By the end of this year, four U.S. tech giants will have spent more on AI infrastructure than the entire annual economic output of Belgium.

Microsoft, Google, Amazon, and Meta are on track to spend more than $700 billion this year alone on the physical backbone of the AI economy.

That’s more than it cost America to land 12 men on the moon in inflation-adjusted dollars. And that’s not the total bill for the AI buildout. That’s just one year.

Microsoft is committing $190 billion. Google $185 billion. Meta $135 billion. Amazon $200 billion. It’s nearly double what these companies spent in 2025. And Wall Street analysts now expect the figure to top $1 trillion in 2027.

It is the largest sustained surge in industrial infrastructure spending the U.S. has seen since the World War II.

The money isn’t going into software, marketing, or research. It’s pouring into the physical buildout of the AI economy — industrial buildings the size of football fields, power systems that draw gigawatts of power, cooling loops that move liquid through thousands of chips, and power transmission lines that don’t yet exist.

And it’s running into a wall.

Demand has outrun the world’s ability to manufacture, build, and approve fast enough to keep up.

You can see these chokepoints across the entire AI supply chain.

In Cheongju, South Korea, the world’s most advanced memory chips are being built one wafer-thin layer at a time. They’re called high-bandwidth memory chips — HBM for short — and they’re the only memory architecture fast enough at scale for the most demanding AI training workloads. Inside SK Hynix’s M15X fabrication plant, machines stack 12 of these layers on top of each other, each thinner than a sheet of plastic wrap, threaded together by thousands of microscopic copper connections.

Cooling is another constraint. The chips inside AI servers run so hot that air conditioning can’t keep up. A single rack of next-generation Nvidia chips produces as much heat as a dozen home furnaces. Without specialized liquid cooling, the chips overheat in minutes — destroying millions of dollars’ worth of hardware. The precision-engineered piping used to deliver coolant directly to the chips, called manifolds, is now on a waitlist that runs into 2027. A data center operator who orders today is buying a slot in a queue that someone else got into 18 months ago.

Then there’s the grid. In Loudoun County, Virginia — the world’s largest data center hub — the local utility, Dominion Energy, has warned customers that some new projects could wait up to seven years for full power hookup.

Even after the grid connection is approved, the transformers that step electricity down to the levels a building can use are on backorder for two to five years.

These aren’t isolated supply hiccups. They’re symptoms of the same problem — a $700 billion spending wave that’s outrun the physical world’s capacity to deliver.

That’s the bad news for the tech giants. For investors, it’s a golden opportunity.

When supply can’t expand to meet demand, the companies that already own the supply are in a rare position. They don’t have to chase customers. The customers are chasing them. The orders come in faster than they can fill them. And the barriers to new competition — capital, expertise, regulatory approval — are too high for anyone else to clear quickly.

That’s where fortunes get made in industrial booms. Not from invention, but from owning the chokepoint.

In the 19th-century railroad boom, the biggest fortunes weren’t made by the railroad operators — many went through repeated booms and busts. They were made by Andrew Carnegie, who sold the steel that went into the rails.

In the electrification era, the biggest winners weren’t the utilities. They were General Electric and Westinghouse — the suppliers of generators, transformers, and turbines.

In the personal computer boom of the 1990s, the biggest fortunes weren’t made by PC makers. They were made by Microsoft, which owned the operating system, and Intel, which owned the chip.

In every case, the same pattern emerged. The largest fortunes weren’t made by the most visible companies. They were made by the ones supplying what those visible companies couldn’t function without.

In this report, we’ve identified five AI Chokepoints that threaten what’s on track to be a multitrillion-dollar AI buildout. The companies that can solve them — by building the equipment, filling the orders, and clearing the backlogs — stand to capture an outsized share of that spend.

We’ll walk through each chokepoint in detail. We’ll look at how it formed, why it’s so difficult to fix, and why the companies solving them stand to be among the biggest winners in the stock market over the coming years.

Let’s start with the first chokepoint — high-bandwidth memory.

Chokepoint No. 1: High-Bandwidth Memory

The Wall the AI Industry Just Hit

For decades, the chip industry has lived by a simple rule: every two years, transistors get smaller, chips get faster, and computing gets cheaper.

That rule has held since the 1960s. It’s the reason your phone is more powerful than the computer that put a man on the Moon.

But there’s a second rule the industry doesn’t talk about as much. It’s called the memory wall.

Here’s the basic idea. A computer chip can only do useful work if it has data to work on. The processor does the math. The memory holds the numbers. And every calculation requires the chip to fetch numbers from memory, do something with them, and put the answer back.

For years, processors got faster much more quickly than memory did. The gap widened. By the early 2000s, top-end processors could perform calculations 100 times faster than memory could feed them data.

It was like having a Formula 1 engine with a garden hose for a fuel line.

