Investing.com — Bernstein believes that trading for compute markets most closely resembles the current structure for commodities, particularly electricity. The brokerage also highlighted a three-step process to build a potential benchmark for the space and provide traders with a mechanism to hedge the trillion dollar artificial intelligence capex cycle.
Computational power, or compute, refers to the resources needed for training and deploying AI. This includes the data center processing power, hardware, and infrastructure that goes into building and running machine learning models. Demand for such power has shot through the roof since the AI boom in late 2022, with leaders in the space recently inking deals worth billions of dollars.
The New York Times on Friday reported that was in early talks to provide compute to Anthropic in a blockbuster $10 billion agreement. Meanwhile, the Wall Street Journal reported that Elon Musk’s was in talks to provide compute to the U.S. Department of Defense. Last month, the rocket and exploration company signed a $920 million per month cloud deal with Alphabet’s Google.
Against this backdrop, major exchange operators such and earlier this year announced plans to launch tradeable futures contracts tied to spot prices for Graphics Processing Unit (GPU) compute, pending regulatory review.
“Spot compute markets include transactions for immediate access to available GPU capacity, where compute resources are rented at prevailing market prices for near-term workloads and delivery,” Bernstein’s Gautam Chhugani and Madison Rezaei said on Friday.
“We think the closest parallel for compute is power. Electricity was once dismissed as untradeable as it cannot be stored, varies by location and quality – until standardized hubs and reference prices turned forward procurement into a financial market,” they said.
“Compute is following the same sequence, because it shares the same physics – GPU hours cannot be stored, capacity varies by location, chip generation and tenant. So the forward curve reflects expected scarcity rather than storage cost, and hedging is the only way to carry price risk,” they noted.
“Platforms are now trying to solve the problem of standardizing compute (benchmarks and per GPU hour approach), develop both cash settled and physical delivery rails for forward contracts and build structured financial products on underlying compute. Our view is that benchmarks and distribution are the key factors for growth of compute markets – not different from any commodity financial markets,” the analysts added.
Bernstein said that index providers would need to standardize compute power, similar to how electricity market operators standardize power. The first step would be to define a single unit for compute, but the brokerage noted that the problem was raw compute prices were not comparable across the market, as the same AI chip could be priced differently depending on whether it was rented from so-called traditional hyperscalers or specialized neoclouds.
“The index provider needs to define the ‘unit’, for instance, H100 SXM 80GB (includes server, memory, bandwidth etc.). Further specifications include tier (neocloud/hyperscaler), type of lease (on-demand/committed term) and normalization steps for region and configuration,” Chhugani and Rezaei said.
The standardized data would then become a reference index, while forward curves would be derived from market observations, the analysts added.
