(Early) AI Compute Asset Pricing

Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.

Ruth: We're looking at the paper: "(Early) AI Compute Asset Pricing".

Warren: Here's the gist of the paper. Compute (computing power) is a scarce, capital-intensive input at the center of the AI economy Compute (computing power) is a scarce, capital-intensive input at the center of the AI economy The impending arrival of a compute futures market will make compute tradable, thereby creating a financial instrument for hedging and price discovery This paper provides an early asset-pricing framework for compute by analyzing the underlying rental market and its implications for future futures contracts The Core Problem of Compute Pricing Compute is a central commodity of the AI economy, reflecting uncertainty over the scale of AI adoption The current price of compute reflects, e.g., the cost of producing AI itself The economic importance of this problem stems from the position of compute in the AI economy, where its price contains information about both sides of the AI economy The combination of uncertainly and scale is precisely what makes compute price risk an economically important object for trading and pricing Theoretical Pricing Restrictions The theoretical analysis makes three points regarding the pricing of compute futures 1) the (storage-based) no-arbitrage logic to derive forward/futures prices from current spot prices for storable commodities should not directly apply to compute 2) existing term-rental contracts provide a useful no-arbitrage reference point, suggesting that the traded futures price Ft(T) should be close to the synthetic futures price inferred from the term rental curve 3) compute futures should carry a risk premium, linking Ft(T) to the conditional expectation of the spot rental price at delivery, ST: Ft(T) = Et

ST: − λt(T) Synthetic Futures Construction The authors construct the first compute futures return panel by utilizing term-rental data made available by Silicon Data to construct synthetic futures prices The first working contracts will likely be designed around a standardized GPU rental index, such as H100, A100 or B200 The settlement price is the spot index value ST at delivery time T, and the corresponding futures price is Ft(T) = ST − Ft(T) Empirical Evidence of Risk Premia Preliminary empirical results using both hold-to-maturity (excess) returns and continuously-rolled constant-maturity (excess) returns are suggestive of a positive risk premium The average realized long hold-to-maturity returns are positive for all three GPU generations The annualized averages show positive compute risk premia in the high single to double digits, which is interpreted as the outcome of relatively more aggressive hedging on the part of compute-service providers Market Design and Manipulation The authors suggest that a transparent, representative, and manipulation-proof index is a critical foundation for a successful futures market They propose safeguards against manipulation, including volume-weighted trimmed means and trimming extreme price inputs in index construction The model suggests that the pre-penalty threshold for positive marginal gains from specified manipulation is QM i = γVi/Wi Compute futures prices are Q-expectations of compute spot prices at delivery, Ft(T) = Et

ST: − λt(T) In the current economy, this will likely be the case for compute providers, i.e., the primary hedgers expected to drive futures prices upon launch of the new market The preliminary empirical results point to a (largely) positive risk premium which is interpreted as the outcome of relatively more aggressive hedging on the part of compute-service providers The analysis aims at providing a framework for a new market and is, therefore, intentionally forward-looking References Acemoglu, D.

Warren: The simple macroeconomics of AI Akerlof, G.

Ruth: A.

Title and authors: Ruth: So we're looking at this paper now titled (Early) AI Compute Asset Pricing by Bandi Yinan Su et al. It’s talking about how to price compute capital because it's becoming a really important thing in the economy.

Warren: Exactly, Ruth. The authors are proposing an early asset-pricing framework for compute, focusing on the underlying rental market and what that means for future futures contracts as that market starts trading later on.

Iris: It seems like the core issue they're tackling is that compute is this central commodity of the AI economy, and its current price reflects uncertainty about how much AI will actually be adopted.

Ruth: Right, and they frame it this way: the combination of uncertainty about scale and scale itself makes compute price risk an economically important object for trading. It’s not just a simple cost calculation anymore.

Warren: The paper sets up some theoretical restrictions right away, saying that standard no-arbitrage logic from storable commodities doesn't directly apply to compute because it isn't something you can just store in a warehouse.

