Chinese AI models spark global concern amid cost and strategic shifts

Author: Stratechery

Compiled by Deep潮 TechFlow

Shenchao Overview: Kimi K3 is challenging the cognitive model market—why are Wall Street and Silicon Valley overreacting? This article breaks down how, since the transition from the ChatGPT era to the Agent era, tokens are no longer commodities; the true commodity is intelligence itself. In this war where marginal cost reigns supreme, Chinese open-source models may appear free, but their service costs are far from low. The panic among frontier labs stems largely from their failure to adapt to the new rules of the game, where inference scale, not training cost, is now the dominant factor.

I once shared a story about my first day in STRT-431 at the Kellogg School of Management, a required course for every MBA student. As I reviewed the course materials and case studies, I was disappointed to find no technology companies on the syllabus. I went to ask the professor why, and his response was: The goal of this course isn’t to study specific industries, but to uncover universal principles applicable to any company in any industry.

As I often say, I was deeply unsatisfied with that 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 in zero to the formula often disrupts everything. But I quickly realized that this was precisely my opportunity. The core insight of Aggregation Theory is that zero marginal cost leads to a value chain entirely unlike what people originally expected from the internet—in a world where controlling demand matters more than distributing supply, you see centralization and scale.

However, what’s interesting about AI is that those ancient universal principles are returning to the forefront. This was most evident last weekend, when a heated debate erupted on X about Kimi K3—a newly released open-weight model from China whose capabilities are now approaching state-of-the-art levels. In short: marginal cost has returned in a significant way, whether viewed through the lens of the short-term impact of free state-of-the-art models or the long-term structure of the industry.

One of the most common misconceptions about open-source weight models is that they are cheaper—or even free. After all, you can directly download the weights and skip the time, cost, and expertise required to build your own model. This is certainly true, but the term "free" here refers only to the amount you spend on research and development (R&D). R&D is a fixed cost, unrelated to your revenue. If you spend $1 million on R&D, you still spent $1 million on R&D, whether your revenue is $100,000 or $100 million (though this, of course, impacts your profitability).

Related to revenue is COGS—cost of goods sold—and for AI, COGS is real, something the software industry hasn’t seen in a long time. Specifically, running inference on models—whether Kimi or Fable—costs money, and AI providers spend on inference in direct proportion to revenue, at least in most business models. Using the earlier example, generating $100 million in revenue versus $100,000 in revenue may require 1,000 times the COGS. Specifically, if the token cost to generate $1 in revenue is $0.50, then $100 million in revenue would entail $50 million in COGS, while $100,000 in revenue would have only $50,000 in COGS.

The key point about open-weight models is that they are not free services. Kimi K3 costs $3 per million input tokens and $15 per million output tokens. This is cheaper than Sol’s $5 per million input tokens and $30 per million output tokens, but this may not be the right metric to use.

NVIDIA CEO Jensen Huang has described what NVIDIA is building as a "token factory," and from NVIDIA’s perspective, this characterization makes sense. NVIDIA’s GPUs are not designed for specific models—they generate tokens, and they do so as quickly and efficiently as possible. This has led to metrics such as tokens per second, time to first token, tokens per watt, and token cost, which Huang believes will become the basis for decision-making.

This framework made sense in the first paradigm of AI—the ChatGPT era—when tokens were directly delivered to end users. However, the second paradigm of AI, the reasoning era, disrupts this measurement. Reasoning entails an explosive growth in chain-of-thought tokens, and different models require varying numbers of reasoning tokens to arrive at the correct answer. For instance, Kimi is reported to use significantly more tokens than Sol, nullifying its price advantage. Agents introduce a similar dynamic: certain models are more efficient in the number of tokens required to execute agent workflows.

This means the tokens are not commodities. The defining characteristic of a commodity is that it is interchangeable: one gallon of oil is the same as another gallon of oil, one ton of copper is the same as another ton of copper, one bushel of wheat is the same as another bushel of wheat. However, tokens from one model are different from tokens from another model. What is interchangeable is the intelligence constructed from the tokens—that is, the output. In other words, if Kimi and Sol both generate the correct answer, then that answer is interchangeable. The difference in the number of tokens required to generate the correct answer leads to differences in COGS.

Smart COGS is a function of several different factors:

Model footprint: Weights and runtime state determine how much expensive memory and accelerators are required to host each service replica.

Inference efficiency: Architecture choices (such as mixture-of-experts models) reduce the computational cost per generated token.

