How does GPU pricing work on a compute exchange?

Silicon Data tracks forward curves for H100 and B200 capacity and reports that every term on those curves got more expensive between 2025 and this year. GPU pricing therefore looks less like a buyer's market and more like a commodity term structure. Compute Exchange sells reserved and forward contracts on that basis, with provider verification before a deal closes.

What Is Compute Exchange and How Does GPU Pricing Work There?

Compute Exchange is a neutral GPU marketplace that owns no hardware and matches verified buyers with verified providers. It started in 2024, runs reserved and forward contracts plus token packages, and processes requests for quote rather than published list prices, so GPU pricing comes back as competing bids from providers.

The company's stated reason for owning zero GPUs is conflict of interest. A marketplace that also held inventory could favor its own capacity over a provider's, or cherry-pick buyers. Removing that position is meant to keep the quote process comparable across roughly 100 verified providers that the CEO said were live on the platform in October 2026.

Carmen Li, who runs Compute Exchange and the data firm Silicon Data, described the model at an AI Engineer conference session on 23 October 2026. Her argument is that capital markets solved this problem with exchanges: liquidity, settlement and price discovery sit in one venue instead of a stack of phone calls.

Before quoting, Compute Exchange checks existing reserved capacity at the hardware level. Li said the team verifies UUIDs and inventory and runs a GPU benchmark suite, including MLPerf, on machines that are already operational. That step separates a verified listing from an unverified broker offer.

How Compute Exchange Verifies Providers, Buyers and Forward Capacity

Verification runs in both directions because forward contracts create credit risk on each side. Providers with future capacity must show purchase orders and permits, while buyers must pass business and identity checks before a scheduled payment plan is accepted.

The verification model borrows from exchange practice. Li described her background as an energy options market maker at DRW Trading in Chicago, where physically delivered and cash-settled Claude contracts served different purposes. Compute Exchange aims at the physically delivered side, with due diligence standing in for a clearinghouse.

• Provider side: existing capacity is checked for working state, and future capacity is checked against supplier paperwork.
• Buyer side: business verification confirms the customer can pay on schedule, not just sign.
• Benchmarking: operational servers are tested so the delivered configuration matches the quoted one.
• Turnaround: the platform commits to a 24-hour response on requests for quote.

Which Contract Types Set GPU Pricing on the Platform?

Three product shapes carry GPU pricing on Compute Exchange: reserved GPU contracts, forward contracts on GPU capacity or tokens, and a secondary market for used hardware. Each serves a different buyer, and the first two are quoted per node per hour across a term.

Reserved contracts start at one node for one month and can extend to several years, depending on provider availability. Li said the platform does not sell on-demand capacity; its customers are planning ahead. If a buyer overcommits on quantity or duration, the platform allows a reconfiguration or resale so the commitment is not stranded.

Token forwards are the newer concept. A buyer reserves bare-metal clusters, then gets a dedicated model deployment on top with fixed quantization, precision and service-level agreements for a set period. The aim is to remove price volatility from token API billing.

Contract type	Buyer need	Term flexibility	Verification focus
Reserved GPU	Workloads starting soon	One month to multiple years	UUID, inventory, benchmark on live machines
GPU forward	Capacity not yet available	Custom future start date	Purchase orders and permits
Token forward	Stable API cost	Fixed period per deployment	Cluster plus deployment configuration
Used server	Capital purchase	Immediate	Hardware spec and condition

What the H100 and B200 Forward Curves Show About GPU Pricing

Silicon Data's term structure data shows a clear shift: in 2025 the curve sloped downward, meaning longer contracts cost less per hour, and during 2026 it moved up and flattened, so the discount for committing long term shrank or reversed.

The red and yellow curves in Li's presentation compare those two states. On the H100 curve, the whole line moved higher and lost its downward slope. On the B200 curve, every single term became more expensive than the prior year. Both readings point to a market consensus that prices rise over the next 36 months, and Li explicitly noted that consensus can be wrong.

Implied forward rates were derived from the term structure using a method the company compares to the Treasury futures curve. There is no exchange-traded GPU future yet, so these rates are a model output rather than traded prices.

• 2025 shape: downward sloping, longer commitments cheaper per hour.
• 2026 shape: shifted up and flattened, term discount largely gone.
• Implied view: providers and buyers expect higher prices over 36 months.
• Caveat: forward rates come from observed term structure, not executed futures.

Why Used H100 and B200 Prices Beat Straight-Line Depreciation

Secondary server pricing shows the same upward pressure: used H100 and B200 accelerators are holding value above the standard five-year straight-line depreciation that finance teams typically apply to GPU assets.

Li described the secondary market starting in 2025, when operators began offering large inventories of H100 servers for resale. A marketplace for used servers followed. By October 2026 it covered Nvidia and AMD hardware, and B200 units were trading above manufacturer list price in some transactions.

She compared it to used cars during the pandemic, when near-new vehicles sold above sticker because supply could not keep up. The analogy matters because resale value feeds directly into financing decisions: banks holding GPU-backed loans use appraisal estimates, and Silicon Data said it supplies those estimates for several banks.

