Nvidia generated $89 billion in Data Center revenue in a single quarter, while Meta expects to spend $135 billion–$145 billion on capital investments this year. The infrastructure behind AI has become an enormous business.
Arthur Hayes wants to create a new way to trade one of its essential inputs: computing power.
In his October 26 essay, “The Trifecta,” the BitMEX co-founder outlined a plan for trading contracts tied to the cost of AI computing. The proposal builds on Flop Network, his project for connecting AI agents with providers of computing power.
The aim is to give buyers and suppliers a way to plan around future prices. The proposed market is still being developed.
The Money Behind the Idea
Recent company disclosures show the scale of the infrastructure buildout:
| Company | Figure | What it measures |
|---|---|---|
| Nvidia | $89 billion, up 117% year over year | Data Center revenue for the quarter ended July 26, 2026. |
| Microsoft | $41 billion | Capital expenditures in the quarter ended June 30, 2026; roughly two-thirds went toward shorter-lived assets, primarily CPUs and GPUs. |
| Meta | $135 billion–$145 billion | Full-year 2026 capital expenditure guidance, including principal payments on finance leases. |
These measure different parts of the business, cover different periods and include spending beyond AI inference; they should not be added together as Flop’s potential market size.
They do explain why pricing matters. Buyers want to know what it will cost to run their services. Suppliers need enough revenue from their equipment to justify buying it.
What Would People Actually Buy?
Flop’s focus is inference: running an existing AI model to produce an answer, rather than training a new model from scratch.
Under its proposed design, an AI agent requests work, a provider runs the model, and the network checks evidence that the requested work was performed. Payment is settled in FLOP, the network’s planned currency.
That distinction matters. A buyer needs the right model, a usable response time and confidence that the provider delivered what it promised. A cheap quote alone does not establish those things.
Flop Labs described its ambition for contracts that could either settle financially or deliver computing power in an October 22 post:
Post from @flop_labsView the original post on X↗
Making the AI Bill Easier to Plan
For a smaller business, the bill is easier to picture. Lambda’s published pricing, checked October 27, lists an eight-GPU Nvidia H100 SXM instance at $3.99 per GPU-hour.
Running all eight continuously for 30 days would cost approximately $23,000 before tax. A 10% change in that hourly rate would change the monthly bill by roughly $2,300. This is an illustration using one provider’s rental price, not a Flop quote or a market-wide average.
A reservation could give the buyer and supplier an agreed price ahead of time. For a business selling an AI service at a fixed monthly fee, more predictable computing costs could make it easier to protect its profit margin.
Prices can fall as well as rise. Epoch AI’s September research estimates that the cost of achieving a fixed level of AI performance fell about 47% per quarter since 2023. That measures the cost of producing results at a given quality, rather than GPU rental rates.
For customers, cheaper AI creates opportunities to do more. For suppliers, it raises the pressure to keep equipment busy and costs competitive. Agreeing to future prices could help them plan, although either side can miss out if the market later moves in its favor.
Flop’s draft whitepaper describes reservations covering future windows of up to 30 days, with suppliers posting collateral against their commitments. It marks that functionality as planned after the core protocol.
Hayes’s essay takes the idea further, proposing futures and perpetual contracts linked to a benchmark drawn from network transactions. Futures settle at a specified time; perpetual contracts have no expiry. They would let participants take positions on computing prices without necessarily buying computing power for their own use.
His starting argument is simple:
“There is no global price for compute.”
— Arthur Hayes, “The Trifecta”
Whether Flop could supply a widely accepted benchmark remains an open question. For price protection to work, the benchmark would need to track the costs businesses actually face, with enough trading activity for buyers and sellers to enter and exit positions.
Where the Project Stands
Flop’s official status page says its development network is running, but there is no live, tradable FLOP token. Its working targets are November 3 for the testnet and the first quarter of 2027 for mainnet and token issuance. Those are targets, not confirmed launches.
On October 26, Hayes said the development network would open to wider participation in the coming days. He also said participation at that stage would not count toward a token allocation.
Post from @CryptoHayesView the original post on X↗
A Practical Connection Between AI and Crypto
As Blockster explored in its coverage of AI spending and business results, computing costs help determine whether an AI service can turn customer demand into profit.
That makes Hayes’s proposal interesting beyond crypto trading. Our view: clearer pricing and dependable reservations could make it easier for businesses to budget for useful AI services. Flop still has to prove it can deliver that experience, but the problem it is trying to solve is one businesses can readily understand.
Reporting by Lidia Yadlos
