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

If you need reserved compute within forty-eight hours, calling three hundred salespeople is an expensive way to discover what is available. Carmen Li, CEO of Compute Exchange and Silicon Data, opens with that mismatch: a growing GPU ecosystem gives buyers more choices while making the transaction harder to organize. She describes a market with more than three hundred neocloud and private providers globally, alongside token factories and API routers.

The stock-market analogy is about reducing that coordination cost. A buyer can call brokers and banks individually, or use an exchange to find liquidity and complete a transaction through an established process. An exchange does not necessarily produce the lowest price on every trade. Its value is making the search, agreement and settlement more efficient. Compute Exchange applies that idea to finding verified compute resources.

Neutrality determines the platform’s business design: Compute Exchange owns no GPUs. Li’s concern is that owning capacity would give the marketplace a reason to favor its own inventory, cherry-pick customers or keep valuable provider relationships to itself. Without that inventory position, the intended job is to find the resources that suit each transaction. The focus is reserved compute, forward contracts and token packages.

At the time of the recording, Li reports more than one hundred verified providers and more than eighty thousand GPUs in supply, covering regions outside China, with a twenty-four-hour turnaround commitment. Verification changes with the thing being sold. An operating cluster can be inspected and benchmarked; a future cluster requires evidence that it can become available on schedule.

The checks address three different uncertainties:

• Existing capacity: Verify GPU identifiers, inventory and performance, including benchmarking such as MLPerf. The question is whether the functioning resources match what the buyer was promised.
• Forward capacity: Examine purchase orders and permits. For a buyer seeking GPUs from January 1, 2027, these documents support due diligence on capacity that cannot yet be tested as a working deployment.
• The buyer: Perform know-your-business and know-your-customer checks. Providers also need to establish who is requesting compute and assess whether that customer can pay at the agreed rate.

That two-sided verification supports a market aimed at actual delivery. Li calls the forward business a precursor to physically delivered futures. Her reference point comes from four or five years as an energy options market maker at DRW Trading in Chicago. In her analogy, WTI represents procuring the physical resource, while cash-settled Brent represents managing financial price exposure. “We’re creating the WTI” describes Compute Exchange’s intended role: arrange access to compute that a customer can use.

A GPU forward contract schedules access to capacity starting in the future. A token forward adds a deployment layer to that reservation: reserve a GPU cluster, then run a dedicated model deployment on top. The resulting agreement can fix the model’s quantization and precision for a period and provide specified service-level agreements. The intended benefit is more predictable service and pricing than a buyer gets from volatile on-demand token APIs.

The distinction matters because a token contract commits to more than a hardware model. Two deployments on similar GPUs can differ in the model served and its numerical configuration. Fixing those choices makes the promised service more specific: the buyer is reserving a defined inference deployment rather than assuming that any available token endpoint will behave the same way.

The third market grew out of a much more tangible request: someone had twenty thousand H100s sitting in a data center and wanted help moving them. Li’s response was practical—“I can’t store in my home, obviously”—and Compute Exchange developed a marketplace for secondary server sales. She describes roughly twelve months of active trading, including B200 servers and transactions sometimes above the manufacturer’s suggested retail price. Her analogy is used cars during COVID: an already available asset can command more than its nominal new price.

Reserved GPU contracts start at one node for one month and can extend to three, four or five years, depending on provider availability. Compute Exchange does not offer on-demand GPU procurement. Li also describes selling back capacity when a customer has overcommitted on quantity or duration, though the recording leaves the rollout timing and commercial terms unclear. That flexibility should therefore be understood as a described option, without assuming a guaranteed exit price or an unconditional buyback.

Buyers can browse listings and discuss delivery directly with providers. When the listed inventory does not fit, a request for quote, or RFQ, expresses the requirement. Li describes requests ranging from one node to twenty thousand nodes. Behind the form, inventory tracking records which of the hundred-plus providers has which resources and roughly when future capacity should become available.

The concrete example is a request for 3,300 B200 GPUs in US East, starting January 1. Compute Exchange identifies three providers it believes can meet the requirement and routes the request to all three. Each responds through the platform with pricing, specifications and service-level agreements. Existing servers are benchmarked. The buyer then compares the offers; the platform does not reduce the decision to the lowest price.

Where does the matching happen, and where does the buyer retain the decision? The flow below separates inventory-based routing from the providers’ independent responses. The three offers converge into a comparison of price, specifications, service commitments and, for existing capacity, benchmark results. This is the practical efficiency gain: one stated requirement reaches plausible suppliers, while the customer still chooses among the returned terms.

The records remain in one place, and the platform introduces the buyer and provider to each other. That follows the earlier neutrality decision: Compute Exchange says it does not keep those relationships hidden. Its used and refurbished inventory spans 4090s through B200s and AMD hardware, with communication through Slack, email or chat. It also tracks transactions end to end and scores both customers and providers, although those scores were not yet visible to users. Li describes the service as a negotiated exchange, with a middleman helping clarify terms and share practices from other transactions.

