Buying GPUs Should Work Like Buying Stocks — Carmen Li, Compute Exchange

Highlights

• Compute Exchange operates as a neutral marketplace for GPU procurement, featuring over 100 verified providers and more than 80,000 GPUs in supply.

• The platform verifies existing reserve capacity by checking UIDs, inventories, and performance metrics using tests like NVIDIA DGX, InferenceX, and MLPerf.

• Secondary server transactions include hardware such as H100 and B200 servers, with some secondary units trading above MSRP.

• Silicon Data provides compute indices used by CME to settle futures contracts and serves as an asset appraiser for banks holding GPU servers on their balance sheets.

• Implied forward curves for GPUs indicate that term structures behave similarly to traditional commodities, carrying distinct term structure risks that enable hedging through block trading and futures.

Timeline

1. 00:00:00 - 05:30:00

   Introduction to Compute Exchange and Neutral GPU Procurement

   Established in 2024, Compute Exchange addresses the rapid proliferation of the GPU ecosystem by providing an efficient exchange framework similar to the NYSE or NASDAQ. By maintaining a design that avoids owning GPUs, the marketplace prevents cherry-picking and preserves neutrality with global providers. Operations cover reserve compute, forward contracts, and token packages with a 24-hour turnaround commitment.

   • Compute Exchange functions as a neutral marketplace for GPU procurement without owning any physical GPUs to avoid conflicts of interest.
   • The platform coordinates transactions across over 300 new cloud providers, token factories, and API routers globally.
   • Verification processes include checking UIDs, inventories, and benchmark performance using NVIDIA DGX, InferenceX, and MLPerf for active capacity, alongside KYB and KYC checks for clients.
2. 05:30:00 - 08:00:00

   Core Products: Reserve Contracts, Token Forwards, and Secondary Hardware

   Buyers can secure resources through structured contract layers designed to mitigate pricing volatility. Reserve contracts range from one node for a month up to multi-year commitments with flexibility to sell excess capacity back to the exchange. Secondary server transactions accommodate large volumes of enterprise hardware previously deployed in data centers.

   • Compute options include reserve contracts, forward contracts on GPUs and tokens, and an open secondary hardware market.
   • Token forward contracts provide dedicated on-premise cluster deployments with guaranteed SLAs, quantization, and precision over fixed periods.
   • Secondary server transactions cover hardware ranging from H100 units to B200 servers, with high-demand equipment occasionally trading above manufacturer suggested retail prices.
3. 08:00:00 - 12:20:00

   Platform RFQ Workflow and Inventory Management

   The digital interface allows users to filter vendor type, model, deployment configuration, interconnect preferences, and geographic regions. Once an RFQ is submitted, matching providers respond with pricing, specs, and SLAs. The exchange acts as a neutral facilitator, introducing parties directly without holding ongoing customer relationships.

   • Request for Quote (RFQ) volume exceeds one billion requests monthly, spanning scales from single nodes up to 20,000 nodes.
   • Inventory tracking matches incoming customer specifications against a global network of over 100 providers.
   • Provider and customer performance is tracked end-to-end through a scoring system that evaluates responsiveness and reliability.
4. 12:20:00 - 18:03:00

   Market Dynamics, Implied Forward Curves, and Silicon Data Valuation

   Analyzing term structures reveals that GPUs behave as physical commodities with clear forward risks that benefit from hedging and block trading. Silicon Data's forward-looking pricing models calculate residual asset values across A100, H100, and B200 servers, providing crucial asset appraisal metrics utilized by financial institutions and referenced in NVIDIA investor presentations.

   • Silicon Data provides index pricing for CME futures contracts covering H100 Neo Cloud on-demand rates and B200 prices.
   • Implied forward curves demonstrate that GPU term structures have shifted from downward sloping to contango, indicating market expectations of rising future prices.
   • Residual value calculations discount future cash flows from long-term leases, showing asset valuations often tracking above standard straight-line accounting depreciation.