Foreign Research Perspective: Who Is Afraid of Chinese AI Models?
TechFlow Selected TechFlow Selected
Foreign Research Perspective: Who Is Afraid of Chinese AI Models?
The panic among frontier labs stems more from not yet adapting to the new game rules shifting from training-cost dominance to inference-scale dominance.
Navigating Web3 tides with focused insightsAuthor: Stratechery
Translated by: TechFlow
TechFlow Editor's Note: Kimi K3 impacts the cognitive model market; why are Wall Street and Silicon Valley overreacting? This article breaks down how after AI entered the Agent era from the ChatGPT era, tokens are no longer the commodity; the real commodity is intelligence itself—and in this war where marginal cost is king, Chinese open-source models seem free, but service costs are not low at all. The panic in frontier labs is more because they haven't adapted to the new game rules shifting from training cost dominance to inference scale dominance.
I once told a story about my experience on the first day of STRT-431 at the Kellogg School of Management, a required course for every MBA new student. I flipped through the course materials and case studies, disappointed to find no technology companies on the syllabus. I ran to ask the professor why, and the answer I received was: the goal of this course is not to study specific industries, but to uncover universal principles applicable to any industry and any company.
As I often say, I was very unsatisfied with this answer at the time: in my view, the essence of technology, especially the zero marginal cost (and zero transaction cost) of software and distribution, is fundamentally different. Plugging zero into the formula often messes everything up. But I quickly realized this was exactly my opportunity. The core insight of Aggregation Theory is: the value chain caused by zero marginal cost is completely different from what people once expected of the internet—in a world where controlling demand is more important than distributing supply, you will see centralization and scaling.
However, the interesting thing about AI is that those old universal principles are coming back to the forefront. This was most evident last weekend, when a fierce debate about Kimi K3 erupted on X. Kimi K3 is another open-weight model launched from China, and its capabilities are approaching the state-of-the-art. Long story short: marginal cost is back in a significant way, whether from the short-term impact of state-of-the-art free models or from the perspective of the industry's long-term structure.
COGS vs R&D
One of the most common misunderstandings about open-weight models is that they are cheaper—even free. After all, you can download the weights directly, skipping the time, cost, and capability required to create your own model. This is certainly true, but the "free" here refers to the amount you need to spend on R&D. R&D is a fixed expense, unrelated to the revenue you generate. If you spend 1 million dollars on R&D, regardless of whether your revenue is 100,000 dollars or 100 million dollars, you still spent 1 million dollars on R&D (of course, this affects your profitability).
What is related to revenue is COGS—Cost of Goods Sold—and COGS is real for AI, which is something the software industry hasn't seen in a long time. Specifically, running inference on models—whether Kimi or Fable—costs money, and the money AI providers spend on inference, at least in most business models, is directly related to revenue. Reusing the example above, generating 100 million dollars in revenue versus 100,000 dollars in revenue might require 1000 times the COGS. Specifically, if the token cost to generate 1 dollar of revenue is 50 cents, then 100 million dollars in revenue will have 50 million dollars in COGS, and 100,000 dollars in revenue will have only 50,000 dollars in COGS.
The point about open-weight models is that they are not free services. Kimi K3 costs 3 dollars per million input tokens and 15 dollars per million output tokens. This is cheaper than Sol's 5 dollars per million input tokens and 30 dollars per million output tokens, but this may not be the correct metric.
Tokens vs Intelligence
Nvidia CEO Jensen Huang described what Nvidia is building as a "token factory," and this statement makes sense from Nvidia's perspective. Nvidia's GPUs are not targeted at specific models: they generate tokens, and do so in the fastest and most efficient way possible. This leads to metrics such as tokens per second, time to first token, tokens per watt, token cost, etc., which Huang believes will become the basis for decision-making.
This framework indeed made sense in AI's first paradigm—the ChatGPT era—when tokens were delivered directly to end users. However, AI's second paradigm, the inference era, disrupts this measurement. Inference means an explosion of chain-of-thought tokens, and different models require different numbers of inference tokens to get the correct answer. For example, reportedly Kimi uses significantly more tokens than Sol, which nullifies its price advantage. Agents introduce similar dynamics: certain models are more efficient in the number of tokens required to execute Agent workflows.
This means tokens are not a commodity. The defining characteristic of a commodity is that it is interchangeable: one gallon of oil is one gallon of oil, one ton of copper is one ton of copper, one bushel of wheat is one bushel of wheat. But one model's tokens are different from another model's tokens. What is interchangeable is what is built from the tokens, which is intelligence. In other words, if Kimi and Sol both generate the correct answer, then this answer is interchangeable. The difference in tokens required to generate the correct answer leads to differences in COGS.
