"Over the past 12 months, NVIDIA’s market cap added roughly the entire GDP of Sweden. Meanwhile, Microsoft’s Azure cloud margins barely budged. The numbers don't lie: this asymmetry is a bug, not a feature."
When I audited my first smart contract in 2017, I learned that no protocol can sustain a 70%+ fee structure if its downstream users are bleeding. The same logic applies to the AI supply chain. JPMorgan's recent note on the AI trade is not a weather forecast. It is a coded warning about the redistribution of pricing power.
Context: The Great Divergence
The report draws a line between two groups. Group A: semiconductor suppliers—NVIDIA, SK Hynix, AMD. Group B: hyperscalers—Microsoft, Google, Amazon, Oracle. Since the AI narrative ignited, Group A’s stock performance has dramatically outpaced Group B’s. On the surface, this makes sense: the "picks and shovels" narrative is seductive. But JPMorgan’s core thesis is that this divergence is structurally unsustainable. They project hyperscaler CapEx growth dropping from +100% in 2026 to just +7% in 2028. If true, the implied demand cliff for AI chips and HBM memory would be severe. This is not a bearish call on AI. It is a realistic recalibration on ROI.
Core: Why the Cartel is Fragile
Let me break down the economics using a DeFi analogy. Imagine a lending protocol where the supplier gets 20% APY, but the borrowers are paying 30% and getting no yield on their collateral. That protocol will soon face a liquidity crisis. Here, the hyperscalers are the borrowers. They are spending billions on NVIDIA’s chips, but their own cloud AI services are still commoditizing. The margin differential—NVIDIA’s 70%+ gross margin vs. Azure’s ~40%—is the structural tension. Based on my audit experience, when a single layer in a stack captures disproportionate value, the system seeks equilibrium through either price compression or fork. The fork here is hyperscaler custom ASICs—Google TPU, Amazon Trainium, Microsoft Maia. These chips may not beat H100s in raw benchmark, but they give the hyperscalers what they need most: leverage. They can say no. That changes the game.
Contrarian: The Self-Fulfilling Prophecy
Here is the counter-intuitive angle most miss. JPMorgan’s report itself is a market signal. When a top-tier bank publishes a widely distributed note predicting a CapEx slowdown, it influences boardroom discussions at Microsoft and Amazon. CFOs who were planning to order 10,000 H200 units might now order 8,000, just to hedge. This is not a conspiracy. It is reflexive market behavior. The report, in essence, accelerates its own thesis. So the danger is not that CapEx growth will slow. The danger is that the fear of it slowing will cause an earlier, sharper deceleration than the numbers project. I saw this exact pattern in the 2022 liquidity freeze: once the narrative shifted from "infinite demand" to "ROI scrutiny," the herd moved faster than any DCF model could predict.
Takeaway: The Rotation is Rational
The smart move is not to bet against AI. It is to bet on who ultimately owns the customer relationship. In DeFi, we say code is law. In AI, the law is capital discipline. The hyperscalers, with their massive cloud revenue bases and captive user ecosystems, can afford to be patient. The semiconductor players, riding on hyper-optimistic multiples, cannot afford a single earnings miss. As the pricing power shifts from the seller to the buyer, the market will reward the latter. The question is not if this rotation happens, but when the next earnings call plants the flag.
In a world of noise, code is the only quiet truth.