By Dr. Gregory S. Carmichael · On the intersection of cryptographic infrastructure, energy, and national economic sovereignty

The Bottom Line. The bottleneck on artificial intelligence is no longer chips, algorithms, or money. It is electricity — specifically, firm power delivered on a date certain. Every serious solution on the table, from gas turbines to small nuclear reactors to a million solar-powered satellites, is a different answer to the same question: How do I get reliable megawatts faster than the utility can give them to me? But solving the power is only the first move. The winners also have to put the machine where the grid can reach it and its fuel can find it — and build it quiet enough that the neighbors don't litigate it out of existence. The companies that master all three — power, place, and consent — will own the next decade of computing. The ones that wait in the interconnection queue will not.
 

For two years, the story of artificial intelligence has been told in chips. Who has the most GPUs. Who can buy the most compute. Who controls the most advanced silicon. That story is over. The new story is about something older and far more stubborn: electricity, and the physical infrastructure required to move it.

The numbers are staggering. U.S. data center electricity demand now sits near 41 gigawatts — roughly the combined output of every nuclear power plant in the country. A single large AI training cluster can pull 100 megawatts, enough to light a small city. By 2030, data centers could consume between 9 and 17 percent of all electricity generated in the United States. That is not a sector growing. That is a sector colliding with the limits of the grid.

And here is the part the headlines miss: the constraint is not how much power exists. It is how fast you can connect to it. In the hottest markets — Northern Virginia, Phoenix, Dallas — the wait to plug a new facility into the grid runs three to seven years. A recent industry survey found that "time-to-power" has stretched roughly eighteen months to two years longer than developers expected even a year ago. In Northern Virginia and Dallas–Fort Worth, the amount of data center capacity under construction has actually declined — not for lack of demand or capital, but because there is no way to power what everyone wants to build.

Imagine the world's fastest restaurant kitchen — top chefs, unlimited ingredients, a line of customers around the block. Now imagine the building has a single extension cord running to it, and the power company says you can have a proper electrical hookup… in five years. That is the AI industry right now. The talent is there. The money is there. The demand is there. The wire is not. Everything else is a footnote to that one problem.

There are only two fundamental strategies, and almost every announcement you have read in the past year is a variation of one of them.

Strategy one: bring the power to the computer. Build the generation right next to the data center, "behind the meter," and skip the grid queue entirely. This is the logic behind the wave of self-build deals: Google's $4.75 billion acquisition of a power developer in late 2025; Meta's two-gigawatt single-site agreement where the company directly funds the new infrastructure; the rush toward on-site gas turbines, fuel cells, and waste-to-energy plants. The political winds now favor it — at a March 2026 White House meeting, the major cloud companies pledged to shoulder their own generation costs rather than push them onto ordinary ratepayers.

Strategy two: bring the computer to the power. If you cannot build power where the compute is, move the compute to where the power already is — stranded gas fields, underused hydro dams, and, in the most audacious version, orbit, where the sun never sets. This is Elon Musk's bet, and we will come back to it, because it is the most revealing move on the board.

So what? These two strategies are not really competing today — they run on different clocks. Bringing power to the computer is a business with revenue this decade. Bringing the computer to orbit is a bet on the 2030s with a trillion-dollar wall in front of it. The mistake is letting the flashier idea make the boring, bankable one look unambitious. The boring one is where the money is for the next ten years.
 

If you need power now, here is the honest menu — scored the way an operator actually ranks them: speed first, firmness (can it run 24/7?) second, cost third.

Natural gas is the option to beat — but it hides a brutal chokehold. Only three companies build large turbines at scale: America's GE Vernova, Germany's Siemens Energy, and Japan's Mitsubishi. They are sold out for years. And a large share of that manufacturing capacity is entangled in joint ventures inside China, built to serve a Chinese gas market the trade war has now severed from U.S. fuel. China has the factories; America has the gas; and the two are in open conflict. Layer on China's near-monopoly over the rare-earth metals turbine generators need, and the "safe, mature" option turns out to be squeezable at the factory, the fuel, and the raw material at once.

Small modular reactors (SMRs) are the most over-narrated option in the conversation. The cloud giants have signed more than ten gigawatts of nuclear deals — yet there is not a single commercial SMR operating in the United States today, and meaningful capacity is a 2030s story. Worse, most advanced reactors run on a special fuel called HALEU, and only Russia and China can produce it at scale. America builds the reactors; Russia makes the fuel they eat. The U.S. banned Russian uranium but has not yet learned to make the replacement at home.

Fuel cells are the quiet winner on speed. A solid-oxide fuel cell is neither an engine nor a battery — it converts fuel directly into electricity through a chemical reaction, with no combustion and no moving parts, which is why it is quiet and fast to deploy (months, not years). Crucially, it contains no large turbine forging, no enriched uranium, no scarce platinum — just engineered ceramic, with manufacturing led by the U.S., Japan, Korea, and Europe. It is the rare firm option that is both the fastest to build and the least geopolitically exposed.