For most computing tasks, engineers found clever ways around it. They built small “caches” of fast memory next to the processor. They wrote software that minimized how often the chip needed to reach for new data. The garden hose was a problem, but a manageable one.

Then AI arrived.

Modern AI models don’t just crunch numbers. They move oceans of them. A large language model like the one behind ChatGPT has hundreds of billions of parameters — essentially, settings that the model adjusts as it learns. Every time the model generates a sentence, it has to pull billions of those settings out of memory and feed them to the processor.

A garden hose won’t cut it.

That’s where high-bandwidth memory — HBM — comes in.

Stacked Like Pancakes

Instead of laying memory chips flat on a circuit board next to the processor, HBM stacks them vertically – twelve layers high, in the current generation. Each layer is connected to the layers above and below by thousands of tiny copper pillars that pass straight through the silicon.

The result is a memory chip that sits inches away from the processor and delivers data through thousands of channels in parallel. It’s the difference between drinking through a straw and drinking from a firehose.

Without HBM, today’s AI chips simply wouldn’t work. Nvidia’s flagship Blackwell GPU pairs each processor with up to eight stacks of HBM3E memory. The next-generation Rubin chip will use HBM4, with even more capacity and bandwidth.

That’s why the M15X factory in Cheongju matters so much. It’s where the next generation of HBM is being built.

And it’s why every chip coming off the line is being spoken for before it’s even produced.

Three Companies. Two Years. No More Supply.

The HBM market is a near-monopoly. Three companies make it:

• SK Hynix — controls roughly 62% of global HBM shipments and an estimated 70% of Nvidia’s HBM orders
• Micron — the only American producer, and now the fastest-growing of the three
• Samsung — has fallen to third place in HBM share as Micron has scaled

That’s it. No one else can make HBM at scale. The manufacturing process is so demanding that it’s taken these three companies more than a decade of R&D to get yields high enough to ship in volume.

In late October 2025, SK Hynix’s head of DRAM marketing told the Financial Times that the company’s entire production capacity for 2026 was already sold. Customers had locked in orders for the year ahead and were now negotiating for 2027 slots. Hyperscalers including Alphabet, Meta, and Microsoft have since approached SK Hynix with proposals to pre-fund new manufacturing capacity in Yongin just to lock in HBM allocation. In April 2026, Goldman Sachs raised its 2026 DRAM supply-demand gap forecast to 4.9% — calling it the worst memory shortage in 15 years.

Industry analysts at SemiAnalysis estimate that memory will account for roughly 30% of total hyperscaler capital spending in 2026 — up from an average of about 8% across 2023 and 2024. That’s a near four-fold shift in how the world’s biggest tech companies are deploying their hardware budgets. And SemiAnalysis expects the share to climb further in 2027.

In other words: the AI buildout isn’t just about chips anymore. It’s about feeding those chips.

And the world’s HBM factories are running at full tilt.

Our Pick to Profit From the Memory Chokepoint

1. Micron Technology (MU)

Micron is the only U.S. producer of high-bandwidth memory, and the fastest-growing of the three global suppliers. Its 2026 HBM output already sold out by December 2025, with production ramping up for 2027.

Every AI accelerator built by Nvidia, AMD, or the hyperscalers needs HBM.  And Micron sits in a near-monopoly with SK Hynix and Samsung — the only three companies in the world that can produce it at scale.

As HBM moves from the current HBM3E generation to HBM4 — used in Nvidia’s next-generation Rubin chips — Micron’s position only strengthens. The company has guided to capital spending of over $25 billion in fiscal 2026, with HBM capacity being a top priority.

For investors, Micron is the cleanest U.S.-listed way to play the most acute supply shortage in the AI economy.

Chokepoint No. 2: Photonics and Interconnects

The Cables That Can’t Keep Up

Solving the memory problem is only half the battle.

Once data leaves a chip, it has to travel somewhere. In a modern AI data center, that “somewhere” is another chip. And another. And another. A single AI model might be split across tens of thousands of processors, all working in parallel, all needing to talk to each other constantly.

The wiring that connects them is now one of the biggest constraints in the industry.

For decades, computers have used copper to move data. Copper is cheap, abundant, and easy to work with. The wires inside your laptop are copper. The cables behind your TV are copper. The lines that connect servers in a data center are copper.

But copper has a problem. The faster you push data through it, the more electricity it wastes as heat.  At low speeds, this is barely noticeable. At AI speeds, it’s a disaster.

A modern AI cluster — say, 200,000 Nvidia GPUs working together to train a single model — needs to move staggering amounts of data between chips. We’re talking petabytes per second,  the equivalent of streaming millions of HD movies simultaneously, with no buffering,  for as long as the model is running.

Push that much data through copper, and two things happen.

First, the cables get hot. Really hot. Hot enough that a meaningful percentage of the data center’s electricity bill goes to moving data instead of computing on it.

