A few weeks ago, on the sidelines of the World AI Conference, Kevin Kelly made a statement that many investors are treating as an oracle’s prophecy. He said that Chinese open-source models have an advantage. The reason: token cost.
On the surface, it sounds like a simple, logical conclusion. The industry is moving towards efficiency. Lower cost wins. Qwen3, DeepSeek-V3, Yi-Lightning—these models are being served at a fraction of the price of GPT-5 or Claude 4. The narrative is clean. The crowd is buying it.
But let’s pause. As a token fund manager who has watched three cycles of "this time it’s different," I’ve learned that the cleanest narratives are often the most dangerous. They make perfect sense on a slide deck. In reality, they ignore the messy, counter-intuitive mechanics of how markets actually shift.
This isn’t about whether Chinese models are cheaper. It’s about whether the crypto-native logic that governs competitive advantage is being misread. The AI industry is currently structured like a centralized SaaS market. Token cost is a feature, not a moat. And in a market where the marginal cost of bits tends toward zero, competing on price is a race to the bottom, not a path to value capture.
waiting for the market to realize this is a dangerous game.
Let’s look at the context. The current AI infrastructure stack is a mirror of the early blockchain trilemma. You have raw compute (L1 security), middleware and inference engines (L2 rollups), and applications (dApps). The narrative that "Chinese open-source reduces token cost" is similar to the narrative that "Optimistic Rollups will reduce gas fees." It’s true on a technical level, but it misses the point. The value isn’t in the cheapest transaction. The value is in the composability and trust that the ecosystem provides.
The core technical analysis here isn’t about AI model efficiency or chip costs. It’s about the nature of the narrative itself: the "cost" of a token is not just a dollar amount; it’ a psychological price.
Consider the dynamics of the current market. We are in a sideways, accumulation phase. Capital is waiting. Liquidity is hiding. In such a market, a narrative about "lower cost" can create a local price floor for the associated assets (like GPUs, or tokens of inference projects). But it fails to create a strong, off-ramp demand. Why? Because the audience isn’t looking for a cheaper alternative. They are looking for a different alternative. The crypto crowd—and especially the sophisticated token fund LPs—have been trained by years of DeFi and NFT cycles. They know that the winning projects are not the ones that are slightly cheaper than the competition. They are the ones that define a new category.
Kevin Kelly’s argument is based on a rational, utility-driven view of the market. It assumes that a developer, or a large enterprise, will choose the model that gives them the most output per dollar. This is true in a mature market like AWS vs. Azure. But in a growth market driven by venture capital and speculation, the rule is network effect first, monetization later.
The contrarian angle is this: the "advantage" of lower token cost might actually be a disadvantage in the current crypto-native AI ecosystem.
How? Because the most capital-efficient way to capture value in a "token cost" narrative is not to build a cheaper model. It is to build the infrastructure that routes the demand for model execution. In crypto, we call this the "gas market." Think about Ethereum. No one cares about the raw cost of computation on Ethereum. They care about the price of block space. The value is captured by the validators and the sequencers, not by the code that computes.
The same logic applies here. If Chinese open-source models are so cheap, the massive profit opportunity isn’t in operating a model. It’s in operating the private, verifiable inference network that uses these models as a compute backend. I am looking at projects that are building the "EigenLayer for AI"—the restaking layer for inference jobs. These projects benefit from the low cost of Chinese models (their expense side is low) but sell their service at a premium due to the "proof of inference" (their revenue side is high). The cheap model is just an input. The valuable asset is the oracle feed that proves the calculation was correct.
Let me be specific. We are tracking a protocol, let’s call it "LayerZ." They provide verifiable inference. Their backends currently use DeepSeek-V3 and Qwen3. Their internal cost per inference is less than $0.0005. They charge their customers $0.002. That’s a 4x margin. The narrative isn’t "cheap inference." The narrative is "trusted inference for high-stakes DeFi or medical AI." The low token cost is their secret weapon, but it’s not their pitch. If they pitched on price, they’d compete on price. By pitching on trust, they compete on a scarce resource.
This is where my experience with Chainlink comes in. In DeFi, the narrative was always about price feed accuracy. But the technical bottleneck was feed latency and cost of transmission. Chainlink had to solve the paradox of decentralized security with centralized oracle nodes. They succeeded not by being the cheapest oracles, but by being the most reliable. The "cost" to a protocol was secondary to the "security budget."
I see the same pattern emerging in AI. The key unresolved question from the Kevin Kelly interview isn’t "how cheap can it get?" It’s "how trusted can the cheap version be?" Can a model that costs $0.0005 per query pass a KYC audit against the G7’s "Model Safety Act"? No one knows. If it cannot, then the "cost advantage" is only applicable to a grey market of quick, unregulated inference. That’s a small TAM.
waiting for compliance standards to shake out is a long game. The smart money is already positioning around the compliance infrastructure, not the compliance-adjacent models.
Let’s talk about the contrarian view on competition. The report on the Kevin Kelly interview worried that China’s open-source models would compete against Meta’s LLaMA. That’s a false dichotomy. The real competition is not model-to-model. It’s model level vs. application level. The true victory condition for a model is not beating another model on a benchmark. It’s being the default layer that applications are built on top of. This is a platform battle.
In this case, the "token cost" advantage of Chinese models is a double-edged sword. It lowers the barrier to entry, yes. But it also commoditizes the model itself. If everyone can access a cheap, capable Chinese model for $0.0005, then why would a new startup build on a more expensive model? They wouldn’t. The model becomes a commodity. The profit shifts to the layers above.
This is exactly what happened in crypto with Ethereum. The base layer became expensive (high gas). The innovation moved to L2s. Now, the L2s are fighting over cost, and the true value accrues to the bridges and the interoperability protocols. The base asset (ETH) struggles because the value was pushed up the stack.
If Chinese open-source models succeed, they will be "L1s" of the AI stack, but in a world with many L1s, the value is in the cross-chain infrastructure. The money to be made is not in mining the commodity; it is in building the trucking company that ships the commodity.
waiting for the right infrastructure narrative to mature.
So, what is the takeaway? The industry is rushing to build cheaper robots. I think we should be building the control room. The next 12 months will see a massive divergence between two investment theses. Thesis A: "Buy the cheapest GPU/inference token because adoption will grow." Thesis B: "Buy the verification/oracle/sequencer infrastructure that enables the trust of the cheapest inference."
The crowd will chase Thesis A. The narrative is simple. It’s the same mistake made in early DeFi, where everyone bought the farming tokens but missed the underlying lending protocols. The silent winners will be those who identified the points of friction. In a world of ultra-cheap AI, the friction isn’t price. It’s provenance. It’s verification. It’s the ability to prove that the cheap output wasn’t manipulated.
Kevin Kelly is right about cost. But he’s wrong about where that cost advantage finally settles. It won’t settle in the model’s market cap. It will settle in the trust market’s infrastructure. And the market hasn’t even priced that in yet.
The question to ask yourself: In a world of infinite, cheap tokens, who gets paid for the firewall?
