Who will control the pricing power of compute in the AI era?

Introduction

In the first half of 2026, the “compute capital markets” concept rapidly evolved from a niche idea into a new track that both Wall Street and Silicon Valley are betting on together. CME and Silicon Data announced the launch of the first compute futures; the New York Stock Exchange parent company ICE teamed up with Ornn and NATIVX in sequence to develop GPU compute futures; Architect, founded by Brett Harrison, the former U.S. president of FTX US, is trying to bring the perpetual contract structure—mature in the crypto market—into regulated compute trading. Meanwhile, CoreWeave’s financing scale using GPUs as collateral has surpassed $20 billion and has completed its first investment-grade rating for a GPU-backed financing.

Compute is evolving along the classic path of commodity finance: from capital expenditure assets used for internal corporate purposes, gradually moving toward spot trading, price indices, and futures hedging, and ultimately entering credit and structured financing markets.

Why We Need Compute: The Value Waterfall of the AI Industry

Before understanding the compute market, you need to understand compute’s position in the AI value chain. The entire chain can be broken down into a nine-layer waterfall: from a commercial value and cash flow perspective, demand starts at the bottom application layer and propagates upward; compute sits in the middle, connecting underlying hardware and data center infrastructure as well as the upper model and application layers.

Layer 1｜Chips and Hardware: NVIDIA, AMD, HBM/DRAM vendors. This is the bottom-most raw material of compute. GPUs determine the basic supply of usable compute, and storage resources like HBM/DRAM are also beginning to be financialized by the market.

Layer 2｜Electricity and Land: Whether a data center can be built depends not only on having GPUs, but on having suitable land and sufficient power access. A large portion of compute’s marginal cost comes from electricity bills, so in terms of commodity characteristics it is more like electricity than oil.

Layer 3｜Neocloud and Independent Data Centers: CoreWeave, Nebius, Lambda, GMI Cloud, Crusoe, etc. They buy GPUs, build clusters, and then rent compute to AI companies—essentially the “mines” and “oil fields” in the compute market.

Layer 4｜Aggregation and Brokerage Platforms: Mithril, Andromeda, SF Compute, etc. They may not own GPUs themselves, but help buyers find supply, standardize SLAs, match trades, and even run market making. They are more like traders in commodity markets, such as Glencore and Vitol.

Layer 5｜Indices and Benchmarks: Silicon Data, Ornn (OCPI), NATIVX (COIL). Without a credible pricing benchmark, it’s hard for the market to develop futures and derivatives. So this layer’s role is to turn previously opaque compute prices into market prices that are trackable and verifiable.

Layer 6｜Derivatives and Credit: CME, ICE, Architect, on-chain perpetual DEXs, as well as GPU-collateralized loans, compute ABS, and other tools. This layer’s function is to allow the market to hedge compute price risk, and to make GPU capacity an asset that can be financed.

Layer 7｜Inference Development Platforms: Fireworks, Baseten, Modal, etc. They package underlying GPUs, model deployment, and inference APIs, so developers don’t have to manage complex compute infrastructure themselves and can use model inference capabilities like accessing cloud services.

Layer 8｜LLM / Model Layer: OpenAI, Anthropic, xAI, DeepSeek, etc. They transform underlying compute into model capabilities and intelligent output. The model itself is the core middle layer connecting the underlying infrastructure and the upper-layer application experience.

Layer 9｜Application Layer: Cursor, Perplexity, Suno, Rime, etc. This layer directly faces end users, packaging model capabilities into specific products and use cases. It is an important entry point for the spread of AI demand and user payments.

This nine-layer waterfall illustrates a core fact: compute is a middle good in the AI economy. Below it connects chips, electricity, land, and capital expenditures; above it connects inference platforms, model companies, and the application layer.

Every time an AI application calls a model, it essentially consumes a small portion of upstream compute. And because compute sits in the middle of the value chain, with suppliers holding GPU and data center assets on one side and model companies, inference platforms, and application companies needing stable compute on the other, when price fluctuations become large enough and both sides’ risks move in opposite directions, compute naturally starts to be financialized.

Why We Need a Compute Market: Hedging Demand and Market Structure

Who Needs Hedging

Source: X @0xfishylosopher

The compute market’s hedging demand comes first from industry participants that hold real compute exposure, not from financial institutions. This follows the same logic as airlines hedging fuel prices, and power plants hedging electricity prices.

The underlying Neocloud and independent data centers, such as CoreWeave, Nebius, and Lambda, hold physical GPU assets. Their revenue comes from future rent. They worry about GPU rental declines, so they are natural sellers/shorts who need to sell forwards to lock in revenue.

The mid-layer inference development platforms, such as Fireworks, Baseten, and Modal, purchase compute upward and provide inference APIs and model deployment services downward. Compute is their key cost.

The upper-layer application companies, such as Cursor, Perplexity, Suno, and Rime, also need continuous procurement of inference capabilities. Inference costs directly affect gross margins. Therefore, the mid-layer and upper-layer are natural buyers/longs who need to buy forwards to lock in costs.

Ultra-large-scale cloud providers, such as Google, Amazon, and Microsoft, are somewhat special. They simultaneously own data centers, cloud platforms, models, and applications, and internally have already formed a degree of natural hedging.

Why Compute Is More Like Electricity, Not Oil

Compute is not a fully homogeneous commodity.

Even for the same one-hour H100/H200 capacity, value differs due to chip specifications, region, latency, network interconnects, cluster size, reservation windows, SLAs, data security, and the specific workload.