Second, the signals start to degrade.  Send a stream of ones and zeros through enough copper, and they start to blur into each other. The receiving chip can’t tell whether it’s getting a one or a zero, so it has to ask for the data again. That asking-again creates a delay (“latency”). And in an AI cluster, latency kills.

The fix is to stop using copper.

Light Instead of Electrons

The solution is called silicon photonics. Instead of pushing electrons through copper wires, you send pulses of light through tiny glass fibers — much like the fiber optics that bring internet to your home.

Light is faster than electricity and can be transmitted without heat. (Think of a cooler LED versus an incandescent bulb.) Plus, optical systems can carry much more data through a much smaller cable.

The problem is making it work at industrial scale.

To send data as light, you need a tiny laser that can flicker on and off billions of times per second — turning electrical ones and zeros into pulses of light. At the other end, you need a sensor that can detect those pulses and convert them back into electricity. And you need to do this with near-perfect reliability, because if even one in a trillion pulses gets misread, the data is corrupted.

Building these components — they’re called optical transceivers — is genuinely hard. The lasers have to be calibrated to wavelengths measured in nanometers. The sensors have to filter out background noise.  The components have to survive in a data center that runs 24/7 for years at a time.

And demand for them is exploding.

Cloud providers like Amazon, Microsoft, Google, and Meta are pouring billions into optical networking equipment.  The components are so essential — and so hard to make — that the companies that supply them have become

major winners of the AI buildout.

If memory is the bottleneck inside the chip, latency is the bottleneck between chips.

And photonics is the way forward.

Our Picks to Profit From the Photonics Chokepoint

2. Broadcom (AVGO)

Broadcom is the dominant designer of networking silicon for AI data centers. Its switching chips can route traffic between hundreds of thousands of GPUs inside a single cluster. Without this architecture, even the fastest processors can’t communicate fast enough to train a modern AI model.

As clusters move from copper to optical, Broadcom’s co-packaged optics initiative puts it at the center of the photonics transition.  It also works with Google, Meta, and other hyperscalers to design custom AI accelerators — chips manufactured by TSMC at the most advanced nodes available.

Where Nvidia leads the GPU market, Broadcom dominates when it comes to the connective tissue that links them together. And every gigawatt of AI compute being built today requires both.

3. Coherent (COHR)

Coherent is a leading supplier of optical transceivers — the tiny components that convert electrical signals into pulses of light and back.

Every fiber-optic connection inside an AI cluster will need them. Every connection between racks. Every connection between data centers. As this new AI infrastructure begins to replace copper, the demand for these components has gone vertical.

Coherent has the manufacturing scale,   the laser expertise, and the relationships with hyperscalers to be a primary beneficiary of the photonics buildout. In March 2026, Nvidia made a $2 billion strategic investment in the company — buying 7.8 million shares at $256.80 per share through a private placement, alongside a multi-year purchase commitment. It was a signal of how central Coherent is to the AI networking layer.

4. POET Technologies (POET) — Speculative

POET Technologies is a pure-play photonics company, focused solely on optical engines, modules, and light-source products, all in high demand for AI data centers.

Most optical transceivers today are built from separate components — a laser, a modulator, a driver, a receiver — assembled together on a board. POET’s “Optical Interposer” technology integrates all of these onto a single module.

If it works at scale, the savings are substantial. Smaller form factor. Lower power consumption. Higher reliability. Cheaper to manufacture.

The company is now taking orders for early production units and has signed development agreements with major partners. Shipments start later this year.

A note on risk: POET is genuinely speculative.

• While it’s growing rapidly, it’s doing so off a tiny base: The company has quarterly revenue of about half a million dollars at this point.
• It competes against giants like Broadcom and Coherent that have decades of optical manufacturing experience.
• It has historically raised money by issuing new shares, diluting existing shareholders.   
• The stock has traded in a wide range over the past year, giving it a “Sky High” Volatility Quotient (VQ%) of 93.01%.

This is the kind of investment that could deliver outsized returns if the technology scales, or could lose most of its value if larger competitors catch up first. Position sizing matters more here than with the larger recommendations in this report.

Chokepoint No. 3: Thermal Management

The Heat Problem No One Saw Coming

A few years ago, designing a data center was an electrical problem. Today, it’s also an HVAC problem.

The reason comes down to a number that’s been climbing relentlessly for the past decade.

It’s the amount of power a single AI chip draws.

In 2018, a top-end data center processor pulled around 200 watts of electricity — about the same as a desktop PC. By 2022, Nvidia’s flagship AI chip, the H100 SXM5, was drawing 700 watts. The latest Blackwell generation pushes to 1,200 watts per chip. And the next generation will go higher still.

Every one of those watts becomes heat.

