Goldman Sachs just dropped a quiet bomb on the AI trade narrative. The August 14 note, buried in institutional terminal feeds, essentially declares the end of the 'AI basket' era. No more indiscriminate buying of anything with a GPU or a data center lease. The market is now splitting winners from losers by profit cycle, valuation reality, and fundamental differentiation.
For those of us who have spent the last decade watching capital rotate through narratives—from ICOs to DeFi Summer to NFT mania—this feels eerily familiar. The 'AI trade' was never a single asset class. It was a liquidity pool with a collective delusion. And now, the dam is cracking.
Context: The Narrative Cycle of AI Hype
Remember the summer of 2020? Every DeFi protocol with a yield farming campaign saw its TVL inflate by 10x in weeks. The market treated all of them as equal—until the incentives dried up. Then, the real fundamentals emerged. Uniswap survived. SushiSwap clung on. A hundred others vaporized.
AI is following the same script. The first phase (2022-2024) was a liquidity tsunami: every startup with 'AI' in its pitch deck raised millions. The second phase (2024-2025) was the infrastructure buildout: data centers, optical interconnects, memory chips, and 'neocloud' providers all rode the same wave. But now, we are in phase three: the great unbundling.
Goldman Sachs observes that during July's correction, sectors like Memory, AI semiconductors, optical communications, data centers, and Neocloud were sold off in lockstep—a classic liquidation of correlated positions. But the August rebound revealed brutal divergence. Optical communications bounced 32% from lows. Neocloud: 20%. AI data centers: 17%. Memory: only 12%. AI Power: a pathetic 6%.
This is not a random walk. It is a signal that capital is now reading the fine print.
Core: The Inference Economy vs. The Memory Trap
Let me be blunt: the Memory narrative is dead. It was always a story about price hikes and profit revisions—a commodity play dressed in AI clothing. When you strip away the hype, what remains? Samsung and Hynix are still beholden to DRAM/NAND cycles. The AI boom gave them a temporary pricing umbrella, but that umbrella is now leaking. Goldman Sachs notes that the market is shifting focus from 'price increases and profit revisions' to 'price stability, long-term agreements, and capital returns.' Translation: the only thing that matters for memory companies now is whether they can lock in contracts at stable prices, not whether they can keep raising prices.
Based on my experience auditing smart contracts for DeFi protocols, I've seen this pattern before. When a project's revenue model depends on a single volatile variable (like token price or transaction fees), it inevitably collapses under the weight of its own hype. Memory is no different. The real value accrual in AI is not in the silicon—it's in the software that orchestrates the inference.
This is where the 'Inference Economy' enters. Goldman Sachs points to software as the emerging new mainline. Why? Because inference is the bottleneck. Training models is a one-time capital expenditure. Running inference is a recurring operational expense. The winners in the next phase will be those who own the inference layer—the middleware, the orchestration, the API gateways that connect AI models to real-world applications.
I have been tracking this shift since 2026, when I collaborated with a small team to prototype an autonomous economic agent that negotiated micro-transactions for data access. That agent, built on a simple Ethereum-based smart contract, consumed more gas in inference than in training. The lesson was clear: the cost of thinking is higher than the cost of learning. The market is now pricing that asymmetry.
Contrarian Angle: The Neocloud Mirage and the Power Paradox
Everyone is bullish on 'neocloud' providers—the decentralized compute networks that claim to undercut AWS. But the numbers tell a different story. The 20% rebound in Neocloud stocks is a dead cat bounce, not a recovery. Why? Because the unit economics are broken.
Most neocloud platforms rely on token incentives to attract GPU miners. Sound familiar? It's the same model as DeFi liquidity mining. You subsidize supply, you get supply. But the moment you stop the incentives, the supply vanishes. I saw this firsthand in 2020 when I analyzed Uniswap's front-running bots. The same dynamic applies here: the neocloud 'decentralization' is a myth when the majority of compute power comes from a handful of large miners who are only there for the token.
The power sector (AI Power) is the biggest loser, with only a 6% rebound. The narrative that 'AI will consume all the world's energy' is true, but it's also a slow burn, not a quick profit. Utilities are regulated, capital-intensive, and slow to pivot. The market is realizing that power generation is a multi-year infrastructure play, not a speculative trade.
Takeaway: The AI Trade is Not Over, But the Label is Worthless
Goldman Sachs' conclusion is stark: 'The era of achieving a unified valuation premium solely based on the AI label is coming to an end.'
For crypto-native investors, this is the moment to stop chasing narratives and start building tools that actually capture value. The 'Inference Economy' is the new frontier. The winners will be those who control the software stack—the AI agents, the middleware, the data provenance layers.
Liquidity flows like water, but greed builds dams. The dam of blanket AI hype is breaking. The water is now flowing to those who can prove they can convert inference into revenue.
Trust is not a feature, it is a failed audit. The neoclouds, the memory fabs, the power plants—they all need to pass the reality check of unit economics. Most will fail.
The market corrects what the mind refuses to see. The divergence in August is not a random event. It is a correction of the collective delusion that all AI is equal.
So, what's the next narrative? I'll bet on the agents. AI agents that execute on-chain transactions without human intervention—that's where the real value lies. But that's a story for another article.
