Arthur Hayes Strikes for a "Three-Peat": Building the AI Computing Power Derivatives Market with Flop Network

Core Viewpoint

Problem

There is no global price for computing power. Various indices attempt to track the sales and rental prices of certain types of CPU, GPU, and TPU chips, but the datasets are incomplete. Additionally, there are various indices that try to track the prices of AI tokens model by model. Chip prices and token prices are both derivatives of computing power. If I know the price of computing power (FLOPS), I can derive the value of chips and model tokens. Instead of focusing on collecting imperfect information about chip and token prices, we should focus on obtaining information about computing power prices.

Solution: Flop Network Computing Power Market

Every second, Flop Network produces a block. Each block contains receipts proving that miners have completed a reasoning task for a certain agent. For each reasoning task, we know three things that help calculate the $FLOP/computing power exchange rate.

Block information objects:

Delay in seconds ------ Time taken to complete the reasoning request
Amount of work completed ------ Total number of floating-point operations completed
Transaction fees ------ How much $FLOP the miner earns

$FLOP/(computing power/second) = [3] / [2] / [1]

Spot prices only become true prices when everyone can trade freely. On the supply side, if you have computer chips (GPU, CPU, or TPU), you can mine FLOP by completing reasoning requests for agents. On the demand side, as long as you hold FLOP in your wallet, you can submit reasoning requests. Flop Network is an open-source network accessible to everyone. Therefore, the spot computing power price becomes the global benchmark for computing power.

Flop Computing Unit (FCU)

Before discussing cash-settled derivatives, let’s continue to focus on the "physical" computing power market. One feature of Flop Network is the ability to reserve computing power from miners and deliver it within a maximum of thirty days in the future. The network will standardize FCU to make it more tradable.

Imagine you know you have a task to complete in seven days, and you want to lock in the computing power price today. You can bid to reserve a certain amount of computing power from miners for seven days later. After seven days, the miner has twenty-four hours to deliver the computing power. If the reservation is accepted, a tradable seven-day FCU object will be created, similar to a deliverable forward. If the spot computing power price rises, and you no longer need or want that computing power, you can natively sell the FCU to another participant on Flop Network.

Here’s a specific example of how FCU works in practice. Please pay attention to changes in the official Flop documentation, which will contain all this information.

1 FCU object:

Delay: 1 second
Computing power: 1 PetaFLOP
AI model: Specified by the buyer
Delivery date and time: UTC 12:00 seven days later

Note: FCU covers reasoning, not training. Training capacity is more valuable than the sum of its parts. Reasoning work is divisible, which is precisely why a single reference unit or FCU is possible.

For each AI model, there will be up to 30 tradable FCUs, one FCU per day. If you need to reserve 1 ExaFLOP of computing power, you will purchase 1,000 FCUs. As a buyer, to attract miners, you post how much FLOP you are willing to pay for that FCU. To have your bid accepted by the network, you must fully fund it. As a seller or provider of computing power, you post how much FLOP you wish to receive for that FCU. To have your bid accepted by the network, you must fully fund it. As a seller or provider of computing power, you post how much FLOP you wish to receive for that FCU and must stake a weighted amount of FLOP as collateral: if you fail to deliver, another miner can profitably deliver the computing power on the spot market. The amount of staked weighted FLOP as collateral is determined by the network based on the total reasoning volume processed recently. It is the $FLOP cost plus a cheating multiplier to prevent dishonest mining and delivery failures. This creates a fully funded bilateral order book without junk information. The funded buy and sell prices settle on-chain, while continuous bilateral order books and matching occur off-chain, just like the spot market.

There is a risk of delivery failure. To mitigate this, as mentioned above, FCU suppliers must stake a large amount of FLOP collateral, far exceeding the reasoning FLOP fees obtained from buyers. Miners who fail to deliver computing power within twenty-four hours after the due date will lose their entire FLOP collateral. Backup miners will bid to complete the order; they do so because the FLOP fees earned per PetaFLOP are much higher than the current spot price. If the spot price is higher than the FLOP collateral, and no miner is willing to deliver computing power, a delivery failure occurs, and the buyer receives a full refund plus the FLOP collateral.

FRCR Futures Contract

FRCR ------ Flop Reference Computing Rate

Hourly, volume-weighted $FLOP/GigaFLOPS print

I discussed how those with real demand use $FLOP to purchase computing power at spot or forward basis. Now let’s discuss how to construct a cash-settled financial derivative that can be traded on traditional financial and/or crypto exchanges. I will outline a proposed contract and discuss how market makers hedge their long or short futures exposure.

