Since the day I read the first Market Wizards book in 1989, I have loved the ever-changing puzzle of the macro world. History does not repeat, but it often rhymes. As Andrew Lo later argued in Adaptive Markets, markets are similar to complex biological systems. They are always changing. The puzzle recycles clues from prior periods, but the setup is never exactly the same. Sometimes markets warn you in strange ways about potential regime shifts. I think that time is now.
For more than a year, I have felt comfortable staying with the AI capex trade and looking through the constant fears around AI demand and return on invested capital. That was because, as a power user of the tools, the bubble argument never matched what I was seeing in the real world model progress. The models were getting better. The use cases were expanding. The coding tools were improving. With those compounding capabilities, enterprise workflows were becoming more obvious. As someone who had managed people since my 20s and currently has multiple digital employees, the adoption of AI was never a question.
Since late November, when Claude Opus 4.5 was released and the agentic gate opened wider, the demand-side fears have largely disappeared. Although you still hear the occasional bubble talk, agents changed the conversation. They moved AI from a chatbot story to an action story and, with it, a compute story. A chatbot answers questions. An agent does work. It writes code, tests workflows, reconciles data, searches documents, updates systems, produces content, and eventually will initiate transactions. Its token demand is insatiable.
As someone who was fortunate to long Micron since January of last year, in some ways it was much more comfortable to stay long when people were only arguing with you. The market kept debating whether AI demand was real while the physical evidence of that future demand kept showing up in memory, compute, networking, power, and cloud commitments. That is the strange way the market is warning me now. Investors have NOW embraced the structural shortage in memory after a 700% price move, and retail investors in the U.S. and South Korea, the memory capital of the world, are now having a capex party.
It was not that long ago, before Opus 4.5 helped drive the capex trade into the parabola of choice, that trader attention was focused on Palantir, gold, and silver. All three now feel like distant parabola memories that completely lost retail attention. The risk in a parabolic market is that there is no day when a bell rings to tell you the regime has shifted and that it is time to get out at the highs. The risk is usually that too much money, too much enthusiasm, and too much urgency all try to pass through the same narrow opening at the same time. That is where AI is now for me.
To be clear, I still believe the AI capex trade is in the early innings and has many more years to go. But just like with the Strait of Hormuz, the choke-point risk for the capex trade has become more clear for me. The near-term supply chain is struggling to keep up with insatiable and rising demand. AI skeptics of 2025 have now embraced AI demand and ROIC at a time where the risk seems to be about how quickly supply can meet demand. AI demand may now be so strong, so urgent, and so strategically important to so many buyers that the supply chain cannot handle the rush. This is a different type of risk. It is a boom built on real demand colliding with limited physical capacity. That is how shortages begin.
At the same time as the surprise arrival of AI agents, the policy and accounting environment have made capex even more attractive. The One Big Beautiful Bill and the return of more favorable depreciation treatment created another tailwind for capital spending. The accounting system already allows companies to capitalize infrastructure and depreciate it over time. The cash leaves today, but the expense is spread out over years. If tax policy accelerates deductions, the incentive to pull capex forward becomes even stronger.
That matters because hyperscalers are spending against massive remaining performance obligations, cloud demand, and revenue visibility. Microsoft is the cleanest example. The market sees rising capex, rising cloud revenue, rising RPOs, and a strategic race to catch up to frontier AI leaders. That makes the spending loop self-reinforcing. Capex rises, supplier revenue rises today, the hyperscaler depreciates the cost over time, RPOs show future demand, tax treatment improves the cash-flow math, and the race to secure frontier AI capacity accelerates.
This is why the demand-bubble argument may now be missing the bigger risk. Hyperscalers, enterprises, sovereigns, and AI labs are all treating AI capacity as strategically essential and seeing prices rise and scarcity of supply show up across the infrastructure buildout.. They are trying to secure that capacity immediately, even before the near-term ROI can be precisely measured. That is where hoarding enters the story.
In a normal cycle, customers buy what they need. In a shortage cycle, customers buy what they might need. In a strategic shortage cycle, customers buy what they fear someone else will get first and before prices rise further. In this case, with close to $1.4 trillion of backlog and orders for Microsoft, Google, and Amazon from the recent earnings season, they have every reason to hoard, especially as prices move higher. That is the psychology now building across the AI supply chain.