Silicon Data's residual value model discounts the cash flows of a hypothetical lease back to today, taking assumptions about utilization and operating cost. The company says it applies the method across A100, H100 and B200 servers worldwide. The Nvidia ecosystem is where most of that secondary volume sits.

How CME Futures Would Put GPU Pricing on a Public Gauge

CME Group intends to settle compute futures against Silicon Data indices, according to the session, which would give GPU pricing a public reference price for the first time. Li said trading would begin in a few weeks from the talk and would cover H100 neocloud on-demand rates and a second GPU benchmark.

The transcript names that second index in a way that does not match any current accelerator name, so it cannot be reported precisely here. The H100 neocloud on-demand index is the confirmed component of the plan.

Settlement against an index changes the hedging conversation. An operator with GPU servers on its balance sheet and financing obligations over five to six years, but reserved contracts only for the first three, carries price exposure in years four through six. A swap or block trade tied to the index can offset that. Silicon Data publishes the underlying indices.

Li also described block trading already happening around the term structure, including trades on the first three years of revenue where counterparty default risk remains. Index-based settlement makes those trades easier to price.

What Happens When You Submit a Request for Quote?

A request for quote turns a specification into comparable provider bids within the platform's 24-hour response commitment. The buyer fills in GPU model, deployment type, node count, interconnect, CPU platform, storage, software stack and region, then receives offers side by side.

Li showed the interface filling in a hypothetical order for 300 B200 GPUs in US East beginning 1 January 2027. The platform matched it against provider inventory and routed it to three providers that could plausibly serve it. Each responded with price, specification and service levels.

• State quantity and start date.
• Select model, deployment type and region options.
• Add head nodes and describe the workload.
• Submit and receive routed provider responses.
• Compare price, specs and service levels before deciding.

Where GPU Pricing Sits Among Compute Exchange, Silicon Data and Skala Blog

Compute Exchange runs the marketplace and Silicon Data produces the market data, two companies with the same chief executive but separate roles in GPU pricing. Keeping them distinct matters, because the marketplace's trade flow informs the indices and the indices inform the marketplace's forward quotes.

An academic bootcamp or a Brazilian developer community following hardware costs can use the same public data. Writers covering the topic, including the channels tracked by Skala Blog, now cite a term structure instead of a single hourly rate. Tools such as Crazystack Typescript and courses like Bootcamp do Dev Doido reach developers who would otherwise only see the retail price.

Gustavo Dev Doido has covered Brazilian developer tooling on YouTube, and the compute cost curve is now part of that conversation. The secondary market for used accelerators, the forward curve and the futures plan all exist because GPU capacity became a traded asset.

Frequently Asked Questions About GPU Pricing

• What is GPU pricing on Compute Exchange based on? Compute Exchange collects requests for quote and routes them to verified providers, so pricing comes from competing provider bids for a specific configuration, region and term. The platform does not publish a single list price.

• Why did the H100 forward curve flatten in 2026? Silicon Data's term structure moved from downward sloping in 2025 to higher and flatter during 2026, which means long-term buyers no longer receive the hourly discount they once did. The company attributes this to a market view that prices will rise.

• Does Compute Exchange own GPUs? Compute Exchange states that it owns no hardware by design. Ownership would create a conflict of interest with both providers and buyers, and the neutral position is what lets the platform route orders across competing suppliers.

• Are GPU futures trading on CME yet? As of the October 2026 session, Silicon Data said CME futures tied to its H100 neocloud on-demand index were scheduled to begin trading within weeks. Confirm the current listing status with CME before assuming a live contract.

• Why are used H100 and B200 servers worth more than depreciation implies? Silicon Data's residual value model discounts future lease cash flows and compares the result with five-year straight-line depreciation. Demand for capacity kept market values above the accounting curve, similar to used vehicle prices during supply shortages.

• What is the difference between a reserved GPU contract and a token forward? A reserved contract books GPU nodes for a term. A token forward books bare-metal clusters plus a dedicated model deployment with fixed precision and service-level agreements, so API cost stays stable.

• How fast does a compute request receive a response? Compute Exchange commits to a 24-hour turnaround from request to provider response. Buyers then compare bids, specifications and service levels on the platform before choosing.

• What verification does a provider go through? Live capacity is checked for UUIDs, inventory and benchmark performance, and future capacity is checked against purchase orders and permits. Buyers go through business verification so providers can assess payment risk.

• Is the forward curve a prediction of future GPU prices? The implied forward rates are derived from observed term structure, not from traded futures. They reflect what market participants currently expect, and Silicon Data's own presentation notes that consensus can be wrong.

From a Conference Talk to a Written Article

The argument in this article came from an 18-minute session about compute markets, credit risk and forward curves. The same material works as a written explainer, where a reader can check the H100 curve, the secondary market numbers and the financing logic at their own pace.

If your own YouTube videos carry that kind of explanation, Skalablog turns a video into a transcribed draft article you can review and publish. Paste a YouTube URL, get the transcript, and shape it into a structured piece.

Source video