The COGS of intelligence is a function of several different factors:
Model footprint: Weights and runtime state determine how much expensive memory and how many accelerators are needed to host each service replica
Inference efficiency: Architecture choices (such as Mixture of Experts models) reduce the computation per generated token
Memory efficiency: Architecture choices can reduce KV cache requirements, allowing more concurrent requests and better GPU utilization
Service efficiency: Batching, scheduling, prefix caching, and other inference optimizations maximize utilization and share work between requests
Token efficiency: The fewer tokens needed to reach the correct answer, the lower the inference cost
This is important because we are rapidly approaching a state where for many economically beneficial tasks, intelligence is actually a commodity. For example, anyone building basic CRUD applications might use models from multiple providers to complete them. And in a commodity market, the path to profitability is not by charging higher prices—you can (or soon will be able to) use multiple models to make the exact same application—but by having a superior cost structure.
Understanding Commodity Markets
It is worth sorting out the mechanisms here, because as I pointed out in "Amazon's Durability" a few months ago, the dynamics of commodity markets are not generally familiar to people in the technology industry:
In a commodity market, everyone charges the same price, because everyone is selling the same thing. This price is determined by supply and demand
Demand for a commodity is a function of price elasticity: the cheaper the commodity, the greater the demand, and vice versa
The supply of a commodity is a function of the marginal cost of producing the commodity
The key to understand is that different suppliers have different marginal costs for producing the commodity. In practice, this means the supplier with the worst cost structure ends up selling the commodity at their marginal cost (if they can produce at all). Everyone else's profit depends on how much better their cost structure is than the marginal supplier.
For example:
Supplier A can produce 10 units of commodity at a cost of 10 dollars per unit
Supplier B can produce 10 units of commodity at a cost of 15 dollars per unit
Supplier C can produce 10 units of commodity at a cost of 20 dollars per unit
Assuming price elasticity creates demand for 25 units of commodity at 20 dollars. This means:
Supplier A will sell 10 units at 20 dollars, earning 10 dollars per unit
Supplier B will sell 10 units at 20 dollars, earning 5 dollars per unit
Supplier C will sell 5 units at 20 dollars, earning 0 dollars per unit
This is not entirely accurate: the reason Supplier C bears the shortage is that Suppliers A and B are able to undercut them slightly on price, which of course affects demand (elastic), but this illustrates the point. Supplier A has a good business, Supplier B has an okay business, Supplier C is going bankrupt.
Bankruptcy risk is where fixed costs come back to the forefront: Supplier C has both fixed costs (such as possible R&D expenses) and may have taken on debt to finance the equipment needed to produce the commodity. It cannot price the commodity based on these costs—remember, the market clearing price is close to the marginal cost of the highest cost unit needed to meet demand—but these costs can absolutely drive the supplier out of the market. If the supplier goes out of business, prices will rise until another supplier decides to enter (or other suppliers expand).
Intelligence Market
Let's go back to models. Now, all the analysis above does not apply, because demand for frontier models exceeds supply, and supply is limited by the lack of compute. This compute shortage not only means compute suppliers like Nvidia can achieve very high profit margins, but also means Nvidia's customers, like SpaceXAI, can resell compute to companies like Anthropic at high profit margins. Meanwhile, Anthropic can pay the markup because they can sell tokens at a higher markup.
However, it is not just excess demand that gives Anthropic huge profit margins: Anthropic and OpenAI may have the lowest cost per unit of frontier quality intelligence, thanks to model capabilities, service scale, and token efficiency. They serve models of specific capability levels months ahead of competitors, while applying the best models to optimize these costs.
It is also worth noting that the market does not yet view intelligence as a commodity: demand is specifically for Anthropic and OpenAI, and demand for less good models is much less (hence SpaceXAI and Meta sell capacity to Anthropic). One way to think about the drive to optimize costs is that this is a function of work to be done defined by intelligence level, so intelligence buyers can create a market where intelligence is commoditized. However, in the long run, whoever is at the frontier is most capable of dominating the non-frontier markets, which are just the frontier minus n months, which are the months frontier model manufacturers have been optimizing their service costs.
All this is to say, I think the general reaction to Kimi and Chinese models is quite excessive, at least from an economic perspective. There is now a price umbrella, which is a downstream result of the lack of compute. I am very skeptical whether Chinese models are cheaper to serve on a marginal cost basis; they just look cheaper because Anthropic and OpenAI's supply is so constrained that they are charging prices far higher than what they would be if there were enough supply to meet intelligence demand.
Frontier Lab Panic
So, why do model manufacturers seem so panicked about Chinese models?
First, I think frontier labs are anchored in a world where training costs dominate their financial modeling. As long as training consumes more GPUs than inference, maximizing inference revenue to help fund the next training run is crucial, which means charging very high prices for inference.
