Over the past 12 months, the cost to train a state-of-the-art large language model has dropped by roughly 40%. Meanwhile, AI tokens in crypto have seen market caps inflate by over 200% on narrative alone. This divergence is not a coincidence—it is a structural signal. I have seen this pattern before. In 2020, when DeFi yields began to converge, the same divergence between narrative and fundamentals led to a violent rebalancing. Smart money doesn’t trade the headline; it trades the block time.
Zhu Su, co-founder of Three Arrows Capital, recently drew a parallel between AI and oil. His thesis: AI will eventually commoditize, becoming a capital-intensive, low-margin infrastructure play similar to the energy sector. The analogy has been retweeted thousands of times, but as a DeFi yield strategist who has spent years dissecting liquidity flows and capital allocation, I see deeper implications for how we should value crypto-AI projects. The analogy is not flawless, but it provides a framework that cuts through the hype. Let me walk you through the data, the mechanics, and the trade.
Context: The Analogy in Plain Terms
Zhu Su argues that AI, like oil, will follow a path from scarcity and high margins to abundance and commoditization. Oil started as a speculative boom—think of the Pennsylvania oil rush in the 1850s—then consolidated into a capital-intensive industry dominated by a few players with massive infrastructure. AI is on a similar trajectory. Training a frontier model today costs hundreds of millions of dollars; inference costs are dropping but still tied to hardware. Zhu Su says that eventually, the value will shift from the model itself to the infrastructure beneath it: compute, energy, and distribution networks.
From a trader’s perspective, this is a bet on capital efficiency. The oil analogy suggests that AI companies will not sustain software-like margins. Instead, they will face the same pressure as refiners: constant capital expenditure, thin margins, and exposure to input costs (chip prices, electricity). This is not a new idea. I saw it play out in DeFi during the summer of 2020. When Compound and Uniswap launched liquidity mining, yields were 200%+. Within six months, they collapsed to 20%. Commoditization happened fast. The same will happen to AI.
Core: Order Flow Analysis and Capital Allocation
Let me break down the analogy using the same framework I apply to DeFi protocols: capital efficiency, liquidity depth, and yield sustainability. First, capital efficiency. Training a large model requires massive upfront capital. According to public estimates, training GPT-4 cost around $100 million. Inference costs for a single query are still higher than a typical software transaction. This is analogous to building an oil refinery: the initial outlay is enormous, but once built, the marginal cost per barrel is low. The problem is that AI models are not like oil. They improve with scale, but they also face diminishing returns. The cost to push accuracy from 90% to 99% grows exponentially. This means that the capital required to maintain a moat increases over time, not decreases. That is a bearish signal for token holders.
Second, liquidity depth. In crypto-AI projects, liquidity is often artificially inflated by token incentives. Take Render Network or Akash Network. Their token prices surged on the narrative of decentralized compute. But when I cross-reference their on-chain usage against the market cap, the numbers do not match. Render’s active node utilization hovers around 30%. Akash’s monthly revenue is less than $500,000. Compare that to centralized GPU providers like CoreWeave, which reported $500 million in revenue in 2023. The market is pricing AI tokens for a future that may never arrive. Sentiment buys the dip; data fills the position.
Third, yield sustainability. I design yield strategies for a living. The most profitable plays are ephemeral. In 2021, I ran a script that captured arbitrage between DAI lending rates and stablecoin peg deviations. It yielded 45% APY for six months, then the opportunity vanished. The same will happen to AI tokens that claim to share revenue. Most lack a real yield mechanism. They are governance tokens with a narrative wrapper. If AI commoditizes, the yield on these tokens will compress to zero. The real yield will come from owning the infrastructure—the GPUs, the energy contracts, the data pipelines. Not from holding a token with an inflationary supply.
Contrarian: The Retail vs. Smart Money Divergence
The contrarian angle is this: most retail traders believe AI tokens are the next big thing, a bet on technological disruption. That is the narrative. The data tells a different story. Large holders of AI tokens are not buying to accumulate. They are buying to distribute. Look at the top 100 wallets for any major AI token. The concentration is high, and the flow is overwhelmingly from exchanges to new wallets. This is a classic distribution pattern. Smart money does not accumulate when the narrative is loudest; it accumulates when panic sets in. Right now, there is no panic. There is FOMO.
Zhu Su’s analogy hits a nerve because it implies that the winners in AI will be the same as in oil: those with the cheapest access to capital, the largest scale, and the best control of supply chains. For crypto, this means decentralized compute networks face an uphill battle. They cannot match the economies of scale of AWS or Azure. They cannot subsidize hardware through cloud service bundling. The analogy suggests that centralization will win in AI infrastructure, not decentralization. That is a bitter pill for the crypto-native community, but it aligns with the data. The largest GPU clusters are owned by hyperscalers, not DAOs.
Another blind spot: the analogy ignores the role of data as a moat. Unlike oil, which is homogeneous, data is differentiated. The best models are trained on high-quality, proprietary data. This creates a barrier to commoditization. However, the trend in open-source models (e.g., Llama 3, Mistral) suggests that base-level capabilities are commoditizing. The differentiation will shift to fine-tuning and vertical applications. That is where crypto can play—not in base compute, but in data marketplaces and verification layers. The contrarian trade is to short broad AI tokens and long specific data infrastructure plays.
Takeaway: Actionable Price Levels and Positioning
Based on my framework, I see two clear trades. First, short the broad AI token index. The market cap of the top 10 AI tokens stands at $15 billion, with an average P/E (if they had earnings) of infinity. Over the next 12 months, as AI commoditization accelerates, these tokens will reprice to zero relative to their infrastructure peers. Second, accumulate concentrated positions in protocols that control physical compute or energy assets. Look for projects with auditable asset registers, proven revenue, and transparent operations. Yield comes from structure, not sentiment.
Price levels? If Render breaks below $5 on a weekly close, the next support is $3.20. That would represent a 60% drawdown from current levels. Akash below $1.50 is a sell signal. Conversely, if any token-backed GPU network can demonstrate 60%+ utilization and positive cash flow, it becomes a buy. But that is not today.
Zhu Su’s analogy is not a prediction; it is a structural lens. Use it to filter noise. In a bear market, survival matters more than gains. I have lived through 60% drawdowns. I know that the portfolio that stays flat in a downturn beats the one that chases alpha. Position accordingly. And remember: smart money doesn’t trade the headline; it trades the block time.