Memory efficiency: Architecture choices can reduce KV cache requirements, enabling more concurrent requests and better GPU utilization.

Service efficiency: Batch processing, scheduling, prefix caching, and other inference optimizations maximize utilization and share work across requests.

Token efficiency: The fewer tokens required to reach the correct answer, the lower the reasoning cost.

This is important because we are rapidly approaching a state where intelligence itself is becoming a commodity for many economically beneficial tasks. For example, anyone building a basic CRUD application may use models from multiple providers to accomplish it. In a commodity market, the path to profitability is not through charging higher prices—you can (or soon will be able to) build identical applications using multiple models—but through having a superior cost structure.

It’s worth clarifying the mechanism here, because as I noted several months ago in “The Durability of Amazon,” the dynamics of commodity markets are not generally familiar to those in the tech industry:

In commodity markets, everyone charges the same price because everyone is selling the exact same product. This price is determined by supply and demand.

Demand for a product is a function of price elasticity: the cheaper the product, the higher the demand, and vice versa.

The supply of a good is a function of its marginal cost of production.

The key point to understand is that different suppliers have varying marginal costs of producing goods. In practice, this means that the supplier with the least favorable cost structure ends up selling goods at their marginal cost (if they are still able to produce). The profits of all other suppliers depend on how much better their cost structure is compared to the marginal supplier.

For example:

Supplier A can produce 10 units of the product at a cost of $10 per unit.

Supplier B can produce 10 units of the product at a cost of $15 per unit.

Supplier C can produce 10 units of the product at a cost of $20 per unit.

Assuming price elasticity results in a demand for 25 units of the good at $20, this means:

Supplier A will sell 10 units at $20 each, earning $10 per unit.

Supplier B will sell 10 units at $20 each, earning $5 per unit.

Supplier C will sell 5 units at $20 each, earning $0 per unit.

This isn't entirely accurate: Supplier C is suffering the shortfall because Suppliers A and B are able to slightly undercut them on price, which naturally affects demand (which is elastic), but it illustrates the point. Supplier A is doing well, Supplier B is doing okay, and Supplier C is going bankrupt.

Bankruptcy risk is where fixed costs reenter the spotlight: Supplier C has fixed costs (such as potential R&D expenses) and may also carry debt to finance the equipment needed for production. It cannot price its goods 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 force the supplier out of the market. If the supplier goes bankrupt, prices will rise until another supplier decides to enter (or existing suppliers expand).

Let’s return to the model. None of the above analysis applies anymore, because demand for frontier models exceeds supply, which is constrained by a shortage of computing power. This computing shortage not only enables computing providers like NVIDIA to achieve very high profit margins, but also allows NVIDIA’s customers, such as SpaceX AI, to resell computing power at high margins to companies like Anthropic. Meanwhile, Anthropic can afford these markups because they can sell tokens at even higher markups.

However, it’s not just excess demand that has given Anthropic such high profit margins: Anthropic and OpenAI likely have the lowest unit cost for frontier-quality intelligence, thanks to model capabilities, service scale, and token efficiency. They have been serving models at specific capability levels months ahead of competitors, while applying their best models to optimize these costs.

It is also worth noting that the market has not yet treated intelligence as a commodity: demand is specifically directed toward Anthropic and OpenAI, with significantly less demand for less capable models (which is why SpaceXAI and Meta sell their capacity to Anthropic). One way to think about the driver of cost optimization is that it is a function of the level of intelligence required to complete a task, allowing intelligent buyers to create a market where intelligence becomes commoditized. However, in the long term, whoever leads at the frontier will be best positioned to dominate non-frontier markets—markets that are simply the frontier minus n months, the very months during which frontier model manufacturers have been optimizing their service costs.

All of this is to say that I think the general reaction to Kimi and Chinese models has been somewhat overblown, at least from an economic perspective. There’s currently a price floor, which is a downstream consequence of compute scarcity. I’m highly skeptical that Chinese models are actually cheaper to serve on a marginal cost basis—they only appear cheaper because Anthropic and OpenAI’s supply is so constrained that their prices are far higher than they would be if there were sufficient supply to meet demand for intelligence.

So why do model manufacturers seem so alarmed by Chinese models?

First, I believe frontier labs operate under the assumption that 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 critical, which means charging very high prices for inference.

