By Jacob Johnson | Zero-Knowledge Researcher
The warning arrived without a timestamp, without a code repository, without a single line of Solidity to verify. It came as something far more dangerous: a veteran economist's calm assertion that the market is building on uneven foundations.
Abby Joseph Cohen, the Goldman Sachs strategist who called the 1990s bull market with unsettling precision, has issued a dual warning that reads like a security audit for the global economy: the economy is uneven, and AI investment is unsustainable.
She didn't mention blockchain. She didn't mention crypto. But as someone who has spent the past six years auditing smart contracts, verifying zero-knowledge proofs, and watching capital flow through immutable ledgers, I can tell you this: her warning translates directly into the language of the blocks.
This is the technical breakdown she didn't provide. And it's the one you need.
The Hook: When Valuation Outpaces Verification
In 2018, I spent six weeks compiling Gnosis Safe contracts on a local testnet. I found three signature malleability vulnerabilities that had been missed by the initial audits. The lesson I took away wasn't about Gnosis—it was about the gap between what projects claim and what they demonstrate.
Cohen's warning is essentially the same lesson, applied to AI.
The market is currently pricing AI companies as if their growth curves are invariant—mathematically fixed, unchangeable, eternal. The S&P 500's AI-driven concentration has reached levels that only appear once or twice a generation. And the actual economic data? Uneven. Fragmented. Divergent.
She's telling you the invariant is broken.
Here's what I mean by that. In decentralized finance, the constant product formula xy=k is the backbone of automated market makers. It's elegant, deterministic, and mathematically enforced. The invariant holds because the code enforces it. But when Uniswap V2 launched, I noticed something: the formula assumed liquidity providers would behave rationally. It assumed the fee structure would incentivize balanced pools. The code was correct, but the economic model* had hidden vulnerabilities.
The AI economy has the same problem. The technology is real. The proofs-of-concept are verifiable. But the capital allocation model—the human incentive layer—is not sustainable.
Cohen isn't questioning the mathematics of large language models. She's questioning the economic invariants that are supposed to keep the system balanced.
The Context: From Interest Rates to Zero-Knowledge Proofs
You can't understand Cohen's warning without understanding the macro context, which requires a quick audit of the policy environment.
Since 2022, the Federal Reserve has been navigating a course that would make a yield-farming protocol blush. They've hiked rates faster than any cycle in four decades, then held them at restrictive levels while inflation data whipsawed. The result? A peculiar form of liquidity fragmentation—but this time, not across chain ecosystems. It's fragmented across the real economy.
Small businesses face borrowing costs that would make a leveraged DeFi position look conservative. Meanwhile, mega-cap technology companies have access to a global capital market that's effectively printing money for AI infrastructure. The result is a two-speed economy.
This is exactly the kind of structural inefficiency I analyze when I look at cross-chain bridges. When the invariant breaks, you get issues. You get exploits. You get reorgs.
Cohen's analysis confirms that the economy is living through a forked state: one chain is at full validation, enjoying high transaction volumes and network effects. The other chain is stuck in finality, with low participation and shrinking blocks.
Core Analysis: The AI Investment Cycle as a Ponzi Scheme with Extra Steps
Let me be very precise about this, because it's a claim I don't make lightly.
I've audited smart contracts that turned out to be honeypots—code that looks legitimate but is designed to extract funds from unsuspecting users. I've also audited legitimate protocols that simply had poor incentive structures. The AI investment cycle that Cohen is warning about exhibits characteristics of the second category, with a dangerous potential to become the first.
The Problem of Capital Lockup
When I examine a blockchain protocol, I look at its tokenomics. How is value locked? How does it flow? What are the mechanisms that create incentives?
The AI capital cycle looks like this: - Front-loaded capital expenditure: Companies are spending billions on GPU clusters, data centers, and specialized hardware. - Deferred value realization: The revenue models are largely speculative. The majority of AI companies are burning cash on compute. - Infinite token supply: The "token" here is new AI models, released at an ever-increasing pace, each requiring more compute.
From a protocol perspective, this is a high inflation model. The "total supply" of AI capabilities is increasing exponentially, but the "market cap" of AI-generated revenue is not keeping pace.
Cohen's "unsustainable" warning maps precisely to a sustainability metric I use when analyzing smart contract tokenomics: the ratio of value creation to value extraction. Right now, AI is extracting enormous value from capital markets but creating value at a much slower rate.
The Uneven Economy as a Liquidity Pool Problem
In a decentralized exchange, if one asset becomes too dominant and the other gets drained, the invariant breaks. The AMM becomes unstable. You get extreme slippage. And if it's bad enough, the entire pool gets exploited.