Picture a 1,200-watt space heater. That’s roughly the amount of heat a single AI chip throws off when it’s running flat out. Now picture an AI server rack with 72 of those chips packed into a unit the size of a refrigerator. It’d be enough heat to warm a small apartment building.

But a data center has hundreds of those racks lined up shoulder to shoulder.

Soon the amount of heat we’re talking about gets pretty crazy.

Why Air Cooling Hit a Wall

For most of the history of computing, data centers cooled their equipment with air. Big industrial air conditioners pumped chilled air through the raised floor of the server room, up through the racks, and back out the top.

It worked, because the chips weren’t that hot. A typical server rack drew 5 to 10 kilowatts of power. The air conditioners could handle it.

Today’s AI racks draw 10 to 50 times that amount. And you can’t blow enough air through them to keep the chips cool. The physics simply won’t allow it.

You’d need to move air at hurricane speeds — and even then, the fans would start getting overheated, too.

The only solution is to switch fluids.

Water Cools 3,500 Times Better Than Air

Water carries heat away from a hot surface roughly 3,500 times more efficiently than air at the same flow rate. That’s due to its basic physics. Water has higher density, higher specific heat, and better thermal conductivity. It’s why your car’s engine is cooled with liquid, not just a fan.

For decades, most of the industry still preferred air cooling due to the logistics: Water and circuit boards don’t mix. One leak, and you’ve destroyed millions of dollars of equipment.

But by 2023, there was no other choice. The chips had gotten too hot.

The new generation of AI data centers is being built with what engineers call direct-to-chip liquid cooling. A specially designed coolant — usually a mix of water and additives — is pumped through small metal plates that sit directly on top of each chip. The coolant absorbs the heat, carries it away, and dumps it into a larger system that releases it outside the building.

It’s a fundamentally different way of building a data center. And it requires a whole new supply chain of specialized equipment.

The Hardware Shortage No One Talks About

The components that make liquid cooling work are obscure. You probably won’t hear about them at a tech conference keynote. They’re unlikely to make the front page of The Wall Street Journal.

But without them, the AI build-out grinds to a halt.

The most critical piece is something called a coolant distribution unit, or CDU. It’s the heart of the system — a box about the right size to fill your walk-in closet. It pumps coolant through the racks, monitors temperature and pressure, and routes heat away from the chips.

Each CDU has to be precision-engineered. The pumps have to run continuously for years without failing. The seals have to hold up to constant pressure. The monitoring sensors have to catch a leak before it reaches the electronics.

Then there are the manifolds — the metal piping that distributes coolant to each individual chip. They have to be machined to exact tolerances, because uneven flow means uneven cooling, and uneven cooling means some chips run hotter than others and burn out faster.

And then there are the quick-disconnect couplings — small valves that let technicians swap out a server without spilling coolant onto the electronics. Each one has to be rated for zero leakage under continuous pressure. A single bad coupling can knock out an entire rack.

These aren’t glamorous components. When most people think about the AI boom, they don’t think about valves and pumps and pipes.

But the AI boom is running on them…

And we’re starting to run short on them.

The companies that make them are working through multi-year backlogs. Data center operators are placing orders for equipment that won’t be delivered for 12 to 18 months. And as the next generation of AI chips arrives — with even higher power draws — the demand is only going to intensify.

Our Picks to Profit From the Cooling Chokepoint

5. Vertiv Holdings (VRT)

Vertiv is the leading supplier of coolant distribution units for AI server racks. As the industry transitions from air to liquid cooling, Vertiv’s order book has exploded.

The company already supplies products at the heart of many direct-to-chip cooling systems being installed today. And as next-generation AI chips push power consumption past 1,800 watts per chip, the demand for Vertiv’s gear only grows.

For investors, Vertiv is a direct way to play the shift to liquid cooling — a shift that isn’t optional. The chips have gotten too hot for anything else.

6. Modine Manufacturing (MOD)

Modine is a specialist in coolant distribution and heat exchangers, with deep experience from the automotive sector now extended to data centers.

Like Vertiv, it has CDU products to handle the heat transfer between the chip-level cooling loops and the building-level systems that release heat outside. Without that handoff, the coolant absorbs heat from the chips and has nowhere to send it.

Modine is smaller than Vertiv, but its data center segment has been one of the fastest-growing parts of the business. For investors willing to risk higher volatility, Modine offers more direct exposure to the cooling boom.

Chokepoint No. 4: Energy Generation

Cooling is a hard problem. But it’s solvable. Engineers know how to move heat. The supply chain is catching up.

The next problem is much bigger:

Creating enough electricity.

A single large AI data center often consumes as much power as an entire city. A 500-megawatt facility — and several of these are now being built — uses about as much electricity as 350,000 American homes. Some of the new mega-campuses being planned will use a full gigawatt or more. That’s the output of an entire nuclear reactor, dedicated to one building.

And these aren’t theoretical. They’re under construction right now.