Iris: They suggest that existing term-rental contracts are actually a useful reference point for no-arbitrage pricing, implying that the future futures price should be close to what you’d get from inferring prices based on those rental curves.

Ruth: And they propose this specific formula for compute futures prices: Ft(T) equals the conditional expectation of the spot rental price at delivery minus some term related to time, lambda t(T).

Warren: That brings us to their synthetic futures construction. They build a panel using term-rental data from Silicon Data to create these synthetic futures prices before the actual market even launches.

Iris: The first working contracts they suggest will likely be based on standardized GPU rental indexes, like an H100 or an A100, and the settlement price would be the spot index value at delivery time T.

Ruth: And this leads to a simple formula for those synthetic futures prices: Ft(T) equals ST minus Ft(T), which basically just means you’re looking at the difference between what you expect the spot price to be and what you expect it to be in the future.

Warren: They then move into some empirical evidence, showing preliminary results from both hold-to-maturity and continuously rolled constant maturity returns that suggest a positive risk premium exists.

Iris: The annualized averages they found show compute risk premia in the high single to double digits, which they interpret as the result of relatively more aggressive hedging on the part of compute service providers.

Ruth: That implies that if you're providing compute services, your hedging strategy is likely quite active because there’s a noticeable premium for taking on that uncertainty.

Title and authors: Warren: They also suggest some things about market design, arguing that a transparent, representative index is critical for the futures market to work successfully without manipulation.

Iris: They propose safeguards against manipulation by using volume-weighted trimmed means and trimming extreme price inputs in the index construction, suggesting a pre-penalty threshold for positive marginal gains from specified manipulation.

Ruth: That points toward a specific model where the quantity of compute is measured by something like qi equals gammaVi divided by Wi, which helps monitor potential issues.

Warren: In terms of wrapping up, they state that compute futures prices are expectations of compute spot prices at delivery net of a risk premium: Ft(T) equals EtST minus lambda t(T).

Iris: They conclude that for compute providers, who are expected to be the primary hedgers when the market launches, this formula is likely how they’ll price those futures.

Ruth: And that their preliminary evidence points toward a largely positive risk premium driven by aggressive hedging on the side of compute service providers. This framework is forward-looking because it's built around anticipating that new market launch.

Warren: So, to summarize, the paper (Early) AI Compute Asset Pricing lays out a synthetic futures construction based on rental contracts and suggests that compute futures will be priced as expectations minus a risk premium.

Iris: It also provides empirical evidence showing those premiums are positive for different GPU generations and maturities, which is important data for testing their model.

Ruth: The improvements they suggest are quite detailed—they propose using the synthetic futures formula Ft syn m t equals (M - t) Pi t(t -> M) minus (M - t - one) Pi t(t -> M-one) to create that reference point <ref:2607.12156#pg1>.

Warren: And they also discuss quantifying basis risk between physical term rentals and financial futures by calculating the physical access wedge, Delta Pi t(t -> T), which they expect to be non-negative and increasing with time to delivery.

Iris: They also suggest a way to estimate that risk premium by calculating rt->M equals SM minus Fsyn t(M) divided by Fsyn t(M), which gives you an empirical measure of the return relative to the synthetic price.

Ruth: Overall, they’re giving us a framework for a new market, but they are being intentional about it because their analysis is forward-looking, anticipating what happens when compute futures actually launch.

Warren: So we’ve looked at how they use rental data to build these synthetic prices and how those prices relate to the actual spot price expectations with that risk premium included.

Iris: It seems like the paper's limitation is that it's focused on setting up this framework and estimating the premium, but it doesn't fully model what happens when everyone tries to run these futures contracts simultaneously.

Title and authors: Ruth: That’s a fair point, Iris. They are building the structure for pricing and hedging before they dive into the actual dynamics of market interaction at scale.

