GPU Pricing: A Free-for-All Without a Referee
- Core View: The current GPU computing power market is far from mature, lacking unified price benchmarks and standardized contracts. This leads to market information chaos, inefficiency, and fosters behaviors like hoarding and fragmented secondary markets.
- Key Elements:
- Significant Divergence in Price Indices: The quotes for H100 from four index providers on the Bloomberg Terminal show huge discrepancies, with inconsistent directions and rhythms, indicating a lack of consensus benchmark in the market.
- Divergent Price Trends for Different Contract Types: The on-demand, spot, and one-year reserved pricing for H100 show different magnitudes and timing of increases. Aggregate indices mask these structural differences.
- Larger Price Spreads in the B200 Market: The on-demand price for B200 surged sharply in the short term, but the starting points of different indices differed by over $2, with varying price patterns, highlighting market opacity.
- Dual Deficiency in Infrastructure: The market lacks both a reliable physical distribution network and financial infrastructure such as standardized financial contracts and transparent benchmarks.
- Scarcity Not Effectively Transmitted to Price: Despite on-demand capacity being practically sold out, price increases have been slow. This indicates suppliers lack real-time pricing intelligence, and the market cannot clear effectively.
- Seven Core Market Problems: These include no consensus benchmark, non-standardized contracts, unguaranteed delivery quality, illiquid contracts, and the absence of a forward curve, among others.
- Spawning Informal Secondary Markets: Market opacity leads to hoarding. Tenants have already begun subleasing their GPU clusters, much like subletting apartments.
Original Author: David Lopez Mateos
Original Compilation: Shenchao TechFlow
Introduction: The media loves to summarize the rise and fall of GPU computing power prices with a single number, but the reality is: the quotes from four index providers on the Bloomberg terminal deviate from each other by over $2, with inconsistent directions and rhythms. The author of this article is David Lopez Mateos, founder of the GPU computing power trading platform Compute Desk. Using firsthand transaction data, he deconstructs the real pricing structure of H100 and B200, revealing a primitive market with no consensus benchmark, no standard contracts, and no forward curve—computing power is being hoarded and sublet like short-term rental apartments.
Media headlines would have you believe GPU computing power prices are skyrocketing. This narrative is comfortable, perfectly fitting into the macro framework of "supply crunch + bottomless AI demand," and it implies something reassuring: we have a well-functioning market with clear, readable price signals.
But we don't. This narrative is almost entirely built on a single index, implying something it shouldn't: that the GPU rental market has become efficient enough to be represented by a single number for its global state.
The supply crunch is real, but the crunch felt by different people is completely different—depending on who you are, where you are, what contract you're trading, and what computing power asset. Faced with this opacity, the market's natural reaction is not orderly price discovery, but hoarding: locking in GPU time you might not even need yet, because you're not sure if you can buy it at any price next month. Where there is hoarding and no transparent benchmark, fragmented secondary markets emerge. At Compute Desk, we have already facilitated tenants subletting their clusters like apartments during major events. This is not a hypothesis; it is happening.
Indices Not Converging
In mature commodity markets, indices built on different methodologies tend to converge. Brent crude and WTI have a few dollars of spread due to geography and crude quality, but they move in sync directionally (Figure 1). This convergence is a hallmark of an efficient market.
Caption: Comparison of Brent and WTI crude oil price trends, showing high directional consistency
Currently, there are three GPU pricing index providers on the Bloomberg terminal: Silicon Data, Ornn AI, and Compute Desk. SemiAnalysis just publicly released a fourth—a monthly H100 one-year contract price index based on survey data from over 100 market participants. Silicon Data and Ornn publish daily H100 rental indices, Compute Desk aggregates data at the Hopper architecture level, and SemiAnalysis captures negotiated contract prices rather than listed or scraped prices. Different methodologies, different frequencies, different perspectives on the same market. Overlaying them reveals clear divergence (Figure 2).
Caption: Overlay comparison of four GPU indices, showing significant divergence in price levels and trends
Where Exactly is the Price Increase Happening
Using Compute Desk data, we can break down H100 price movements by supplier type and contract structure, and overlay Silicon Data's SDH100RT index (Figure 3). All indicators show prices rising, but the starting points and magnitudes vary greatly depending on the index and contract type.
