Author: Stratechery

Translated by: Deep Tide TechFlow

Introduction by Deep Tide: Kimi K3 impacts the cognitive model market; why are Wall Street and Silicon Valley overreacting? This article dissects how AI has transitioned from the ChatGPT era to the Agent era, where tokens are no longer commodities; the real commodity is intelligence itself— and in this war where marginal cost is king, Chinese open-source models seem free, but the service costs are not low at all. The panic of frontier laboratories is more due to the fact that they have not yet adapted to the new game rule where inference scale dominates over training cost.

I once told a story about my experience on the first day of STRT-431 at Kellogg School of Management, a required course for every MBA freshman. I flipped through the course materials and case studies, only to be disappointed to find that there were no tech companies on the schedule. I ran to ask the professor why, and the answer I got was: the goal of this course is not to study specific industries, but to discover universal principles that can be applied to any industry and any company.

As I often say, I was not satisfied with this answer at that time: in my view, the essence of technology, especially the zero marginal cost of software and distribution, is fundamentally different. Filling zero into a formula often muddles everything. But I soon realized that this was precisely my opportunity. The core insight of Aggregation Theory is that the value chain resulting from zero marginal cost is completely different from what people once expected of the internet—in a world where controlling demand is more important than distribution supply, you will see centralization and scaling.

However, the interesting aspect of AI is that those old universal principles are returning to the forefront. This became most apparent last weekend when a fierce debate erupted on X regarding Kimi K3. Kimi K3 is another open-source weight model launched in China, whose capabilities are nearing the state-of-the-art level. To make a long story short: marginal cost has returned in a significant way, whether in terms of the short-term impact of cutting-edge free models or the perspective of long-term industry structure.

COGS and R&D Spending

One of the most common misconceptions about open-source weight models is that they are cheaper— even free. After all, you can download the weights directly, bypassing the time, cost, and capabilities required to create your own model. This is certainly not wrong, but here "free" refers to the amount you need to spend on R&D (Research and Development). R&D is a fixed cost, unrelated to the revenue you generate. If you spend $1 million on R&D, whether your revenue is $100,000 or $100 million, you still spent $1 million on R&D (of course, this affects your profitability).

What is related to revenue is COGS—Cost of Goods Sold— and COGS is very real for AI, which has not been the case in the software industry for a long time. Specifically, running inference on models— whether Kimi or Fable— costs money, and the money AI providers spend on inference is, at least in most business models, directly related to revenue. Reusing the example above, generating $100 million in revenue may require 1,000 times the COGS compared to generating $100,000 in revenue. Specifically, if the token cost to generate $1 of revenue is 50 cents, then $100 million in revenue will have $50 million in COGS, while $100,000 in revenue will only have $5,000 in COGS.

The key point about open-source weight models is that they are not free services. The cost of Kimi K3 is $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 correct benchmark.

Tokens and Intelligence

NVIDIA CEO Jensen Huang describes what NVIDIA is building as a "token factory," which makes sense from NVIDIA's perspective. NVIDIA's GPUs are not targeted at specific models: they generate tokens, and they do so in the fastest and most efficient manner. This has led to metrics such as tokens per second, first token time, tokens per watt, and token cost, which Huang believes will become the basis for decision-making.

This framework indeed makes sense in AI's first paradigm—the ChatGPT era—when tokens were directly delivered to end-users. However, AI's second paradigm, the inference age, has disrupted this measure. Inference implies an explosive growth in the chain of thought tokens, where different models require different amounts of inference tokens to arrive at the correct answer. For example, reports suggest that Kimi uses significantly more tokens than Sol, nullifying its price advantage. Agent introduces a similar dynamic: certain models are more efficient in terms of the number of tokens needed to execute Agent workflows.

This means tokens are not commodities. The defining characteristic of a commodity is that it is interchangeable: a gallon of oil is just a gallon of oil, a ton of copper is just a ton of copper, a bushel of wheat is just a bushel of wheat. But a token from one model is different from a token from another model. What is interchangeable is the intelligence built from tokens. In other words, if both Kimi and Sol generate the correct answer, then that answer is interchangeable. The difference in the tokens required to generate the correct answer will lead to differences in COGS.

