How long will it take for the GPU futures market when computing power is commoditized?

Author: Caleb Shack, Alana Levin

Compiled by: Jiahua, ChainCatcher

At Variant, we are passionate about exploring emerging markets. Emerging asset classes, financial products, asset issuance, expanded market access, and novel participation methods are deeply rooted in our founding DNA.

Recently, we have been contemplating the market built around computing power.

Acquiring computing power is a vast and growing field, which can be said to have the conditions for further financialization.

However, the supply and demand dynamics of computing power are highly complex, opaque, and constantly evolving. There are still many unresolved mysteries regarding market timing, structure, and even the specific assets being traded.

In the process of debating and discussing these issues, we hope to share an emerging analytical framework as a window for thinking about the computing power market.

The birth of a new futures market typically requires the following five prerequisites:

• Fragmented supply side
• Continuous price volatility
• Some form of physical settlement infrastructure
• Standardized, tradable units
• Lack of alternatives for price discovery or hedging

Our framework examines the current landscape of the computing power market across these five dimensions. We use historical analogies to explain the importance of each dimension and predict when the market might reach a critical point of explosion.

Summary of Key Points

A quick glance at the framework reveals that the current computing power market lacks the maturity needed to sustain a robust futures market.

(That said, the market is vibrant, with many startups actively working to change the status quo; if you are doing this, please reach out to us!)

Here are our current scores for the computing power futures market across the five dimensions:

• Fragmented supply side: 🔴 Supply is highly monopolized by ultra-large cloud service providers
• Price volatility: 🟢 GPU prices are highly volatile
• Physical settlement infrastructure: 🟢 OTC brokers have established physical settlement infrastructure
• Standardization: 🔴 Computing power lacks standardized, tradable units
• Lack of alternatives: 🟡 Vertically integrated suppliers can hedge internally, while other participants are forced to go long

1. Fragmented Supply Side (Computing Power Score: 🔴)

The futures market is a mechanism for price discovery.

Under monopolized supply, price discovery loses its necessity, as prices are determined by a few large suppliers, eliminating any pricing uncertainty.

Historically, this situation is not uncommon.

Oil futures only grew after the power of supply-side cartels (like the "Seven Sisters," the seven multinational giants that dominated global oil in the mid-20th century) weakened.

The electricity market formed only after the government relaxed regulations, broke monopolistic pricing, and allowed independent producers to enter the market. The fragmentation of the supply side drove the futures market to become an important venue for price discovery.

Looking at today's computing power dynamics, the supply side appears relatively concentrated.

The four major cloud giants (such as AWS, Azure, GCP, Oracle) control about 78% of the global self-built critical IT power capacity and about 69% of the H100 supply (calculated based on the assumption that there are 12.4 million H100s by Q4 2025).

From this, we infer that they also dominate the global supply of computing power hours. The supply is not fragmented.

Nevertheless, we are still considering factors that might change this dynamic.

New cloud providers are emerging. New chip architectures create opportunities for other suppliers to gain market share.

Some long-term contracted capacities from major labs may ultimately go underutilized, meaning these labs could eventually turn into suppliers or sellers of computing power in the market.

Therefore, while we are uncertain about the future level of concentration, the current judgment is that the development direction of the market supply side will become more fragmented than it is now.

2. Price Volatility (Computing Power Score: 🟢)

Ornn H100 Index on Bloomberg Terminal

Another prerequisite for the futures market is that the underlying asset must have high volatility.

Without significant price uncertainty, hedgers lack the motivation to guard against volatility risk.

Volatility also attracts speculators, who can profit from large price swings. If the market is stable or predictable, speculators will turn their attention to other markets.

We saw this in the oil market of the 1950s.

At that time, due to an oversupply of oil, the price set by the Soviet Union was lower than that of the "Seven Sisters." The "Seven Sisters" then lowered prices in the region without notifying the Middle Eastern oil-producing countries.

The resulting chain reaction led to the nationalization of Middle Eastern oil, the formation of OPEC, and increased global oil price uncertainty. Subsequently, the volatility of oil triggered fluctuations in the electricity market in the 1970s.

The pricing of computing power has been volatile in the past and will continue to be so in the future.

The speed at which new supply enters the market is uncertain. New chips or data center architectures may improve token efficiency for specific tasks. Demand continues to surge and expands in unpredictable ways.

We are very confident that this prerequisite is now in place.

3. Physical Settlement Infrastructure (Computing Power Score: 🟢)

For the market to operate efficiently, buyers must be confident that they can receive and consume the underlying instrument on the specified date and time.

This requires supporting infrastructure: aggregating supply, ensuring reliable delivery, clearing transactions, handling collateral, and managing settlements. These tasks are typically undertaken by intermediaries or brokers.

