GPU Pricing: A No-Holds-Barred Free-for-All

Original author: David Lopez Mateos

Original compilation: Deep Tide TechFlow

Intro: Media often likes to use a single number to capture how GPU compute power pricing moves up and down, but the reality is this: on the Bloomberg terminal, the quotes provided by four index providers for the same segment deviate from each other by more than $2, and they don’t agree on direction or timing. The author of this article is David Lopez Mateos, founder of the GPU compute trading platform Compute Desk. Using firsthand trading data, he breaks down the real pricing structure of H100 and B200, revealing a raw market with no consensus benchmark, no standardized contracts, and no forward curve—compute power is being stockpiled and sublet the way short-term apartments are.

Media headlines will make you think GPU compute power prices are surging. This narrative is comforting—it fits neatly into the macro framework of “supply tightness + AI demand with no bottom,” and it also implies something reassuring: that we have a smoothly functioning market where price signals are clear and readable.

But we don’t.

This narrative is built almost entirely on a single index, and it implies things that shouldn’t be implied: that the GPU leasing market is efficient enough to be represented by one number as a global state.

Supply scarcity is real, but the scarcity people feel is completely different—depending on who you are, where you are, what contract you trade, and what compute asset you hold. In the face of such opacity, the market’s natural response is not orderly price discovery, but hoarding: locking in GPU hours you might not need yet, because you don’t know whether you’ll be able to buy them next month at any price. Where there is hoarding and no transparent benchmark, a fragmented secondary market emerges. At Compute Desk, we’ve already enabled tenants to sublet their clusters the way people sublet apartments during major events. This isn’t a hypothesis—it’s happening.

Indexes Don’t Converge

In mature commodity markets, indexes built on different methodologies tend to converge. Brent crude and WTI, for example, may differ by a few dollars in price because of geography and crude quality, but they move in sync in terms of direction (Figure 1). This convergence is a hallmark of an efficient market.

Caption: Comparison of Brent and WTI crude oil price trends—direction is highly consistent

On the Bloomberg terminal, there are currently three GPU pricing index providers: Silicon Data, Ornn AI, and Compute Desk. SemiAnalysis has just released a fourth—an H100 one-year monthly contract price index built from survey data of more than 100 market participants. Silicon Data and Ornn publish daily H100 rental indexes; Compute Desk aggregates data at the Hopper architecture level; SemiAnalysis captures negotiated contract prices rather than list prices or scraper prices. Different methodologies, different frequencies, and different angles on insights into the same market. When you stack them together, the disagreements are obvious (Figure 2).

Caption: Overlay comparison of four GPU indexes—the price levels and trends are clearly different

Where Exactly Did Prices Rise?

Using Compute Desk data, we can break down the H100 price movements by provider type and contract structure, and overlay Silicon Data’s SDH100RT index (Figure 3). All metrics show prices rising, but the starting points and magnitudes vary dramatically depending on the index and contract type.

Caption: Price trends for H100 split by contract type overlaid with SDH100RT

Compute Desk’s H100 neocloud data tells a more specific story than the aggregated index. On-demand pricing is relatively stable throughout the winter—around $3.00 per hour—then sharply jumps in March to $3.50. Spot pricing is noisier and lower, with only a slight upward trend until March. Silicon Data’s SDH100RT shows a smoother, steady climb, rising from $2.00 to $2.64 over the same period. The two indexes remain at different price levels and also describe the timing differently: Compute Desk says there was a jump in March, while Silicon Data says it was a slow climb.

One-year reserved pricing stays basically flat before February, then jumps from $1.90 to $2.64 from the end of March—this isn’t a gradual catch-up, but a sudden repricing. This looks more like providers adjusting contract fee rates in a bunch after tightening the on-demand market, rather than a sustained, structurally driven demand story.

The March story for B200 is even more intense (Figure 4). Compute Desk’s on-demand index rockets from $5.70 to above $8.00 within a few weeks. Silicon Data’s SDB200RT spikes from $4.40 to $6.11, then falls back to $5.47. Both indexes capture this rally, but the starting points differ by more than $2, and the shapes of the rise and the pullback are also different. With B200 having less than five months of data, fewer providers, and wider price spreads, the two indexes are effectively viewing the same event through very different lenses.

Caption: On-demand versus reserved price trends for B200—overlay of Compute Desk and Silicon Data

Infrastructure Issues, Not Just Regional Differences

Commodity markets have basis differentials. Appalachian natural gas is a textbook case: vast reserves sit atop structurally constrained pipeline capacity. Utilization rates in the Pennsylvania–Ohio corridor often exceed 100%, and new projects like Borealis Pipeline don’t come online until the late 2020s.

The GPU market has something similar: an H100 in Virginia and an H100 in Frankfurt are not the same economic commodity. But you can’t explain the magnitude of divergence among indexes that measure the “same market” based on geographic differences alone. The misalignment in the GPU market runs deeper than in Appalachian natural gas. The problem with natural gas is a single missing link: pipeline capacity connecting supply and demand. In compute markets, the infrastructure gap exists on both the supply and demand sides. Physical infrastructure—consistent networking required for reliable compute delivery, predictable configuration, and predictable availability—is still immature, and sometimes simply doesn’t work. Financial infrastructure—even though physical differences can be compressed via standardized contracts, transparent benchmarks, and arbitrage mechanisms—also doesn’t exist yet.

The data told a story. Early 2026’s real experience of trying to procure compute power told an even sharper one.

All on-demand capacity for every GPU type is effectively sold out. Getting 64 H100s is difficult: Compute Desk shows that for on-demand clusters, 90% of providers have zero available capacity, and the reserved market isn’t much better. In a well-functioning market, this level of scarcity would have already pushed prices to a new equilibrium. But it hasn’t. This suggests that providers also lack real-time pricing intelligence to adjust. Prices are rising, but rising too slowly to clear the market. The gap between list prices and actual willingness to pay is being filled by hoarding, subletting, and informal secondary market trading.

What Needs to Change

The current GPU compute market has seven core problems:

No consensus benchmark. Multiple indexes coexist, each with different methodologies and contradictory conclusions.

Aggregation narratives hide structure. A single number for “H100 prices” masks huge differences across provider types and contract tenors.

Lack of trade-level data. In bilateral markets, the gap between list prices and actual executed prices is very large.

No contract standardization. Most GPU leases are bilateral negotiations with different terms. Shorter, more standardized contract tenors can improve liquidity and price discovery.

Delivery quality isn’t guaranteed. Enormous differences exist in interconnect topology, CPU pairing, network stack, and runtime. Buyers need to know what quality of compute they’re purchasing before making a commitment.

Contracts lack liquidity. If demand changes during the reserved period, choices are very limited: either eat the cost or sublet informally. The market needs a way to transfer or resell already committed compute infrastructure so capacity flows to where it’s needed most.

No forward curve. Without the ability to price forwards, you can’t hedge. That’s why lenders apply a 40%-50% discount to GPU collateral, keeping financing costs high.

Building a normally functioning market for the most important commodities of this century can’t be done by pushing on just one line. Measurement, standardization, contract structure, delivery quality, and liquidity must move forward in sync—until then, no one can truly say how much a GPU hour is worth.