Warren: So for listeners who just want to know what this means: compute is moving from being just a cost input to something that can be hedged financially, and that financialization comes with a measurable risk premium based on hedging behavior.

Iris: It changes how you look at AI infrastructure because it introduces a formal pricing mechanism for uncertainty in the scale of adoption.

Ruth: Indeed, this paper (Early) AI Compute Asset Pricing gives us a starting point to discuss how this new asset class might be structured before the CME Group launches its first-in-class market later in two thousand twenty-six <ref:2607.12156#pg1,Early) AI Compute Asset Pricing>.

Warren: We’ll keep an eye on those empirical results and see if those positive risk premia hold up as more data comes in about the actual adoption curves of AI.

Iris: It’s a solid theoretical and empirical setup for getting the conversation started on pricing compute risk in this new asset class.

Ruth: We've covered the title, what they are trying to model, their specific formula constructions, and where they land on empirical risk premia. This paper (Early) AI Compute Asset Pricing is definitely a foundational piece for anyone interested in the financialization of AI infrastructure.

Warren: It’s important for us to remember that this whole discussion is based on the assumption that these theoretical links hold up when actual market participants start trading these contracts.

Iris: It’s a framework, not a prediction of what will happen next, which is a key distinction for us as econometricians looking at identification.

Ruth: Exactly. We’ve got the structure now and some early numbers on that risk premium based on hold-to-maturity and rolling returns.

Warren: So we’re leaving the details of how they constructed those specific indices and moving toward what these findings imply for compute providers trying to manage their exposure.

Iris: It really does suggest that aggressive hedging, as they noted, is a key driver behind the observed premium in this early stage.

Ruth: We’ve seen how they use term rental data to build synthetic futures prices and how those relate back to the spot price expectations with that lambda term.

Warren: So, this paper (Early) AI Compute Asset Pricing gives us a tangible way to think about pricing compute before the official market opens its doors in two thousand twenty-six <ref:2607.12156#pg1,Early) AI Compute Asset Pricing>.

Iris: It’s a strong analysis of how financial instruments can be adapted to an asset that is inherently non-storable like compute capital.

Ruth: That’s where we wrap up this discussion on the paper. We've looked at the theory, the synthetic constructions, and those early empirical signals regarding positive risk premia in (Early) AI Compute Asset Pricing.

The paper's summary: Ruth: So, to recap, this paper is about how they are trying to price compute capital because it's becoming an asset that can be traded, not just something you buy for production.

Warren: Right. They’re using a rental market as their starting point and trying to build a way to price future contracts before the whole thing gets fully tradable.

Ruth: Essentially, they are saying compute has this inherent uncertainty about how much AI adoption will actually happen, and they're trying to bake that into a price formula.

Warren: And the core idea is using this formula Ft equals EtST minus lambda t(T), where lambda t(T) is this risk premium.

Ruth: That risk premium comes from the covariance between how much compute you expect to rent later and some measure of how much that uncertainty affects you.

Warren: It’s linking the futures price directly to the expected spot rental price at delivery, minus this term related to time.

Ruth: What this means for us is that if someone is providing compute services, their hedging strategy—how they try to manage that uncertainty—is what drives this premium up or down.

Warren: The empirical results they showed suggest that providers are being relatively aggressive with their hedging, which translates into a positive risk premium in the high single to double digits for those GPU generations.

Ruth: That’s interesting because it suggests the market isn't just pricing the cost of the hardware; it's pricing the behavior of people who rely on that hardware.

Warren: Exactly, and they also talked about how you build a synthetic futures price using term rental data to create a no-arbitrage reference point before any actual trading starts.

Ruth: So they’re building this structure for pricing and hedging now, anticipating when the official market launches later in two thousand twenty-six.

Warren: And that sets up the next question for us: what happens to this framework when thousands of people are trying to run these futures contracts at once?

The paper's improvements: Tom: So, to recap, the authors aren't just stopping at the initial model; they’re suggesting several ways to make this compute pricing framework more robust and usable in a real market setting.