Caption: H100 price trends broken down by contract type, overlaid with the SDH100RT index
Compute Desk's H100 neocloud data tells a more specific story than the aggregate index. On-demand pricing remained relatively stable throughout the winter, around $3.00/hour, then spiked sharply to $3.50 in March. Spot pricing was noisier and lower, showing only a slight upward trend until March. Silicon Data's SDH100RT shows a smoother, steady rise, increasing from $2.00 to $2.64 over the same period. The two indices remain at different price levels and describe different time rhythms: Compute Desk talks about a March jump, Silicon Data talks about a slow climb.
One-year reserved pricing was mostly flat until February, then surged from $1.90 to $2.64 at the end of March—not a gradual catch-up, but a sudden repricing. This looks more like suppliers collectively adjusting contract rates after the on-demand market tightened, rather than being driven by sustained structural demand.
The March story for B200 is even more dramatic (Figure 4). Compute Desk's on-demand index exploded from $5.70 to over $8.00 within weeks. Silicon Data's SDB200RT soared from $4.40 to $6.11 before retreating to $5.47. Both indices recorded this move, but the starting points differed by over $2, and the shapes of the rise and fall were different. With less than five months of data, fewer suppliers, and larger spreads for B200, the two indices are observing the same event through very different lenses.
Caption: B200 on-demand vs. reserved price trends, Compute Desk and Silicon Data data overlaid
Infrastructure Problems, Not Just Geographic Differences
Commodity markets have basis differentials. Appalachian natural gas is a textbook case: massive reserves sit atop structurally constrained pipeline capacity, with utilization in the Pennsylvania-Ohio corridor often exceeding 100%, and new projects like the Borealis Pipeline not coming online until the late 2020s.
The GPU market has a similar situation: an H100 in Virginia and an H100 in Frankfurt are not the same economic good. But geographic differences alone cannot explain why indices measuring the same market diverge so much. The dislocation in the GPU market is deeper than in Appalachian natural gas. The natural gas problem is a single missing link: pipeline capacity connecting supply and demand. The infrastructure gap in the computing power market exists on both the supply and demand sides. Physical infrastructure—the consistent networks, predictable configurations, and predictable availability needed for reliable compute distribution—is immature and sometimes simply doesn't work. Financial infrastructure—standardized contracts that compress spreads despite physical differences, transparent benchmarks, arbitrage mechanisms—also does not yet exist.
The data tells one story. The real, stinging experience of trying to procure computing power in early 2026 tells another. On-demand capacity for all GPU types is virtually sold out. Finding 64 H100s is difficult: Compute Desk shows 90% of suppliers have zero on-demand cluster availability, and the reserved market isn't much better. In a well-functioning market, this level of scarcity would have pushed prices to a new equilibrium long ago. But it hasn't. This suggests suppliers themselves lack real-time pricing intelligence to adjust. Prices are rising, but too slowly to clear the market. The gap between listed prices and real willingness to pay is being filled by hoarding, subletting, and informal secondary market trading.
What Needs to Change
The current GPU computing power market has seven core problems:
No consensus benchmark. Multiple indices coexist with different methodologies and contradictory conclusions.
Aggregate narratives mask structure. A single "H100 price" number masks huge differences between supplier types and contract durations.
Lack of transaction-level data. In bilateral markets, the deviation between listed prices and actual transaction prices is very large.
No contract standardization. Most GPU rentals are bilateral negotiations with varying terms. Shorter, more standardized contract durations would improve liquidity and price discovery.
No delivery quality guarantee. Interconnect topology, CPU pairing, network stack, and uptime vary enormously. Buyers need to know the quality of the compute they are purchasing before committing.
Contracts lack liquidity. If demand changes during the reservation period, options are limited: either eat the cost or sublet informally. The market needs infrastructure to transfer or resell committed computing power, allowing capacity to flow to those who need it most.
No forward curve. Without the ability to price forward, there is no hedging. This is why lenders discount GPU collateral by 40%-50%, keeping financing costs high.
Building a properly functioning market for the most important commodity of the century cannot be advanced on just one front. Measurement, standardization, contract structure, delivery quality, liquidity—these must progress in sync. Until then, no one can truly say how much a GPU hour is worth.