However, looking ahead, I expect the growth of the inference market to far outpace training costs (assuming training costs continue to rise sharply), meaning they can truly compensate with volume. Until eight months ago, it was unclear whether this would become reality, but the unlocking potential of the agent paradigm is so immense that leading labs should now have greater confidence—not only that they can survive at lower prices, but that they can thrive once they have sufficient compute.

Second, intelligence is not a perfect commodity, partly because using intelligent systems makes them smarter. Specifically, whoever runs inference is also collecting data, and this data is used to improve the next iteration of the model. On one hand, this is why leading labs have reason to lower prices and increase usage as more computing power comes online. On the other hand, this is why companies like Microsoft are increasingly focused on helping businesses run their own models—if Chinese models are viable alternatives, this becomes even more feasible.

Third, another way frontier labs can distinguish themselves not only from Chinese models but also from each other is by continuing to integrate upward into the customer experience. The remarkable stickiness of Claude Code and Codex is striking: whichever tool you start with, you’re likely to stick with it—especially for non-technical users. In the long term, this upward integration makes frontier models a genuine threat to software providers, including Microsoft. Conversely, the extent to which software companies currently owning the customer experience can acquire competitive models determines their ability to resist encroachment by frontier labs.

Finally, Anthropic’s ideological stance is particularly noteworthy. It is a company that believes only it can be entrusted with AI, and the existence of open-source weight alternatives deals a fatal blow to this assumption.

Kimi is not the only new Chinese model. Bloomberg reports:

Alibaba Group Holding Ltd.'s stock rose 5.4% on Monday after the company unveiled a preview of its flagship Qwen3.8 Max model, claiming it is second only to Anthropic’s Fable 5. This Sunday release came just days after startup Moonshot AI launched a powerful new product, shaking the market and fueling U.S. concerns that China is narrowing the gap with global leaders like Anthropic and OpenAI. The Qwen3.8 Max, with 2.4 trillion parameters, joins Moonshot AI’s Kimi K3 in the heavyweight category; the K3, with 2.8 trillion parameters, rivals top-tier models, and Alibaba has set similarly high expectations.

Developers can now access Qwen3.8 Max through Alibaba’s coding platform, including Qoder. Alibaba plans to soon open-source the model’s weights, extending access beyond the preview version. Interest in these Chinese-made AI systems and models has been so high that Moonshot AI was forced to temporarily suspend new subscriptions late Sunday evening to manage overwhelming demand.

Alibaba Group Holding Limited's stock rose as much as 5.4% on Monday after the company unveiled a preview of its flagship Qwen3.8 Max model, claiming it ranks just below Anthropic PBC’s Fable 5. This Sunday release came just days after Moonshot AI launched a powerful new product that shook the market and sparked U.S. concerns about China’s progress in catching up with global leaders like Anthropic and OpenAI. With 2.4 trillion parameters, Qwen3.8 Max joins Moonshot’s Kimi K3—a model with 2.8 trillion parameters that rivals top-tier products—in the heavyweight category, as Alibaba sets equally high expectations.

Developers can now access Qwen3.8 Max, including Qoder, through Alibaba’s coding platform. Alibaba plans to soon release the model’s weights, expanding access beyond the preview version. Demand for these Chinese-made AI systems and models has been so high that Moonshot was forced to temporarily suspend new subscriptions late Sunday to manage overwhelming demand.

It is noteworthy that Qwen3.8 Max will also open its weights. Alibaba halted the release of weights for its leading models earlier this year, but appears to have reversed this decision; I suspect this shift is related to President Xi’s speech on AI last week, which strongly emphasized an open-weights approach:

We should uphold the principle of openness and mutual benefit to drive innovation-led development. As a new engine for global economic growth and an accelerator for transitioning growth drivers, artificial intelligence is moving from the digital world into the physical world. We should seize this rare historical opportunity by encouraging open-source, openness, collaboration, and sharing. We should promote technological innovation, industrial development, and practical applications of artificial intelligence. We should coordinate the transformation and upgrading of traditional industries, the cultivation and expansion of emerging industries, and the forward-looking planning of future industries, ensuring that all sectors benefit from artificial intelligence.

China’s strategy is clear: commoditize your complements. Note that Xi Jinping explicitly links openness with AI’s transition “from the digital world to the physical world”; the physical world is one dominated by China, and its leadership in fields such as robotics will benefit immensely from widely available AI models.

Similarly, China does not want the United States to gain an asymmetric advantage in AI; it would be even better if China could weaken U.S. frontier labs while strengthening any and all potential U.S. competitors, and it can also benefit from innovations tied to open ecosystems.

