After my last article on the $500 billion AI infrastructure buildout, I received several comments and private messages raising some very interesting questions.
I thought my next article might address a couple of them.
Then I came across something even more intriguing:
The financialization of AI compute.
Stay with me, because this gets interesting very quickly.
Imagine a world where AI compute itself becomes something that can be bought, sold, financed and eventually traded — not entirely unlike other commodities.
Oil. Natural gas. Even pork bellies.
Compute.
That sounds strange at first. But pieces of this market are already beginning to emerge.
First, securitize the infrastructure
Think back to the $500 billion AI infrastructure financing initiative I wrote about previously.
Enormous amounts of private capital could be raised to build data centers filled with GPUs, networking equipment, cooling systems and power infrastructure.
Traditionally, we might think of all of this simply as technology CapEx.
But what if we instead treat it as a productive infrastructure asset?
The data center generates cash flow by selling compute capacity. The physical infrastructure and GPUs have residual value. Those assets and their future cash flows can potentially support debt — and that debt can ultimately be packaged into securities that investors can own and trade.
Suddenly, some of that enormous pool of relatively illiquid private infrastructure investment can potentially become a financial asset.
That alone is interesting.
But take it one step further
What if we don't just securitize the data center?
What if we create a market for the compute itself?
If a GPU cluster can provide a defined amount of computational capacity over a defined period, perhaps that capacity can eventually be priced, contracted and traded.
And that is where the commodity analogy starts becoming much more interesting.
A company expecting to need enormous amounts of compute six months from now might want to lock in its price today.
A data-center operator with future capacity might want to lock in a buyer.
Investors might want exposure to rising or falling compute prices without owning a data center at all.
We suddenly aren't very far conceptually from a futures market.
But here's where I get skeptical
Oil doesn't become obsolete because someone releases Oil 2.0 next year.
GPUs do.
A data center built around today's leading chips may be enormously valuable today. But what is the economic life of those chips?
Three years? Five years? Longer?
What happens when a new generation delivers dramatically better performance per watt?
What happens to the collateral value of the previous generation?
The building may still be there. The power connection may still be incredibly valuable. The cooling and networking infrastructure may still have value.
But the technology inside it can depreciate very differently from a traditional infrastructure asset.
And that creates an interesting problem for anyone trying to turn AI infrastructure into a long-duration financial asset.
And then there is my original question
Even if we solve all of that, will the customers come?
Financial engineering can change who provides the capital.
Securitization can create liquidity.
Futures can transfer price risk.
None of them creates the underlying demand for compute.
Ultimately, somebody still has to buy all those GPU hours.
Which brings me back to the question from my previous article:
Can enterprises turn AI into enough real economic value to justify the extraordinary amount of infrastructure we're building to support it?
I find the idea of compute becoming an asset class fascinating.
But perhaps the most interesting thing about creating a market for AI compute is that eventually the market itself may tell us what all of this infrastructure is really worth.