Who's Afraid of Chinese Models?

The Return of Universal Principles

I often recount an experience from my first day in STRT-431 at the Kellogg School of Management. This was the mandatory introductory course for every first-year MBA student. As I scanned the assigned case studies, I was struck by a glaring omission: there wasn't a single technology company on the list.

When I questioned the professor about this absence, the response was that the course aimed to teach "universal principles" applicable to any business, regardless of the specific industry.

At the time, I found this logic entirely satisfying deeply flawed. From my perspective, the tech sector—specifically software and its distribution—operated on a different plane because it enjoyed zero marginal costs and zero transaction costs. In the world of mathematics, when you plug zeros into a formula, the results are often chaotic.

Aggregation Theory and the Zero-Cost Era

The core of Aggregation Theory is the idea that when marginal costs vanish, the value chain shifts. We moved toward a world of extreme centralization and scale, where the primary competitive advantage was the ability to control demand rather than the ability to distribute supply.

However, the emergence of Artificial Intelligence is fascinating because it is bringing those "old" universal economic principles back into play.

This shift became a focal point of intense debate on X recently, sparked by the release of Kimi K3. This open-weights model from China is rapidly approaching state-of-the-art (SOTA) capabilities, prompting a realization: marginal costs have returned, impacting both immediate pricing and the long-term industry architecture.

---

The Great Misconception: R&D vs. COGS

There is a pervasive myth that open-weights models are free. The logic is that since you can download the weights, you bypass the massive investment required to train the model. While true for the initial creation, this confuses two very different types of expenses.

Fixed vs. Variable Costs

1. Research & Development (R&D): This is a fixed cost. Whether a company generates \100,000$or$$100,000,000$ in revenue, the initial cost to build the model remains the same.
2. Cost of Goods Sold (COGS): This is a variable cost. In AI, this refers to inference—the actual compute cost of running the model to generate an answer.

Unlike traditional software, AI has significant COGS. We can express this relationship using a simple linear function:

$\text{Total COGS} = \text{Cost per Token} \times \text{Total Tokens Generated}$

For example, if the cost to generate tokens for \1$of revenue is$$0.50$, the scaling looks like this:

Therefore, Kimi K3 is not "free" to operate. While its pricing is lower than Sol (see below), the cost of serving the model is still a critical factor.

---

Tokens vs. Intelligence

Nvidia's CEO, Jensen Huang, views the future as a series of "token factories." From a hardware perspective, this is logical. GPUs are agnostic; they simply produce tokens as efficiently as possible. This leads to a focus on metrics such as:

• tokens-per-second
• time-to-first-token
• tokens-per-watt

This "token-centric" framing worked during the first AI paradigm (the ChatGPT era), where tokens were delivered directly to users. However, we have entered the Reasoning Era, which complicates the math.

The Reasoning Paradox

Reasoning requires "chain-of-thought" processing, which generates a massive volume of internal tokens before reaching a final answer. Not all models are equally efficient; some require far more tokens to arrive at the same conclusion.

If Kimi uses significantly more tokens than Sol to solve the same problem, its lower price-per-token is an illusion—the total COGS for the answer may actually be higher.

The key insight: A token is not a commodity. A gallon of oil is a gallon of oil, but a token from one model is not the same as a token from another. What is actually fungible is Intelligence.

def calculate_intelligence_cost(token_price, tokens_required):
    # The cost of the 'answer', not the cost of the 'token'
    return token_price * tokens_required

# Kimi might have lower token_price but higher tokens_required
# Sol might have higher token_price but lower tokens_required

The Components of Intelligence COGS

The actual cost of producing intelligence is determined by several technical factors:

• Model Footprint: The size of weights and runtime state, which dictates the amount of expensive HBM (High Bandwidth Memory) and accelerators needed.
• Memory Efficiency: Architectural decisions that optimize the KV cache, allowing for higher concurrency.
• Serving Efficiency: The use of prefix caching, scheduling, and batching to maximize GPU utilization.
• Token Efficiency: The ability of the model to reach the correct answer using the minimum number of tokens.