However, looking forward, I expect the inference market to grow much faster than training costs (this includes the assumption that training costs will continue to skyrocket), which means they can really make up for it with volume. Until eight months ago, it was unclear whether this would become reality, but the unlocking of the Agent paradigm is so huge that frontier labs should be more confident that they can not only survive at lower prices but thrive (once they have enough compute).
Second, intelligence is actually not a perfect commodity, partly because applying intelligence makes itself smarter. Specifically, whoever is running inference is also collecting data, and this data will be used to make the next iteration of the model better. On one hand, this is even more reason for frontier labs to lower prices and increase usage when more compute comes online. On the other hand, this is why companies like Microsoft are increasingly obsessed with helping companies run their own models. If Chinese models are viable alternatives, this becomes even more feasible.
Third, another way frontier labs can differentiate themselves from Chinese models, and from each other, is to continue integrating upwards into the customer experience. Claude Code and Codex have proven to be quite sticky, which is shocking. Whichever tool you start using is likely the one you stick with, especially for non-technical users. In the long run, this necessity of moving upwards does mean frontier models are definitely a threat to software providers, including Microsoft. On the other hand, software companies that currently own the customer experience can resist erosion by frontier labs to the extent that they can obtain competitive models.
Finally, Anthropic's ideological angle is particularly impossible to ignore. This is a company that believes only it can be entrusted with AI, and the existence of open-weight alternatives is a fatal blow to this assumption.
China's Motivation
Kimi is not the only new Chinese model. Bloomberg reported:
Alibaba Group Holding Ltd stock price rose 5.4% on Monday after the company launched a preview version of its flagship Qwen3.8 Max model, claiming it is second only to Anthropic's Fable 5. This Sunday launch came just days after the startup Moonshot AI launched a powerful new product, which shook the market and triggered US concerns about China narrowing the gap with global leaders like Anthropic and OpenAI. Qwen3.8 Max has 2.4 trillion parameters, entering the heavyweight category along with Moonshot's Kimi K3. K3 has 2.8 trillion parameters, comparable to top products, and Alibaba also set equally high expectations.
Developers can now access Qwen3.8 Max through Alibaba's coding platform (including Qoder). Alibaba plans to open-source the model weights soon, extending access beyond the preview version. Interest in these Chinese-made AI systems and models is so high that Moonshot was forced to pause accepting new subscriptions late Sunday to cope with overwhelming demand.
Alibaba Group Holding Ltd stock price rose 5.4% at one point on Monday after the company launched a preview version of its flagship Qwen3.8 Max model and claimed it is second only to Anthropic PBC's Fable 5. This Sunday launch came just days after Moonshot AI launched a powerful new product, which shook the market and triggered US concerns about China catching up with global leaders like Anthropic and OpenAI. Qwen3.8 Max has 2.4 trillion parameters, entering the heavyweight ranks along with Moonshot's Kimi K3. K3 has 2.8 trillion parameters, comparable to top products, and Alibaba also set equally high expectations.
Developers can now access Qwen3.8 Max through Alibaba's coding platform, including Qoder. Alibaba plans to open the model weights soon, extending access beyond the preview version. Demand for these Chinese-made AI systems and models is so high that Moonshot was forced to pause accepting new subscriptions late Sunday to cope with overwhelming demand.
The fact that Qwen3.8 Max will also open weights is noteworthy. Alibaba stopped releasing weights for its leading models earlier this year, but seems to have changed this decision; I suspect this shift is related to Xi Jinping's speech on AI last week, which doubled down on the open-weight approach:
We should adhere to the principle of open win-win cooperation and promote innovation-driven development. As a new engine of world economic growth and an accelerator of growth momentum conversion, artificial intelligence is moving from the digital world to the physical world. We should seize this rare historical opportunity and encourage open source, openness, collaboration, and sharing. We should promote technological innovation, industrial development, and scenario-based application of artificial intelligence. We should coordinate the promotion of transformation and upgrading of traditional industries, cultivation and expansion of emerging industries, and forward-looking planning of future industries, so that all industries can benefit from artificial intelligence.
China's strategy is obvious: commodify your complements. Note Xi Jinping explicitly linked openness with AI "moving from the digital world to the physical world"; the physical world is a world dominated by China, and China's leading position in fields like robotics will benefit immensely from widely available AI models.
Similarly, China does not want the US to gain an asymmetric advantage in AI; if China can weaken US frontier labs while enhancing any and all potential US opponents, that would be even better, and it can benefit from innovation attached to the open ecosystem.