Ruth: They are proposing a way to price those synthetic futures prices using term rental data through a specific formula involving Pi t(t -> M).

Warren: That formula looks like Fsyn m t equals (M minus t) Pi t(t -> M) minus (M minus t minus one) Pi t(t -> M-one). It’s trying to create that clean no-arbitrage reference point.

Iris: And they also show how you can estimate that risk premium by calculating rt->M equals SM minus Fsyn t(M) divided by Fsyn t(M). That gives you a concrete empirical measure of the return relative to the synthetic price.

Ruth: It’s about making those formulas more specific so that they aren't just abstract ideas, but actual calculable inputs for a trading system.

Warren: They also suggest a way to design an index for trading by using volume weighting where Wi equals Vi divided by VAGG, which makes bigger providers more susceptible to manipulation if you look at the quantity Qi equals qi divided by seven hundred thirty GPU-months.

Iris: And they introduce a way to measure basis risk between the physical term rentals and those financial futures by calculating that physical access wedge, Delta Pi t(t -> T). They expect that wedge to be non-negative and increasing with time to delivery T.

Ruth: That means if you’re looking at the actual rental contract versus the theoretical futures price, the difference should only get bigger as you wait longer for delivery.

Warren: It’s a nice check on your model, because it tells you exactly where your physical access is relative to the market price.

Iris: So, these are all methods to verify if their identification holds up—checking if the variation they are measuring actually drives the results and what would falsify that whole setup.

Ruth: It’s a lot of machinery to make sure that when this market launches, it’s built on solid ground rather than just theoretical constructs.

Warren: So we’ve seen the initial structure, the empirical risk premium signals, and now we see how they plan to refine those formulas for better trading and verification.

Iris: It really shows they are thinking about the practical implementation details of an asset class that doesn't exist yet in a formal way.

Ruth: This paper (Early) AI Compute Asset Pricing is building the toolkit for this new market, but we still have to figure out how much actual trading volume will actually materialize around these rules.

Conclusion: Ruth: So we’re wrapping up this discussion on (Early) AI Compute Asset Pricing by Bandi Yinan Su et al., looking at how they built a framework for pricing compute as an asset.

Warren: Right, it’s about taking the uncertainty around AI adoption and trying to turn that into a tradable financial instrument using rental data.

Iris: They did this by constructing synthetic futures prices based on term rental curves, which serves as this no-arbitrage reference point for what the actual futures price should be.

Ruth: And they showed that when you look at the empirical results, there is a positive risk premium in the high single to double digits for those GPU generations.

Warren: That premium comes from providers hedging more aggressively, and it suggests that if you’re giving compute services, your hedging strategy matters financially.

Iris: I just want to mention that they flag that their model doesn't fully capture what happens when everyone tries to trade these contracts at the same time.

Ruth: That is true; this paper gives us a great starting structure for pricing and risk, but it’s not a full market simulation yet.

Warren: It’s definitely a foundation for how we think about AI infrastructure moving from just being an operational cost to something that carries financial risk.

Iris: It sets up the identification challenge well, showing exactly what variation in the data is doing the work, which helps us know if their conclusion holds up.

Ruth: We’ve seen how they use rental data to build those synthetic prices and how they relate back to spot price expectations with that lambda term.

Warren: So, this paper (Early) AI Compute Asset Pricing gives us a tangible way to think about pricing compute before the official market opens its doors later in two thousand twenty-six.

Iris: It’s a strong analysis of how financial instruments can be adapted to an asset that is inherently non-storable like compute capital.

Ruth: That’s where we wrap up this discussion on the paper. We've looked at the theory, the synthetic constructions, and those early empirical signals regarding positive risk premia in (Early) AI Compute Asset Pricing.

Warren: It’s important for us to remember that this whole discussion is based on the assumption that these theoretical links hold up when actual market participants start trading these contracts.