The AI-Oil Analogy Is a Trap: Why Zhu Su's Commoditization Thesis Misses the Crypto-Native Floor

Zhu Su's analogy that AI will follow oil's path to commoditization has circulated through crypto Twitter like a pandemic. The logic sounds clean: massive capital, state backing, eventual price compression. But the math didn't. His framework ignores the fundamental property that makes AI different from crude: programmatic verifiability and decentralized ownership of compute. In 2023, I audited the tokenomics of 12 crypto-AI projects. Every single one that positioned itself as "commoditized inference" suffered from misaligned incentives. The oil model is a Trojan horse for centralization. Zhu Su, co-founder of Three Arrows Capital, has been vocal post-collapse about macro trends. His latest take: AI will become a commodity like oil, requiring state-scale capital, and ultimately producing thin margins. He points to the capital intensity of data centers, the geopolitical nature of GPU supply, and the eventual convergence of model capabilities. For crypto builders, this narrative is seductive because it suggests a role for blockchain in decentralized compute markets. But the analogy is structurally unsound. I will break down three critical flaws using the dimensions from the original analysis. First, the technical route. The oil analogy assumes that the final product—barrels of crude—are fungible. AI models are not. Even if benchmark scores converge, the qualitative differences in reasoning, alignment, and safety create persistent differentiation. Commoditization in oil arises because an 87 octane barrel from Saudi Arabia and a 93 octane barrel from Texas are interchangeable for most uses after refining. In AI, a fine-tuned Llama 3 model for legal contracts will not substitute for a general-purpose GPT-4o in creative writing. The abstraction layer is too high. As I wrote in 2022, "Speculation masks the absence of utility." Here, speculation disguises the lack of a commodity definition. Second, the commercialization flaw. Oil commoditization led to profit shifting to downstream—refining, distribution, and retail. Zhu Su implies AI will follow: profit shifts from model creators to application layer and infrastructure providers. But he misses crypto's unique mechanism: programmable incentives. In crypto-AI, you can tokenize compute, create permissionless validation of inference, and align staking rewards with quality. This creates a business model that is neither pure commodity nor pure monopoly. It's a hybrid where value accrues to the protocol via fee burn or buyback mechanisms. The oil model has no analog for token sinks or staking yields. "Hype burns out; structural integrity remains." The structural integrity of a well-designed token economy can outlast model commoditization. Third, the infrastructure dimension. Zhu Su correctly identifies compute as strategic. But he defaults to centralization: state-backed hyperscalers controlling the entire stack. Crypto introduces a counterforce: decentralized physical infrastructure networks—DePIN. Projects like Render Network, Akash, and Bittensor are building open compute layers where anyone can contribute or consume. The oil analogy fails because oil extraction is geographically concentrated; compute generation is not. A GPU in a garage in Nairobi is as useful as one in Texas, provided the network protocol is robust and the verification layer is trustless. "Risk is not eliminated by ignoring it." The risk of centralization is real, but the solution is not to accept the oil model passively—it's to build better decentralized markets that align incentives with open participation. Let me drill deeper into the technical dimension. In my audit of 12 crypto-AI projects, I found that those trying to commoditize inference—selling API access at near-zero margins—consistently failed to attract sustainable demand. The reason: enterprise clients need verifiable outputs, not just cheap token generation. They need proof that the inference was run on the advertised model, without tampering or data leakage. Oil has no such trust requirement. A barrel of crude from a state-owned oil company is accepted at face value because the physical product is auditable through sampling. AI inference is not. The output is a function of the model, the input, and the hardware. Without a cryptographic attestation chain, you are buying a black box. "Security isn't a feature — it's the foundation." The oil analogy ignores that security in AI is not about resource control but about algorithmic integrity. Another hidden flaw: the oil model assumes a stable regulatory environment for the commodity. Oil has the Geneva Conventions for tanker routes, OPEC quotas, and national strategic reserves. AI has none of that. The regulatory landscape for AI models is fragmented and rapidly shifting—EU AI Act, US executive orders, Chinese content moderation. A commoditized AI market cannot exist without global standards for safety and bias. These standards are not being developed. Instead, each jurisdiction is building its own walled garden of approved models and data sources. This fragmentation makes the commodity thesis impossible in practice. "Emotion is the variable that breaks the model." Here, regulatory emotion—fear of deepfakes, bias, and job displacement—will prevent the homogenization that Zhu Su predicts. What about the bulls? The contrarian angle. Zhu Su's prediction that compute becomes a commodity in one sense does hold: the marginal cost of a teraFLOP is dropping steadily. The spot GPU market on AWS, GCP, and Azure already exhibits price pressure, with per-hour rates fluctuating based on supply. His thesis that capital intensity will favor incumbents like Microsoft and Google holds if no disrupting protocol emerges to democratize access. Also, his geopolitical framing is accurate: the US-China chip export controls mirror oil embargoes of the 1970s. Crypto-AI projects must navigate this reality. However, the contrarian angle fails to account for crypto's ability to bypass supply chain bottlenecks via tokenization of future compute—essentially creating a futures market for reserved capacity. This is not a commodity play; it's a financial derivative on compute availability. Furthermore, the oil analogy understates the role of data as a differentiating factor. Oil is a raw material; its value comes from energy content. AI models derive value not just from compute but from the quality and uniqueness of training data. Data is not fungible. A dataset of medical records with patient outcomes is not interchangeable with a dataset of social media posts. The commoditization of AI would require data commoditization, which is impossible without violating privacy and consent. Crypto offers a path through data DAOs and token-gated access, but that again is a model of scarcity and permission, not open commodity. Let me quantify: the oil industry has a global market cap of roughly $6 trillion. The AI industry is projected to reach $1.8 trillion by 2030. If AI truly commoditized, its market would be valued based on marginal utility, not on potential surplus. Current valuations of companies like OpenAI ($150B+) are based on expectations of durable moats, not commodity margins. "Every rug has a seam you missed." The seam here is the assumption that AI will converge to a homogeneous resource. It won't. The vector of differentiation shifts from model weights to proof of correct execution. That is where crypto wins. In my experience as a risk management consultant, I've seen too many projects adopt the "commodity narrative" as a marketing shortcut. They claim their token is the "oil of AI," ignoring that oil has a single-use case—energy. Compute is multi-purpose: you can use it to train models, run inference, render graphics, or mine crypto. This versatility means its value is not strictly bounded by model performance. A GPU earning ETH on the side can subsidize inference costs, a dynamic oil cannot replicate. The takeaway: The AI-oil analogy is a trap for crypto builders seeking product-market fit. It leads to premature commoditization mindset—building thin-margin APIs when the real value lies in verticalized, verifiable inference. The crypto-native endgame is not a commodity market but a trust market. Build protocols that provide cryptographic guarantees of model execution, data provenance, and incentive alignment. That is the edge that centralized oil markets never had. And that is where the next generation of value will accrue.