Goldman's AI Trade Is Deleveraging: Proofs Over Promises in the Infrastructure Layer

Trust is a bug. And for the past 18 months, the market has been running a buggy program called the AI Trade—a high-beta momentum loop that priced in every possible future earning before the present even compiled. The first real patch to that system was applied last week. Goldman Sachs' data shows the high-beta momentum basket fell 12% in a single week. The AI hedge fund portfolio dropped 10% in five days. Leverage is coming out of the system, and it is coming out fast. This isn't a crash narrative. It's a code review. And if you're not looking at the git history, you're going to miss the most important commit in this cycle: the capital rotation from compute to storage. Proofs over promises. The market is finally asking for receipts. I've spent 28 years in this industry, most of them dissecting protocols and balance sheets with the same forensic tools. I've audited smart contracts that drained millions, and I've watched infrastructure layers get priced like they were immune to physics. The current AI market is no different. The narrative is shifting from the "pick and shovel" hardware phase to the "data storage and retrieval" phase. Goldman is telling us that the AI trade isn't dead—it's just entering a new execution mode. The era of buying the whole sector and watching it go up is over. The era of granular, verifiable profit analysis has begun. This is a fundamental shift in how we must evaluate the AI stack. Let's get into the numbers. The drawdown is significant. A 12% weekly drop in a high-beta momentum basket is not a normal pullback; it's a forced deleveraging event. The AI hedge fund portfolio's 10% loss in five days suggests that even the most sophisticated, targeted bets are getting hit. This is the sound of a crowded trade unwinding. The leverage that fueled the 2023-2024 H1 rally is being withdrawn, and the market is repricing risk. The critical question is not whether the AI trade is over—Goldman explicitly states it isn't—but rather where the residual alpha is hiding. The answer, according to the firm's data, lies in the storage and data center sectors. They are calling these the "most tactically attractive" sectors, with the most significant valuation gaps. The logic is simple: profit recovery has not yet been fully reflected in the stock prices. This is where my "Infrastructure Skepticism" kicks in. In my 2021 NFT metadata audit, I demonstrated that 40% of top NFT collections relied on centralized servers, creating single points of failure. The market was paying for the illusion of ownership while ignoring the fragility of the underlying storage. We are seeing the same pattern now in the AI trade, but on a macro scale. The market is paying for the promise of AI output without adequately pricing the physical and digital infrastructure required to store and deliver it. The GPU narrative is well understood; the memory and storage narrative is not. When Goldman points to storage and data centers, they are pointing to the unglamorous, unsexy back-end of the AI revolution. And that's exactly where the inefficiencies lie. The context is critical. Goldman's analysis is a direct response to a market that is saturated with AI narratives. The semiconductor and AI complex have been moved into short portfolios. This is a major signal. The most beloved hardware names of the last two years are now being used as hedges or outright shorts by sophisticated momentum traders. Meanwhile, software has replaced semiconductors as the largest weight in the three-month momentum long portfolio. This is a rotation of biblical proportions. The market is saying that the value capture is shifting from the "picks and shovels" (hardware) to the "gold miners" (software applications). It is moving from the cost of compute to the value of the output. But let's be precise about what this means technically. It's not just about "software good, hardware bad." It's about the maturation of the AI stack. We are moving from the training phase to the inference phase. Training requires massive, dense compute clusters. Inference requires distributed, low-latency, and high-bandwidth access to data. This is where storage and data centers come in. The model weights, the training data, the KV caches—they all need to live somewhere. The demand elasticity for storage and data center capacity is far higher in the inference phase than in the training phase. Goldman is essentially telling us that the AI value chain is extending, and the profit pools are deepening in areas that were previously ignored. The core analysis here is about the "profit recovery" in storage and data centers. This isn't a speculative bet on future revenue; it's a bet on existing revenue that hasn't been priced in. The memory industry is a classic oligopoly—Samsung, SK Hynix, and Micron control the vast majority of the market. They have been through years of consolidation and supply discipline. Now, with the explosion in HBM (High Bandwidth Memory) demand for AI accelerators, they have a new growth curve with significant