developer ecosystem dependent on Chinese models, expand its technological influence, and simultaneously retain the most valuable training secrets. However, this strategy has another frequently overlooked beneficiary: although model files can be downloaded for free, the machines that actually make them run are by no means cheap.

Just how expensive are the machines that actually make Kimi run? Let's do the math. Kimi K3 has 2.8 trillion parameters; you can think of parameters as the experience records left behind after model training. Even with high compression, where each record takes up only half a byte, the entire model still requires about 1.4TB of space, equivalent to nearly 300 high-definition movies. But being able to store a file on a hard drive does not mean the model can function properly. A hard drive is like an underground warehouse: it has large capacity and is cheap, but moving things is relatively slow. The high-speed memory on a graphics card is like a workstation at a chef's fingertips; space is expensive, but the materials the model needs immediately must be placed there. If most of the model's content remains on the hard drive, and it has to stop and move data every time it generates a few words, it might be able to start, but it will be too slow to have any practical value. Moreover, the model also needs to remember the ongoing conversation; the longer the files you upload and the more tasks you give it simultaneously, the more high-speed memory it requires. Kimi claims to be able to process 1 million tokens. However, if you really want to utilize this capability, the hardware pressure will continue to rise.

Some might ask, if it only activates about 104 billion parameters to answer a question each time, why does it need to load the entire 2.8 trillion parameters? We can imagine it as a company with hundreds of specialized departments; when solving a problem, perhaps only a dozen departments are actually working. But before a problem arises, the system does not know whether it needs programming, images, or mathematics this time. Therefore, all departmental data must be placed in a position where it can be accessed rapidly. This is why successfully starting a model and using it as smoothly as GPT are two different things. Some people online have run DeepSeek or even Llama on ordinary computers. Technically, this might be true, but the method usually involves further compressing the model or placing part of the content on the hard drive to be moved slowly. In the end, the model does speak, but the speed is slower, the materials it can handle are shorter, and it cannot compare to the user experience of commercial websites. Furthermore, many of the models running on personal computers are not the complete flagship models with 671 billion parameters mentioned in the news, but smaller versions distilled from the large models. It's like a car company applying a few racing technologies to a family car; the family car does gain some racing performance, but that doesn't mean you bought a professional race car for the price of a family car.

If you want to run the complete model normally, the hardware bill will be completely different. Because the complete DeepSeek and high-performance solutions showcased are servers equipped with eight high-end graphics cards, each costing around 200,000 yuan. Another solution previously announced by DeepSeek used a total of sixteen high-end graphics cards across two machines. The graphics cards and servers alone could cost millions of yuan. Add in high-speed network storage, power supply, and cooling, and 3 million RMB is not an exaggeration. The overall scale of Kimi K3 is even larger than DeepSeek's flagship version. Although it saves on computation by only calling a portion of the parameters each time, if you want to achieve GPT-level response speeds locally, the actual investment is likely even higher. Therefore, we take 3 million as a very conservative starting point. Now, let's convert this money into subscription fees familiar to ordinary people. ChatGPT Plus and Claude Pro currently both cost $20 per month. Based on an exchange rate of 6.76 RMB to $1, that's about 135 RMB per month. How many years could 3 million RMB pay for a person's subscription? Over 1,800 years. 3 million is just the starting price to buy the machine; what about the future?

You still need to continue paying for electricity and maintenance costs. Therefore, for ordinary people, moving a complete Timmy home for 3 million to get a daily experience similar to GPT or happiness is a losing proposition. Of course, not everyone can afford that 3 million. This is like someone giving you the blueprints for a large passenger plane for free—it truly costs nothing, but you cannot fly sitting on a blueprint. You still have to buy engines, build the fuselage, rent a hangar, and be responsible for maintenance. Subscribing to GPT or CLUB is like buying a plane ticket. You do not need to own an airplane; you only need to pay the price of a ticket to use the entire system already built by the airline. So-called open-source and open weights provide the right to build your own plane, not a free plane handed to you that can take off immediately. Therefore, saying China gave away world-class A3 for free to everyone is only half right. What is free is the model files, not the computing power to run the model. What can be downloaded is the training result; what cannot be downloaded are graphics cards, server rooms, and electricity. However, as long as governments, banks, and large institutions decide to move models back to their own server rooms for data security, they must purchase more graphics cards and servers.

