Back in the heyday of pirating, the Spanish treasure fleets rode low and slow, weighed down with New World silver. Every sailor dreaded this one particular leg of the voyage home… a narrow blue corridor between Florida and the Bahamas.
There was no sailing around it.
The wind and current here squeezed every homebound ship through the same few miles of water. The pirates who understood this never chased galleons across an ocean. Rather, they found the one harbor those galleons were forced to pass through, then dropped anchor and let the wealth of an empire come to them.
The harbor was Nassau.
This winter, divers fighting the ripping currents off New Providence Island raised the first pirate wrecks ever found there. Forty feet beneath the surface, in a current strong enough to drag a diver out toward the bull sharks, an archaeologist closes his hand around a worn stone. The kind of stone once used to put an edge on a cutlass.
Somewhere in the silt of Nassau harbor nearby lie the charred ribs of what may be the Fancy, the 46-gun frigate that robbed a Mughal emperor blind and then sailed off into legend.
Divers brought her up this winter, the first pirate wreck ever recovered in these waters. And she was empty. Stripped to the waterline… torched… every fitting of value was carried off long before she went down.
The famous ship was worthless. It always has been. Why? Because the treasure was never in the vessel everyone remembers but in the harbor that vessel was forced to pass through… the chokepoint where an empire’s wealth was forced to slow down long enough to be taken.
Now, here we are, three centuries past the Fancy’s expiration date, with a new treasure map in hand. Yet, investors are making the same old mistake the less-seasoned pirates made… they’re chasing galleons.
Today’s galleons aren’t flush with gold. They’re AI models, like ChatGPT, or Gemini, or Claude. And too many are betting on which model wins. The actual “treasure map,†however, points to less-travelled locations.
Because, no matter which model wins, it still must cross the same narrow path of the proverbial seas… I’m talking about the chips, the power, the rare-earth magnets, the launch capacity, and the manufacturing that every AI model counts on to run.
Every quarter, Wall Street asks the same question: is the AI spending boom real, or is it a story companies are telling to justify their valuations?
Follow the capital, and that narrow path lights up. Microsoft (MSFT), Alphabet (GOOGL), Amazon (AMZN), and Meta (META) have now committed more than $700 billion to AI infrastructure – the physical layer, not the models. Demand still outruns supply at every step. Customers aren’t negotiating on price so much as asking for volume.
These numbers tell us that the AI economy is not a bubble waiting to pop… it is a structural buildout that is still accelerating. Demand still exceeds supply everywhere in the stack. Customers aren’t negotiating price. They’re just asking for volume.
That $700 billion in capital is landing across five fronts at once – humanoid robots taking real factory shifts, data centers headed for orbit, AI glasses arriving from every major tech name in a single year, custom silicon loosening Nvidia’s (NVDA) hold, and nuclear reactors rising to power the whole thing. Inside, we walk through all five and name the 13 companies sitting on the supply, including Tesla (TSLA), Taiwan Semiconductor (TSM), and the lone U.S. rare-earth miner the Pentagon has already taken a stake in.
Own the water, don’t chase the ships. We lay out our complete treasure map below.
Let’s go through it one megatrend at a time.
What’s Inside This Report
Megatrend 1: Humanoid Robots Clock In
At the Consumer Electronics Show (CES) in January 2026, Hyundai brought out Boston Dynamics’ Atlas robot — an engineering marvel that, for years, was fun to watch but offered no obvious commercial value.
Then they showed it doing something boring: parts sequencing.
Atlas walked into a disorganized staging area in a mock factory, identified heavy car components using a Google DeepMind-powered reasoning brain, and placed them precisely onto an assembly line feeder. When it needed to turn around, its torso spun 180 degrees while its legs stayed planted — then it bent its knees backward to walk away.
It was unsettling. Not because it was scary. Because it was efficient. These things are designed for pure throughput, uninhibited by human biology.
And then Hyundai announced Atlas would be deployed to its Georgia Metaplant this year.
