"Compute is the new oil" has become one of the AI era's favorite lines. It shows up in investor pitches, conference keynotes and, increasingly, in paid newsletter advertising. One campaign has even given AI computing power a commodity-style name, "computium," and built a stock-picking subscription around it.
The comparison has something real in it. Computing power is becoming a strategic resource that nations compete for, companies stockpile capacity of, and governments write policy around. But the analogy breaks down in a way that matters to anyone who works in energy. Oil is an energy source. Compute is a product made from energy. If anything in the AI economy deserves to be called the new oil, it's the electricity, and the grid, generation and equipment that deliver it.
Where the analogy works
Start with the scale. In its widely cited analysis, "The cost of compute: A $7 trillion race to scale data centers," McKinsey estimates the world will need about $6.7 trillion in data center capital spending by 2030, about $5.2 trillion of it for AI workloads. In McKinsey's base case, global data center capacity demand nearly triples by 2030, with roughly 125 GW of new AI capacity added between 2025 and 2030.
Governments are treating it as strategic infrastructure too. In July 2025, Executive Order 14318, "Accelerating Federal Permitting of Data Center Infrastructure," directed federal agencies to fast-track data center projects that add more than 100 MW of load or involve at least $500 million in capital spending. It opened federal land to qualifying projects and directed the Commerce Department to line up loans, grants, tax incentives and offtake agreements.
Its implementation in 2026 has been steady. The Department of Energy has solicited proposals for data centers at Idaho National Laboratory, Oak Ridge and Savannah River. The Air Force has offered about 3,100 acres across five bases. Federal regulators have issued guidance to speed environmental reviews, and the first data center project received fast-track FAST-41 status in April 2026.
Large capital flows, geopolitical competition and government backing for supply are all features of the oil economy. On those points, the comparison holds.
Where it breaks down
Look at what Executive Order 14318 actually covers. Its list of "covered components" reads like an energy-industry inventory: transmission lines, substations, switchyards, transformers, switchgear, gas pipelines, gas turbines, nuclear, geothermal and other firm power, and backup generation. A policy meant to speed up AI infrastructure is mostly a policy to speed up energy infrastructure.
McKinsey's numbers point the same way. Of the $5.2 trillion in AI-related spending, it assigns about 25%, roughly $1.3 trillion, to what it calls "energizers," the companies supplying power generation, electrical equipment, cooling and grid connections. That's before counting utility transmission upgrades.
Where McKinsey expects AI data center spending to go by 2030
Source: McKinsey & Company, The cost of compute (2025). Figures rounded.
The deeper difference is physical. Oil can be pumped, shipped, stored in tanks and sold on a futures exchange. Compute can't be stored at all. It exists only while electricity flows through a chip, which makes it more like a power plant's output than a barrel of crude. As I explored in From Data Centers to AI Factories, an AI data center is essentially a machine that converts megawatts into model output. If the power isn't there, there's no compute, no matter how many GPUs are sitting in the warehouse.
Compute is electricity in another form
The math is simple. Each generation of AI chips does more work per watt, but draws more power per chip. With cooling and facility overhead at a PUE of 1.3, here's what running AI hardware around the clock looks like at an industrial power price of 8 cents per kWh:
| Hardware | Continuous draw incl. cooling | Energy per year | Power cost per year |
|---|---|---|---|
| One H100-class GPU (about 700 W) | 0.91 kW | about 7,970 kWh | about $640 |
| One Blackwell-class GPU (about 1,000 W) | 1.3 kW | about 11,400 kWh | about $910 |
| 100,000 Blackwell-class GPUs | 130 MW | about 1.14 billion kWh | about $91 million |
That last row uses about as much electricity in a year as roughly 107,000 average U.S. homes. You can run your own numbers in the GPU & Compute Power Load Calculator.
This is why the real bottleneck in the AI economy so often turns out to be electrical. The chips can be bought. Powering them is the hard part.
The bottleneck is electrical
The numbers on the grid side are moving fast. In its October 2026 outlook, the U.S. Energy Information Administration projected record U.S. electricity use in both 2026 and 2027, driven largely by data centers dedicated to AI and cryptocurrency. It expects wholesale power prices to average about $52/MWh in 2026, 11% above 2025, with prices in the PJM region up about 41%.
Behind those numbers are the constraints that make up the AI power wall: transformers with multi-year lead times, long interconnection queues, capacity shortfalls and a growing political fight over who pays for new generation, as I covered in Who Should Pay for the AI Power Boom and in the debate now playing out in North Carolina. State legislatures filed more than 300 data center bills in 2026 alone.
Compute can only grow as fast as the power system behind it. That's the clearest sign of which one is the real commodity.
The Jevons problem applies here too
The oil comparison holds in one more uncomfortable way. When engines got more efficient, the world didn't use less oil. It used more, because cheaper energy found new uses. Economists call this the Jevons paradox, and it's playing out in AI. Chips keep getting more efficient per calculation, but falling costs push AI into more products and more queries, so total electricity demand keeps rising.
That's why the more useful metric isn't raw compute but intelligence per watt: how much useful work comes out of each unit of electricity. A commodity framing that only counts how much compute gets produced misses the point. In a power-constrained world, the winners won't be the ones who consume the most megawatts, but the ones who get the most value from each one.
A note on the "computium" pitch
The term "computium" appears to come from paid advertorials by Weiss Ratings promoting an investment subscription and a proprietary "Computium Daily Surge-Index." By the promoters' own description, computium is a branded name for AI computing power, not a tradable asset. There's no exchange, futures contract or published price for it, and Executive Order 14318 doesn't use the word or create any national compute reserve.
None of that makes AI infrastructure a bad investment theme. The build-out is real and enormous. But readers should be clear about what they'd be buying: individual stocks and funds exposed to chips, power, data centers and materials, chosen by a paid service. Paid investment promotions deserve the same scrutiny as any other sales pitch. This article is not investment advice.
Where the value actually sits
For anyone trying to understand the AI infrastructure economy, the physical constraints, and much of the durable value, sit on the energy side:
- Generation: gas, nuclear, geothermal, renewables and the firm capacity needed to serve large loads around the clock.
- Grid equipment: transformers, switchgear, substations and transmission, many with lead times measured in years.
- Battery storage: peak management, backup, renewable firming and the flexibility that can get a data center connected sooner.
- Energized sites: land with a substation and interconnection already in place, which is why Bitcoin mining campuses are being repurposed for AI, as I covered in From Hashrate to Megawatts.
- Efficiency: cooling, power electronics and software that raise the useful output of every megawatt.
What this means for developers and investors
If you're evaluating an AI infrastructure project, start with the electricity, not the chips:
- Where is the power coming from, and is it contracted or still in a queue?
- What will the site pay under today's tariff, and under a tariff that adds generation costs?
- What PUE will the facility actually achieve, and how does that change total grid draw?
- How exposed is the project to rising capacity prices in markets like PJM?
Then model the full power cost and total cost of ownership over the life of the hardware. The projects that still work under conservative power assumptions are the ones that will get built.
Conclusion: follow the electrons
"Compute is the new oil" makes a good headline but a weak model. Oil was valuable because it was energy. Compute is valuable because it's something energy can be turned into. The AI race is ultimately a race to produce, deliver and use electricity at a scale and speed the grid has never had to handle.
Before buying into any new commodity story, check where the electricity is coming from. That's where the real constraint, and much of the real value, will be.
Chris Kalowes
Founder, WattThe?!
Energy intelligence. Simplified.