Similarly, don’t expect China to take any action against distillation attacks on frontier labs. I think it’s wrong to attribute all of China’s lab successes solely to distillation, but it’s equally wrong to pretend that distillation hasn’t given Chinese labs a significant advantage. This advantage has become truly evident over the past year, as post-training reinforcement learning has grown increasingly critical to model performance. Chinese labs don’t need to build reinforcement learning environments from scratch—they can simply use frontier labs’ models as teachers to achieve rapid improvements at a much lower cost (this is not the only reason Chinese model development is cheaper, but it is a major 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 an open-weight model, relies on Chinese models to address the cold start problem in 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 compared to Chinese ones:

Distillation does not fully explain China's overall lead in open models. Chinese labs possess world-class researchers, vast computational resources, powerful pre-trained models, tight hardware-software co-design, and rapidly improving post-training capabilities. But distillation compresses the costly final gap between strong foundational models and near-state-of-the-art systems. Even if distillation represents a small share of China’s overall model capabilities, it accounts for a significant portion of its relative advantage over U.S. open models.

The 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. Thus, every frontier advancement in the West creates another teacher for Chinese labs. Western builders must either 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 bears repeating: Because U.S.-based open-weight model developers must comply with the terms of service of leading labs, they are (1) less competitive than their Chinese counterparts and (2) ultimately distill distilled models that merely pass through Chinese labs. Wouldn’t it be better if Western open-weight model developers could access the source directly?

For this reason, there’s an even more interesting question surrounding distillation: why is it actually bad? After all, aren’t large language models simply a distillation of all knowledge available on the open internet, gathered and distilled by cutting-edge labs into models that are themselves being distilled? Who is truly being harmed here?

In fact, this paradox is the solution. I believe open-weight models foster innovation (and, based on the above, I think leading labs will be fine), but relying on China is problematic. The U.S. should pass a law that (1) explicitly recognizes collecting data for model training as fair use, and (2) at minimum, prohibits U.S. companies from enforcing terms of service that ban distillation. Blocking distillation—essentially querying APIs—is nearly impossible; the U.S. should take another path, favoring a new copyright policy that protects labs while ensuring the knowledge they acquire fuels further innovation by others.

The entire article has been addressing the overreaction to Kimi K3 and Chinese open-weight models; however, there is a legitimate concern: cybersecurity. Consider this story from The Stack:

Hugging Face stated that its production infrastructure was compromised last week by an "autonomous" AI agent system. The platform’s security team initially encountered obstacles in incident response (IR) due to unnamed U.S. LLM frontier model safeguards, which "could not distinguish between incident responders and attackers," they said. As a result, Hugging Face’s defenders turned to the open-source GLM 5.2 model from China’s Z.ai Lab, running it on their own infrastructure to analyze over 17,000 logs or digital footprints left by the attackers.

For Hugging Face, headquartered in New York, this is a notable public acknowledgment as the company, which enables users to collaborate on models, datasets, and applications, reached the milestone of $100 million in annual recurring revenue (ARR) this summer. In an incident report, the company advised defenders to "prepare a capable model that can run on your own infrastructure [our emphasis] beforehand, to avoid guardrail lock-in and prevent attackers from exfiltrating your data and credentials."

It is difficult to overstate how misguided the Trump administration’s panicked response to Anthropic’s release of Fable was, especially since it reinforced Anthropic’s worst tendency—the assumption that only they can be trusted to use powerful AI. In a world with only one AI, it might make sense to reserve the most powerful cybersecurity capabilities for the U.S. government and trusted allies; however, this is not the world we live in.

Models capable of launching cyberattacks on existing infrastructure already exist and will become widely available—indeed, they already are. The best—and in fact, the only viable—defense will be to ensure defenders also have access to the best models. Yet, due to directives from the Trump administration, defenders are effectively prohibited from using Fable or Sol for cybersecurity; this means the best available alternatives are models from a country that has spent years trying to undermine our cyber defenses. This is insane!

A better approach is to be clear: first, relax restrictions on Fable and Sol regarding cybersecurity; second, ensure that U.S. open-weight model manufacturers compete on equal footing with China. Yes, frontier labs will make a lot of noise, but the government must recognize that yielding to their hysteria has led the U.S. into a position where American companies rely on China for defense. Let frontier labs win by doing better—not by defining security or safety, nor by pulling up the ladder of humanity’s collective knowledge. China already faces enough challenges; allowing them to champion openness and innovation is essentially surrendering our greatest advantage.