The Shift From Science Project to Workforce
The inflection point I called at the end of 2025 has moved beyond the theoretical. Consider what is already deployed in the real world:
- Figure AI‘s robots completed an 11-month program at BMW’s Spartanburg facility, running 10-hour shifts Monday through Friday and contributing to the production of more than 30,000 X3 vehicles.
- Tesla‘s Optimus Gen 3 is operating autonomously inside the Austin Gigafactory, moving boxes and sorting parts, powered by the same Full Self-Driving neural nets that run Tesla vehicles. Elon Musk has said Optimus will eventually be worth more than Tesla’s car business and FSD combined.
- Agility Robotics‘ Digit is moving through the full commercial arc: pilot, paid deployment, multi-site expansion. GXO logistics is using a robots-as-a-service model. Mercado Libre signed a commercial agreement. Amazon is testing Digit for tote handling and item consolidation.
- 1X‘s NEO home robot is accepting $20,000 early-access pre-orders with U.S. deliveries starting this year.
This is not the robotics story of 2021, when every company was announcing partnerships and no one was shipping anything. This is deployment.
Industry forecasts now call for more than 500,000 humanoid units operational by the mid-2030s.
And there is a tailwind that most investors are missing entirely.
The Government Is Getting Into Robotics
The Trump administration has identified at least seven industries it wants to build up using direct federal investment — and robotics is among them. We have already seen what happens when the government decides an industry is strategically critical.
Earlier in this administration, the Pentagon took a 15% equity stake in MP Materials to secure the rare earth supply chain. The stock surged 111% in the week following the announcement. Shortly afterward, the government poured $9 billion into Intel — and the stock nearly doubled over the following three months. Then came Lithium Americas (+194% in two weeks after an equity investment announcement) and Trilogy Metals (+211% in a single trading session).
When Washington writes a check, assets re-price overnight.
The “One Big Beautiful Bill Act” earmarked $7.5 billion specifically for critical strategic assets. I believe a meaningful portion of that is heading toward the robotics and rare earth supply chains. If I am right, you want to be positioned before the announcement — not after.
The Stocks to Buy
My approach to the humanoid opportunity? Own the supply chain bottlenecks that every robot must pass through regardless of which company manufactures it.
| Tesla (TSLA) |
| Tesla is the only publicly traded company that is simultaneously building humanoid robots at scale AND using them in its own factories — which means it has a captive testing environment that no other company can replicate. |
| – Optimus Gen 3 is running autonomously in Austin Gigafactory today — not a demo, an active worker – Tesla is both the manufacturer AND the first customer — internal demand gives it a runway to perfect the technology before selling externally – The same FSD neural net architecture powering Autopilot also powers Optimus — massive R&D leverage – Musk’s stated long-term thesis: Optimus will eventually exceed the value of Tesla’s car business and FSD combined |
| Qualcomm (QCOM) |
| Qualcomm’s Snapdragon platforms are the edge AI architecture of choice for both humanoid robot perception systems and the next generation of AI-powered smart glasses — making it one of the most powerful cross-megatrend plays in this report. |
| – Snapdragon chips power real-time visual processing and AI inference at the edge — exactly what robots need to perceive and navigate their environment – Qualcomm is the chip architecture underlying both the Google/Warby Parker AI glasses platform AND OpenAI’s AI device project – Cross-megatrend exposure means Qualcomm wins whether robots or wearables scale faster – Edge AI is structurally advantaged over cloud AI for latency-sensitive real-world applications |
| MP Materials (MP) |
| Every humanoid robot uses rare earth magnets in its motors and actuators. MP Materials is the only rare earth mining and processing company of scale in the United States — and it is already government-backed. |
| – Pentagon took a 15% equity stake in MP in 2025 — the stock surged 111% in the following week – Humanoid robots require large quantities of neodymium-iron-boron magnets for their joints and drive systems – The U.S. has essentially zero domestic rare earth production outside of MP — strategic importance is baked in – A second government investment signal related to the robotics supply chain would likely reprice MP again |
Megatrend 2: The AI Grid Is Leaving Earth
In 1900, if you wanted to power a factory, you built your own generator.
You hired engineers to design it, bought fuel to feed it, and built your entire operation around the assumption that power was something you produced yourself, on-site, because there was no other option.
