Trillion-Dollar Compute Trading Still Matched Over Text and Phone Calls
Compute contracts hit $10 billion per deal, yet the market lacks unified pricing and hedging tools—buyers and sellers still negotiate privately through relationships, phone calls, and text messages.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Large compute contracts cannot be hedged
GPU rental and purchase is the largest cost in AI economics, with single contracts from $100 million to $10 billion. Neocloud and hyperscaler operators absorb all price risk—losing money when prices drop, missing gains when they rise—with no instruments to hedge. This is why building a compute derivatives exchange is essential: establish price transparency first, then hedging and speculation become possible.
— Brett HarrisonFrom three vendors to five hundred: compute has become a commodity
The compute market once had only hyperscalers like Google and AWS. Neocloud specialists like FluidStack and Lambda appeared next. Now ComputeDesk tracks approximately 500 GPU service providers. Andrawes' view: anyone with power, a data center, and GPU access can enter. This low barrier plus extreme demand is the signature of commodity markets—and commodities lose pricing power. Hyperscaler GPU margins are already visibly narrowing in financial statements.
— Andrawes BahouPhantom capacity: contracts exist before hardware does
Data centers borrow money to build clusters by having lenders require customers to sign multiyear offtake agreements—commitments to buy all or part of not-yet-existing capacity. Only then do lenders fund hardware from Dell, Supermicro, and NVIDIA, with delivery taking 4 to 16 weeks. Andrawes calls this phantom capacity: this vaporware gets resold through layers of subcontracting, forcing buyers to spend enormous effort distinguishing real production from air.
— Andrawes Bahou10% daily growth makes every capacity decision potentially ruinous
Take Instinct, a company growing 10% daily. Buy only what you need today and you'll run out in a week. But compute contracts lock in 3 to 5 years—buy extra and you're stuck with unused capacity for years. Founders face a brutal choice between price and speed. To secure current capacity, they'll pay up to 2x list price, because the revenue opportunity cost of being short far exceeds the unit economics.
— Andrawes BahouThe same chip costs wildly different prices
Compute pricing doesn't converge globally like WTI crude. The same H100 or H200 can have vastly different quotes. Reason: it's peer-to-peer negotiation with no unified reference price. ComputeDesk's solution: sign exclusive agreements with large neocloud operators, access their actual transaction invoices, aggregate them into a price index published on Bloomberg—replicating the price-reporting-agency model from crude oil exactly.
— Andrawes BahouNon-investment-grade companies can only access subprime compute lending
Compute lending is hard: most borrowers lack years of credit history and can't forecast GPU value five years out. Below-investment-grade companies face interest rates of 20-25% or can't borrow at all. A subprime compute lending market has emerged—a comparison that troubles Andrawes, who remembers the subprime crisis from twenty years ago. Lenders meanwhile experiment with contract insurance and derivatives hedges to manage risk and lend more.
— Andrawes BahouHyperscaler cash flows turning negative trigger compute financialization
Lenders once underwrote on creditworthiness—Google earns from ads, pays GPU loans reliably. A widely-circulated chart now shows hyperscaler free cash flow crushed toward zero or negative by AI capex. This is the real trigger for compute financialization becoming mainstream: capital markets must shift from "trust Google's credit" to "price compute as an asset, hedge it, settle it." This is risk no single hyperscaler can absorb alone.
— Andrawes BahouPhysical compute trading has reached half the scale of crude oil
Few people track this comparison: global crude oil physical trade is roughly $2.8 to $3 trillion. Compute's physical trade has already reached one to two trillion dollars—approaching half of crude's scale. Crude's derivatives market is 40x the physical market. If compute follows the same path, the derivatives market could be one to two orders of magnitude larger than today—before even accounting for the tenfold growth in compute volumes forecast ahead.
— Andrawes BahouIn their own words · checked verbatim
People are, you know, spending anywhere from 100 million to billion to 10 billion at a time on large-scale compute contracts.
Brett Harrison2:12
Whenever you have a very large number of participants in the market, it is usually the hallmark of something that is a commodity.
Andrawes Bahou7:33
There's a complete bimodal distribution of prices between what the hyperscalers offer for compute versus what the neoclouds offer for compute.
Andrawes Bahou9:38
There's a big problem right now. I call it phantom capacity.
Andrawes Bahou15:49
Like, it's not just a vehicle for speculation or for kind of gambling or betting on the prices of things. It's a way to transfer risk and to deal with things like duration mismatch where, like, you know, I have a particular need now and it might be for what my computes can be over the next year.
Brett Harrison23:29
I think another watershed moment for me was NVIDIA coming out and announcing their 25% backstop on these contracts.
Brett Harrison51:44
This is kind of like if you have a car, this is like the handbrake. What's funny is if you have a car and you only have the gas pedal and the handbrake, you're not going to go very fast. If you have a little brake, you can go way faster because you can stop pretty quickly. This is what a hedge provides.
Andrawes Bahou57:10
The physical trading of crude oil is something like a $2.8 trillion market, maybe $3 trillion market. The physical trading of compute is 1 point something trillion. We're basically halfway there.
Andrawes Bahou1:03:20
Figures
| Single compute contract size | $100 million to $10 billion | 2:12 |
| GPU service providers tracked | Approximately 500 | 8:36 |
| Compute cluster delivery lead time | 4 to 16 weeks | 16:52 |
| Premium for expedited access | Up to 2x list price | 20:13 |
| NVIDIA minimum price guarantee | 25% | 51:44 |
| Crude oil physical trading market | $2.8 to $3 trillion | 1:03:20 |
| Compute physical trading market | $1 to $2 trillion | 1:03:20 |
| Crude oil derivatives-to-physical ratio | 40x | 1:04:23 |
Glossary
- Offtake agreement
- A long-term contract where customers commit to purchase all or part of a data center's GPU capacity for multiple years, typically required by lenders before funding hardware acquisition.
- Phantom capacity
- GPU production capacity marketed and resold before it actually exists—promised in contracts but not yet built or available.
- Neocloud
- Emerging GPU cloud providers that sell raw compute capacity without the full stack of cloud services hyperscalers provide.
- DCM (Designated Contract Market)
- A CFTC-issued license permitting legal operation of a futures and options exchange.
- Node
- A unit of compute capacity consisting of eight GPUs clustered together for measurement and billing.
How to listen
Founders and investors tracking AI infrastructure financing, compute procurement, and commodity derivatives—plus neocloud and data center operators.
Skip the Token 2049 ad at the start and the first few minutes of terminology definitions.