Article on AI compute
Why compute is the new oil
People keep calling compute the new oil. Here is what that means, where the comparison is useful, and where it falls apart, in plain language with real data.
Key takeaways
- Compute has become a strategic resource that shapes who can build and run advanced AI.
- Like oil, it is scarce, concentrated, and costly to produce, which gives it strategic weight.
- Unlike oil, compute is not consumed once. The same hardware can serve many workloads over time.
- The analogy is useful for understanding scarcity, but ownership and operations work differently.
What people mean by the new oil
Over the last few years, compute has become the resource that decides what is possible in AI. Train a larger model, serve more users, run more agents: each one comes down to access to compute. When a single input gates so much of an industry, people reach for the oil comparison, because oil once played a similar role for the industrial economy.
The comparison captures something real. Compute is scarce, expensive to produce, concentrated in a few hands, and strategically important. Those are the same traits that made oil a source of leverage and competition throughout the twentieth century.
The phrase is also a way of saying that compute has moved from a back-office cost to a front-line strategic asset. Boards, governments, and investors now talk about access to it the way they once talked about energy security.
Where the oil analogy holds up
Scarcity
Advanced compute cannot be produced quickly. It depends on complex supply chains, much like oil depends on hard-to-reach reserves and refining capacity.
Concentration
Both resources cluster in a few regions and a few large players, which gives those holders outsized influence over access and pricing.
Strategic value
Whoever controls more of the resource can do more with it. For oil it was industry and transport. For compute it is AI capability.
The scarcity in numbers
4 to 5x
Annual growth in training compute for frontier AI models since 2010, according to Epoch AI.
Source: Epoch AI, May 2024
415 TWh
Electricity used by data centres worldwide in 2024, the physical fuel behind compute, according to the IEA.
Source: International Energy Agency (IEA), April 2025
Where the analogy falls apart
The comparison is not perfect, and the differences matter. Oil is consumed once. You burn a barrel and it is gone. Compute hardware is durable. The same GPU can run many different workloads over its lifetime, and its value comes from being used repeatedly rather than spent in a single act.
There is also the question of obsolescence. Oil in the ground does not get worse over time, but a GPU competes with newer, faster hardware as it ages. That is why holding compute is less about hoarding a commodity and more about keeping capable hardware productively at work while it is current.
A third difference is substitutability. Oil is a fairly uniform commodity, but compute comes in generations and architectures that are not interchangeable. The newest hardware can do things older hardware cannot, which keeps the market moving rather than static.
A shift in what powers the economy
The deeper reason the analogy resonates is that it marks a shift in what an economy runs on. Where industrial growth once tracked access to energy and oil, a growing share of economic activity now tracks access to compute. That is the real claim behind the phrase, even where the details differ.
What the framing means for access
Treating compute as a strategic resource helps explain why access is contested and why early, organized buyers tend to win. It also reframes the question for everyone else: not how to win a supply race, but how to hold a useful position in a resource that underpins a growing industry.
Because compute gains its value through use, the way it is operated matters as much as the fact of owning it. Idle hardware is just a depreciating asset, while well-run hardware is productive capacity. That distinction is where the oil analogy quietly breaks down and where operations become central.
The framing also carries a warning. Calling compute the new oil can make it sound like a guaranteed store of value, the way people once spoke of land or precious metals. It is not. Hardware ages, demand can shift, and the resource only pays off when it is kept working. The analogy is a lens for understanding scarcity and strategic weight, not a promise about returns.
If not oil, then what
Because the oil comparison breaks down on durability and obsolescence, some people reach for other analogies. Compute has been likened to electricity, to real estate, and even to industrial machinery. Each captures part of the truth. Compute is an enabling input like electricity, a located asset like real estate, and a depreciating productive tool like machinery.
The machinery comparison may fit best. Like a factory machine, a GPU is valuable only while it runs useful work, it ages against newer models, and it needs skilled operation and maintenance to stay productive. That framing keeps the focus on use and operations rather than on hoarding a commodity.
No single analogy is perfect, and that is the real lesson. Compute is a new kind of strategic resource with its own rules. Borrowing language from oil or machinery helps explain the scarcity and the stakes, but the smartest way to think about it is on its own terms.
Holding a position in a strategic resource
If compute is the resource that matters, one straightforward response is to own the hardware that produces it, while a professional team keeps it operating. The managed ownership model is built on exactly that split: you own the physical GPU hardware, and an operations team handles hosting, cooling, monitoring, and connecting it to demand.
This responds directly to the way compute differs from oil. Because the asset has to be kept productive rather than simply held, the operations layer is not optional. Pairing ownership of a durable, depreciating asset with professional operation is an attempt to capture the strategic value the analogy points to while respecting the ways it does not hold.
Our service on managed GPU compute explains how that works in practice. Owning hardware does not guarantee any result, and hardware does age against newer chips. Operational benefits are not guaranteed and depend on utilization, uptime, demand, costs, hardware performance, and market conditions.
References and data
- Training compute of frontier AI models grows by 4 to 5x per year. Epoch AI. May 2024.
- Energy and AI. International Energy Agency (IEA). April 2025.
Common questions about compute as a resource
Because access to compute now decides what is possible in AI, the way oil once gated the industrial economy. Compute is scarce, costly to produce, concentrated, and strategically important, which gives it similar leverage.
Oil is consumed once, while compute hardware is durable and can run many workloads over time. Hardware also ages against newer chips, so holding compute is about keeping capable hardware in productive use, not hoarding a commodity.
Demand has outrun supply for years. Epoch AI reports training compute for frontier models has grown 4 to 5 times per year since 2010, and the IEA tracks the electricity behind it rising sharply, both signs of a tightly constrained resource.
Unlike oil, compute comes in generations and architectures that are not interchangeable. The newest hardware can do things older hardware cannot, so the market keeps moving rather than trading a uniform good.
It reframes the goal from winning a supply race to holding a useful position in a resource that underpins a growing industry. Because compute gains value through use, how the hardware is operated matters as much as owning it.
Want to hold a position in the resource behind AI?
Talk through what owning managed NVIDIA GPU hardware would look like, with no pressure and straight answers.
Operational benefits are not guaranteed and depend on utilization, uptime, demand, costs, hardware performance, and market conditions.