The COGS of intelligence is a function of several different factors:

Model footprint: Weights and runtime states determine how much expensive memory and how many accelerators are required to host each service instance.

Inference efficiency: Architectural choices (such as mixture of experts models) reduce the computational load for each generated token.

Memory efficiency: Architectural choices can reduce KV cache requirements, allowing for more concurrent requests and better GPU utilization.

Service efficiency: Batching, 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 inference cost.

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

Understanding Commodity Markets

It is worth untangling the mechanics here, because as I noted a few months ago in "The Durability of Amazon," the dynamics of commodity markets are not something that tech industry insiders are generally familiar with:

In a commodity market, everyone charges the same price because everyone is selling the same thing. This price is determined by supply and demand.

The 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 of producing commodities. Practically, this means that the supplier with the worst cost structure ultimately sells commodities at their marginal cost (if they can still produce them). The profits of everyone else depend on how much better their cost structure is compared to the marginal supplier.

For example:

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

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

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

Assuming price elasticity leads to a demand for 25 units of the commodity at $20, this means:

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

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

Supplier C will sell 5 units at $20, earning $0 profit.

This is not completely accurate: the reason Supplier C suffers from a shortage is that Suppliers A and B can slightly undercut them on price, which will affect demand (which is elastic), but this illustrates the point. Supplier A has a good business, Supplier B has a decent business, and Supplier C is going bankrupt.

The risk of bankruptcy is where fixed costs come back into play: Supplier C has both fixed costs (such as possible R&D expenses) and may have incurred debt to finance the equipment necessary to produce the commodity. They cannot price the commodity based on those costs—remember, the market-clearing price is close to the marginal cost of the highest-cost unit required to meet demand—but those costs can certainly push suppliers out of the market. If suppliers go out of business, prices will rise until another supplier decides to enter (or other suppliers expand).

The Intelligence Market

Let’s return to the models. Now, all of the analysis above doesn’t apply because the demand for frontier models exceeds supply, which is limited by a lack of compute power. This lack of compute power not only means that compute suppliers like NVIDIA can achieve very high profit margins but also that NVIDIA's customers, such as SpaceXAI, can reseller compute power to companies like Anthropic at a high profit margin. Meanwhile, Anthropic can afford to pay a markup because they can sell tokens at a higher markup.

However, it is not just excess demand that gives Anthropic huge margins: Anthropic and OpenAI may have the lowest unit costs for cutting-edge quality intelligence, thanks to model capabilities, service scale, and token efficiency. They are deploying models with specific capability levels to clients months ahead of competitors, while also applying the best models to optimize those costs.

It's also worth noting that the market has not yet viewed intelligence as a commodity: demand is specialized towards Anthropic and OpenAI, with far less demand for less capable models (hence SpaceXAI and Meta selling capacity to Anthropic). One way to think about the drive to optimize costs is as a function defined by the level of intelligence needed to get the work done, allowing intelligence buyers to create a market where intelligence is commodified. However, in the long run, whoever is at the frontier is most capable of dominating the non-frontier market, which is merely the frontier minus n months—that is, the months during which frontier model manufacturers have been optimizing their service costs.

All this to say, I believe 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 power. I have serious doubts that Chinese models offer cheaper service on a marginal cost basis; they merely appear cheaper because the supply from Anthropic and OpenAI is so constrained that the prices they charge are far higher than they would be if there were sufficient supply to meet intelligence demand.

The Panic of Frontier Laboratories

So why do model manufacturers seem so panicked about Chinese models?

First, I believe frontier laboratories are anchored in a world where training costs dominate their financial modeling. As long as the GPUs consumed for training outnumber those for inference, maximizing inference revenue to help fund the next training run becomes critical, which means charging very high prices for inference.

However, looking forward, I expect the growth of the inference market to far outpace training costs (which includes the assumption that training costs will continue to surge), meaning they can genuinely make up for it with volume. Until eight months ago, it was unclear whether this would become a reality, but the unlocking of the Agent paradigm is so huge that frontier laboratories should be more confident that they not only survive at lower prices but also thrive (once they have sufficient compute power).