In the electricity market, these tasks are handled by independent system operators, which act as neutral third parties and function as quasi-government entities.

Currently, there is no fully equivalent role in the computing power market, but our hypothesis is that computing power brokers or OTC desks are beginning (and increasingly inclined) to take on many of these functions.

Today, brokers are building indices and data aggregation tools around computing power purchase and leasing agreements to anchor market prices.

Ornn and Silicon Data have begun publishing price data for data center-grade GPUs.

The broker community is also forming consensus on contract agreements, similar to how SAFE agreements standardized early financing terms. These tools enhance the underlying physical settlement infrastructure—before this, much of this coordination was largely confined to group chats.

We give the physical settlement infrastructure a green score because it lays the foundation for price discovery.

However, compared to mature spot markets, it is far from perfect. These purchasing activities occur at the infrastructure level, and not all market participants have the right to publicly resell after purchase. We are closely monitoring the progress of new market creation at this level.

4. Standardization (Computing Power Score: 🔴)

A major challenge faced by new commodities is often the uniqueness and irreplacability of their units.

Too many variables can disperse liquidity across numerous markets or lead to excessive basis risk, making it impossible to meet most hedging and delivery needs.

For example, crude oil is measured based on density and sulfur content, which varies by origin.

NYMEX found a product market fit with its WTI index (light sweet crude) because it locked in a standard that could serve the global upstream market and was even used by downstream markets (like airlines) for hedging.

Electricity is standardized regionally, taking into account supply and demand fluctuations due to factors like temperature and population density.

The computing power market lacks a level of standardization that can meet general hedging needs.

The challenge is that one H100 instance is not always equivalent to another H100 instance.

Factors such as region (and local power input), full machine configuration (i.e., hardware and network components), and duration (i.e., contract length) exacerbate the pricing differences of GPU instances.

However, we have seen early signs of standardization, especially when demand comes from the long tail (i.e., non-frontier labs) for inference.

Unlike training, the subtle differences required for inference workloads are much fewer, and they can run in distributed rather than colocated deployment environments.

If inference supply is dispersed among many suppliers, for example, open-source weight models increasing market share, standardization may emerge.

5. Lack of Alternatives (Computing Power Score: 🟡)

This is a subtle yet often overlooked point in the market formation process.

The establishment of the futures market is to serve hedgers. If there are alternatives with sufficient liquidity and negligible basis risk, then alternative contracts will go unnoticed.

A textbook example is the lack of adoption of aviation fuel futures—because WTI and other upstream indices have fully met the demand.

In the electricity-related field, temperature-based futures failed because market participants found that hedging the price fluctuations (electricity) was more efficient than hedging their cause (temperature).

Today, model providers hedge computing power risks through long-term leasing agreements or joint ventures, which often take the form of "pay-as-you-go," exchanging spot price risk exposure for counterparty risk.

Ultra-large cloud service providers typically physically own the GPUs they deploy.

On the other hand, long-tail suppliers lack the contractual leverage to secure favorable leasing terms and lack the capital to build their own vertical infrastructure, thus bearing the brunt of spot market volatility.

From a market perspective, there are no alternatives; however, participants controlling supply can hedge internally through vertical integration.

Overall Judgment

From a comprehensive scorecard perspective, it may be too early for computing power to support a robust futures market.

This market has the volatility to attract speculators and the early settlement infrastructure to support trading, but it lacks the supply fragmentation and standardization necessary for true price discovery on a large scale.

Most trading occurs on the OTC side.

Brokers are building price sources, Ornn and Silicon Data are publishing indices, and group chat trading is being standardized into contract templates.

This is not without significance, but it has yet to mature into a market like WTI or PJM. Trading volumes are too small, contracts are too customized, and supply is too concentrated, leading to existing infrastructure being unable to clear on a large scale.

The correct way to interpret this framework is as a diagnostic tool rather than a final conclusion. It tells us what is missing rather than what is impossible.

Unresolved Mysteries

The market will evolve in ways we are currently uncertain about.

We have many unresolved mysteries and some preliminary hypotheses. These hypotheses are tentative and require further validation or refutation. Below, we will outline the strongest arguments for these hypotheses.

▍Will the supply side of the market become more fragmented or more concentrated in the next 1-2 years?

We expect moderate fragmentation.

New cloud providers are bringing new capacity online faster than any other category.

As electricity becomes a core constraint, new regions are being activated, benefiting operators who can establish capacity near cheap electricity (rather than near existing ultra-large cloud service provider footprints).

Fortune 2000 companies are even supporting small-scale data centers. Expansion in this area seems inevitable.

However, standard business models rely on large, long-term contracts with reliable counterparties (such as ultra-large cloud service providers and frontier labs).

Cloud brokerage service providers like Hyperbolic and SF compute are going against the grain, offering capacity billed by the hour.