The platform demonstration at 11:22 begins with available inventory. Li describes signup and browsing as free, with no requirement to purchase. The RFQ form provides a way to ask for capacity that the buyer does not find among the listings.

What must a provider know beyond “I need GPUs”? The form collects vendor type, GPU model, deployment type and GPU count, then interconnect preferences, CPU platform and storage. Optional requirements include system and software stacks, regions, head nodes and the intended use case. These fields turn a quantity into a deployment specification: enough detail to distinguish offers that might otherwise look comparable on GPU count alone.

Li encourages buyers to select multiple regions if the workload permits it. Region is a matching constraint in the same way that hardware and deployment type are constraints. Allowing more than one location gives the platform more possible inventory to match, while keeping a strict location requirement narrows the set. Once those preferences and requirements are entered, the buyer submits the RFQ for routing.

The final part moves from procuring compute to pricing it through time. Silicon Data supplies compute indices; at the time of the recording, Li says CME futures using its H100 neocloud on-demand and B200 price indices were expected to begin trading within a few weeks. That was a forthcoming market. The curves she then discusses are drawn from reserved-contract term structures, with forward rates inferred from those observations.

How did the reward for signing a longer contract change? In the term-structure discussion at 13:00, the horizontal axis is contract duration and the vertical axis is price. The earlier yellow curve slopes downward: committing for longer bought a lower rate. Early in the following year, the curve moved upward and became upward-sloping, which Li calls contango. Later observations moved higher again while becoming flatter. Price level and slope changed separately: capacity became more expensive, and the old duration discount weakened.

An observed multi-year contract price and an implied forward rate answer different questions. The contract price describes the quoted commitment over its duration. The implied forward curve uses the term structure to infer rates for future periods within that horizon. Li describes a methodology similar to Treasury-curve calculations and interprets the result as a market expectation of rising prices over the next thirty-six months. These are calculated rates rather than traded GPU futures prices, and a market expectation can be wrong.

B200 begins from a different shape: Li says its earlier curve was already in contango. The later curve is higher at every term and flatter, with implied forward rates again pointing toward higher future prices. The useful comparison is between the price level, the reward or penalty for a longer commitment, and the future rates inferred from the collection of contract prices.

Term structure matters because a provider’s financing and customer contracts can end at different times. Li’s example has servers on the balance sheet with financing obligations extending five or six years, while reserved customer contracts cover only the first three. Years four through six still depend on the rates the provider will be able to earn after the current reservations expire.

That creates two distinct exposures:

• Uncontracted later years: Debt obligations continue after the reserved revenue ends. Li describes swaps and block trades as ways market participants seek to hedge the resulting future price exposure.
• Default during contracted years: The first three years also carry risk if the reserved customer contract defaults. Li reports trading around indices used as floating rates, raising the question of how dependable the promised contract revenue is.

The earlier physical-procurement analogy now has a financial counterpart. A reservation arranges a period of access or contracted revenue; a hedge addresses exposure to changing prices. Neither the existence of a reservation nor a favorable-looking forward curve removes the mismatch between a server’s financing horizon and its contracted earning period.

The same term structure supports another calculation: valuing the servers themselves. Li says Silicon Data provides appraisals for several banks holding GPU servers on their balance sheets and that NVIDIA cited its data in a September investor presentation. The valuation method she explains is forward-looking discounted cash flow: estimate the cash an asset can earn, then translate those future receipts into a present value.

The example is an A100 server with a new three-year lease. The lease rate supplies an earning assumption; utilization and costs affect how much cash the server produces; discounting brings that cash back to time zero. The resulting value depends on those assumptions. This is an estimate of earning power, so a change in rental rates can change the valuation even as the hardware gets older.

How does that estimate compare with depreciating hardware over five years? Silicon Data compares normalized residual-value estimates for A100, H100 and B200 servers with a five-year straight-line depreciation curve. Li describes three different outcomes:

• A100: Estimated value declines, but remains above the accounting depreciation curve. Getting older and retaining more value than the accounting schedule predicts can both be true.
• H100: The described value has risen. Li calls the pricing “going bananas” and returns to the analogy of used cars during COVID.
• B200: The described value has moved above MSRP, consistent with the earlier discussion of secondary transactions above the nominal new-hardware price.

The ending makes the procurement market consequential for anyone financing or holding the machines. A five-year depreciation schedule reduces book value with time; a cash-flow valuation responds to what compute can earn. In the examples Li presents, those paths diverge: A100 value declines but remains above the accounting depreciation curve, while H100 and B200 values rise. The rates negotiated for access to compute help explain why the servers supplying it can retain—or gain—economic value.