The AI Infrastructure Trade: The Anatomy

A market that isn't afraid of the highest amount of capital expense spending in history is saying something important.

The Anatomy

The AI Infrastructure Trade: Where the Money Goes, Where the Bottlenecks Form, and Who Gets Paid

A market that should be scared of AI infrastructure spending, and visibly is not, is saying something important.

This week gave investors plenty of reasons to retreat. CPI came in hot. PPI came in hotter. Oil pushed above $113 as the Strait of Hormuz reminded everyone that “global supply chain” is often just a polite phrase for hoping the chokepoints behave. Rate hikes, somehow, wandered back into the conversation like an unwelcome dinner guest who still has your number.

The broad market sold off. Advances and declines across the AlphaApes universe split nearly three-to-one in favor of sellers. New 52-week highs fell. Bull Snort, our measure of stocks making new highs on elevated volume, dropped by a third.

Semiconductors went up anyway. Micron gained nearly five percent. Marvell rose more than eight. Corning added more than four after its newly expanded relationship with Nvidia put it closer to the center of the buildout. Semiconductor Equipment remained near the top of the universe by RS Master, our composite ranking of momentum, growth, and market leadership.

That is not how a fragile theme behaves. A market worried about rates should punish long-duration capex stories. A market worried about oil should punish energy-intensive buildouts. A market worried about macro pressure should not be rewarding the companies supplying one of the most capital-intensive infrastructure cycles in modern history. Unless the trade underneath is stronger than the macro pressure around it.

The AI infrastructure trade is not one trade. It is a hierarchy of scarcity. Pricing power moves to whichever constraint the system is currently hitting, and right now those constraints are moving in sequence through compute, memory, efficiency, and physical infrastructure. Copper runs through all of them.

Compute: The First Bottleneck

The first layer is compute: GPUs, accelerators, custom silicon, and the chips that make AI training and inference possible.

For three years, Nvidia absorbed almost all the oxygen in the room. It earned the role. Nvidia built the dominant platform, controlled the software ecosystem, and became the toll booth for the first wave of AI infrastructure spending. If the AI buildout had a front door, Nvidia charged admission.

But secular trades mature. The first winner proves the market, then capital begins searching for the next constraint. That rotation is now visible.

AMD’s data center business is accelerating. Marvell is being rewarded for custom silicon, networking, and optical connectivity. Micron has become one of the clearest beneficiaries of AI-specific memory demand. Silicon Motion and other semiconductor-adjacent names are showing elite relative strength behavior.

The question has changed. It is no longer only, “Who makes the best AI chip?” It is now, “Who supplies the parts of the system that cannot scale without them?”

That is a more useful question for investors. It moves the discussion away from the obvious winner and toward the supply chain beneath it, where the next round of scarcity begins.

Memory: The Current Chokepoint

The current chokepoint is memory.

AI accelerators do not perform in isolation. They need high-bandwidth memory to feed them data fast enough to justify the cost of the chip itself. A faster engine still loses the race if the fuel line is too narrow.

That is why memory has moved from cyclical afterthought to strategic constraint. Samsung, SK Hynix, and Micron control nearly all of the global DRAM market. They are now reallocating capacity toward HBM because the economics are better and the demand is tied directly to AI accelerators.

Every wafer moved into HBM is a wafer not available for conventional memory. That tradeoff tightens supply elsewhere and gives the memory producers something most commodity-adjacent businesses spend their lives dreaming about: pricing power with a structural demand tailwind.

The moat here is not a glossy product demo. It is capacity. Cleanrooms, process expertise, yields, packaging, and customer qualification cannot be summoned because a hyperscaler wants more supply by Tuesday.

When customers are receiving only part of what they need, suppliers are no longer simply filling orders. They are allocating scarcity. That is why memory leadership matters. It is not merely riding the AI trade; it is one of the places where the AI trade is currently constrained.

Efficiency: The Next Chokepoint

Speed is the constraint. Efficiency is the chokepoint.

Compute gets the headlines and memory gets the scarcity premium, but as AI clusters scale, moving data becomes almost as important as processing it. At large scale, the cost of transferring data between chips, switches, racks, and facilities becomes a financial issue, not just an engineering issue.

Copper connections have limits. Power budgets have limits. Cooling has limits. Physics, annoyingly, still attends earnings calls.

This is why optics and photonics have become more than a science project. Co-packaged optics, optical interconnects, lasers, and high-speed connectivity are attempts to solve the economic problem of data movement at AI scale. Nvidia’s own framing around co-packaged optics emphasizes the reduction in power consumption per bit moved. The lead argument is not glamour. It is efficiency.

The point is not simply more speed. The point is less waste, less heat, less power loss, and more usable output from every dollar spent on infrastructure.