The Distillation Issue
Similarly, do not expect any measures against China's distillation attacks on frontier labs. I think it is wrong to attribute all of Chinese labs' success to distillation, but it is equally wrong to pretend distillation has not given Chinese labs a huge advantage. This advantage truly became apparent in the past year as post-training reinforcement learning became increasingly important for model performance. Chinese labs do not have to build reinforcement learning environments from scratch, but can simply use frontier labs' models as teachers to achieve rapid improvement at much lower costs (this is not the only reason Chinese model development costs are lower, but it is an important one).
Interestingly, one of the most important use cases for Chinese models in the West is distillation itself. For example, Thinking Machines, which just released open-weight models, relies on Chinese models to solve the cold start problem of reinforcement learning. Dean Meyer and Konstantine Buhler wrote an excellent article on X explaining that distillation means Western open-weight models are fundamentally at a disadvantage relative to China:
Distillation does not explain all of China's lead in open models. Chinese labs have world-class researchers, massive compute, powerful pre-trained models, software-hardware co-design, and rapidly improving post-training capabilities. But distillation compresses the costly last gap from powerful foundations to near-frontier systems. Even if distillation accounts for a small share of total capability in Chinese models, it accounts for a significant share of their advantage relative to US open models.
New enforcement mechanisms will make large-scale distillation harder, slower, and more expensive for Chinese companies. However, enforcement cannot eliminate distillation supported by state actors. Therefore, every frontier progress in the West creates another teacher for Chinese labs. Western builders must independently replicate these capabilities, or wait to learn from Chinese models. This gap provides Chinese labs with a recurring structural advantage relative to Western companies.
This point is worth repeating: because US open-weight model manufacturers must comply with frontier labs' terms of service, they (1) are inferior to Chinese alternatives, and (2) ultimately distill distillates, just taking a detour through Chinese labs. Would it not be better if Western open-weight model manufacturers could access directly from the source?
For this, there is an even more interesting question around distillation: why exactly is it bad? After all, aren't large language models just a distillation of all knowledge on the open internet, scraped and distilled into the model being distilled by frontier labs? Who exactly is being harmed here?
In fact, this paradox is the solution. I believe open-weight models are good for innovation (and, based on the above, I think frontier labs will be fine), but relying on China is a problem. The US should pass a law that (1) explicitly classifies collecting data for training models as fair use, and (2) prohibits terms of service that ban distillation, at least for US companies. Stopping distillation—actually querying APIs—is almost impossible; the US should go the other way, leaning towards a new copyright policy that protects labs while guaranteeing that the knowledge they learned provides momentum for further innovation by others.
Reasons to Worry
The whole article has been eliminating overreaction to Kimi K3 and Chinese open-weight models; however, there is one reason to worry, which is cybersecurity. Consider this story from The Stack:
Hugging Face stated that its production infrastructure was breached by an "autonomous" AI agent system earlier last week. The platform's security team was initially hindered in Incident Response (IR) by unnamed US LLM frontier model guardrails, "which could not distinguish between incident responders and attackers," they said. Therefore Hugging Face defenders turned to using the open-source GLM 5.2 model from China's Z.ai laboratory—running it on their own infrastructure to analyze over 17,000 logs or footprints left by the attackers.
For New York-based Hugging Face, this is a striking public admission that the company, which lets users collaborate on models, datasets, and applications, reached a milestone of 100 million dollars ARR this summer. In an incident report, the company advised defenders to "prepare a capable model that can run on your own infrastructure before an incident occurs [our italics], both to avoid guardrail lock-in and to prevent attacker data and credentials from leaving your environment."
It is hard to exaggerate how wrong the Trump administration's panicked reaction to Anthropic releasing Fable was, especially because it exacerbated Anthropic's worst tendency, which is assuming only they can be trusted with powerful AI. In a world with only one AI, it might make sense to reserve the most powerful cybersecurity capabilities for the US government and trusted allies; however, this is not the world we live in.
There are already and will be models fully capable of launching cybersecurity attacks on existing infrastructure, and these models will be—already are—widely available. The best defense—in fact the only viable defense—will be to ensure defenders also have access to the best models. Now, due to the Trump administration's directive, defenders are effectively prohibited from using Fable or Sol for cybersecurity; this means the best alternative is to use models from a country that has been trying to undermine our cyber defenses for years. This is insane!
The better approach is clear: first, relax restrictions on Fable and Sol for cybersecurity, and second, ensure US open-weight model manufacturers are on a level playing field with China. Yes, frontier labs will complain loudly, but the government should realize that listening to their hysteria has already led to a situation where US companies rely on China for defense. Let frontier labs win by doing better; do not let them define security or safety, and do not let them pull up the ladder of human collective knowledge. China is hard enough to compete with; letting them carry the banner of openness and innovation is simply giving up our biggest advantage.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
@stratechery