Second, intelligence is not actually a perfect commodity, partly because applying intelligence makes itself smarter. Specifically, whoever is running inference is also collecting data, which will be used to improve the next iteration of the model. On one hand, this gives frontier laboratories more reason to lower prices and increase usage as more compute power comes online. On the other hand, this is why companies like Microsoft are becoming increasingly obsessed with helping companies run their own models. If Chinese models are viable alternatives, this becomes even more feasible.

Third, another way for frontier laboratories to differentiate themselves from Chinese models is to continue to integrate upwards into the customer experience. Claude Code and Codex have proven to be quite sticky, which is striking. No matter which tool you start using, it could very well be the one you stick with, particularly for non-technical users. In the long term, this necessity to move upwards indeed means that frontier models are a serious threat to software vendors, including Microsoft. On the other hand, to what extent do software companies that currently have customer experiences get hold of competitive models, to the same extent they can fend off the encroachment of frontier laboratories?

Finally, Anthropic's ideological perspective is particularly noteworthy. 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 that assumption.

China's Motivation

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

Alibaba Group Holding Ltd. saw its stock rise 5.4% on Monday, following the preview launch of its flagship Qwen3.8 Max model, which it claims is second only to Anthropic's Fable 5. This Sunday release comes just days after the startup Moonshot AI launched a powerful new product that shook the market and sparked U.S. concerns about China's narrowing gap with global leaders like Anthropic and OpenAI. Qwen3.8 Max has 2.4 trillion parameters, entering the heavyweight category alongside Moonshot's Kimi K3, which has 2.8 trillion parameters and can compete with top products; Alibaba has set equally 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, expanding access beyond the preview version. The interest in these China-made AI systems and models is so high that Moonshot was forced to pause accepting new subscriptions late Sunday evening in response to overwhelming demand.

Alibaba Group Holding Ltd.'s stock briefly rose 5.4% on Monday after the company launched the preview version of its flagship Qwen3.8 Max model, calling it second only to Anthropic PBC's Fable 5. This Sunday release came just days after Moonshot AI unveiled a powerful new product that shook the market and raised concerns in the U.S. about China catching up with global leaders like Anthropic and OpenAI. Qwen3.8 Max has 2.4 trillion parameters, entering the heavyweight category alongside Moonshot's Kimi K3, which has 2.8 trillion parameters, comparable to top products, 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 the model's weights, extending access beyond the preview version. The demand for these China-made AI systems and models is so high that Moonshot was forced to pause accepting new subscriptions late Sunday to respond to overwhelming demand.

It is noteworthy that Qwen3.8 Max will also have open weights. Alibaba had stopped releasing the weights of its leading models earlier this year but seems to have reversed that decision; I suspect this shift is related to Xi Jinping's remarks on AI last week, which doubled down on the open-weights approach:

We should adhere to the principle of openness and win-win, and promote innovation-driven development. As a new engine for world economic growth and an accelerator for transformation in growth dynamics, artificial intelligence is moving from the digital world to the physical world. We should seize this rare historical opportunity, encourage open source, openness, collaboration, and sharing. We should promote the technological innovation, industrial development, and scenario-based applications of artificial intelligence. We should coordinate the transformation and upgrading of traditional industries, support the cultivation and growth of emerging industries, and prospectively plan for future industries, so that all sectors can benefit from artificial intelligence.

China’s strategy is clear: commodify your complements. Note that Xi Jinping explicitly links openness with AI "moving from the digital world to the physical world"; the physical world is a world dominated by China, which will benefit immensely from widely available AI models in areas like robotics.

Similarly, China does not want the U.S. to gain an asymmetric advantage in AI; if China can weaken U.S. frontier laboratories while strengthening any and all potential U.S. opponents, that is even better, and it can benefit from the innovations attached to an open ecosystem.

