How AI Is Reshaping Chips, Power, and Data Centers: All the Way to Space-Based Data Centers

Hello, I'm Ryuta Hamamoto from TIMEWELL.

In the previous articles, we looked at the structure of the semiconductor industry, geopolitics, and various national regulations. But behind all of these moves, there is one enormous force blowing up demand itself: AI. In this article, I'll trace how AI is transforming demand for semiconductors, how it then runs into an unexpected bottleneck called "power," and how this has finally led to something that sounds like science fiction becoming reality, building data centers in space.

Climb the staircase one step at a time, and it goes like this. AI demands GPUs. Running GPUs requires power. There isn't enough power, so companies pull nuclear reactors back into service. Even that isn't enough, so they finally set their sights on space. It may sound far-fetched, but follow it one step at a time and every step is connected to the ground beneath it.

AI Pushed Semiconductors into a "You Can Never Have Enough" State

Roughly speaking, how smart a generative AI is comes down to "how much computation it has done." It learns from vast amounts of data and answers users' questions. Both of these require tens or hundreds of thousands of computation-specialized chips called GPUs.

This demand is no longer normal by any measure. NVIDIA's CEO publicly stated that "we can see $500 billion of orders through 2026," then revised that upward, adding "if you look out to 2027, we see at least $1 trillion." Beyond that, just four companies, Microsoft, Google, Amazon, and Meta, are expected to spend a combined total of over $700 billion on data centers in 2026, roughly 77% more than the year before, nearly doubling. These companies are competing to "grab every GPU you can find." As a result, HBM, the memory used for AI, has become so scarce that SK Hynix, the largest maker, says it is "sold out for the next several years." No matter how much they make, they can't keep up. AI has turned the entire semiconductor industry into a seller's market.

One thing worth keeping in the back of your mind here is that an investment rush this large always comes with a worry that "this may be going too far." Will the money invested actually pay off in profits? There's still no clear answer to that question. The demand is unquestionably real, but you can't rule out the possibility that expectations have inflated too far. The more you're caught up in the frenzy, the more important it is, I think, to keep both sides in view.

The Real Bottleneck Turned Out to Be "Power"

You might assume that grabbing enough GPUs is the end of the story, but it's nothing of the sort. Here, a wall that many people overlooked rises up: power. Imagine buying thousands of the latest sports cars, only to find there isn't a single gas station nearby where you can fill them up. That's exactly the situation unfolding now.

The numbers are staggering. A single large AI data center now consumes anywhere from a few hundred megawatts to several gigawatts of power. One gigawatt (GW) is enough to supply roughly 750,000 to 1 million US homes, equivalent to the output of one large nuclear reactor. And a single building swallows all of that. The International Energy Agency (IEA) estimates that global data center power consumption will nearly double from about 415 TWh in 2024 to roughly 945 TWh in 2030. That is a scale that, as of 2030, almost matches the amount of electricity all of Japan uses in a single year. Think of it this way: the world's data centers alone will eat a whole Japan's worth of electricity, and the sheer abnormality of it comes through.

Let's line up a few more figures. "Stargate," which OpenAI is building in Texas, will reach about 1.2 gigawatts when complete; "Colossus," which xAI has set up in Memphis, runs around 555,000 GPUs for roughly 2 gigawatts; and "Hyperion," which Meta is planning in Louisiana, is on a scale of up to 5 gigawatts and around 2 million GPUs. A single data center draws as much power as an entire regional hub city. By one researcher's estimate, training GPT-4 alone, the foundation of ChatGPT, consumed power equivalent to what roughly 40,000 US homes use in a year. A single round of an AI's "studying" consumes a whole town's worth of electricity. This is why tech companies around the world are frantically scouring the map for power plants.

Naturally, the grid can't keep up. In the US power market PJM, a capacity auction fell short of its target for the first time ever, by a full 6,625 megawatts. In Texas, large-scale grid interconnection requests have ballooned roughly fourfold in a year, to 233 gigawatts. That's precisely why tech companies have turned to a startling measure: nuclear power. Microsoft is restarting the Three Mile Island plant (835 megawatts), famous for that accident, in 2027 to secure it for its own use; Amazon contracted for up to 1,920 megawatts from the Susquehanna plant; Meta for up to 6.6 gigawatts; and Google signed a 500-megawatt deal with Kairos. The new nuclear capacity that major tech companies moved to lock in over the past year totals more than 10 gigawatts. Even so, it falls far short of what's needed, and the lead time for gas turbines has reached as long as seven years. After the chip shortage came the power shortage.