The US economy is now a highly concentrated liquidity pool. A handful of mega-cap AI companies are absorbing the vast majority of capital inflows. The rest of the economy is the low-liquidity side of the pool, facing extreme slippage when it tries to borrow or invest.
Cohen's warning is essentially: the AMM model hides its truth in the invariant. The invariant is under attack.
The Technical Case for AI Overinvestment
I want to give a more concrete technical analysis of what "unsustainable" looks like when you trace the actual supply chain.
Hardware bottleneck: There is a finite amount of high-bandwidth memory. There is a finite supply of advanced packaging capacity. The manufacturing capability for AI accelerators is not infinite—it's governed by physics and the geopolitical constraints of the semiconductor supply chain.
The economics of scale: AI models are designed to scale. But the cost of scaling is not linear—it's superlinear. More parameters, more data, more compute, more energy. The returns on scaling, on the other hand, are diminishing. We're approaching what engineers call the "scale ceiling."
The energy limit: AI infrastructure requires enormous energy. The power grid in many parts of the world cannot support the demand that AI is projected to generate. This is not a model constraint; it's a physical constraint.
When I audit a project, I check whether the assumptions are valid under realistic market conditions. The AI investment narrative is built on assumptions that fail under realistic energy constraints, hardware constraints, and economic constraints.
The Imbalance: A Consensus Failure
In blockchain, we have a term: failure to finalize. It's when the network cannot reach consensus on the state of the ledger.
Cohen is essentially saying that the economy is failing to reach consensus on its own state. The AI sector is at an all-time high, while the broader economic data shows a slowdown. The market is not synchronized.
This divergence is an artifact of "two markets": 1. The AI market: Priced to perfection, with unlimited risk appetite, driven by FOMO, sovereign concerns, and a technology narrative. 2. The rest of the economy: Priced for a potential recession, with tight liquidity, high debt, and a fragile consumer.
In DeFi, when there's a large divergence between the oracle price and the real price, it's an arbitrage opportunity. In the real economy, it's a systemic risk.
The Contrarian Angle: The Blind Spots in Cohen's (and the Market's) Vision
Cohen's warning is valuable, but it's incomplete. And the missing pieces are exactly what I work on professionally.
First, the AI investment might not be as "unsustainable" as it seems. AI is a general-purpose technology, similar to electricity or the internet. The infrastructure buildout may be oversized, but the long-term returns could be extraordinary. We're not in a bubble; we're in a "density" phase.
Second, the "uneven economy" is not necessarily a bad thing. A blockchain, by design, has "uneven" performance across different use cases. You don't measure the health of the ecosystem by the transactions per second of a single side-chain. You measure the security of the entire protocol.
*Third, the key risk isn't the AI investment itself; it's the coordination failure.* The central problem is not that AI is overfunded, but that the economy's "consensus layer"—the policy tools that allocate resources across the entire economy—has failed to adapt to the new reality.
The Security Audit of the Economy
Let me apply the security audit checklist I've developed over years of auditing smart contracts to the current economy:
Check 1: Protocol Logic (GDP growth) — The economy is growing, but unevenly. The logic is valid, but the implementation is flawed.
Check 2: Access Control (Capital allocation) — Capital is not being allocated equally. The AI sector has privileged access to capital, while other sectors are denied.
Check 3: Reentrancy (Inflation) — Inflation is a recursive problem. The more AI is funded, the more assets need to be issued, the more inflation occurs, the more assets are required to fund the AI.
Check 4: Integer Overflow (Debt levels) — The debt levels are a risk of overflow. A default at a major institution could trigger a cascade.
Check 5: Oracle Manipulation (Market data) — The market data is still being controlled by a few dominant players, creating a manipulation risk.
The economic ledger is not simply "uneven." It has a critical vulnerability that is being exploited.
The Takeaway: The Protocol Will Be Redeemed, But Not the Way You Think
I've been through cycles. I've seen the 2018 Ethereum crash, the 2022 LUNA crash, the 2022 DeFi crash. I've learned one thing: the technology survives, but the "certificates" change.
The same will happen here. The AI infrastructure will survive. It's too useful, too fundamental. But the investments that will survive will be the ones that are built on protocols—clear, transparent, verifiable. The ones that are built on "narratives" will be liquidated.
Cohen's warning is not a call to abandon AI. It's a call to verify the invariants. Check the assumptions. Audit the infrastructure. Look at the math. Zero knowledge isn't magic; it's math you can verify.
The AI economy is a proof-of-work system. The proof is in the profit.
What will the next block reward look like? It depends on whether the market can find a new consensus.
The oracle data is clear: the market is at an extreme. The question is whether the protocol will correct the error, or whether the error will correct the protocol.
I'm auditing. You should be too.