Then the grid arrived. The constraint — not the technology, the capital, or the ambition — disappeared overnight. What followed was an entirely new industrial era.
AI is living through its generator moment right now. And the grid that’s coming isn’t on the ground.
It’s in orbit.
The Real Bottleneck Nobody Is Talking About
Most investors think the AI infrastructure race is about chips. And chips matter — Nvidia’s GPUs, Broadcom’s custom accelerators, Marvell’s networking silicon. These are real, important, and worth owning.
But the constraint that is becoming the actual ceiling on the AI buildout is watts, not silicon.
A single next-generation AI training cluster — the kind that trains frontier models — can consume one to two gigawatts of electricity: enough power to run a mid-sized American city. The hyperscalers are each trying to build dozens of these clusters simultaneously around the globe.
The U.S. electrical grid was not designed for this. Utility interconnection queues — the waiting list to plug a new facility into the grid — now stretch five to 10 years in many regions. Bloomberg estimates that nearly half of all AI data center projects in the U.S. will be delayed this year due to power constraints alone. Microsoft, Alphabet, and Amazon have all found themselves in the same absurd position: having the chips and the capital, and still not being able to get a power connection in time.
The binding constraint is land, water, and power — physical resources subject to physical limits. And physical limits don’t yield to larger budgets.
Why Space Solves What Earth Cannot
Space has three things Earth’s data centers are running out of: energy, cooling capacity, and proximity to where AI data originates.
Solar panels in low Earth orbit receive 1,400 watts per square meter of raw energy when in direct sunlight. At Earth’s surface, the same panels average just 20 to 60 watts per square meter once night, clouds, and shifting sun angles take their toll. That is not a marginal difference — it is a structural one.
Cooling is even more straightforward. Space is a near-perfect vacuum, sitting just a few degrees above absolute zero. That means heat can be dumped straight off of GPUs and into the void through large radiator panels; no fans, water, or intensive cooling infrastructure required. Terrestrial data centers, by comparison, burn through billions of gallons of water per year just to keep systems from overheating.
Then there is the third advantage, which may become commercially significant faster than anyone expects: data proximity. An enormous volume of AI workloads — satellite imagery analysis, defense surveillance, weather telemetry, maritime tracking — originates in space. Today, that data travels down to Earth, gets processed in ground-based data centers, and results get distributed back out. Move the compute to where the data originates and you eliminate that bottleneck entirely. For an Earth observation satellite, it is the difference between transmitting 100 gigabytes of raw imagery and transmitting one megabyte of actionable intelligence.
The Proof Is Already In the Timeline
This might sound like a decade-away thesis. The deployment timeline says otherwise.
November 2025: Starcloud launches a satellite containing an Nvidia H100 GPU — the first chip of that class ever sent into orbit.
December 2025: That same satellite trains a language model in space for the first time in history.
January 2026: SpaceX files with the FCC for authorization to launch up to one million satellites as orbital AI data centers — not a press release, a regulatory filing.
February 2026: SpaceX acquires xAI in the largest private merger to date, valuing the combined entity at $1.25 trillion, with orbital compute as the stated rationale.
March 2026: Nvidia announces the Vera Rubin Space-1 module at GTC — a chip platform purpose-built for orbital data centers. Jensen Huang’s exact words: “Space computing, the final frontier, has arrived.”
April 2026: SpaceX confidentially files for an IPO targeting a $1.75 trillion valuation — the largest in market history — centered around orbital compute.
And Meta struck a commercial agreement with orbital energy startup Overview Energy to power its AI data centers using solar energy collected in space and beamed back to Earth, with commercial delivery targeted for 2030.
Meta is the world’s fifth-largest company by market cap. It is currently spending $60 to $65 billion on AI capital expenditure annually.
It signed this contract because its engineers ran the numbers, looked at the global power grid, and concluded that Earth cannot supply enough electricity to support their AI ambitions.
The Economics: Expensive Today, Inevitable Tomorrow
Running one H100-equivalent GPU-hour in a terrestrial data center can cost only a few dollars — and potentially lower at hyperscaler scale — once hardware, power, cooling, facilities, and operations are spread across high utilization. In orbit, current economics are still far more expensive. Our estimate puts the cost near $142 per GPU-hour, with roughly 60% of that tied to launch cost alone.
