The AI Boom Is Running Out of Earth. Here’s Who Benefits.
The AI boom’s next chapter may depend on who can build the infrastructure fast enough to keep up.
Listen to the audio version of this article (generated by AI).
For the past two years, we’ve spent a lot of time talking about the AI Boom as a semiconductor story.
First, the world needed more GPUs.
Then it needed more memory.
Then networking.
Then cooling.
Then power.
Then data centers.
And now? Apparently, AI needs space. Literally.
Google (GOOGL, GOOG) has published new research exploring what may sound like science fiction at first glance: putting AI computing infrastructure in orbit. The company’s Project Suncatcher envisions fleets of solar-powered satellites equipped with Google’s Tensor Processing Units, or TPUs, communicating through high-speed optical links and performing machine-learning workloads in space.
And Google isn’t alone.
Redwire (RDW) recently announced a partnership with Sophia Space to develop orbital data-center infrastructure. Honda is working with Redwire on robotic systems for future commercial space stations. Meanwhile, terrestrial data-center developers are increasingly confronting power shortages, grid constraints, community opposition, and demands for enormous new sources of electricity.
At first glance, all these developments look unrelated.
We don’t think they are. We think they’re different manifestations of the exact same underlying trend:
The AI infrastructure boom is getting so big that the industry’s biggest challenge is increasingly becoming physical scarcity.
Not a lack of demand. But a lack of everything needed to satisfy that demand. And that has enormous investment implications.
The AI Boom Keeps Getting Bigger
Let’s start with the demand side, because the numbers there have become almost difficult to comprehend.
Bloomberg Intelligence recently raised its forecast for the annual AI accelerator market in 2030 from about $560 billion to $1.4 trillion.
That’s a 150% increase.
The firm now expects accelerator revenues to grow roughly 38% annually through the end of the decade, with the market approaching $1 trillion as early as 2028. Importantly, that forecast isn’t simply assuming Nvidia keeps selling more expensive GPUs. Bloomberg increased its expected accelerator unit shipments by 50% while also raising its expected blended selling prices by 65%.
In other words:
We’re going to need far more AI chips, and each AI system is going to contain far more valuable hardware.
And we’re already seeing evidence that demand is outrunning supply.
Amazon Web Services is raising pricing on reserved GPU capacity by about 15%.
Micron (MU) just said more than 75% of its fiscal 2027 shipments are already committed, with allocation discussions moving into 2028 and some customer agreements stretching beyond 2030.
Management still can’t identify when DRAM supply will catch up with demand.
And Micron is responding by dramatically increasing investment. After spending $27.37 billion in fiscal 2026, its current plans imply more than roughly $50 billion of capex in fiscal 2027.
Meanwhile, Anthropic has reportedly committed roughly $518 billion to compute and infrastructure over approximately the next decade, with about 80% of those obligations reportedly noncancelable or payable regardless of usage.
This doesn’t look like an infrastructure cycle that’s nearing its peak. On the contrary, it looks like the cycle is running into the physical limits of its supply chain.
Power Is the New Bottleneck
The AI Boom started with a chip shortage. But as more accelerators came online, another bottleneck emerged: memory.
Today’s leading AI systems require enormous amounts of high-bandwidth memory, or HBM, because the chips need to constantly move massive datasets into and out of the processors.
That requirement is growing rapidly.
Bloomberg expects HBM contract pricing to rise 70% to 80% in 2027, while next-generation accelerators are incorporating significantly more memory per chip. The result is that memory is becoming an increasingly large portion of the value of an AI system.
Then comes power delivery.
Vicor (VICR) recently raised its quarterly growth outlook twice in a matter of weeks, most recently to more than 30% sequential revenue growth, specifically because royalties tied to its Vertical Power Delivery technology increased faster than expected.
Why?
Because supplying enormous amounts of electricity efficiently to increasingly powerful AI processors is itself becoming a valuable technology problem.
And once that problem is solved inside the server, the same issue appears one level higher.
The data center needs electricity.
A lot of it.
That’s why Amazon (AMZN) just signed a long-term agreement supporting more than $3 billion of investment at Maryland’s Calvert Cliffs nuclear plant.
It’s why governments around the world are rethinking how giant data centers connect to their electricity grids.
Finland, for example, is considering rules that would prioritize smaller grid connections while requiring certain large electricity users to bring substantially more local generation and flexibility to the system.
And CoreWeave (CRWV) CEO Mike Intrator recently made what we think is one of the most important comments of this entire debate.
Asked about growing regulatory pushback against data-center development, he said the company has seen “absolutely not” any reduction in orders.
His point was simple: Restrictions aren’t reducing demand. Rather, they’re changing where the infrastructure gets built.
That is the key.
AI Infrastructure Is Becoming a Geography Problem
For decades, computing became progressively less constrained by geography.
Software moved to the cloud.
Applications became accessible everywhere.
Data moved instantaneously.
AI is partially reversing that trend. Because the intelligence may be digital, but the infrastructure producing it is extremely physical.