Thus, a seemingly contradictory result appears: Chinese companies make their models open and free, yet the global demand for purchased computing equipment might actually increase. What is given away for free by the dark side are model weights, while NVIDIA sells the machines that make those weights truly run. Looking back at why the Chinese community is so enthusiastic, it is easier to understand. After frantic releases, some investors worried that NVIDIA would be impacted. If Chinese companies use fewer resources to make cheaper models, does it mean there will no longer be a need for so much expensive allocation? When Deepak appeared, the market experienced the same panic. However, his brother's answer was: you got cause and effect backward. When the cost of using a commodity drops, total demand sometimes does not decrease, but increases. Lighting is a classic example. Lightbulbs have become increasingly energy-efficient, yet this has not caused people to use less lighting; because the cost of use dropped, city streets, stores, and homes actually installed more lights. AP might be the same—if an inference becomes ten times cheaper, enterprises will not be satisfied with saving on their original bills, but will instead integrate APIs into more services, such as customer service, programming, design, healthcare, automobiles, and robotics.

In the end, the total computing energy consumed might be even greater. For OpenAI and loved ones, models are usually products. For NVIDIA, models are first and foremost towering demands. Closed-source APIs concentrate computing in the hands of a few cloud giants, while open weights spread the urge to deploy to every organization hoping to master its own data. Governments want sovereignty, banks want local servers, hospitals are unwilling to upload patient medical records, and the military is even less likely to hand over files to foreign companies. Every such decision can translate into a batch of orders for GPUs, networking equipment, and enterprise software, right? Fret is especially in NVIDIA's interest. It has 2.8 trillion parameters and 1 million peek-context emphasizing programming and agent services. Although model weights are free, the more popular it becomes, the more high-end hardware worldwide will likely be used to run it. Because data center revenue for fiscal year 2026 grew 68% year-over-year, the company's profit core is no longer gaming graphics cards, but computing platforms. It sells not only GPUs, but also high-speed interconnected server systems and software ecosystems. Chinese advocates of open models certainly have technical philosophies, but they also advocate for a world composed of more models, more data centers, and more towering consumption.

His advocacy for continuing to sell U.S. chips to China follows the same dual logic. Strategically, he believes that if Chinese developers continue to use cry and NVIDIA platforms, U.S. technology can maintain global standards. If the U.S. completely withdraws, China will be forced to foster alternatives like Huawei. Commercially, China was originally a market potentially worth tens of billions of dollars. Lifting restrictions can maintain the U.S. ecosystem while also restoring revenue for NVIDIA. Therefore, emphasizing that China is only slightly behind, that Chinese models are extremely excellent, cannot be taken as a final verdict given by a neutral technical merchant. His observations are indeed more profound than most, but he also has a clear direction of interest. This does not mean everything he says is false. Open weights can indeed lower software monopolies, protect local data, and promote research and competition. Chinese engineers are also truly excellent. The problem is not that the Chinese scholar told lies, but that the part of the truth he repeatedly emphasized happens to lead toward the future NVIDIA most hopes will emerge. OpenAI hopes the whole world will continue to buy its intelligence, while NVIDIA hopes that everyone wanting to possess intelligence will buy a computing factory.

When we see this clearly, the fact that Darcy also admits China is about to surpass is no longer a decisive piece of evidence. It only shows that the Chinese model is already powerful enough to generate global computing demand, and this is precisely because of Darcy's business. However, those who support the theory of China's surpassing still have a stronger card. Time's API price is lower than leading models in the United States. DPC has even shaken the market with low prices, which is a real competitive advantage for Chinese APIs. But the word deviation hides at least four different accounts. The first is the API price seen by the user, which is how much it costs to generate a certain amount of text. The second is the direct cost of running a model each time, including GPUs, electricity, and server depreciation. The third is the cost of training a model. The fourth is finding the right technical path. The total cost of all previous failed experiments, talent, data, and infrastructure. The first is the easiest to promote at a press conference, and sometimes the third is announced. But what truly determines whether a laboratory can continue to innovate is often the fourth. A new restaurant can temporarily sell a dish that costs 100 for 60 to attract customers, it can also be subsidized by investors, or it might only offer low prices on its most advantageous dishes.

Prices are real, and customers do save a little money, but it cannot prove that the kitchen really only spent 60. There are multiple possibilities for Chinese patriotic companies to adopt low prices to coexist. Mixture-of-experts architectures reduce the amount of computation per run, quantization reduces memory pressure, and engineering optimization improves efficiency. China's market competition is fierce, companies are willing to accept lower profits, investors may subsidize prices, and enterprises may also use low prices to compete for developers and build ecosystems before looking for revenue. Therefore, scrambling to be cheap can prove its product competitiveness, but it cannot alone prove that all its costs are lower than those of the United States. The low-cost story of the fourth piece has also experienced similar misinterpretations. The company once announced that the V3 formal training stage used about 2,800,000 and 800 beer hours. The media converted this stage into millions of dollars, and then wrote that DPC only spent a few million dollars to replicate OpenAI. But this figure is not a company's total investment. It does not include early research, failed experiments, data preparation, existing server procurement, personnel costs, nor does it represent the cost of building a cutting-edge laboratory from scratch.