That is the crux of the orbital compute thesis: it is a bet on dollars per kilogram to low Earth orbit collapsing. Google’s Project Suncatcher team has already modeled a future in which launch prices below $200/kg could make space-based AI infrastructure roughly comparable to terrestrial data-center energy costs on a per-kilowatt-year basis. SpaceX’s Starship is designed to attack that same variable through full reusability and massive payload capacity. Historical launch costs have already fallen sharply in the reusable-rocket era. Though whether Starship can push the curve low enough for orbital compute to cross from science project to economic reality remains to be seen.
Our analysis puts the economic crossover around 2038. At that point, orbital compute likely flips from premium to low-cost option for AI inference. The reason: orbital compute costs are technology-bound, subject to engineering learning curves that reliably drive prices lower. Terrestrial compute costs are increasingly resource-bound, subject to physical scarcity in power, land, and water that reliably drives prices higher. Both forces are already in motion. Neither is reversing.
The Stocks to Buy
Every significant space investment cycle has ended in violent repricing. Capital markets cannot distinguish “thesis correct, timing uncertain” from “thesis wrong.” The SpaceX IPO will price in the 2035 bull case on day one. Everything in the supply chain will spike. And then, 12 to 18 months later, the first execution disappointment will arrive — a Starship delay, an FCC ruling, a technical setback. None of it will kill the thesis. All of it will reprice the sector.
The playbook: build positions now, in the pre-IPO window. Trim at the spike. Buy the correction with the proceeds. This sequence has played out in cloud, fiber, mobile, and shale. The thesis isn’t going anywhere. The only way to miss it is to chase the narrative instead of the price.
| Rocket Lab (RKLB) |
| Rocket Lab is the most established public-market pure-play on the orbital infrastructure buildout. |
| – Electron rocket has a strong track record of successful commercial and government missions – Larger Neutron rocket in development to compete for medium-payload contracts as orbital compute demand grows – Vertical integration across launch and satellite manufacturing creates a recurring revenue base beyond individual launch events – $805 million Space Development Agency contract provides revenue visibility; $1 billion-plus total backlog |
| Tema Space Innovators ETF (NASA) |
| For investors who want broader orbital compute exposure — including the closest public-market proxy for SpaceX — the Tema Space Innovators ETF offers 12% direct SpaceX exposure via a Forge/Schwab SPV alongside Rocket Lab, Planet Labs, and AST SpaceMobile |
| – ~12% SpaceX exposure via SPV — the most direct public-market access to the orbital compute keystone – Diversified across the orbital stack: launch, Earth observation, satellite broadband – Natural trim candidate once SpaceX IPO prices and the narrative spike arrives |
Megatrend 3: AI Glasses Go Mainstream
The traditional smartphone era runs on a ritual: pull out your phone, open an app, type something, tap something, scroll something. Repeat that hundreds of times a day.
But the AI era breaks that ritual entirely. It’s built around ambient intelligence — technology that meets you in the real world, sees what you see, hears what you hear, understands context, and acts on your behalf.
That requires a different kind of device. And that device is launching this year — from multiple companies at once.
The Year Everyone Ships
I said at the start of 2026 that this would be the year AI wearables — specifically glasses — hit enough product-market fit to become a legitimate tech category again.
The validation has been arriving steadily since.
- Google and Warby Parker (one of the most important AI hardware stories investors missed): Reuters reported that the companies plan to launch lightweight AI-powered smart glasses in 2026, built around Google’s Android XR platform and Gemini AI. Google has outlined two categories for Android XR glasses: screen-free AI glasses and versions with an optional in-lens display. Expected features include real-time translation, navigation, contextual visual assistance, and hands-free access to Google apps. Google brings the AI and operating system; Warby Parker brings the eyewear design. That makes this a serious attempt to solve the smart-glasses problem from both sides: intelligence and wearability.