In Abilene, Texas, OpenAI and Oracle are building a complex called Stargate that’s projected to consume 1.2 gigawatts — two of its eight buildings are already operational, with the remaining six due to come online by mid-2026. In Mt. Pleasant, Wisconsin, Microsoft is building a data center campus designed to draw similar amounts. Meta has broken ground on a 2 gigawatt facility in Louisiana — called Hyperion — that’s been engineered to eventually scale toward 5 gigawatts. Mark Zuckerberg has described it as ‘so large it would cover a significant part of Manhattan.’

Add up all the AI data centers planned for the next five years, and the numbers get genuinely alarming.

The International Energy Agency estimates that global electricity demand from data centers will roughly double between 2022 and 2026, and projects total data-center electricity demand of 945 terawatt-hours globally by 2030. In the United States, the Electric Power Research Institute now projects that data centers could account for 9% to 17% of all electricity consumption by 2030 — up from about 4% today. Analysts have called it the largest sustained surge in power demand the country has seen since the post-World War II industrial boom.

And the electricity has to come from somewhere.

Why Wind and Solar Aren’t Enough

For the past decade, the alternative energy conversation has been dominated by renewables. Wind and solar are now the cheapest forms of new electricity generation in most of the world. The cost curves have been spectacular.

But there’s a problem.

Solar panels don’t produce at night. Wind turbines don’t spin when the air is still.

AI data centers can’t run on intermittent power. They need electricity 24 hours a day, seven days a week, with no interruptions. An AI training run that takes weeks can be ruined by a single power blip.

So, to run a data center on renewables alone, you need batteries — lots of them — and the math on storing gigawatt-hours of electricity in batteries hasn’t yet penciled out at the scale AI requires.

What data centers need is baseload power — generation that runs constantly, regardless of weather. There are three sources that can deliver it at the scale AI demands: natural gas, hydroelectric, and nuclear.

Hydro is fairly tapped out — most of the good dam sites have been built.

That leaves gas and nuclear.

The Nuclear Renaissance

For decades, nuclear power was a story of decline. After the Three Mile Island accident in 1979, Chernobyl in 1986, and Fukushima in 2011, public support collapsed. Plants got more expensive to permit and build. Many existing reactors were shut down before the end of their useful lives, written off as a relic.

AI has started to change that.

In September 2024, Microsoft (MSFT) signed a 20-year power purchase agreement with Constellation Energy (CEG) to restart the shuttered Three Mile Island Unit 1 reactor. Its entire 835-megawatt output is committed to Microsoft, matching the power its PJM-region data centers consume. The deal is projected to add roughly $16 billion to Pennsylvania’s GDP over its 20-year life. Constellation is investing $1.6 billion to bring the reactor back online.

A few months later, Amazon Web Services bought a data-center campus directly adjacent to Talen Energy’s Susquehanna nuclear plant in Pennsylvania, along with a 10-year deal with the plant. In June 2025, the deal was restructured and expanded — AWS now has a 17-year, roughly $18 billion power purchase agreement with Talen for up to 1.92 gigawatts.

Google has signed a deal to buy power from a fleet of small modular reactors being developed by Kairos Power. Meta has put out a request for proposals for up to 4 gigawatts of nuclear power.

Three of the four largest U.S. tech companies are now buying decades of nuclear output up front — locking up entire reactors, in some cases before they’re even back online.

It’s hard to overstate how much the conversation has shifted. A decade ago, no one would have predicted that the most aggressive buyers of nuclear power would be software companies. But software companies are now the only buyers in the market large enough — and rich enough — to fund the multi-billion-dollar capital costs.

Gas has its own boom underway. GE Vernova’s order book for natural gas turbines has stretched to multi-year wait times. The company has said publicly that it’s sold out of certain turbine models through 2028. And uranium prices — flat for a decade — have roughly doubled since 2022, with peaks nearly tripling the price as utilities scramble to lock in fuel supply for an expanded nuclear fleet.

For the first time in nearly half a century, the United States is preparing to build new power plants at scale.

And every gigawatt of that new capacity has effectively been pre-sold to the AI industry.

Our Picks to Profit From the Energy Chokepoint

7. Constellation Energy (CEG)

Constellation is the largest operator of nuclear plants in the United States — with 21 directly operated reactors and ownership interests in several more after its Calpine acquisition closed in early 2026. The company also signed the landmark September 2024 deal with Microsoft to restart Three Mile Island for AI data center power.

Every gigawatt of this baseload demand from AI flows directly to Constellation’s bottom line. And in a world where hyperscalers are locking up nuclear output decades in advance, Constellation owns more of the existing fleet than anyone else.

Nuclear power was written off a decade ago. Today, it’s the most strategic energy asset class in the country. Constellation sits at the center of it.

8. GE Vernova (GEV)

GE Vernova is the leading U.S. manufacturer of gas turbines, with an order book reportedly selling out through 2030 for certain models. It’s also positioned in nuclear services and small modular reactor development.