What it truly proves is that DPC completed an important training session with high engineering efficiency, rather than that a cutting-edge API can detach from large-scale force. Time is the same. 2.8 trillion parameters, long context, and large-scale reinforcement learning cannot emerge out of thin air. Technological progress can increase the output of each allocated resource, but it will not make the training machine disappear. There is also a more sensitive issue here: distillation. The so-called distillation is having one model learn the output of another strong model. You can think of it as a student repeatedly reading a master teacher's problem-solving demonstration to train their own problem-solving habits. The industry generally uses distillation, and it is not a cultish technique in itself. The controversy lies in whether the student violated terms of service, massively obtained answers, copied protected achievements, and concealed the source upon release. The U.S. government has accused the Vietnam dark line of potentially using model outputs to train Timmy, and is investigating whether violations were involved. Currently, public materials are not enough to turn this accusation into a final verdict, but it reveals an advantage of the chasers. The pathfinder paid the cost of exploring uncharted territory, and latecomers can follow up faster along papers, model outputs, and public tools.

This capability is certainly worthy of respect, but it is not the same concept as independently pushing forward all technological frontiers. A student can master the knowledge system that a teacher spent 10 years building in less time, which proves the student is smart and also proves the effectiveness of knowledge dissemination. But it does not prove that the teacher's initial 10 years of exploration are no longer needed. Therefore, algorithm efficiency and computing power are not an either-or choice. Better algorithms make each experiment more valuable, and more resources allow laboratories to try more directions simultaneously. China is good at catching up with fewer resources, while the United States still has the ability to pay for more exploratory failures. To judge whether this gap truly exists, the most worth listening to is not the declaration of victory on the public stage, but how China's cutting-edge laboratories describe this calmness themselves. In May 2026, DPC founder Wen Feng allegedly held a nearly four-hour closed-door communication with potential investors. Meeting minutes were later leaked, and DPC did not officially confirm their authenticity. But the discussion of technical and operational details therein was quite specific. In public opinion, DPC is often used to prove that computing power is no longer important, but this record conveys the exact opposite message.

According to the records, DPC is estimated to be about 2 years behind the cutting-edge level in the United States, with hopes of narrowing the gap to six months to a year. The company's computing power scale is only about 1/20th that of top U.S. laboratories. The real bottleneck is not a lack of excellent talent in China, but rather that researchers lack sufficient computing resources to verify their ideas. So, why does computing power limit innovation? Because training frontier models is not like manufacturing according to a mature blueprint. Researchers might have ten new architectures, twenty data schemes, and dozens of sets of training parameters. Most attempts will not succeed. Laboratories with sufficient computing power can conduct experiments simultaneously, quickly eliminate failed paths, and then invest resources into the most promising directions. Teams with insufficient computing power must choose more cautiously, or even abandon high-risk schemes because they cannot afford a single failure. This is like drug development; the intelligence of scientists is certainly important, but if one company can only conduct ten experiments while another can conduct 200, the latter has a higher probability of discovering an effective drug. When facing investors, Liang Wenfeng also has reason to emphasize the need for funding and computing power, but precisely because investors ultimately decide whether to pay, he cannot simply repeat the propaganda that China is already leading in all aspects.

Investors care about what is missing in the next step and why the company needs so much money. The most valuable part of this record is not the 2-year figure; AI development is too fast, and any fixed time gap could change after a single model release. What is truly important is that Wenfeng acknowledges that while structural engineering efficiency can alleviate computing power shortages, it cannot eliminate the need for computing power. Domestic chips can handle some tasks, but they have not yet completely replaced Nvidia's position in frontier training. This also explains why the success of Time 3 cannot prove that U.S. chip restrictions have failed. Restrictions were never intended to make China suddenly stop training models. China already possessed a large number of Nvidia chips before the restrictions were implemented, and it can still obtain some equipment through third-party clouds and gray channels. Chinese teams also use software optimization to improve the efficiency of older chips. What the restrictions are really meant to do is make it more difficult for China to obtain next-generation chips, make building ultra-large clusters more expensive, and reduce the number of frontier experiments it can conduct simultaneously. The fact that China can still produce impressive results shows that the restrictions are not an airtight wall, but it does not mean the wall does not exist at all.

