Goldman's AI De-Leveraging Playbook: The Rotation Nobody's Watching

The AI trade isn't dead. It's just no longer a trade for everyone. You don't get paid for buying the sector anymore. You get paid for knowing which part of the stack the smart money is quietly rotating into. Goldman's latest note on the AI trade isn't a eulogy. It's a technical post-mortem on a market structure shift that most retail portfolios haven't priced in yet. The headline numbers are stark. The high-beta momentum basket fell 12% in a week. The bank's own AI hedge basket dropped 10% in five days. That's not a correction. That's a forced unwind. Leverage is coming out of the system, and it's coming out fast. But here's the part that matters: Goldman explicitly says the trade isn't over. The method of extraction is what's changing. The beta is gone. The alpha is just getting started. To understand why this is happening, you have to look at the mechanics of how this market got built. Over the past 18 months, the AI complex was a liquidity sponge. It absorbed every marginal dollar looking for a narrative. Semiconductors, power, infrastructure—it didn't matter. If you could spell GPU, you got a bid. That created a feedback loop. Rising prices attracted more leverage, which pushed prices higher. This is standard momentum dynamics, but it was amplified by the sheer size of the AI narrative. When you have a sector that is essentially the entire market's growth story, it becomes a crowded trade. Crowded trades don't end with a whimper. They end with a structural break in the volatility surface. The real signal is in the factor shifts. Goldman's data shows that software has replaced semiconductors as the largest weight in the three-month momentum long basket. Simultaneously, semis and the AI complex have moved into the short book. Think about that for a second. The asset class that defined the last two years of market leadership is now a crowded short. This is not a casual rebalancing. This is a systematic repricing of where value accrues in the AI stack. My own experience auditing proof-generation circuits taught me that value concentrates where the bottleneck is. For the last two years, the bottleneck was compute. Everyone needed GPUs to train models. But training is a finite problem. Inference is infinite. Once you deploy a model, the compute requirement for each query is small, but the storage and retrieval infrastructure required to serve millions of users is massive. This is why Goldman is pointing to storage and data centers as the tactical sweet spot. The profit recovery in those sectors hasn't been fully priced into the equities yet. Let's get specific about the valuation gap. Goldman flags storage and data centers as having the most significant discrepancy between current price levels and the earnings recovery trajectory. The market is still pricing these names as if they're cyclical hardware plays. They're not. They're the toll roads of the inference economy. Every AI query requires model weights to be loaded from storage. Every conversation requires KV cache memory to maintain context. This isn't speculative demand. This is metered usage. The HBM (High Bandwidth Memory) market is a perfect example. It's controlled by three suppliers, and demand is outstripping supply. That's a pricing power environment that most equity markets haven't fully modeled. The data center REITs and operators are similarly positioned. Their utilization rates are climbing because AI workloads are persistent. They don't spike and fade like a training run. They run 24/7. This is the kind of recurring revenue that the market loves, but the market is still treating these stocks like they're at the mercy of the next CapEx cycle. Here is where the contrarian angle kicks in. Everyone is watching Nvidia's Q2 earnings as the catalyst for the next leg. They're treating it like a binary event. I see it differently. The earnings report is a lagging indicator. The leading indicator is the capital flow data I'm seeing in the options market. The put-call ratios on the semiconductor complex are building. Open interest is clustering on the downside strikes for the next two months. Smart money is hedging against a specific outcome: a beat on revenue that is accompanied by a conservative guide. If Nvidia guides down, or even guides flat, the market will interpret that as the end of the training supercycle. It won't matter that inference demand is exploding. The narrative will shift, and the de-leveraging will resume. The retail side is still positioned long, waiting for a pop. They're not seeing the structural shift in the order flow. The institutional players are already rotating. They're not selling AI. They're selling the crowded parts of AI and buying the ignored parts. That's the game. This is a classic second-derivative trade. The first derivative was buying everything AI. The second derivative is buying the parts of the stack that benefit from the AI that already exists. The software rotation is a signal that the market is looking for actual revenue generation. The AI agent narrative is moving from concept to deployment. Companies that can show a per-seat price or a per-query price are going to get the premium that semiconductors used to get. The gold miners, copper miners, and European banks that Goldman mentions are not random diversifiers. Copper is the physical layer of the power grid. AI data centers consume enormous power, and that power needs to be transmitted. Copper is the bottleneck there. European banks are a value play against a backdrop of regulatory clarity. The capital that left the AI complex is looking for asymmetry. It's not leaving the market. It's leaving the consensus. Let me give you the actionable framework. The de-leveraging cycle has a duration. You don't catch a falling knife by buying the dip on day one. You wait for the volume profile to show exhaustion. Look at the AI hedge basket. It fell 10% in five days. That's a fast bleed, but it's not a capitulation. Capitulation is a 20% drawdown in a single session on massive volume. We haven't seen that yet. So, the risk is that we see another leg down before the structure stabilizes. For the storage and data center names, the entry point is not now. It's after the Nvidia report. If the report confirms the demand shift to inference, these names will gap up. If the report disappoints, they'll get dragged down with everything else, and you'll get a better entry. The key metric to watch is the weekly change in the momentum factor. Software is the long weight now. If that reverses, if semis start to reclaim the top spot in the momentum basket, then the rotation thesis is broken, and you need to reassess. The gold and copper miners are a different play. They're not correlated to the AI earnings cycle. They're correlated to the physical reality of building infrastructure. I would treat them as a hedge against the AI narrative failing to deliver on time. The market is pricing in a smooth transition from training to inference. If that transition is lumpy, if there are supply chain hiccups or regulatory surprises, the physical commodity trade will outperform the digital one. This is a barbell strategy. You have the high-beta tech rotation on one side, and the hard asset inflation hedge on the other. The middle, the consensus AI trade, is where the risk is concentrated. The takeaway is not about being bearish or bullish on AI. It's about being right about the structure. The era of indiscriminate buying is over. The era of forensic stock selection has begun. I've seen this pattern before in the crypto markets. The 2021 NFT mania was the same setup. The infrastructure narrative (Layer 1s) peaked first, then the application layer (marketplaces) had a brief moment, and then everything collapsed when the liquidity dried up. The difference here is that AI has actual revenue. The question is whether the revenue is growing fast enough to justify the valuations that are still embedded in the high-flying names. My bet is that the storage and data center names are the ones with the most asymmetric upside. The market hasn't woken up to the recurring revenue profile of inference infrastructure. When it does, the re-rating will be violent. Watch the Nvidia guide. Watch the momentum factor. And watch the capital flows into the physical commodity names. The clues are all there. The market is just too busy watching the same screen as everyone else to see them. Code is law, but gas fees are the reality. Arbitrage is just efficiency with a heartbeat. The same principle applies to the AI trade. The efficiency is being re-optimized. The heartbeat is shifting from the data center to the software that runs on it. You don't fight the de-leveraging. You position for the re-allocation. The smart money isn't leaving the table. It's changing seats. You should too. But do it with a plan, not a prayer. Check the delta, ignore the drama. The math doesn't care about your conviction. It only cares about the execution.