Does Not Compute
On December 2, 1999, Enron announced the first forward trade of bandwidth, a monthly contract for a line between New York and Los Angeles. Its executives were projecting a $500 billion market by 2005. Bandwidth was going to trade "like pork bellies."
Two years later, to the day, Enron filed for bankruptcy. Global Crossing, which sold Enron that first contract, followed eight weeks later, and it came out that it had been swapping capacity with Qwest so both could book the revenue.
Now fast forward. CME has filed to launch futures on GPU rental prices and is waiting on the CFTC, whose chairman says "America cannot win the AI race without a robust derivatives market for compute." Kalshi's CEO says compute derivatives could be 10-20x a spot market he expects to hit $7-10 trillion by 2030. And we've all seen the headlines on the dozens of startups building compute exchanges, indices, and brokers over the last year or so.
Meanwhile the neoclouds that would be the natural players in those markets are raising and borrowing tens of billions of dollars, and everyone in the AI infrastructure world is booking circular revenue on deals they’re doing with each other and investments they’re making into each other.
We all remember the internet bubble bursting, we remember the names of the overlevered companies that went down, we remember the frauds. But I bet almost no one remembers our brief flirtation with bandwidth trading. Because it died a quiet and unremarkable death, along with nearly every technology exchange that has tried something similar.
Telecom ended up consolidating into a handful of large carriers that owned their own networks and set their own prices, and it turns out a handful of giants doesn't need a market.
I’m a student of the history of technology and financial markets. And the more I learn, the more I see that everything old is new again. The neocloud market is frothy, highly levered, and benefiting from current scarcity, but history has been pretty consistent in that mega infrastructure buildouts tend to end in consolidation.
And the compute markets and financial products that people want to create around this buildout face the same risks that every capacity market has faced - if (when?) the industry consolidates, the market becomes moot.
First, scarcity
Every tech wave involves a sudden ramp in demand, while supply of some scarce resource lags and becomes a gold rush. In AI right now, the scarcest resource is data centers that are already plugged into power. In a world where building things is nearly illegal, and the power interconnection queue takes over five years, anyone with existing energized capacity can rack Nvidia GPUs and rent them by the hour to labs and hyperscalers who can't build fast enough.
In the rush to build more and more, everyone needs to borrow. This year alone, hyperscalers, data center developers, and neoclouds have raised $346B of debt and equity, double 2025. CoreWeave carries about $35B of debt, Nebius has raised about $15B, IREN about $8B, and former bitcoin miners like TeraWulf, Cipher and Applied Digital another $18B between them. And this all largely rests on incestuous circular deals between the hyperscalers, neolabs, and their various suppliers.
We've seen this before. By 1889, railroad bonds made up a huge amount of the securities quoted in the country's leading financial paper, many of them handed to British ironmakers as payment for the rails and to insiders' own construction companies like Crédit Mobilier. Telecom was built on junk bonds and $25.6B of vendor financing from equipment makers like Lucent, Nortel and Cisco.
And in modern times, when people rush to fill the capital gaps, the exuberance brings talk of financialization - a new asset class, an index to price it, an exchange to hedge it. Telecom had Enron, Band-X and RateXchange. DRAM had six attempts at futures between 1989 and 2004. Cloud had a Deutsche Börse exchange in 2015. Today it's futures, indices, and exchanges for H100s.
Then, frenzy
What gets missed in the exuberance is that the more leverage goes into a buildout, the more tenuous it gets. CoreWeave holds about 16 cents of cash for every dollar of debt, and Cipher about 12. The loans are backed by contracts, the contracts by labs, and the labs by their next fundraise. Trace it far enough and much of the debt ends up relying on OpenAI and Anthropic, whose compute commitments top $1.2 trillion. The whole thing is a one-way bet that demand keeps growing at the same pace and the scarcest resources stay that way indefinitely.
The more leverage piles on, the less it takes to upset the equilibrium. What if demand slows for any reason, or supply catches up and inference races toward commodity pricing? What if premium GPU pricing only lasts about one chip generation, despite data center leases with 10+ year terms? What if regulation kills the buildout, and there’s no new data centers to make those loan payments?
Bubbles don’t slowly bleed out, they implode. Lenders make collateral calls, bids for everything disappear, and everyone is forced to sell the same assets at the same time, into a vacuum.