Think of GE Vernova as the picks-and-shovels play across every form of new power generation being built for AI. Whether the next wave of capacity comes from gas, nuclear, or a mix of both, GE Vernova supplies the core equipment.

The stock has been one of the strongest performers in the entire AI infrastructure trade since it was spun off from GE in 2024. The order book suggests there’s a lot more to come.

9. Cameco (CCJ)

Cameco is the one of the largest publicly traded uranium producers. As the nuclear renaissance unfolds, every reactor needs fuel — and uranium prices have roughly doubled since 2022 as utilities scramble to lock in long-term supply.

Cameco is the cleanest way to play the fuel-supply side of the nuclear story. Unlike utility operators, it isn’t exposed to regulated electricity prices. Its revenue scales directly with uranium prices and reactor restarts.

As more nuclear plants come back online — and as new small modular reactors begin operating later this decade — Cameco stands to benefit from a multi-decade uplift in uranium demand.

10. Oklo (OKLO) — Speculative

Oklo is one of the most prominent small modular reactor (SMR) developers in the United States. Backed by OpenAI’s Sam Altman, who served as the company’s chairman from 2015 to 2025, Oklo is developing a new class of nuclear reactor designed to power data centers, industrial sites, and remote locations.

The company’s flagship product is called Aurora — a small reactor designed to produce between 15 and 75 megawatts of electricity. That’s far smaller than a traditional nuclear plant, but ideal for powering a single AI data center campus.

The pitch is compelling. Big nuclear plants take a decade or more to build. SMRs are designed to be smaller, simpler, and faster to deploy. If they work as designed, they could be the only realistic way to add gigawatts of clean baseload power to the grid in time to meet AI demand.

Oklo has signed letters of intent to deploy reactors with Equinix, Diamondback Energy, Meta, and more, and is working through the regulatory approval process with the U.S. Nuclear Regulatory Commission.

A note on risk: With its reactors still waiting approval to operate, Oklo is pre-revenue. Aurora is not expected to be commercially operational until 2027 or 2028 at the earliest — and nuclear projects routinely slip behind schedule. The stock is among the most volatile in the entire energy sector, with a 52-week range that has seen the share price swing by nearly 5x.

The company is burning cash, raising money periodically through new share issuances. If regulator approval is delayed, or if a competitor delivers a working SMR first, Oklo could lose a significant portion of its value.

But if Aurora becomes operational on schedule, and if even a fraction of the hyperscaler interest in SMR power translates into firm orders, Oklo has the potential for substantial returns. This is a high-risk, high-reward bet on the future of nuclear energy. We recommend sizing your position accordingly.

Chokepoint No. 5: Grid and Transmission

Here’s the part that surprises most people.

Even if you have a nuclear plant ready to deliver electricity. Even if you’ve signed a 20-year contract. Even if every megawatt is paid for in advance.

You still might wait five years to actually use it.

The reason is the grid.

Power doesn’t just flow from a generator to a building. It has to travel through a network of transmission lines, substations, transformers, and switches. Every connection point is a piece of physical equipment that has to be designed, ordered, manufactured, shipped, installed, tested, and approved by regulators before electricity can flow through it.

And every part of that supply chain is now backed up by years.

The Transformer Shortage

The single biggest choke point is a piece of equipment most people have never thought about. It’s called a large power transformer.

A large power transformer is a steel-and-copper box, often the size of a house, that steps electricity up or down between voltage levels. They’re essential for moving power long distances. Every power plant needs them. Every substation needs them. Every large industrial customer — like a data center — needs them.

In the United States, there are currently only a handful of factories that can build them at the size and scale the grid requires. And those factories are now booked solid.

Lead times for large power transformers have stretched from about 12 months in 2020 to over 2.5 years today, with some specialized units now approaching five years. Certain heavy gas-turbine frames stretch even further — as far as seven years. The shortage is so severe that utilities have started raiding spare-parts inventories. Some are buying used transformers from decommissioned plants overseas.

The reason for the shortage isn’t complicated. The world stopped building grid-scale transformers in serious quantities during the 2010s, when electricity demand was flat. The skilled workforce shrank. The supply chain for specialty steel — called grain-oriented electrical steel — atrophied.

Then AI showed up. Now, the factories can’t ramp fast enough. Even when new orders are placed today, the equipment won’t arrive until late in the decade.

The Interconnection Queue

Even if you have a transformer ready to install, you still need permission to connect to the grid.

In the United States, this permission is granted by regional grid operators — non-profit entities that manage the flow of electricity across multi-state networks. Before a new power plant or large customer can connect, the operator has to study the impact on the grid, approve the connection, and schedule the work.

That process is called the interconnection queue. And it’s the most quietly broken piece of American infrastructure.