A modern AI data center requires chips manufactured in billion-dollar fabs, HBM produced in highly specialized facilities, advanced networking, transformers, cooling equipment, enormous amounts of land, and gigawatts of reliable electricity.
Those resources aren’t available everywhere. And increasingly, they aren’t available fast enough anywhere.
So companies are adapting.
Some are moving data centers closer to available energy. Some are financing new power generation themselves. Some are pursuing nuclear plants. Some are exploring natural gas. Some are moving infrastructure across national borders.
And now Google is asking an even more radical question: Why build all the AI infrastructure on Earth at all?
Google’s Wild Idea: Put the Data Center in Space
Google’s Project Suncatcher research starts with an observation that is both obvious and surprisingly powerful.
The Sun is the largest energy source in the solar system.
On Earth, solar generation has to deal with nighttime, weather, land constraints, and the cost of transmitting electricity from where it is generated to where it is consumed.
In orbit, sunlight can potentially be accessed much more consistently.
So Google researchers are studying a system in which satellites carry TPUs and solar arrays and communicate with one another using free-space optical links.
There are obviously enormous engineering challenges.
Launch costs.
Heat dissipation.
Radiation.
Networking.
Servicing.
Reliability.
But Google has already begun testing some of the critical assumptions.
Its researchers found that Trillium TPUs survived radiation exposure equivalent to roughly a five-year mission without permanent failure. And Google’s analysis suggests that if launch costs to low-Earth orbit eventually fall below roughly $200 per kilogram — something its learning-curve analysis suggests may become possible in the mid-2030s — amortized launch costs could become roughly comparable to terrestrial data-center energy costs on a per-kilowatt basis. Pasted text
We’re not saying Google is about to replace Virginia data centers with orbital TPU farms next year.
Obviously not.
This is a moonshot.
But the fact that one of the largest computing companies on Earth is seriously studying this idea tells us something very important about the magnitude of the problem.
The industry isn’t searching for less compute.
It’s searching for dramatically more places to put it.
The Biggest AI Winners May Be the Companies Solving Scarcity
This is why we’ve consistently argued that investors shouldn’t think about the AI Boom as simply an Nvidia story. The opportunity keeps broadening.
First there were accelerator winners. Then memory winners. Networking winners. Power-management winners. Equipment winners. Utilities. Nuclear. Data-center developers. And potentially, eventually, space infrastructure.
The common thread isn’t any one technology, but scarcity.
Whenever AI demand runs into a bottleneck, an enormous economic incentive emerges to solve that bottleneck.
Not enough GPUs? Build more GPUs.
Not enough memory? Build more fabs.
Not enough power? Restart nuclear plants.
Not enough grid capacity? Build generation directly beside the data center.
Not enough suitable terrestrial infrastructure? Maybe start looking up.
And as AI agents become increasingly capable, these demands should only intensify.
OpenAI’s Dots and Meta’s (META) Muse point toward a world where AI doesn’t simply answer one prompt and stop. Agents can continue working after users leave, operating browsers, executing code, retrieving information, and performing multistep assignments.
That means one human request can trigger dozens (eventually perhaps hundreds) of computational actions.
More inference, CPUs, and memory. More networking, storage, and electricity. And, last but not least, more infrastructure.
That’s why we remain convinced the AI infrastructure cycle has years left to run. The bottlenecks are evidence of how quickly it’s expanding.
The Bigger Picture
Markets remain noisy.
Rates are high. Oil remains elevated. And rising Treasury yields keep putting pressure on stock valuations.
Those forces have kept AI stocks trapped in the tug-of-war we’ve been discussing for months: earnings pulling stocks higher while rates push valuations lower.
But underneath that noise, the fundamental AI story continues to strengthen.
Infrastructure commitments are stretching into the 2030s. Memory supply is tight. Data-center developers are searching for electricity. Nuclear power is attracting renewed interest.
And Google is investigating whether some of the computing should simply leave the planet.
Put it all together, and a bigger picture emerges.
The AI boom’s next chapter may depend on who can build the infrastructure fast enough to keep up.
More chips. More power. More computing capacity… potentially even in orbit.
That’s where our research is taking us. And it’s why I want you to see what I’m calling “XPANSE.”
It centers on an ambitious shift taking shape around Elon Musk’s businesses… one that I believe could connect several of the opportunities we’ve been tracking, from AI infrastructure and advanced manufacturing to the emerging space economy.
The question that interests me most is… which companies could get paid to help build it?
Even the most ambitious vision needs suppliers. Someone has to provide the specialized equipment, critical materials, and essential components that turn an idea into something that works.
In my XPANSE presentation, I explain the potential shift, why it matters for America’s technological future, and three steps investors can take to prepare. I also share the name and ticker of an investment positioned to participate, free of charge.
The infrastructure constraints we’ve discussed aren’t going away overnight. For the companies that help solve them, that could mean years of opportunity.
Watch my XPANSE presentation to see where I believe that opportunity is taking shape.
