SF Compute: Physical Settlement in AI Infrastructure
SF Compute is pioneering physical settlement for GPU capacity, addressing the lack of fungibility in AI hardware. By operating the full cloud stack without owning hardware, they create a neutral marketplace that reduces financing costs and enables liquidity for sovereign and private compute assets.
The Rise of Physical Settlement in AI Compute
The AI infrastructure market is evolving from a simple hardware procurement model into a complex financial ecosystem. SF Compute has emerged as a critical player by introducing "physical settlement" to GPU trading, a concept borrowed from commodity markets like oil and soybeans. Unlike cash-settled derivatives, which can diverge from reality, physical settlement ensures that the traded asset is actually delivered and operational. This distinction is vital in AI, where hardware heterogeneity—ranging from firmware bugs to storage speeds—makes simple fungibility impossible. Without a neutral operator to manage the physical delivery, the market faces high basis risk, where the price of the derivative does not reflect the cost of actual capacity.
Operational Independence as a Strategic Moat
SF Compute’s core competitive advantage is its structural neutrality. The company operates the entire cloud stack, including backplane management, cluster architecture, and supply chain consulting, but does not own the underlying hardware. This "Switzerland" model is essential for market integrity. If a major player like NVIDIA or a hyperscaler like CoreWeave operated the exchange, they could manipulate prices to benefit their own inventory or demand. By remaining independent, SF Compute creates a trustworthy index that lenders can use to finance forward contracts, thereby reducing the cost of capital for the entire industry. This independence allows them to serve as a neutral venue where buyers and sellers can transact without fear of counterparty bias.
Navigating Heterogeneity and Liquidity
The primary challenge in compute markets is the lack of standardization. GPUs are not identical; early H100s had firmware issues, and different clusters have varying storage and network capabilities. SF Compute addresses this by building a multi-dimensional order book that allows customers to specify exact constraints, such as preemption tolerance or storage type. This approach reduces dimensionality, making the market liquid without forcing a false standard. The forwards curve currently shows a "hump" shape, driven by the NVIDIA release cycle and the scarcity of immediate capacity. This creates significant arbitrage opportunities for those who can navigate the spread between spot and future prices.
Strategic Implications for Sovereign AI
As nations invest in sovereign AI, the need for financeable infrastructure becomes paramount. SF Compute positions itself as a partner for these initiatives, offering the operational expertise to build local clouds while providing the market access to ensure they are commercially viable. By virtualizing the data center and allowing programmable resource allocation, the company is effectively creating a new asset class for AI compute. This shift from capex-heavy ownership to liquid, tradeable capacity is set to reshape how AI infrastructure is financed and deployed globally.
Key insights
-
Physical settlement is required to create a reliable compute market because cash-settled derivatives suffer from high basis risk when the underlying asset is heterogeneous and hard to deliver.
Impact: Reduces financing costs for AI infrastructure by providing a trustworthy index for lenders, enabling larger scale deployments.
-
Market neutrality is a prerequisite for liquidity; major hardware vendors and hyperscalers cannot operate the exchange because their own demand and supply would allow them to manipulate prices.
Impact: Creates a structural moat for neutral operators like SF Compute, as incumbents are structurally unable to replicate the model without sacrificing their core business.
-
GPU clusters are not fungible due to firmware, storage, and network differences, requiring a multi-dimensional order book that matches specific technical constraints rather than a single standardized unit.
Impact: Allows for more precise pricing and risk management, enabling sophisticated buyers to arbitrage spreads between different hardware configurations.
-
The compute forwards curve exhibits a hump shape driven by NVIDIA release cycles and immediate capacity scarcity, creating arbitrage opportunities between spot and future prices.
Impact: Provides a framework for hedging and speculation, allowing companies to lock in future capacity at lower rates than current spot prices.
-
Sovereign AI initiatives require neutral partners to make local infrastructure financeable, as national governments lack the operational expertise to build and operate high-performance clouds independently.
Impact: Opens a new market segment for infrastructure providers, enabling them to capture value from national AI investments while ensuring commercial viability.
Action items
-
Evaluate the basis risk in current AI compute derivatives by comparing cash-settled prices with physical delivery costs, and adjust hedging strategies accordingly.
Impact: Prevents financial losses from diverging derivative prices and ensures that hedges actually protect against physical capacity costs.
-
Partner with neutral operators for sovereign or large-scale AI infrastructure projects to ensure the assets are financeable and have a clear path to liquidity.
Impact: Reduces the cost of capital for infrastructure projects and ensures that the assets can be sold or traded if the initial business case changes.
-
Implement programmable resource allocation in data center operations to allow for dynamic scaling and subdivision of fabrics, enhancing the ability to trade capacity.
Impact: Increases the liquidity of owned infrastructure by making it easier to match with diverse buyer requirements and market conditions.
-
Monitor the compute forwards curve for arbitrage opportunities, particularly around NVIDIA release cycles, to lock in future capacity at favorable rates.
Impact: Reduces long-term infrastructure costs and provides a hedge against future price increases driven by hardware scarcity.
-
Standardize internal reporting on GPU cluster performance metrics, such as MFU and preemption tolerance, to facilitate easier trading and valuation of assets.
Impact: Improves transparency and reduces the friction in buying and selling compute capacity, making the asset class more attractive to investors.
Quotes
“Our impression is that NVIDIA can't make it, but benefits tremendously from it existing. The reason why they can't make it is because they would then not be independent enough for people to trade on it.”
“We call that physical settlement. We're taking a term from the futures market, from futures markets, from other markets, where in, you know, oil or soybeans or something, physical settlement means you get delivered the oil and the soybeans.”
“If you try to build too much financialization before the industry has had time to build out the physical settlement, then I think what you have is you have... a market that's very easy to manipulate and a market that has a lot of basis risk”