Hook: The Cost Signal

The number is $0.94.

For every single task processed by Kimi K3, the model burns through nearly a dollar in compute. Across the table, GPT-5.6 Terra does it for $0.55. That is a 71% premium for playing in the same league.

Wall Street noticed. Gavin Baker, CIO of Atreides Management, published a note on X claiming this new model from Moonshot AI represents a turning point for the entire AI industry. Not because K3 is better. But because its existence proves that the entrance price for the frontier game has just collapsed.

Baker's argument is elegant in its simplicity: The era of model profit is ending. The value will flow elsewhere.

I do not trust the hype. I trust the gas fees. And K3's gas fees are telling a story the market has not fully priced in.

Context: The Crowded Frontier

The AI model market is currently an oligopoly. On one side, OpenAI and Anthropic have dominated both the performance metrics and the public imagination. On the other, Google DeepMind and Meta operate from positions of deep technical strength. The barrier to entry was assumed to be prohibitive: billions in compute, top 0.1% research talent, and years of iteration.

Kimi K3, developed by Beijing-based Moonshot AI, challenges that assumption. Its performance is reportedly competitive with frontier models from OpenAI and Anthropic. But the cost data from Artificial Analysis tells a different story. At $0.94 per task, K3 is significantly more expensive than GPT-5.6 Terra ($0.55) and only marginally cheaper than GPT-5.6 Sol ($1.04).

This is the critical data point Baker seizes. K3 proves that the technical gap is shrinking. New entrants can reach the frontier. But they cannot yet do so efficiently. The token efficiency of K3 is simply not competitive.

The code does not lie. It just costs more to run.

Core: The Efficiency Gap as Systemic Vulnerability

Let me dissect this "token efficiency" problem in cold, forensic detail. It is not just a cost issue. It is a structural vulnerability.

Session 1: The Broken Pipeline

Every production AI pipeline has a throughput constraint. The bottleneck is always in the inference engine. A model that costs $0.94 per task compared to $0.55 means you need 71% more compute hardware to serve the same number of users. That hardware cost is not a one-time CAPEX. It is a recurring OPEX burden that compounds with every query.

For a startup like Moonshot AI, this is existential. They can subsidize the cost temporarily with venture funding. But the core economics do not work. A $0.94 cost per task with no clear path to $0.30 or lower means the unit economics are inverted. They lose money on every task they serve.

Based on my experience auditing the business logic of AI infrastructure companies, I have seen this pattern before. A company builds a technical marvel, but the capital efficiency is broken. The market rewards the company, not the technology. In this case, the company (Moonshot) is burning cash to acquire users that its competitors can serve at a significant discount.

Session 2: The Model Profit Myth

Baker's core insight is that model profit is a myth sustained by a lack of competition. In a duopoly, OpenAI and Anthropic could charge premium prices. They could maintain high gross margins and reinvest those into building moats—proprietary data pipelines, fine-tuning tools, and enterprise relationships. The model profit thesis, popularized by investors betting on frontier AI companies, assumes that the company that builds the best model captures a disproportionate share of the value created.

Baker is rejecting that thesis. He argues that as more models reach the frontier, the pricing power of any single model collapses. The winner does not take all. The winner gets a commoditized product with razor-thin margins.

The data on K3 supports him. A model that is competitive but 71% more expensive does not command a 71% premium. It commands zero premium. Users will switch to the cheaper option immediately. The model profit has already been competed away. It just has not been reflected in the financial statements yet.

Session 3: The Allocation of Value

Where does the captured value go? Baker is explicit: upstream to infrastructure providers (NVIDIA, cloud providers, data centers, power companies) and downstream to application layer software companies (SaaS, vertical AI tools). The middle layer—the model layer—gets squeezed.

This is not a controversial take. It is the same dynamic that plays out in every technology cycle. The layer with the highest barriers to entry initially captures the profits. Then the barriers erode, and the profits flow to the layers that do not depend on a specific provider.

In the cloud computing analogy, AWS, Azure, and GCP initially captured enormous margins. Then the market matured, pricing pressure emerged, and the value flowed to the application companies that built on top. Salesforce, Workday, and Shopify all captured more long-term value than any single cloud provider would have liked.

The same is happening here. The GPU providers (NVIDIA), the power generators, the data center operators are the ultimate beneficiaries. Baker is explicitly betting on them.

Contrarian: What the Bulls Got Right

I am a skeptic by nature. But even I must acknowledge the counter-arguments.

First, the performance gap may matter more than the cost gap. If K3 significantly outperforms GPT-5.6 Terra on critical benchmarks, the $0.94 cost becomes acceptable. Moonshot AI could be targeting a niche where performance is the priority, not cost. Think medical diagnosis, legal contract analysis, or high-stakes financial modeling. In those markets, a 71% premium is a rounding error.

Second, the efficiency problem is fixable. Token efficiency is a function of model architecture, quantization, and inference engine optimization. It is not a fundamental law. With enough engineering, K3's cost could drop significantly. The company could be playing the long game: launch with a premium price, iterate on efficiency, and drop the cost over time.

Third, Baker may be wrong about the time frame. The transition to a commoditized model layer could take years, not months. OpenAI and Anthropic have deep enough pockets to buy time. They can cut prices, launch new products, and maintain their ecosystem advantage. OpenAI has ChatGPT Pro at $200/month. Anthropic has Claude for Enterprise. These are sticky user bases that do not switch based on a 50% price difference.

The rug was pulled before the mint even finished, but sometimes the minting takes a decade.

Takeaway: The Accounting of Reality

The Kimi K3 story is not about a new model. It is about a new arithmetic. The model profit thesis was built on the assumption of scarcity. K3 proves that scarcity is an accident of time, not a feature of the technology.

The question every investor must answer is simple: Do you believe the model layer will be a commodity in five years? If yes, follow Baker's logic. Bet on the infrastructure. Bet on the applications. Avoid the model companies.

If you believe that OpenAI or Anthropic can build a moat deep enough to sustain premium pricing, then ignore K3. Ignore the $0.94 per task. Ignore the efficiency gap.

I know which side the gas fees support.

The code does not lie. The cost data does not lie. The arithmetic is brutal.

The turning point has arrived. It just costs $0.94 per task.