- Apple: Bloomberg has reported that Apple is targeting smart glasses for late 2026, with suppliers expected to produce large quantities of prototypes ahead of launch. The key is Apple’s approach: not bulky AR goggles, but an iPhone-connected AI wearable built around cameras, audio, Siri, and ecosystem integration. That attacks the real failure point for smart glasses so far — they have either been too awkward to wear, too limited to matter, or too disconnected from the devices people already use. When Apple enters a category, it usually means the category is ready to move from tech demo to consumer product.
- Meta Ray-Ban and Orion: Meta has been shipping its current Ray-Ban Meta smart glasses line since late 2023, building on its earlier Ray-Ban Stories partnership and steadily refining the product from camera/audio glasses into AI-first wearables. Its Orion specs are not a consumer product yet, but Meta has shown it as a working AR prototype and is using it with employees and select external testers to build toward a future consumer AR glasses line. With roughly 3.5 billion daily users across its apps, shipping AI glasses already in market, and retail distribution through Meta, Ray-Ban, LensCrafters, Best Buy, and other partners, Meta’s wearable strategy is more advanced than many investors realize.
- OpenAI + Qualcomm: We’ve also learned that OpenAI has tapped Qualcomm and MediaTek to develop chips for an AI-first consumer device, with Luxshare expected to serve as the exclusive system co-design and manufacturing partner. OpenAI is trying to build the next consumer computing interface — something that can see, hear, understand context, and act on the user’s behalf. The device is not expected to reach mass production until 2028 — further out than the other launches in this section, but significant as a signal of where the category is heading.
It’s no coincidence that Apple, Google, Meta, and OpenAI are all launching wearable AI products in the same 12-month window.
These tech giants are all responding to the same thesis: the smartphone’s reign as the center of our digital lives is nearing its end.
The Stocks to Buy
| Apple (AAPL) |
| Apple has a history of entering categories late and defining them permanently. The iPod wasn’t the first MP3 player. The iPhone wasn’t the first smartphone. Apple Glass won’t be the first smart glasses. But if history rhymes, it will be the first one everyone actually buys. |
| – Mass-production of prototypes began with suppliers in late 2025; launch targeted for end of 2026 – Seamless integration with iPhone, iCloud, AirPods, and Apple Watch gives Apple a closed ecosystem advantage no competitor can replicate – Over 1 billion active iPhones worldwide means the addressable install base for an iPhone-dependent accessory is larger than any other platform – Apple’s financial model transforms accessories into recurring revenue — hardware + services + subscriptions |
| Meta Platforms (META) |
| Meta is the only company that already has a commercially available, consumer AI glasses product on the market — Ray-Ban Meta glasses — giving it a real-world data and iteration advantage over every competitor entering the category in 2026. |
| – Ray-Ban Meta glasses are already shipping with AI voice assistant integration and camera-based visual search – Orion AR glasses represent the next generation — a full augmented reality device that layers digital information onto the physical world – Meta’s 3.5 billion daily active users are the most natural distribution channel for a consumer AI wearable – AI infrastructure investment also secures Meta’s position: extended Broadcom custom chip partnership through 2029; $125–145 billion in 2026 capex |
Megatrend 4: Custom Silicon Breaks Nvidia’s Lock
In 1455, Johannes Gutenberg printed his first Bible… and effectively went bankrupt.
His financier, goldsmith Johann Fust, took over the press, the inventory, and the entire business. Within a generation, Fust’s heirs had turned movable type into one of the most profitable industries in Europe.
Gutenberg built the marvel. Fust owned the supply.
Every transformative technology works this way. The inventor gets the credit. The infrastructure owner gets the wealth.
I’ve been making this argument for ages: the biggest gains in the AI cycle will not come from the companies building the AI models. They will come from the companies building what those models must run on.
And right now, the most important shift in AI infrastructure is hiding in plain sight…
Big Tech is quietly dismantling Nvidia’s monopoly — and the companies aiding them are the ones you want to own.
Why the Economics Changed
For the past three years, one company has sat at the center of every major AI buildout on the planet: Nvidia.