pricing power. The data center REITs, similarly, are seeing utilization rates and rental rates improve as cloud providers and enterprises deploy AI infrastructure. The profit is there. The stock prices just haven't caught up to the EPS revisions. This is a quantitative discrepancy that a forensic analyst can sink their teeth into. My own experience with zk-Rollup optimization in 2024 taught me a valuable lesson about the commercialization of infrastructure. We reduced proof generation time by 40% and gas fees by 25% for end-users. The technology was impressive, but the real value was unlocked when we could translate that technical efficiency into a clear business value proposition: lower costs, faster transactions, and verifiable privacy. The same principle applies here. The AI infrastructure layer—storage, memory, data centers—is the "proof generation" for the AI economy. It's the physical and digital substrate that makes the application layer possible. And like ZK-proofs, it's the layer that's hardest to understand and therefore most likely to be mispriced. The contrarian angle is where this gets interesting. The market is treating Goldman's note as a simple "buy storage" signal. That's a mistake. The deeper read is a warning about the fragility of the entire AI infrastructure. The move to put semiconductors in the short book isn't just about valuation. It's about the physical limits of the supply chain. We are hitting the power wall. AI data centers are voracious consumers of electricity. The mention of copper miners in Goldman's note is not a coincidence. Copper is the metal of electrification, and it's a critical input for power transmission. The market is starting to price in the constraints of the physical grid. This is the ultimate infrastructure skepticism: it doesn't matter how much compute you have if you can't power it. Furthermore, the rotation to software as the top momentum weight is a double-edged sword. On one hand, it signals a belief in the commercialization of AI applications. On the other, it's a high-risk bet. Software companies are not immune to the infrastructure bottleneck. If the cost of inference (compute and storage) remains high, the gross margins of AI software companies will be squeezed. The "profit recovery" in storage could actually be a threat to the software layer's profitability. This is a dynamic that the market hasn't fully grasped. The cost of inputs (storage, compute) is rising, which will eventually put pressure on the output (software margins). The current rotation might be rewarding the software layer for future revenue, but it's ignoring the input cost pressure coming from the very sectors Goldman is recommending. This is a classic value chain conflict. Another blind spot is the nature of the "profit recovery" itself. Is the recovery in storage driven by AI-specific demand (HBM, enterprise SSDs for AI workloads), or is it a cyclical recovery in the broader memory market? The distinction is crucial. If it's cyclical, the trade has a limited lifespan. If it's structural, driven by AI's insatiable appetite for data, then the trade has legs. Goldman hasn't clarified this. My experience with the DeFi crash in 2022 taught me that when you can't distinguish between cyclical and structural factors, you are vulnerable to a liquidity trap. The same applies here. The market is betting on structural AI demand, but a significant portion of the current profit recovery could be cyclical, masking the true signal. The catalyst is NVIDIA's Q2 earnings, scheduled for late August. This is not just a single stock event; it's a referendum on the entire AI infrastructure thesis. If NVIDIA delivers a blowout quarter with strong guidance, it will validate the demand side of the equation, which will be positive for storage and data centers. If the guidance is weak, the deleveraging will accelerate, and the "profit recovery" in the storage sector could be delayed. The market is holding its breath. The 9% drop in the AI hedge fund portfolio is a pre-emptive adjustment to this binary event. The positioning is defensive, and the risk is to the downside. Based on my audit of the Optimism testnet in 2020, I know that the most critical vulnerabilities are often found in the "economic sustainability" of the protocol, not just the code. The same is true for the AI trade. The economic sustainability of the entire AI value chain depends on the cost of inference coming down. This requires both cheaper compute (efficiency gains in GPUs) and cheaper storage (increased density and lower cost per bit). The storage companies are making progress on the latter with HBM and advanced DRAM. The compute companies are making progress on the former with new architectures. The key is that the rate of cost decline must outpace the rate of demand growth to create a sustainable profit pool for the application layer. If the cost curve flattens, the entire house of cards collapses. Goldman's advice to focus on "stocks with significant divergences between price and earnings per share" is sound, but it requires a sophisticated understanding of the