Solar and storage is the cheapest energy and the fastest to permit — but it is not firm. It cannot keep a 24/7 cluster online alone. It is a layer that blends down cost, never the anchor.

Waste-to-energy is the option closest to my own work. Municipal and industrial waste is a fuel communities will pay you to take — a negative fuel cost. Stack the tipping fee, the power sold, and the materials recovered, and you have an economic model no single-source plant can match, with a permitting story that inverts the usual fight: you are the operation cleaning up the garbage, not the polluter the neighborhood resists.

So what? No single option wins on all three axes — speed, firmness, and cost. That is the whole point: the winning product is a stack, not a single source. Firm baseload to anchor it, solar to blend the cost, a gas peaker as insurance, fuel cells for speed, and nuclear waiting in the wings. And note the pattern the supply-chain story reveals — the two "serious" firm answers, gas now and nuclear later, both run through Beijing or Moscow. Firm power has quietly become a question of national economic sovereignty, not just engineering.
 

Solve the power, and you are still left with a physical building — and that building is becoming the second front in the war over data centers. Once a facility generates its own power on site, it stops being a quiet warehouse and becomes a small industrial plant that never sleeps.

Noise levels at some sites exceed 105 decibels — comparable to a passenger jet. Worse is the infrasound: low-frequency vibration below the threshold of hearing that does not register on a standard meter but is linked to headaches, nausea, and sleep disruption. It is now being pleaded in court as an environmental harm, residents are filing suit from Lowell to South Carolina, towns are rewriting zoning codes, and there is a proposed federal moratorium bill.

Here is what the industry keeps forgetting: none of this is hard to fix at the design stage. Quieter cooling, enclosed generators, earthen berms, smart building orientation, screening and setback — all mature, off-the-shelf practices that cost a fraction of a single lawsuit. The builders who win the next decade will make their facilities quiet, screened, and welcome — because the cheapest megawatt is the one the community never tries to litigate out of existence.

Where you put the thing is no longer a real-estate decision — it is an energy-infrastructure decision. One analysis found that only about 17 percent of substations near otherwise-buildable land have spare capacity today; the rest is effectively stranded — looks buildable on a map, worthless in practice, because the wire is already full.

The most valuable site in 2026 solves three things at once: a grid node with real headroom (the power can actually reach it), short logistics to the fuel or feedstock that powers it (a waste or gas plant has to sit near its inputs), and enough buffer to keep the community on your side. Miss any leg and the project dies. Whoever learns to read a map for all three is sourcing the real estate everyone else discovers too late.

Now to the most revealing move on the board. After merging his AI company with his rocket company, Elon Musk filed plans for up to one million orbital data center satellites — solar-powered computers in a sky where the sun never sets. No real estate, no neighbors, no noise complaints, no zoning board, no interconnection queue. Read the sections above and you can see exactly what he is trying to escape.

The reality check is sobering: deploying a million satellites would cost more than a trillion dollars, rejecting heat in vacuum is brutally hard, and it is not even established that AI training can be split across satellites. As of just days ago, reporting notes the terrestrial-solar bet has been quietly abandoned in favor of the space narrative.

So what? Here is the takeaway most people get backwards. Elon's leap to orbit is not proof that Earth-based power failed. It is the loudest possible confirmation of how scarce firm power has become — when the richest builder on the planet would rather solve heat rejection in a vacuum than wait in a utility queue, the queue is the real story. That scarcity is precisely what makes down-to-earth, firm, fast, sovereign power the best business of the decade. Elon is solving 2035. The smart money is solving 2027 — with revenue, a permit, and a wire in the ground.
 

Strip away the chip wars, the trillion-dollar valuations, and the satellites, and the AI revolution comes down to a deeply unglamorous set of questions: Can you get firm electrons to the machine, on time? Can you put the machine where the wire can reach it and the fuel can find it? And can you do it without the neighbors driving you into court?

Solar cannot be firm. Nuclear cannot be soon — and when it arrives, its fuel runs through Russia. Space cannot be this decade. And even gas rests on a turbine supply chain partly entangled in China. What remains is a disciplined portfolio of the available, the buildable, and the bankable — engineered as a stack, sited as a triangle, and built as a neighbor.

The next phase of artificial intelligence will not be decided in a lab. It will be decided at the substation, the permit office, the loading dock — and the property line. That is the wall. The builders who learn to climb it own what comes next.

Download the full analysis — the complete 12-page report, including the supply-chain sovereignty breakdown, the fuel-cell deep dive (how a solid-oxide cell actually works and who ships them), and the full scorecard — is available as a PDF above. ↑

Dr. Gregory S. Carmichael is a retired U.S. Air Force Lieutenant Colonel, former U.S. Representative to the G20 Summit in India (2023), and DPA Title III Advisor (2024). He is CEO of Quantum Reserve Capital and writes on the intersection of cryptographic infrastructure, energy, and national economic sovereignty at CryptoSoWhat.

© 2026 CryptoSoWhat. All rights reserved.