This has played out time and time again. Railroads in the 1880s and fiber in the 1990s were both overbuilt, prices collapsed while the debt payments didn't, and the defaults followed. By 1894 a quarter of US railroad capital was in receivership, and a century later Global Crossing sold for a penny on the dollar.
This isn’t happening here yet, but CoreWeave's 2031 bonds, sold in April at under 10%, recently crossed 16%, investors have been pulling record amounts out of the private credit funds that were supposed to buy all this AI-related debt, and supposedly hedge funds are starting to dump their data center loans. Credit markets tend to be the canary in the coalmine, and they’re getting antsy.
Next, consolidation
After an implosion, the few with the strongest balance sheets get to buy the wreckage cheap, so the next step is consolidation. Harriman and his partners bought the Union Pacific at foreclosure, and the railroads came out as a few big systems. Telecom rolled up into a handful of carriers. DRAM went from about 20 makers to three that control nearly 90% of the market.
This time the strong hands are the hyperscalers, which borrow at around 5% while neoclouds pay up around 10%. Banyan counts roughly seven buyers setting demand against more than 30 leveraged neoclouds and ex-miners, and Gartner expects more than 80% of neoclouds to lose their independence by 2030. CoreWeave has already tried to buy Core Scientific for $9B. Whoever is left standing gets the land, the data centers, the power and the chips.
The credit markets already seem to agree. On October 30, five-year default protection stood near 60 basis points for Microsoft and 90 for Alphabet. By November 5, during the post-midterm market rout, CoreWeave’s spread had reached a record 1,268 basis points.
Stocks tell the same story. CRWV, IREN, and NBIS are down anywhere from 60-80% off their highs, and the midterms didn’t help. The day after voters restricted data centers across the country, the hyperscalers rallied, the neoclouds fell again, and CoreWeave's bonds traded as low as ~80 cents on the dollar. Stocks and bonds are pricing the same ending - whoever has the balance sheet wins.
Finally, equilibrium
Ironically, an unwind is exactly when everyone needs the hedge. Lenders want to protect their collateral, neoclouds want to lock in rental prices, and buyers want to know what they'll pay next year. The downside event is exactly when a compute futures market would be most useful.
But a liquid market needs a standardized product and a large, fragmented crowd of buyers and sellers who disagree on price and need to hedge in both directions, all day, every day. Treasury futures, oil, energy futures all have that. An unwind puts everyone on the same side, and then consolidation wipes out the playing field. When all you have left is a few hyperscalers, a few neoclouds, and a couple of labs, there’s no more market, and you can do all the deals with handshakes.
History says the same thing. The railroads never got a market for freight capacity. Rates were set by pools and then, after 1906, by the ICC. The bandwidth exchanges died alongside their sellers. The DRAM futures never traded at all. The first real hedge for bitcoin miners launched about ten weeks before Core Scientific went down.
History rhymes
The fun thing about the world, especially in a time of big change, is that nobody knows anything. I'm wrong as often as I'm right, and there’s a lot of people smarter than me building in this space, and backing those who are. Plenty of things could happen to change how the story plays out.
These days, people love trading stuff. Crypto exchanges list H100 perps, and Kalshi has a compute forward curve constructed from event contracts. So it’s very possible the long term state is trading compute through crypto or prediction markets, if nothing else.
It’s also very possible that demand for intelligence is infinite and we can never conceivably build enough supply to meet it. So that there’s always a scarcity dynamic influencing pricing volatility, there are always arbitrage opportunities in the market to entice new suppliers, and then there’s no need or opportunity to consolidate those suppliers.
On the buyer side, it’s possible that inference spreads to thousands of enterprises and sovereigns, and open-weight models run everywhere, so that seven buyers today end up as hundreds in the future, and hyperscalers aren’t the ones meeting all that need.
But as billions get poured into more and more neoclouds, and as people rush to financialize the assets and liabilities of AI, we should always at least heed the lessons of the past, so we don’t repeat their mistakes in the future. The railroads were real. The internet was real. Cloud was real. AI is real. That was never the question. The question is who ends up owning it.
Scarcity and volatility make people want a market. Consolidation makes them stop needing one. And history favors consolidation.