Its GPUs became the default infrastructure for training and running AI models — and the demand was so overwhelming that Nvidia became a chokepoint. At the peak of the AI buildout frenzy, hyperscalers were waiting eight to 12 months just to receive GPU shipments. Data center operators were redesigning entire facilities around Nvidia’s power and cooling requirements. AI companies were burning through cash just sitting on waiting lists.
So Big Tech is doing what Big Tech does when the economics don’t work: it is building its own chips.
The Avalanche of Custom Chip Deals
As AI has scaled into mass consumer and enterprise adoption, inference has become the dominant — and fastest-growing — compute cost in the entire industry.
That’s the opportunity that custom chips — Application-Specific Integrated Circuits (ASICs), or XPUs — are built to capture. Instead of doing everything, these chips are built for a single task. Less flexible, yes; but better performance-per-watt and significantly lower operating costs at the scale that Big Tech firms operate at.
And the deal flow we’ve seen from this niche has been staggering:
- Broadcom extended its TPU design partnership with Google through 2031 — a five-year commitment that says Google is building its entire next-generation AI infrastructure stack around custom silicon through the end of the decade.
- Anthropic committed to 3.5 gigawatts of TPU-based computing capacity — nearly four times what it was using just months prior. Anthropic’s annualized revenue has crossed $30 billion in 2026, up from roughly $9 billion at the end of 2025. The company chose Google’s custom TPUs over Nvidia GPUs for its $50 billion U.S. computing infrastructure buildout. That is a statement of intent from one of the most important AI companies in the world.
- Meta extended its Broadcom partnership through 2029, with Meta committing to using Broadcom’s custom silicon architecture for its AI workloads.
- Amazon’s Trainium chip — its custom AI accelerator — now has a $225 billion total revenue commitment. Trainium2 is largely sold out. Trainium3 just started shipping and is nearly fully subscribed. Trainium4 — 18 months from broad availability — is already substantially reserved.
Now, I want to be clear about one thing: I am not predicting the death of Nvidia. Nvidia’s data center revenue exceeded $60 billion in Q1 2026 — up 75% year-over-year — and its market cap has crossed $4 trillion. The company is also pivoting with the Grace and Vera CPU platforms as the infrastructure stack evolves.
Nvidia is not going away. But its monopoly is.
The Cisco Analogy That Changed How I Think About This
During the dot-com era, betting on the winning website was extraordinarily hard. Amazon survived. Pets.com didn’t. AOL rose, then faded. Dozens of companies burned through hundreds of millions of dollars and left investors with nothing.
But Cisco made money through all of it — because every byte of internet traffic needed its routers and switches to move across the web. The stock rose roughly 3,400% in five years.
Cisco didn’t have to pick the winning website. It just had to own the infrastructure every website needed.
The same logic applies to the AI model race. OpenAI, Anthropic, Google, Meta, xAI — they are all competing. Some will win. Some will fade. Most investors cannot predict which.
But every single one of those models needs custom silicon.
And right now, two companies design the vast majority of custom AI chips for the hyperscalers.
The Stocks to Buy
| Broadcom (AVGO) |
| Broadcom is the premier designer of custom AI chips — the ‘Builder’ at the center of the custom silicon ecosystem. It does not have to win the model race. It just has to design chips for whoever does. |
| – 5-year Google TPU partnership extended through 2031 — Broadcom has been Google’s custom chip designer since 2016 – Meta custom silicon partnership extended through 2029 – Anthropic’s 3.5-gigawatt TPU commitment runs through Broadcom’s design relationship with Google – Total addressable market for AI chips estimated at $700-plus billion — Broadcom sits at the center of the non-Nvidia share of that market – The business model is powerful: Broadcom designs the chip, earns engineering fees, and collects ongoing royalties — without the capital intensity of a fab |
| Marvell Technology (MRVL) |
| Marvell is Broadcom’s chief rival in the custom silicon design business — and it is winning some of the most important contracts in the industry, including chips for OpenAI and Amazon. |
| – Designing custom AI accelerators for Amazon (Trainium family) and reportedly for OpenAI – Amazon’s Trainium offers 30-40% better price performance than comparable Nvidia GPUs — validating Marvell’s design work – Amazon has $225 billion in total Trainium revenue commitments — a decade of demand floor for Marvell’s architecture – As custom chip adoption expands beyond Google and Amazon, Marvell is positioned to capture the next wave of hyperscaler design contracts |
| Taiwan Semiconductor (TSM) |
| TSM manufactures every leading-edge AI chip in the world — Nvidia’s, Google’s, Amazon’s, Meta’s, Apple’s. It is the single unavoidable tollbooth in the entire semiconductor value chain. |
| – No leading-edge AI chip can be manufactured at scale without TSM — it controls approximately 90% of the world’s most advanced semiconductor capacity – Custom chip boom is additive to TSM’s business: more design diversity means more manufacturing contracts – U.S. government backing via CHIPS Act funding for Arizona fabs reduces geopolitical risk while expanding capacity – As AI chip demand doubles and doubles again, TSM’s pricing power increases — it is one of the few companies in the world where demand structurally exceeds supply for years at a time |
Megatrend 5: Nuclear SMRs Hit Criticality
There’s a number that, in my view, doesn’t get nearly enough attention.