This is where companies like Marvell, Lumentum, Coherent, and Astera Labs enter the anatomy. Their relevance is not that they sound futuristic. Their relevance is that the system is running into a constraint they are built to relieve.

When oil is above $113 and data center power demand is already becoming a grid-planning problem, efficiency stops being a technical improvement. It becomes margin protection. The market appears to understand this. Power, thermal management, optical connectivity, and high-speed interconnect names are being treated less like speculative accessories and more like necessary organs.

Physical Infrastructure: The Layer Everyone Forgets Until It Breaks

The least glamorous layer may be the most important.

Data centers do not float in the cloud. They sit on land. They require steel, concrete, permits, substations, transformers, cooling systems, backup power, fiber, electrical contractors, construction labor, and grid access. The future, as usual, requires more electricians than keynote speakers.

This is where the article stops being about technology and becomes industrial. A hyperscaler can sign a large AI commitment in a press release, but someone still has to build the facility, wire the switchgear, bring in power, manage heat, and keep the whole thing from becoming a very expensive toaster oven.

This layer is also where macro stress hits hardest. A fabless chip designer and an electrical contractor do not experience inflation the same way. Higher rates, labor shortages, materials inflation, and fixed-price project risk can compress margins even when demand is strong.

That tension is already visible in the AlphaApes data. Engineering and Construction has drifted from earlier leadership toward the middle of the universe, while select operators like Comfort Systems USA, MYR Group, and Sterling Infrastructure remain elite. The average is slipping, but the best names are holding.

That is dispersion, not collapse. It tells us the market is no longer rewarding exposure alone. It is separating companies with real earnings leverage from companies merely standing near the theme wearing a reflective vest.

Copper: The Metal Running Through the Whole Trade

Copper threads through the entire system.

Every megawatt of data center capacity requires wiring, power distribution, cooling systems, transformers, substations, transmission lines, and redundant grid connections. Some internal data movement may shift toward optics over time, but getting power to the facility remains stubbornly physical. There is no optical substitute for a substation.

This is why copper deserves more attention than it is getting. The AI premium has been priced aggressively in semiconductors, and reasonably so. Chips are the obvious first-order winner. Copper is still treated more like a traditional industrial cyclical, even though its demand profile is increasingly tied to strategic infrastructure: data centers, grid expansion, electrification, power redundancy, and the physical movement of energy into compute.

The supply side is not generous. Mine development is slow. Refining capacity is constrained. Political risk is real. Energy disruption can raise both operating costs and processing costs, which means the same macro shock that lifts oil can also tighten the metal required to expand the AI grid.

This is the second-order trade. Not “AI needs chips.” Everyone knows that. AI needs power. Power needs copper. Copper needs supply chains that are already tight. That is where the opportunity may still be underpriced.

What Breaks It

The bear case deserves a serious hearing.

Inflation can hurt this trade. Rates can hurt this trade. Energy shocks can hurt this trade. The mistake is assuming those risks hit every layer equally. Higher rates pressure multiples and financing conditions. Energy inflation pressures data center operating costs. Labor and materials inflation pressure contractors and construction margins. Fixed-price contracts can become traps when input costs move faster than pricing.

Those are real risks, but they are not the central thesis killer. The central thesis killer is a sustained reversal in hyperscaler capex guidance.

If Microsoft, Amazon, Google, Meta, and Oracle begin guiding down AI infrastructure spend in a coordinated or lasting way, the anatomy changes immediately. The whole hierarchy depends on forced spending continuing through the system.

That is the signal to watch. Not every oil spike. Not every inflation print. Not every Fed speech delivered in the usual dialect of cautious confusion.

The question is whether the largest buyers of compute still believe they cannot afford to slow down. So far, they do.

The Point

The anatomy of the AI infrastructure trade is not a straight line from GPU to data center. It is a moving hierarchy of scarcity.

First compute. Then memory. Then efficiency. Then power, construction, cooling, and copper. Pricing power concentrates wherever the system hits constraint.

That is why the trade keeps showing resilience even when the macro backdrop looks hostile. The macro environment is not separate from the trade. It is part of the trade. Higher energy prices make efficiency more valuable. Grid pressure makes power equipment more valuable. Supply shortages make memory more valuable. Construction inflation separates the operators from the tourists.

The mistake is thinking AI infrastructure is one giant basket where everything with the right label should go up. It is not. Some companies sell excellent shovels. Some sell overpriced shovels. Some paint a spoon gold and hope nobody asks about margins. The job is to know the difference.

A mania needs belief and egos. This needs copper, high-bandwidth memory, and light moving faster than electricity can carry it.

Jason Bartlett

Jason Bartlett is CEO and President of Veche, Inc, parent company to Avalanche Markets. He works extensively in U.S. energy market finance and economics and is a member of the board of Thinking About Thinking, Inc--a 501c3 research and convening organization dedicated to advancing ideas about intelligence.