Hyperscalers reserve GPUs. AI labs secure compute years in advance. Data-center developers reserve transformers. Enterprises sign cloud commitments before they know their exact usage. Sovereigns build national AI capacity. Buyers pre-order equipment, reserve factory slots, and build inventory buffers because they no longer trust just-in-time supply chains. Everyone moves from just-in-time to just-in-case.
That behavior strengthens the boom, but it also makes the signal noisier. Some of the demand is final demand. Some of it is precautionary demand. Some of it is strategic hoarding. Some of it is double ordering. Some of it is fear. The demand remains real, but the measurement becomes harder.
This is the most important distinction. The old bubble debate focused on whether AI demand was real. The new risk environment focuses on whether companies can meet that demand. A data center is not one thing. It is chips, HBM, advanced packaging, substrates, power semiconductors, transformers, switchgear, substations, fiber, optical transceivers, cooling systems, chemicals, land, water, construction labor, permitting, and grid access. If one piece is missing, the entire project can slow.
That is the real AI risk: correlation inside the supply chain and correlation in the market’s dependence on AI earnings. Because of the agentic inference needs for the next stage of Nvidia chips, GPU demand has expanded the supply chain requirements. The data center needs more cooling. Cooling needs pumps, fluids, heat exchangers, chemicals, and facility redesign. The chips need HBM. HBM needs advanced packaging. Advanced packaging needs substrates, cleanroom chemicals, specialty gases, and precision equipment. This is a correlated system.
That makes bottlenecks more dangerous than weak demand. If demand weakens, companies can slow spending voluntarily. But if bottlenecks emerge while demand remains strong, companies may keep trying to spend and fail to deploy capital productively. That creates inflation, project delays, double ordering, hoarding, margin pressure, and eventually disappointment.
This is where the Strait of Hormuz and the Iran risk now become an important part of the AI story. So far, markets have largely ignored the oil doomers but the reality is, when combined with this unexpected insatiable AI demand, there is no doubt, the risk side of supply has risen. Over the last month, in a game of rock, paper, scissors, AI earnings and profit margins have beaten rising oil and the shutdown of the Strait of Hormuz. The reason market breadth is so bad right now is that oil- and rate-sensitive consumer stocks have been hit hard, while semiconductor and memory names have powered the market. The breadth issue is not just a US reality. The Kospi breadth has broken down over the last few weeks as well. Through May 19th, the Kospi 200 index is up 14% MTD but 8 of the 10 sectors are down MTD and the average for those 8 sectors is down more than 8.5%. Hormuz is having an impact.
The data-center buildout is oil sensitive. It depends on petrochemicals, plastics, insulation, PVC, cooling materials, printed circuit boards, packaging materials, specialty chemicals, and global shipping. Semiconductor manufacturing depends on gases, chemicals, wafers, tools, and precision logistics. Electrical infrastructure depends on metals, transformers, switchgear, and long-cycle industrial supply. A disruption in energy or shipping can tighten the bottlenecks that are already becoming visible, even if AI demand remains strong.
Hormuz is the stress test for the AI supply chain. Markets can handle growth. Markets can handle demand. Markets can even handle shortages for a while. What markets struggle with is when everyone is forced into the same trade, the same suppliers, the same bottlenecks, and the same delivery windows at the same time. That is when a boom becomes more fragile.
AI agents can scale in software time. Jensen Huang recently said that agents will consume 1000x more compute than chat. Transformers, substations, power plants, fab capacity, petrochemical supply chains, and permitting cannot. That mismatch is the center of the story.
The AI capex debate is now centered on whether the physical supply chain can support the capex that AI demand is trying to pull forward. If the answer is no, the next phase of the AI cycle will be defined by shortages, hoarding, project delays, component inflation, correlated supply-chain stress, and periodic speed crashes inside a still-powerful secular boom.
That is why I am not writing this as a call to short the AI trade. I still believe AI demand will continue to power through. I still believe the agentic world is real. I still believe the physical buildout of AI infrastructure remains one of the most important investment themes in the world. But there are moments when the risk changes.
For the last year, investors debated whether the AI capex boom was built on fake demand. Today, the more interesting risk is that the demand is so real that everyone starts hoarding the same scarce pieces of the same fragile supply chain. The AI race is not ending. Everyone is trying to run it through the same narrow bridge at the same time.