If even the internal judgment of the founder of IPSec regards computing power as the primary constraint, then interpreting a single model release as China having bypassed the chip gap is not a technical judgment, but political propaganda. What remains to be seen is just how significant this constraint is. According to Apple II's statistics on public clusters, as of May 2025, the United States accounts for about 3/4 of the global GPS cluster computing power, while China accounts for about 15%. Since many corporate and military facilities are not public, this cannot be a complete map, but the order-of-magnitude gap it reveals is generally consistent with the data center and chip supply situations in both countries. The U.S. advantage is not just one company or one chip; it possesses multiple competing frontier laboratories such as OpenAI, Google, Meta, and Anthropic; it has large cloud platforms like Microsoft, Amazon, and Google; it has the chip design capabilities of Nvidia and AMD; and it has a high-speed interconnect ecosystem. China certainly has its own advantages: a massive number of engineers, fierce domestic competition, rich manufacturing scenarios, the government's ability to concentrate investment, and companies competing for the market with lower prices.

Huawei chips and the domestic software ecosystem are also improving. But domestic substitution is not just a slogan; it requires simultaneously solving issues with advanced processes, yields, high-bandwidth memory packaging, chip-to-chip communication, software compatibility, and large-scale cluster stability. A single chip being close in paper performance does not mean that tens of thousands of chips connected together can achieve the same efficiency. Frontier AI is increasingly resembling a national-level industrial project; the chip is just the engine, the data center is the car, and it also requires power, cooling, network software, and people capable of maintaining the system over the long term. This is why the gap in public model capabilities can be small, while the national patent gap remains large. A chaser can concentrate resources to train a star model and keep pace with the leader in public tests. The leader, however, can simultaneously train language, video, robotics, scientific, biological, and defense models, and let multiple companies explore different routes. The former can win an important race, while the latter possesses the ability to host multiple races. Of course, the United States is not resting on its laurels. China has proven that it can quickly absorb public research, use limited hardware more efficiently, and compete for global developers through open weights.

If the U.S. wastes its lead on internal regulatory chaos, slow energy infrastructure development, and technological blockades against allies, its structural advantages will certainly shrink. However, the possibility of losing the lead is not the same as being on the verge of being comprehensively surpassed. As of today, Kimi proves that China possesses frontier product capabilities, Deepak proves that China has excellent algorithm and system optimization skills, and the millions of models prove that China's open-weight ecosystem has reached a certain scale. Together, they constitute a serious challenge, but they do not prove that China already possesses frontier computing power, an advanced chip supply chain, and multi-path exploration capabilities comparable to the U.S. Now we can reach a final verdict. Exactly how far behind is China's AI compared to the U.S.? If we must compress the answer into a single number, this program would be repeating the very mistake it is criticizing. In terms of public language models and programming capabilities, China is not years behind. Timmy and the millions of models have already entered the world's first tier; in some tasks, they can run neck-and-neck, or even briefly take the lead. The overall product gap is likely measured in months. In terms of price, engineering efficiency, and the dissemination of open weights, China is not just a chaser; in some aspects, it has already taken the initiative.

If U.S. companies continue to provide only expensive, closed models, they may indeed cede a large number of global developers to China. But in terms of total computing power, advanced GPUs, large-scale clusters, semiconductor manufacturing equipment, and the ability to explore multiple frontier paths simultaneously, the U.S. still holds a clear structural advantage. This is not a gap that can be erased by a leaderboard, nor is it a gap that can be bridged by uploading model files to the internet. Therefore, the desperate speed of Kimi and others does not prove that China has comprehensively surpassed the U.S., but rather that China has been able to push certain AI products into the world's first tier using limited computing power, open weights, and fierce competition. This is worth the U.S. being vigilant about, but not worth panicking over. A leaderboard measures a race held today; national strength determines who has the ability to continuously host the next race, bear more failures, manufacture more advanced chips, and explore technological paths that do not yet exist. China has learned to play a good hand with fewer chips, but the U.S. still holds more chips, a larger table, and still controls the machines that manufacture most of the high-end chips. It is true that China is still very strong, and it is also true that the U.S. cannot be complacent.

However, jumping from these two statements to the conclusion that China is about to comprehensively surpass the U.S. is not technical analysis; it is packaging Beijing's propaganda, Silicon Valley's interests, and media anxiety into a non-existent, inevitable outcome. Our position is clear: one should not underestimate Chinese engineers just because one opposes the CCP, nor should one declare a comprehensive victory for the CCP that has not yet happened just because Timmy won a few races. What can truly relieve anxiety is not another set of self-comforting propaganda, but seeing reality clearly. The U.S. lead is being challenged, but it is far from being erased. China has caught up to the product front line, but it still does not control the entire industrial foundation that determines the next generation of AI. Models can be downloaded, but computing power cannot. Leaderboards can be tied, but the ability to manufacture the machines behind the leaderboards has not yet been matched. This is the truth behind the Kimi and AI wave. Alright, that's all for today. If these programs are valuable to you, please like, hit the notification bell, share, and subscribe to this channel. Friends who are able to can join as members to support the development of our small channel. I am the demon from the logic of the exam, and we will see you in the next episode!