In the PJM region — the grid that covers Virginia, Pennsylvania, and most of the mid-Atlantic — the interconnection queue has at times held more than 250 gigawatts of proposed new projects. The average wait time to reach commercial operation has stretched past eight years. Some projects have been in the queue since 2018 and still haven’t been approved.

In Texas, the situation is slightly better, but the volume of new requests has overwhelmed even that grid. In California, the wait times are even worse.

This matters because most of the large data center campuses being built today are sized for hundreds of megawatts, even a gigawatt. Each one requires a substantial grid upgrade. And each upgrade has to go through the interconnection queue.

The result is a strange situation that didn’t exist five years ago. Data center operators are finishing construction on buildings that can’t be turned on, because the transmission lines that would feed them are stuck in regulatory limbo.

In Virginia — the largest data center market in the world — utilities have started warning customers that any new project filed after 2024 may not receive grid hookup until 2028 or later. Several major projects have already been delayed by multiple years, not because of construction problems, but because the grid simply isn’t ready.

This is the final, hardest bottleneck. And it can’t be fixed quickly.

Why This Chokepoint Is Different

The other four bottlenecks — memory, photonics, cooling, and generation — are all engineering problems. They can be solved with money. If you spend enough, you can build more HBM factories. You can manufacture more optical transceivers. You can install more liquid cooling systems. You can build more nuclear plants.

And if you thought those regulatory hurdles were high – the grid is on another level.

You can’t simply build more transmission lines. Each one requires permits from federal agencies, state regulators, and often dozens of local governments. Landowners along the proposed route have to be compensated, and sometimes sue. Environmental reviews can take years. A single major transmission project has been taking a decade or more to approve and build — and that’s after the equipment is ordered.

The result is that the grid bottleneck is the least visible piece of the AI infrastructure crunch, and the most expensive to fix.

Which is exactly why the companies that build grid equipment — the transformers, the substations, the transmission lines, the high-voltage switches — are about to enter a buildout unlike anything they’ve seen in half a century.

Our Picks to Profit From the Grid Chokepoint

11. Eaton (ETN)

Eaton builds the electrical infrastructure that connects data centers to the grid — transformers, switchgear, power distribution systems. The company has flagged data centers as its fastest-growing end market, with order growth running well ahead of its broader business.

Every gigawatt of new data center capacity requires equipment like Eaton’s to connect it to the grid. And every utility upgrade required to expand the grid itself drives more demand for the same products.

This is one of the cleanest indirect AI plays in the market. Eaton doesn’t depend on any single AI company, model, or technology. It just needs the buildout to continue.

12. Quanta Services (PWR)

Quanta Services is the largest builder of electric transmission and distribution infrastructure in North America. It’s the company that physically constructs the new transmission lines, substations, and grid upgrades the AI buildout requires.

When utilities need to expand the grid — for a new data center, a new power plant, or a new substation — Quanta is the company they call. Its multi-year backlog has grown sharply as utilities respond to surging demand.

If the grid is the hardest bottleneck to fix, Quanta is the company doing most of the fixing. And it’s getting paid a lot of money to do it.

A Word on Risk

Now that we’ve laid out our 12 recommendations, it’s important to add a word on risk.

To make the risk profile of each pick clearer, here’s how the companies in this report generally break down by size and volatility.

Large-Cap Companies (Lower Volatility, Established Businesses)

• Broadcom (AVGO)
• GE Vernova (GEV)
• Eaton (ETN)
• Micron (MU)
• Constellation Energy (CEG)
• Quanta Services (PWR)
• Coherent (COHR)
• Vertiv (VRT)
• Cameco (CCJ)

These are large, diversified businesses with established revenue streams. They have exposure to the AI infrastructure theme, but they also have other revenue drivers that smooth out the bumps. Even if you buy at an awkward moment, time has a way of evening things out.

For most investors — especially those new to thematic investing — these stocks are the lowest-risk way to gain exposure to the AI infrastructure buildout.

Mid-Cap Companies (More Direct Exposure, Higher Swings)

• Modine Manufacturing (MOD)

Modine is more tightly focused on data center thermal management than the larger names in this report. That gives it more direct leverage to the liquid cooling theme — but it also means the stock can move more sharply on individual contract wins, capital spending cycles, or shifts in expectations.

This is the middle of the risk spectrum: more potential upside, paired with more volatility.

Speculative Companies (High Volatility, High Risk, High Potential Reward)

• POET Technologies (POET)
• Oklo (OKLO)

These are the highest-risk recommendations in this report. Both are early-stage, both have limited or no revenue, and both have stock prices that have swung dramatically over the past year.

POET is competing against giants in the optical components market. If its integrated photonics technology scales, the upside is substantial. If larger competitors catch up first, the downside is severe.

