GPU has no "price": The four major indices are competing, and the computing power market is more chaotic than you think

Author: David Lopez Mateos

Translation: Deep Tide TechFlow

Deep Tide Introduction: The media likes to summarize GPU compute power price swings with a single number, but the reality is this: on the Bloomberg terminal, the quotes provided by four index providers differ by more than $2, and their direction and timing are also inconsistent. The author of this piece is David Lopez Mateos, founder of the GPU compute trading platform Compute Desk. Using firsthand trading data, he breaks down the true 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 hoarded and sublet like short-term apartment rentals.

Media headlines may make you think GPU compute power prices are surging. This narrative is comfortable—it fits perfectly into the macro framework of “supply crunch + AI demand bottomless pit,” and it suggests something reassuring: that we have a well-functioning market where price signals are clear and easy to read.

But we don’t. This narrative is nearly entirely built on a single index; it implies things that shouldn’t be implied: that the GPU leasing market is efficient enough to be represented by one number for the global state.

Supply scarcity is real, but the scarcity each person experiences is completely different—depending on who you are, where you are, what contracts you trade, and what compute assets you’re dealing in. In the face of this opacity, the market’s natural response isn’t orderly price discovery—it’s hoarding: locking in GPU time you might not need yet, because you’re not sure you’ll be able to buy them next month at any price at all. Where there is hoarding and no transparent benchmark, fragmented secondary markets emerge. At Compute Desk, we’ve already facilitated tenants subletting their clusters the way people sublet apartments during major events. This isn’t a hypothetical—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, can differ by a few dollars due to geography and crude oil quality, but they move in sync directionally (Figure 1). This convergence is a hallmark of an efficient market.

Figure caption: Comparison of Brent and WTI crude price trends, with highly consistent direction

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

Figure caption: Four GPU indexes overlaid show clear differences in both price levels and trends

Where did the price increases actually happen?

Using Compute Desk data, we can break down H100 price movements by vendor type and contract structure, and overlay Silicon Data’s SDH100RT index (Figure 3). All metrics show prices rising, but the starting points and magnitude differ dramatically across indexes and contract types.

Figure caption: H100 price trends split by contract type overlaid with the SDH100RT index

Compute Desk’s H100 new cloud (neocloud) data tells a more specific story than aggregated indexes. On-demand pricing stays relatively stable throughout the winter at about $3.00 per hour, then spikes sharply to $3.50 in March. 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 describe the timing differently: Compute Desk says it jumps up in March; Silicon Data says it climbs slowly.

One-year reserved pricing stayed largely flat before February, then at the end of March it surged from $1.90 to $2.64—not a gradual catch-up, but a sudden repricing. This looks more like vendors adjusting contract fee rates in a batch after tightening in the on-demand market, rather than a continuously structure-driven demand effect.

The March story for B200 is even more intense (Figure 4). Compute Desk’s on-demand index exploded from $5.70 to above $8.00 within a few weeks. Silicon Data’s SDB200RT jumped from $4.40 to $6.11, then fell 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 fall are also different. With less than five months of data, fewer vendors, and larger price spreads, the two indexes are observing the same event through very different lenses.

Figure caption: B200 on-demand vs. reserved price trends, with Compute Desk and Silicon Data data overlaid

Infrastructure issues—more than just geographic differences

Commodity markets have basis differentials. Appalachian natural gas is the textbook case: massive reserves sit on pipelines with structurally constrained capacity. Utilization along the Pennsylvania-Ohio corridor often exceeds 100%, and new projects like the Borealis Pipeline don’t come online until the late 2020s.

The GPU market has a similar situation. A single H100 in Virginia is not the same economic good as a single H100 in Frankfurt. But you can’t explain why indexes measuring the same market diverge so widely just by geographic differences. The misalignment in the GPU market runs deeper than that of Appalachian natural gas. The problem with natural gas is a single missing link: pipeline capacity connecting supply and demand ends. In the compute market, the infrastructure gap exists on both sides of supply and demand. Physical infrastructure— the consistency networks needed to reliably distribute compute power, predictable configurations, predictable availability— is not mature and sometimes simply doesn’t work. Financial infrastructure— even though physical differences can be compressed by standardized contracts, transparent benchmarks, and arbitrage mechanisms— also doesn’t exist.

The data tells a story. Early 2026’s real experience trying to procure compute power tells an even more painful story. On-demand capacity for all GPU types is actually sold out. Trying to find 64 H100 units is difficult: Compute Desk shows that 90% of vendors have zero available on-demand cluster capacity, and the reserved market isn’t much better. In a well-functioning market, scarcity like this would quickly push prices to a new equilibrium. But it doesn’t happen that way. This indicates vendors themselves lack real-time pricing intelligence to adjust. Prices are rising, but rising too slowly to clear the market. The gap between posted prices and actual willingness to pay is being filled by hoarding, subletting, and informal secondary market trades.

What needs to change

The current GPU compute market has seven core problems:

No consensus benchmark. Multiple indexes coexist, with different methodologies; the conclusions contradict each other.

Aggregated narratives obscure structure. A single “H100 price” number hides huge differences between vendor types and contract tenors.

Lack of trade-level data. In a bilateral market, the gap between posted prices and actual transaction prices is very large.

No contract standardization. Most GPU leasing involves bilateral negotiation with varying terms. Shorter, more standardized contract tenors could improve liquidity and price discovery.

No assurance of delivery quality. Differences in interconnect topology, CPU pairing, network stack, and runtimes are substantial. Buyers need to know what quality of compute they’re actually purchasing before making a commitment.

No contract liquidity. If demand changes during the reserved period, choices are very limited: either absorb the cost or sublet informally. The market needs the ability to transfer or resell already committed compute infrastructure so capacity flows to the people who need it most.

No forward curve. If you can’t price forwards, you can’t hedge. That’s why lenders apply a 40%-50% haircut to GPU collateral, keeping financing costs high.

Building a normally functioning market for one of the most important commodities of this century can’t be done by pushing forward on just one line. Measurement, standardization, contract structure, delivery quality, liquidity—these must advance together. Until then, no one can truly explain what one GPU hour is worth.