The market is finally asking the question that should have been asked two years ago: what is the actual return on this capital?
Fu Peng's recent analysis of AI's investment return mismatch carries a subtext that extends far beyond Silicon Valley. It describes a systemic delusion: a trillion-dollar infrastructure buildout predicated on future demand that has not yet materialized. The same narrative pattern is playing out in crypto, especially in the Layer 2 and modular blockchain sectors. The only difference is the timeline. AI's reckoning is happening now. Crypto's is still deferred by a few quarters. But the structural mechanics are identical.
Context: The Narrative of Infinite Scaling
Since 2023, the dominant narrative in both AI and crypto has been infrastructure-first. In AI, the story was that more compute, more data, and larger models would inevitably lead to AGI—and that whoever controlled the compute fabric would control the future. In crypto, the story was that more blockspace, more rollups, and more modularity would unlock mass adoption. Both narratives assumed that supply creates its own demand.
But the data tells a different story. In AI, the aggregate capital expenditure of the top seven tech giants (Microsoft, Google, Meta, Amazon, Apple, Tesla, NVIDIA) on AI-related infrastructure rose from roughly $150 billion in 2023 to an estimated $250 billion in 2025. Yet the incremental revenue directly attributable to AI, excluding inter-company cross-sales (AI companies buying compute from each other), is estimated at only $40–60 billion annually. That's a 5:1 capital-to-revenue ratio, and it's deteriorating. The free cash flow of these companies is being compressed, and the market's tolerance for narrative without proof is shrinking.
In crypto, the equivalent is the hundreds of billions of dollars locked in Layer 2 bridge contracts, sequencer nodes, and modular execution layers—most of which are still subsidized by token emissions rather than genuine user demand. The total value secured (TVS) of the top 10 L2s is around $15 billion, but the actual transaction fees generated (excluding token subsidies) is less than $100 million per year. That's a 150:1 capital-to-revenue ratio. The asymmetry is worse than AI.
Core: The Unit Economics Trap
Fu Peng's key insight is that the AI industry is waiting for a breakthrough in "unit compute cost" and "workflow reconstruction threshold." The same applies to crypto infrastructure. The cost of deploying a transaction on Ethereum L1 is $0.10–$0.50. On an L2, it's $0.01–$0.05. But the value of that transaction—the actual economic activity it enables—is often less than $0.001. The marginal cost of execution is still higher than the marginal value created for most use cases. The 2017 ICO mania generated a brief spike in transaction value, but that was speculation, not utility.
What the market is missing is that the "workflow reconstruction" for crypto requires a different set of primitives: reliable agent execution (current success rate of on-chain agents is below 60% for complex tasks), sub-cent transaction costs for high-frequency operations, and standardized cross-chain interfaces. None of these are close to production grade. The technology is still in the "capability demonstration" phase, not the "cost-friendly" phase.
Structure beats speculation every time. The infrastructure buildout in crypto has been driven by venture capital narratives—"liquidity fragmentation is a problem that needs a new L2 solution"—rather than genuine user demand. The real problem is not fragmentation; it's that the existing solutions don't generate enough value to justify their own existence. The narrative that "liquidity fragmentation is a problem" is a VC-manufactured story to sell new products. The market is waking up to this.
Contrarian: The Value Transfer from Infrastructure to Application
Here is the counter-intuitive angle: a slowdown in infrastructure spending is actually positive for the application layer. In AI, as unit compute costs decline, the gross margins of AI application companies (like ChatGPT, Midjourney, etc.) improve. The same is true in crypto. As sequencer costs drop and L2 transaction fees compress, the margins of decentralized applications (DEXs, lending protocols, gaming platforms) expand. The value is shifting from the "pickaxe sellers" to the "miners."
But this shift depends on the miners finding a rich vein. In crypto, the "rich vein" is the killer app that generates sustainable, non-speculative revenue. The closest candidates are stablecoin settlement (e.g., USDC on Solana) and certain DeFi protocols (like Uniswap's fee generation). However, most of the so-called "web3 social" or "gaming" applications are still burning cash. The ROI verification period for crypto applications is similar to AI: the next 2–3 quarters will determine which projects survive.
2017 called. It wants its lessons back. The ICO bubble taught us that infrastructure built without application demand is a ghost town. The same lesson is being retaught now. The current L2 explosion is a repetition of the 2017 sidechain narrative—only this time the capital is larger, the hype is more sophisticated, and the collapse will be more painful.
Takeaway: The Next Narrative Is Capital Efficiency
The market is moving from a "growth at all costs" framework to a "capital efficiency" framework. Projects that can demonstrate a clear path to unit economic profitability—where the cost of acquiring a user (CAC) is less than the lifetime value (LTV), and where the marginal cost of a transaction is less than the marginal value it creates—will re-rate upward. Those that cannot will be compressed.
In AI, the next major narrative is not a bigger model; it's a cost-effective model that fits into enterprise workflows. In crypto, the next major narrative is not a faster chain; it's a chain that generates real economic output per unit of capital deployed. The ROI question is the same. The answer will determine the winners of the next cycle.
The question is not whether the technology works. The question is whether it can work profitably. And that is a question the market is only now beginning to ask.