Oklo is pre-revenue and pre-commercial. The future of small modular reactors is genuinely uncertain. The stock could deliver outsized returns if Aurora becomes operational on schedule, or could lose most of its value on regulatory delays.

These are speculations, not investments in the traditional sense. They should represent a small portion of any portfolio — money you can afford to see swing sharply in either direction.

Practical Risk Management

A few simple principles can help manage risk when investing in emerging themes like this one:

• Start with the large caps. They provide exposure without requiring the theme to play out perfectly.
• Size small with mid-caps and speculative names. Higher-risk companies should typically represent smaller positions in a portfolio, not the core.
• Expect volatility — and don’t react to it. Sharp price moves are normal in technology and infrastructure sectors. Volatility is the “price of admission” you pay to be in with the chance to make fortune-making gains.
• You don’t need to own everything. This report is a menu, not a checklist. Even one or two well-chosen positions can provide meaningful exposure.

The AI infrastructure buildout is a long-term development, not a short-term trade. Matching your position sizing and volatility expectations to the level of uncertainty associated with each pick is just as important as picking the right companies.

From AI Infrastructure to AI-Powered Returns

The AI revolution isn’t just changing how the world works. It’s also changing how we invest.

At TradeSmith, we’ve built our own AI system — not for designing chips or cooling data centers, but for predicting short-term stock moves.

We call it Predictive Alpha.

It’s a large-scale AI model trained on vast amounts of stock market data. And it uses proprietary machine learning models to forecast the expected price path of thousands of stocks for every trading day over the next month.

We’ve engineered Predictive Alpha on more than 120 million data points. These include…

• 4.2 million historical price outcomes, spanning seven years and more than 2,300+ stocks
• 88.9 million daily forecasts, covering 21 forecast days for every stock on every trading day of the year
• Plus, tens of millions of additional “validation runs,” including target accuracy, pattern recognition layers, and more

Based on this data, Predictive Alpha learns from the past, adapts to the present, and projects the future.

We’ve found that every stock has a specific “profit window” when it can move the most. For example, Tesla may have a 6-day window, but Apple may have a 15-day window.

Every day, our system finds the ideal window to trade a particular stock.

Predictive Alpha can’t predict the future with 100% accuracy. And it won’t get every trade right. But it can forecast, with up to 85% probability, where stocks will be tomorrow.

And the longer it runs, the smarter it gets. Here are the top 10 wins that it has identified so far…

• 9.5% on SOFI in 3 days
• 9.6% on SOFI in 8 days
• 10.3% on UPST in 1 day
• 11.4% on MCW in 3 days
• 11.6% on ANF in 3 days
• 12.0% on ACCO in 26 days
• 12.6% on SLV in 4 days
• 16.4% on APLD in 6 days
• 17.6% on NTAP in 21 days
• 25.5% on CVNA in 2 days

These aren’t annual returns. They’re happening over a matter of days. Repeating these types of gains over these timeframes is like having a “house edge” on a casino-hopping Las Vegas trip.

Put the “House Edge” on Your Side

Each year, the Nevada Gaming Control Board (NGCB) writes a report on the success of the casinos on the Las Vegas Strip.

We all know the house always wins — this report tells us by how much.

In 2019, tens of millions of people flocked to Las Vegas to play casino games like blackjack, poker, and roulette. The NGCB shows that, collectively, those folks went home $6.6 billion poorer.

The house edge is the percentage of a player’s bet that the casino is likely to win. Said another way, it’s the statistical advantage a casino holds in any given game.

Casinos don’t win year in and year out because they get lucky. They have a deep knowledge of probabilities. And they use this knowledge to build an edge into every game they operate.

They make small wins thousands of times a day, millions of times a year. And these small edges, applied relentlessly, pile up.

It’s the same with Predictive Alpha. A 9.5% win in three days is impressive. Sustained across a trading year, those kinds of wins really add up.

The secret to making money in Vegas or in markets is simple: find your edge and apply it over and over. Predictive Alpha gives you that edge.

Want to See Predictive Alpha in Action?

With Predictive Alpha, we didn’t chase the impossible dream of predicting the future or being right 100% of the time. We looked for an edge we could exploit over and over again.

Billionaire casino operators know how powerful that kind of edge is. The world’s best traders know it as well.

And as I mentioned up top, you can now put this edge to work for you.

So, if you haven’t already, check out a live demo.

You’ll see how our AI analyzes five top stocks and predicts where their share prices could land over the next 21 days. It’s one of the most powerful trading tools ever developed — and you can try it free.

Just click the link below. It will take you to the page where you can try out the free demo, live on screen.

When you do, you’ll also discover how you can use this same AI system to forecast price movements on any of more than 2,300 stocks we track. From the AI infrastructure winners in this report to every major stock on Wall Street.

The future belongs to those who embrace AI — as an investment theme and as a trading edge.

Get your free access now.

Keith Kaplan

CEO, TradeSmith