Contract specifications:

Name: FRCR Quanto Quarterly Futures Contract
Expiration: Last Friday of March, June, September, and December at UTC 12:00
Underlying Index: FRCR
Multiplier: $1 per point
Example: FRCR = 100
Contract Value: Multiplier * FRCR = $1 * 100 = $100
Margin Currency: USD

If you go long one contract at an entry price of 100 FLOP and settle at 200 FLOP, your profit and loss are as follows:

Multiplier * Number of Contracts * (Expiration Price - Entry Price) = $1 * 1 * (200 - 100) = $100

If you go short one contract at an entry price of 200 FLOP and settle at 100 FLOP, your profit and loss are as follows:

Multiplier * Number of Contracts * (Expiration Price - Entry Price) = $1 * -1 * (100 - 200) = $100

Quanto derivatives are a type of exotic derivative. It must be a Quanto derivative because I assume most participants want to use USD as margin and earn profits in USD, while the underlying spot market is priced in FLOP. Therefore, we must use a Quanto structure to achieve the returns most market participants expect. To price Quanto derivatives correctly, you must adjust based on the covariance between FRCR and FLOP pricing.

If the market wishes to maintain the margin currency as $FLOP, then:

Multiplier = 1 GigaFLOP
Contract Value = Multiplier * FRCR = 1 * 100 FLOP = 100 FLOP

The key to the success of any derivative contract lies in whether market makers can easily replicate long and short futures contracts from the components.

Long Futures Contract:

To replicate a long exposure, market makers must acquire $FLOP and then purchase a continuously rolling basket of FCUs. Here’s how they synthesize to create a 60-day long FRCR futures contract.

Spend $FLOP and purchase a 30-day FCU. After 30 days, roll it into another 30-day FCU.

On November 1, the 30-day FCU = 100 FLOP. On November 30, you sell your original 30-day FCU for 100 FLOP and purchase a new 30-day FCU for 200 $FLOP, which expires on December 30.

You have recreated a 60-day long futures contract, and your entry price is 200 $FLOP.

Short Futures Contract:

To replicate a short position, market makers must acquire FLOP and then supply a continuously rolling bundle of FCU. This will require more FLOP, as the margin far exceeds the computational costs. Market makers do not need to procure chips, as they will roll the FCU before delivery. Here’s how they synthesize to create a 60-day short FRCR futures contract.

Spend $FLOP and supply a 30-day FCU. After 30 days, roll it into another 30-day FCU supply contract.

On November 1, a 30-day FCU = 100 FLOP. On November 30, you buy back your original 30-day FCU for 100 FLOP and supply a new 30-day FCU for 200 FLOP, which expires on December 30. As long as you roll before the FCU expires or find a backup miner within the 24-hour delivery window, you will not fail to deliver and lose your FLOP supply a new 30-day FCU, which expires on December 30. As long as you roll before the FCU expires or find a backup miner within the 24-hour delivery window, you will not fail to deliver and lose your FLOP margin.

You have recreated a 60-day short futures contract, and your entry price is 200 $FLOP.

FRCR Perpetual Contract

If we understand how futures contracts work, then extending it to perpetual contracts is very simple. Remember, a perpetual contract is just an infinite futures contract strip.

Contract Specifications:

Name: FRCR Quanto Perpetual Contract
Underlying Premium Index: 8-hour TWAP of [Perpetual Price - FRCR]

The perpetual price will be the hourly volume-weighted average, matching the calculation method of FRCR. There will then be eight samples averaged over eight hours.

Multiplier: $1 per point
Example: FRCR = 100
Contract Value: Multiplier * FRCR = $1 * 100 = $100
Margin Currency: USD

The funding mechanism is the same as traditional perpetual contracts. If the premium index is positive, funding is positive, and longs pay shorts. If the premium index is negative, funding is negative, and shorts pay longs. Since this is a Quanto derivative, payments are made in USD.

To synthesize a long perpetual contract, traders will go long on FCU and continuously roll it. For example, if a trader wants to create a synthetic long perpetual exposure for 120 days, they will buy a 30-day FCU and roll it four times. To synthesize a short perpetual contract, traders will supply FCU and continuously roll it. For example, if a trader wants to create a synthetic short perpetual exposure for 120 days, they will supply a 30-day FCU and roll it four times.

Which AI Model?

One issue in this market is that the computational price referenced by derivatives is unrelated to the AI model. However, each model has a different FCU market. Therefore, to truly synthesize the reconstruction of futures and perpetual contracts, traders must purchase and supply a bundle of FCU for each model and match the FRCR with the correct weights.

FRCR releases a categorized bucket by model alongside the comprehensive index, making the basket weights observable. The comprehensive index will drift as the model composition changes, with models retiring rather than hardware becoming obsolete, becoming a risk that hedgers need to manage.

No Friction

To facilitate the growth of the derivatives market derived from the spot and FCU computational prices on the Flop Network, in addition to paying the FLOP fees to validators for processing FCU trades between buyers and sellers, there will be no trading fees for trading FCU. This fee is no different from sending or receiving FLOP from your wallet.

Benefits of Flop Spot and Forward Market

Anyone can participate in buying or supplying computational power on the Flop Network.

Settled trades are executed in a public and transparent market.

The use of all Flop Network data is free.

It’s that simple. If you want to build any type of computational derivatives, you must start by observing the activities and prices in the Flop Network spot and forward markets.

Long live my king!

[1] RWA------Real World Assets (e.g., oil or gold)