The Front Line Finally Moves to "Space"

If there isn't enough power on the ground, what then? Here, an idea on a literally different dimension emerges: building data centers off the Earth, in space.

The one who fired the opening shot was SpaceX, led by Elon Musk. On June 12, 2026, it listed on Nasdaq. The amount raised was about $75 billion, the largest IPO ever, and its valuation at IPO was around $1.75 trillion (the market cap topped $2 trillion only after subsequent share-price gains). Around the same time, the company unveiled its concept for an orbital data center satellite called "AI1," carrying AI's computation hardware. Why space? In space there are no clouds and no weather, and sunlight pours down 24 hours a day. The vacuum of space itself provides the cooling environment. It frees you from the ground's power shortages and grid constraints.

This isn't only about SpaceX. Google has announced "Project Suncatcher" to launch its in-house AI chips (TPUs) into space, with prototype satellites planned for early 2027. And there is already a company with actual hardware running in orbit. Starcloud, an NVIDIA partner (its funding was led by Benchmark and EQT, while NVIDIA is its partner and GPU supplier, not the lead investor), launched an NVIDIA H100 into space in November 2025 and became the first in the world to successfully run an AI model in space. Amazon founder Jeff Bezos has also said that "in 10 to 20 years, gigawatt-scale data centers will be built in space."

To be honest, SpaceX's AI1 and Google's Suncatcher are still at the design and planning stage, and no full-fledged orbital data center has actually been completed. The only thing actually operating in space is Starcloud's single unit. Even so, the very fact that the world's top companies have begun seriously aiming for space speaks volumes about how serious the power constraints on the ground have become.

How This Loops Back to Semiconductor Demand

Let's bring this story back to the perspective of semiconductors. This wave, which spread from AI to power and then to space, comes full circle to push semiconductor demand even higher.

First, demand for GPUs keeps growing. On top of investment on the ground, a new market in space is now emerging. Second, demand for HBM, which determines a GPU's performance, rises in step. And third, easy to overlook, demand for "power semiconductors" explodes. These are chips for efficiently converting and controlling electricity, made from materials such as SiC (silicon carbide) and GaN (gallium nitride). Why have they suddenly become important? Because the power consumption of an AI rack is set to jump roughly fivefold, from about 120 kilowatts in the current Blackwell generation to a full 600 kilowatts in the next generation in 2027. To handle this efficiently, NVIDIA is moving to standardize its power architecture around 800-volt high-voltage direct current (HVDC). Using GaN improves power-conversion efficiency from 94% to 98%. Data center power supplies, new nuclear plants and SMRs, solar generation for space data centers, in every scene surrounding power, these power semiconductors are needed.

NVIDIA's Jensen Huang calls a data center an "AI factory" and says that "raw electricity and data are the fuel, and tokens (AI's output) are the new commodity." He also notes that "just building a 1-gigawatt factory costs about $40 billion." The main battlefield in the fight over AI has shifted from "how to compute more cleverly" to "how to secure physical power." And to secure that power, yet more new semiconductors are required. It has become a structure where demand begets demand.

Now, AI this powerful is no longer merely a convenient tool. Climb one more step up the staircase and you arrive at a question like this: can nations really leave AI models this capable unregulated? In fact, in June 2026, an unprecedented thing happened: AI models themselves became subject to export controls. Next, in The Era When AI Models Themselves Become Subject to Export Controls, we'll look at the front line of that shift.

References: NVIDIA's orders and the hyperscalers' investment figures are based on GTC announcements and each company's earnings; power demand on the IEA's "Energy and AI"; nuclear contracts on each company's IR materials, Utility Dive, and World Nuclear News; and the SpaceX IPO and space data centers on CNBC, NPR, Google's official blog, NVIDIA's official blog, and others (2025–2026). AI1 and Suncatcher are at the design/planning stage.