The AI data centers being built right now — the ones that will process the inference calls, run the agents, and power everything we have talked about in this report — will consume roughly 8% of all electricity generated in the United States by 2030. Some estimates put it higher.
Today, that number is about 2%.
That means we’re going to see a four-fold increase in electricity demand, from a single sector, in roughly five years. And every major tech company building these data centers knows it.
But where will all that power come from?
Solar and wind are variable — they don’t produce electricity when the sun isn’t shining or the wind isn’t blowing. Natural gas produces carbon. Existing large-scale nuclear plants have limited capacity to expand. Which leaves one answer that actually solves the problem:
Small Modular Reactors (SMRs).
What SMRs Are (and Why 2026 Matters)
An SMR is a nuclear reactor designed to be built in a factory, transported in modules, and assembled at the site where the power is needed. Unlike traditional nuclear plants — which cost $10 billion to $30 billion and take 15 years to build — SMRs are designed to be deployed much more quickly, at a fraction of the cost.
The technology has been in development for years. The promise has been enormous. The results, until now, have been… not great.
But 2026 is different for one specific reason: the U.S. Department of Energy’s Reactor Pilot Program.
The DOE has funded approximately $900 million across seven advanced reactor projects — including X-energy‘s Xe-100, Kairos Power‘s Hermes, and TerraPower‘s Natrium — with a specific goal: achieve criticality on at least three of these designs by July 4, 2026; America’s 250th anniversary.
Criticality means the reactor sustains a nuclear chain reaction on its own. It is not the same as generating commercial power — but it is the moment the technology crosses from R&D to reality.
And that changes the investment calculus entirely.
The AI-Nuclear Convergence
AI companies are desperate for power. They are signing 20-year electricity contracts, building private transmission lines, and cutting deals directly with nuclear operators. Microsoft restarted the Three Mile Island nuclear plant — specifically for AI data center power. Amazon, Google, and others are pursuing similar agreements.
Meanwhile, nuclear companies are desperate for customers. They have the technology. They need the revenue certainty to justify building.
These two needs are perfectly matched.
The AI boom has created the first genuinely commercial market for nuclear power in decades — not based on government subsidies or climate policy, but on cold, hard economic necessity. The data centers need baseload power that runs 24 hours a day, 365 days a year, regardless of weather. Only nuclear can reliably provide that.
And here is the government tailwind on top of all of this: the Trump administration has made nuclear a strategic priority. The “One Big Beautiful Bill Act” earmarked $7.5 billion for critical strategic assets, and multiple government signals are pointing toward nuclear as a target for the same type of equity investment that produced 111%–211% stock gains across several other strategic sectors in the past 12 months.
The pattern has repeated four times. I believe nuclear is next.