The Distillation Issue

Also, do not expect China to take any measures against the distillation attack on frontier laboratories. I believe attributing all the success of Chinese labs to distillation is misguided, but pretending that distillation has not given Chinese labs a tremendous advantage is equally misguided. This advantage has truly manifested in the past year as reinforcement learning post-training has become increasingly important to model performance. Chinese labs do not have to build reinforcement learning environments from scratch but can simply use the models from frontier labs as teachers to achieve rapid improvements at far lower costs (this is not the only reason Chinese model development costs less, but it is an important one).

Interestingly, the most important use case 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 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 China:

Distillation does not explain all of China’s leading advantages in open models. Chinese labs have world-class researchers, abundant compute power, strong pre-trained models, software-hardware co-design, and rapidly improving post-training capabilities. But distillation compresses the costly last gap between strong foundations and near frontier systems. Even if distillation constitutes a small share of China’s overall capability, it occupies a significant portion of its relative advantage over U.S. open models.

New enforcement mechanisms will make mass distillation more difficult, slower, and costly for Chinese companies. However, enforcement cannot eliminate distillation supported by state actors. Therefore, every frontier advance 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 is worth repeating: because U.S. open-weight model manufacturers must follow the service terms of frontier labs, they (1) are less favorable than Chinese substitutes, (2) ultimately distill what is distilled, merely taking the detour through Chinese labs. Wouldn’t it be better if Western open-weight model manufacturers could access directly from the source?

To this end, there’s a more intriguing question surrounding distillation: what exactly is wrong with it? After all, aren’t large language models just a distillation of all knowledge on the open internet, scraped and distilled into a model being distilled by frontier labs? Who is actually being harmed here?

In fact, this paradox is the solution. I believe open-weight models benefit innovation (and, based on the above, I think frontier labs will be fine), but reliance on China is an issue. The U.S. should move through legislation that (1) clearly states that collecting data for training models constitutes fair use, and (2) at least prohibits the service terms banning distillation for U.S. companies. Preventing distillation—essentially querying APIs— is nearly impossible; the U.S. should pursue another pathway, leaning towards a new copyright policy that both protects laboratories and ensures that the knowledge they acquire fuels further innovation for others.

Reasons for Concern

The entire article has been aimed at dispelling excessive reactions to Kimi K3 and Chinese open-weight models; however, there is a reason for concern, and that is cybersecurity. Consider this story from The Stack:

Hugging Face reported that its production infrastructure was compromised by "autonomous" AI agent systems earlier last week. The platform's security team was initially impeded in their incident response (IR) by unnamed U.S. LLM frontier model guardrails, "they could not distinguish between incident responders and attackers," they said. Therefore, Hugging Face's defenders resorted to using the open-source GLM 5.2 model from the Chinese Z.ai lab— running it on their own infrastructure to analyze over 17,000 logs or footprints left by the attackers.

For the New York-based Hugging Face, this is a striking public acknowledgment that the company allows users to collaborate on models, datasets, and applications, hitting a milestone of $100 million ARR this summer. In an incident report, the company suggested defenders "prepare a capable model [italicized by us] that can run on your own infrastructure before an incident, to avoid guardrail lockouts and to prevent attacker data and credentials from leaving your environment."

It is hard to overstate how wrong the panic response of the Trump administration to Anthropic releasing Fable was, especially as it exacerbated Anthropic's worst tendencies, assuming that only they can be trusted to use powerful AI. In a world where there is only one AI, it may make sense to keep the strongest cybersecurity capabilities reserved for the U.S. government and trusted allies; however, this is not the world we live in.

Models that are already, and will continue to be, fully capable of launching cyber-attacks against existing infrastructure will be—and already are—widely available. The best defense—the only viable defense—will be to ensure that defenders also have access to the best models. Right now, due to the directives of the Trump administration, 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!

A better approach is clear: first, relax the restrictions on Fable and Sol for cybersecurity, and second, ensure that American open-weight model manufacturers are in a level competitive environment with China. Yes, frontier laboratories will raise a ruckus, but the government should realize that yielding to their hysteria has left the U.S. in a position where American companies are dependent on China for defense. Let frontier laboratories win by doing better; do not let them define security or protection, and do not let them pull up the ladder of human collective knowledge. China faces enough competition as it is; allowing them to wave the flag of openness and innovation is simply surrendering our greatest advantage.