I remember the summer of 2021, sitting in a cramped Seattle coffee shop, watching a friend pitch a 'blockchain AI' startup that promised to mine Ethereum while training models. The whitepaper was a patchwork of buzzwords—neural networks, decentralized compute, tokenized GPUs. The slide deck had a chart that went up and to the right. Everyone nodded. Nobody asked about the math. That was the era of the 'AI basket trade,' where any project with the letters A and I glued to a protocol could raise a fund. Two years later, Goldman Sachs just published a note that reads like a post-mortem for that same mentality—but applied to the real AI supply chain. The era of blanket valuation premiums is ending, and the market is starting to behave like a crypto bear market: selective, technical, and brutally honest about fundamentals.
Context: The Macro Liquidity Tail That Painted All AI Boats
To understand why this shift matters for crypto, you have to look at the liquidity map of the past 18 months. The Federal Reserve's quantitative tightening paused in late 2023, and a wave of institutional capital—driven by the Bitcoin ETF approvals and the narrative of AI as the next industrial revolution—rushed into anything that could be spun as 'AI infrastructure.' Memory chips, optical networking, data centers, Neoclouds, even specialized power grid companies. They all traded in lockstep. The correlation was so tight that a single earnings beat from Nvidia could lift the entire basket by 5% overnight. It was a liquidity-driven beta trade, not a fundamental alpha trade. Sound familiar? That's exactly how DeFi's 'total value locked' narrative worked in 2020: if you slapped a yield on it, capital flowed.
Goldman Sachs's data shows that in July, the correction hit uniformly. Memory, AI semiconductors, optical communications, data centers, and Neocloud all sold off together by 15-20%. It looked like a classic liquidation cascade—the kind we see in crypto when a leveraged long position on Bitcoin gets wiped out and drags every altcoin down with it. But then August happened. The rebound diverged violently. Optical communications bounced ~32% from the lows. Neocloud ~20%. AI data centers ~17%. Memory managed only ~12%. AI power? A measly ~6%. That divergence is not noise. It's the market starting to price in individual profit cycles, valuations, and real-world fundamentals instead of buying the bundle.
Core: The Crypto Lens on the AI Supply Chain Divergence
Let me map this divergence onto the crypto market structure, because this is exactly what happened during the 'DeFi Summer' to 'Lido Winter' transition. In 2020, every DeFi token—Uniswap, Aave, Compound, Yearn—rose together. The 'DeFi basket' was a narrative trade. Then, in 2021, as the market matured, capital started rewarding specific protocols with real revenue, sustainable tokenomics, and actual user retention. Uniswap decoupled from SushiSwap. Aave outpaced Compound. The uniform premium for 'DeFi' collapsed. The AI market is now at that inflection point.
Goldman Sachs identifies software as the new 'mainline' in the 'Inference Economy.' Think about that: inference is the act of running a trained model, not training it. In crypto terms, training is proof-of-work mining—expensive, energy-intensive, and commoditized. Inference is proof-of-stake validation—cheaper, more scalable, and where the network effects accumulate. The funds are flowing into the layer that can actually capture value from usage, not just from building. Optical communications (the cables and switches that move data between GPUs) rebounded 32% because they are the 'Ethereum of AI infrastructure'—the base layer that every application needs, regardless of which model wins. Neoclouds (rentable GPU clusters) rebounded 20% because they are the 'AWS of AI'—a platform play with recurring revenue. Memory chips only managed 12% because they are the 'ASICs of AI'—hardware that faces price erosion and demand cyclicality.
This is where my technical audit experience kicks in. In 2017, I manually audited 15 ICO smart contracts and found three with reentrancy vulnerabilities. I learned that the most hyped projects often had the worst code. The AI market today is similar: the most hyped segment—memory (HBM, DRAM)—is the one with the most fragile fundamentals. Price increases are slowing, profit revisions are decelerating, and the market is shifting from 'price-up' stories to 'price-stability, long-term agreements, and capital returns.' That's a classic sign of an industry moving from growth to maturity. In crypto, we saw this with Bitcoin mining stocks after the halving: the narrative switched from 'hashrate growth' to 'fleet efficiency and hedging.' The same shift is happening in AI memory.
Contrarian: The Decoupling Thesis That Crypto Investors Should Fear
Here's the counter-intuitive angle: the AI-crypto convergence narrative is overhyped, but not for the reasons you think. The prevailing wisdom is that AI and crypto will merge through decentralized compute networks, zk-proofs for model verification, or tokenized data markets. But the Goldman Sachs divergence suggests something else: the AI supply chain is becoming a standard financial market, not a crypto-anarchist playground. Optical communications, Neocloud, and data centers are being valued on traditional metrics like EBITDA, P/E, and free cash flow. They are not trading on 'AI tokenomics' or 'staking yields.' They are being judged by the same standards as a utility company or a chip manufacturer.
This means that the 'AI multiplier'—the premium that crypto projects attach to themselves by claiming AI integration—is evaporating. Projects like Render Network, Akash Network, or Golem that position themselves as 'decentralized AI compute' will soon face the same scrutiny. Investors will demand to see actual inference workloads, not just token emissions. They will ask for unit economics, not just total value locked. The market is waking up to the fact that running a model on a decentralized GPU network is slower and more expensive than using AWS, and the premium for 'decentralization' is a luxury, not a necessity.
My 2024 study on ETF capital flows showed that institutional investors treat crypto as a risk-on macro asset, not a technology bet. The same logic applies to AI: the 'basket trade' was a macro liquidity play, not a tech revolution. Once the liquidity tide turns, only the projects with real adoption will survive. The rest will be like the 2017 ICOs that raised $100 million and then disappeared. The AI market is now entering its 'post-ICO' phase, where the hype fades and the fundamentals are laid bare.
Takeaway: Cycle Positioning in the AI-Crypto Era
Listening to the silence between market cycles, I see a clear path forward. The AI trade is not over—it's just becoming selective. The winners will be the ones that provide infrastructure for inference, not just training. In crypto terms, that means protocols that enable cheap, verifiable computation for AI agents—think zk-proofs for model execution, or decentralized storage for training data. The memory and power segments will consolidate, much like the DeFi lending market consolidated into Aave and Compound. The era of buying a 'basket of AI tokens' is ending. The era of fundamental analysis is beginning.
The question is: are you still buying the basket, or are you ready to audit the code?