accounting. In the data center space, the divergence is often caused by depreciation schedules and interest rate sensitivity. In the storage space, it's about memory pricing cycles. A simple price-to-EPS screen will not capture these nuances. You need to stress-test the balance sheets, understand the capital expenditure cycles, and model the impact of rising power costs. This is not a passive investment strategy; it's an active, forensic approach to stock selection. The capital flow is also telling. The fact that Goldman mentions capital rotating into European and Japanese banks, gold miners, and copper stocks is a clear sign of risk-off behavior. This isn't just a rotation within the AI complex; it's a rotation out of the AI complex into more defensive, value-oriented sectors. This suggests that the "AI trade" is not just being recalibrated; it's being de-risked. The smart money is taking profits from the crowded AI trade and parking it in assets that are uncorrelated to the AI narrative. This is a classic late-cycle behavior. The AI trade is not over, but the easy money has been made. The next phase will be much more difficult and will require a level of technical and financial analysis that the average market participant does not possess. Looking at the technical details, the market is mispricing the shift from training to inference. The narrative has been dominated by the massive training clusters and the "God-like" compute power of NVIDIA. But the real value is in the long tail of deployment. Every enterprise deploying an AI chatbot, every developer using an AI coding assistant, every autonomous vehicle on the road—they all require inference compute and the associated storage and data center capacity. This is a much larger, more distributed market than the centralized training market. Goldman's focus on storage and data centers is a bet on this long tail. The profit recovery in these sectors is the first evidence that the inference economy is scaling. The market is still looking at the headline number of GPU sales and missing the quiet, compounding growth in the back-end. The failure mode is clear. If NVIDIA's earnings show a slowdown in data center revenue, the market will panic. The stock will drop, and it will drag down the entire AI complex, including the storage and data center names that are supposed to be the "safe" tactical plays. The correlation during a sell-off is always 1. The "profit recovery" story will be postponed, not canceled. The fundamentals haven't changed, but the market's risk appetite has. This is the risk of the current environment. The signal is good, but the noise is deafening. The only way to navigate this is to have a clear, quantitative framework and to stick to it, ignoring the emotional swings of the market. The final piece of the puzzle is the regulatory environment. The MiCA regulation in Europe is a good example of how policy can disrupt business models. The compliance costs for CASPs are significant, and the stablecoin reserve requirements are stringent. This is killing small projects and consolidating power in the hands of large, established players. The same thing is happening in the AI market. The cost of compliance, the cost of compute, the cost of data—these are all rising. This creates a barrier to entry that only the largest companies can overcome. Goldman's advice to focus on large, established players in the storage and data center space is a recognition of this dynamic. The small players will be squeezed out, and the oligopolies will get stronger. This is not a market for the faint of heart. So, what's the takeaway? The AI trade is not over, but the playbook has changed. The market is moving from a "beta" trade to an "alpha" trade. The easy money from buying the entire sector has been made. The future profits will come from identifying the specific points in the value chain where the profit recovery is real, verifiable, and not yet priced in. Storage and data centers are the current hotspots, but they come with their own risks. The power wall, the cyclical vs. structural debate, and the potential for a negative NVIDIA catalyst are all significant headwinds. The software rotation is a high-risk bet that could be undermined by rising input costs. The only way to win is to do the work, to stress-test the assumptions, and to treat the market narrative with the same skepticism you would apply to an unaudited smart contract. If you're not looking at the physical constraints—power, water, land—you're not looking at the real infrastructure. If you're not modeling the cost curves, you're not understanding the profit pools. If you're not verifying the EPS revisions against the actual demand signals, you're just guessing. Proofs over promises. The market is finally asking for the receipts, and the companies that can provide them—the storage giants, the data center operators, the software platforms with real revenue—are the ones that will survive the deleveraging. The rest are just bugs in the system, waiting to be patched out. The question is whether you are positioned on the right side of the patch.