These companies serve the long-tail computing power needs of AI-native startups, application layer companies running inference on open-source weights, and research labs without frontier-level budgets.

We believe that the adoption of open-source weights will particularly lead to further fragmentation of computing power capacity—because supply will "de-verticalize" from frontier labs and ultra-large cloud service providers.

▍How will standardization unfold?

Index providers are establishing standards around the hourly cost of GPU instances.

These data sources represent rough estimates rather than precise prices.

Instance prices vary due to numerous factors, including region, full machine configuration, and duration, making standard pricing difficult to achieve.

The differentiation of full machine configurations is particularly pronounced, resulting from data centers customizing for tailored workloads and ultra-large cloud service providers optimizing for ecosystem lock-in rather than market uniformity.

Standards emerge when there is unified market demand.

The WTI standard gained adoption because it served a wide range of downstream refining products like gasoline, diesel, and aviation fuel.

Today, computing power demand is driven by AI training and inference workloads.

Training infrastructure is customized, optimized for long, compute-intensive tasks in large centralized facilities, making the underlying computing power instances nearly irreplaceable.

On the other hand, inference infrastructure requires simpler hardware specifications and less energy consumption; it is optimized for latency, meaning the infrastructure is distributed across different regions rather than colocated.

Inference is homogeneous, expected to account for over 65% of AI computing power demand by 2029. We speculate that optimizations around the computing power infrastructure serving this market will lead to a convergence of computing power requirements among suppliers.

If chip-level instances still have differences, other avenues for standardization may be hardware-level benchmarking.

NVIDIA created the MLPerf benchmark to score inference and training performance across various model architectures.

Under this conception, the basis for trading GPU instances is not their hardware specifications but the quality and efficiency of their output.

▍What might hinder the emergence of standards in the next 1-2 years?

We believe that "walled gardens" and customized workloads will stifle attempts at standardization.

In the next 1-2 years, ultra-large cloud service providers and frontier labs will strive to maintain their dominance in AI infrastructure and model provision.

If the two are not decoupled, they will maintain hardware based on their own needs, which differ from company to company. The adoption of new chip architectures will further shatter hardware specifications, making standardization difficult.

▍How will open-source weights gain meaningful application?

This is the simplest path to the formation of the computing power market.

The two core bottlenecks these markets face today are concentrated supply and lack of standardization.

The widespread adoption of open-source weights democratizes the ability to run inference.

This, in turn, creates incentives for the formation of independent operators and promotes infrastructure optimization tailored to these specific models.

We have seen a similar story in Bitcoin mining: open-source software has spawned numerous miners and driven standardization around hardware configurations.

So far, open-source weights have lagged behind closed-source weight models in performance.

But if this trend continues, open-source weights will soon reach the performance thresholds we see today in closed-source models.

Companies have already begun to embed closed-source models extensively in their systems and have witnessed significant productivity boosts. In three months, models that can similarly enhance productivity may cost only a fraction of what they currently do.

However, most companies may still prefer to choose the highest-performing models.

We believe that one day, frontier closed-source models will become too expensive for the tasks they undertake, and companies will optimize intelligent configurations between different models.

It is important to remember that frontier labs are currently providing inference services at a loss, and they will ultimately have to raise prices to sustain operations. At that point, open-source weights will have their moment.

▍What will be the final unit of pricing for transactions?

Computing power can roughly be broken down into three layers: chip, chip instance hours, Token.

Chip level ------ Supply is highly concentrated.

ASML monopolizes the photolithography machines used by TSMC, TSMC monopolizes the chip foundries used by NVIDIA, and NVIDIA monopolizes frontier chip design.

Moreover, chips are only useful when connected to power and maintaining high uptime. This leads us to believe that individual, deliverable chips will not become the final unit of pricing.

Chip instance hours level ------ Refers to the time period during which a chip can be practically used.

This can be said to be the most valuable state of the chip and is the core layer discussed in this article.

At this level, as long as there is sufficient demand around computing power resources, the performance of computing power as a commodity will resemble that of electricity.

We envision that computing power will be traded in a manner similar to electricity and other utilities: achieving standardization in regional contracts (computing power is a function of electricity) and layering spot and futures markets for hedging on top of that. This is achievable under the "chip instance hours" format.

Token level ------ Is a downstream product of computing power instances and may also become the final unit of pricing.

If Tokens are the primary driver of computing power instances, then the Token market will provide a way for the demand side to hedge costs and allow the supply side to lock in revenue.

The supply side can hedge costs through ongoing long-term contracts or vertical integration while maintaining concentration.

However, Tokens are not standardized across different models. Each model has its own text segmentation standards and produces varying outputs, making them not fully interchangeable between different use cases. Nevertheless, we are closely monitoring developments in this area.