The Stocks to Buy
| Constellation Energy (CEG) |
| Constellation is the largest nuclear fleet operator in the United States — which makes it the most direct, liquid way to invest in the AI-nuclear convergence. |
| – Owns and operates 21 nuclear reactors across 12 plants — the largest nuclear portfolio in the U.S. – Microsoft deal to restart Three Mile Island is the template: AI hyperscalers will sign similar long-term agreements with Constellation as data center demand grows – Nuclear power is 24/7 carbon-free baseload electricity — exactly what AI data centers need and renewables cannot reliably provide – As AI data center power demand grows from 2% to 8%-plus of U.S. electricity by 2030, every nuclear megawatt becomes more valuable |
| Oklo Inc. (OKLO) |
| Oklo is the AI-nuclear convergence stock in purest form: a next-generation nuclear fission company backed by OpenAI CEO Sam Altman — the man most responsible for creating the AI energy demand that Oklo’s reactors will help solve. |
| – Sam Altman is Oklo’s chairman — the symbolic and strategic link between AI energy demand and nuclear supply is literally embodied in the company’s leadership – Building compact, factory-produced fission reactors designed specifically for deployment near AI data centers and industrial sites – No large-scale nuclear project in history has had a pre-built customer base of this quality before breaking ground – Earliest-stage play in this report — highest risk, highest potential reward for investors who understand the thesis |
| NuScale Power (SMR) |
| NuScale is the only publicly traded pure-play small modular reactor company — and if a government equity investment in the nuclear sector follows the pattern we have seen in rare earths, semiconductors, and battery materials, NuScale is the most obvious target. |
| – The only public pure-play SMR stock — if you want direct exposure to the SMR thesis without the conglomerate noise, this is the cleanest expression – NuScale’s SMR design was the first to receive design approval from the U.S. Nuclear Regulatory Commission – DOE Reactor Pilot Program milestones in 2026 could serve as a direct catalyst for re-rating – The government-backed equity pattern (MP Materials, Intel, Lithium Americas, Trilogy Metals) suggests nuclear is a high-probability target for the next equity stake announcement |
2026: The Year Tech Leaves the Screen
Every year, the tech industry promises we are about to “enter the future,” and then delivers a slightly thinner phone, a slightly stronger GPU, a slightly more sophisticated app.
But 2026 is different. The physical evidence is already showing up in the world.
Robots are running factory shifts. Data centers are being launched into orbit. The biggest technology companies in the world are signing billion-dollar chip deals to reduce their dependence on the one company that has dominated AI compute for five years. Nuclear reactors are achieving milestones that government-funded programs have been working toward for a decade.
And through all of it — powering the robots, the data centers, the glasses, the chips, and the reactors — $700 billion in committed capital is flowing into the infrastructure layer of the AI economy.
Gutenberg built the press. Fust owned the supply.
The inventor got the credit. The infrastructure owner got the wealth.
Every transformative technology works this way.
We are living through one of those transformations right now. The AI models — ChatGPT, Claude, Gemini, Grok — will get most of the headlines. Some of the companies behind them will generate extraordinary returns. But the most consistent money in this cycle, as in every technology cycle before it, will be made by the investors who own the physical infrastructure these technologies need to function.
Robots need rare earth motors, edge AI chips, and actuators. Orbital data centers need launch infrastructure and radiation-hardened chips. Smart glasses need a chip architecture. AI models need custom silicon and manufacturing capacity. And all of it — every server rack, every data center, every inference call — needs power.
That is the investment thesis in this report. Own the supply, not just the marvel.
The five megatrends detailed here are my highest-conviction opportunities for the remainder of 2026 and beyond. The plays I’ve highlighted — Tesla, Qualcomm, MP Materials, Rocket Lab, the Tema Space Innovators ETF, Apple, Meta, Broadcom, Marvell, TSM, Constellation Energy, Oklo, and NuScale — are not equally speculative. Some are large, liquid, well-established businesses that happen to sit at the center of these trends. Others are earlier-stage and carry more risk. Size positions accordingly and do your own due diligence.
But if you ask me where I would want to be positioned as the AI boom continues to grow, I’m all-in on infrastructure.
We hope you found this special report useful. Before we go, let us remind you that you’re now also a member of my free Hypergrowth Investing newsletter.
By uncovering early investments in hypergrowth industries, we put you on the ground floor of world-changing megatrends. Keep an eye on your email inbox for my next Hypergrowth Investing article soon. I typically send them every day of the week. In the meantime, you can check out our website by clicking here.