More importantly, compute cannot be stored. GPU hours that are not used today cannot be stored like oil so they can be sold next year. Therefore, in terms of commodity characteristics, compute is closer to electricity: it is time-sensitive, region-dependent, and highly reliant on local infrastructure.

This leads to three outcomes:

First, real compute transactions often require bilateral customization around specific SKUs and delivery conditions.

Second, for now, there is no unified, transparent price benchmark like WTI crude oil.

Third, indices and benchmarks become extremely critical. The core job of teams like Silicon Data, Ornn, and Compute Desk is to convert fragmented compute prices into trackable, hedgeable market signals.

Web3’s Previous-Generation Decentralized Compute vs the New Generation of Compute Traders

The compute market is not a completely new concept. In the previous cycle, Web3 projects like Akash, io.net, and Aethir already used the narrative of a “decentralized compute marketplace,” connecting globally idle GPUs via networks driven by token incentives.

But the question is: why didn’t most of the older-generation projects become the mainstream AI compute procurement layer, while newer players like Andromeda and SF Compute were able to win enterprise customers and generate U.S. dollar revenue in a short time?

The products being sold are different: decentralized supply vs deliverable capacity

The core logic of older Web3 projects was: connect scattered GPU access networks and incentivize supply through token rewards, so users can buy compute at lower cost.

They solve the “where there are GPUs” problem.

But what enterprise buyers truly care about is a different set of questions: is it H100/H200? Is there InfiniBand? Is the cluster size sufficient? Can it run stably for weeks or even months? Who is responsible for the SLA? Who compensates if there is a failure?

In other words, enterprises aren’t buying “GPUs located somewhere”; they are buying GPU capacity that guarantees delivery, is measurable, and is accountable.

Dispersed, heterogeneous, cross-carrier GPU supply may be useful for batch inference, rendering, or low-sensitivity tasks, but for large-model training and production-grade inference, stability, network conditions, and delivery responsibility are what matter.

The four structural problems of the previous generation

First, token incentives can bring supply, but don’t necessarily create real demand.
Token subsidies can quickly produce impressive numbers of nodes, GPUs, and network scale. But if the demand side mainly relies on token narratives rather than natural-paying customers, the result is often low utilization, low-quality revenue, and incentive-driven distortion of price discovery.

According to Messari’s “State of Akash Q1 2026,” Akash’s average GPU usage in the quarter fell 57.4% quarter-over-quarter to 84 units, and its average available GPU capacity fell 57.5% quarter-over-quarter to 249 units, indicating clear contraction on both the supply and demand sides of its GPUs. Under io.net’s early mechanism, nodes earn rewards as long as they stay online, regardless of whether GPUs actually execute effective work. Its token has also declined sharply from its historical highs until it launched a more demand-driven new incentive mechanism in June 2026.

Second, enterprise-grade SLAs are hard to be handled by protocols alone.
Enterprise customers need an invoice, a support channel, standard SLAs, a refund mechanism, compliance reviews, and legal responsibility. All of these require a clear commercial entity to take on—rather than relying purely on the protocol layer.

Third, AI workloads and distributed supply have a natural mismatch.
Large-scale synchronous training and production-grade inference require very high levels of GPU interconnect, NVLink/InfiniBand, cluster scheduling, fault recovery, and data security. A geographically dispersed, hardware-heterogeneous network is difficult to directly satisfy such high requirements workloads.

Fourth, token-denominated pricing doesn’t match enterprise procurement processes.
Enterprises are more accustomed to U.S. dollar contracts, invoices, budget approvals, and vendor management, and they don’t want to bear token price volatility, accounting treatment, and compliance uncertainty.

Important exception: Aethir

Aethir is an exception.

Aethir’s revenue exceeded $127 million in 2025, with more than 150 paying enterprise customers and 430k GPU containers, covering high-end GPUs such as H100, H200, B200, and B300. Based on its self-disclosed figures, its revenue scale already exceeds Andromeda’s approximate $100 million run-rate and is far higher than SF Compute’s.

Aethir’s path is more like placing Web3 tokens and network effects in the capital structure and ecosystem incentive layer, while making the truly customer-facing portion more centralized, more standardized, and more enterprise-grade: centralized or semi-centralized clusters, explicit service commitments, USD-denominated contracts, and enterprise customer support and delivery responsibility.

Tokens can help with early-stage fundraising, incentive supply, and organizing networks, but they shouldn’t become the core interface that enterprises must face when procuring compute.

What’s new with the new generation of traders

The starting point for new players is not “build a decentralized network first,” but to directly target the AI buyers’ procurement pain points.

AI companies often need to sign long-term compute contracts, but their actual demand is volatile. SF Compute’s approach is for customers to buy long-term compute capacity financed by a third party, and then list any unused portions on the order book for resale or sublease. It doesn’t own GPUs itself; instead, it behaves more like a market built around secondary liquidity for compute contracts.

Andromeda is closer to a compute dealer: real-time pricing across 100-plus suppliers, validating performance, standardizing SLAs, and acting as the customer’s sole contract counterparty. Its value is not only matching trades, but also taking on procurement, delivery, and some credit intermediary functions for customers—so it also calls itself a “market maker for compute.”

Andromeda conducts principal transactions, holds or controls inventory, earns bid-ask spreads, and takes on SLA and delivery responsibility. SF Compute is more like a hybrid of exchange/broker: focused on agency matching and secondary liquidity, not necessarily holding underlying GPUs, and earning trading fees and benefiting from market network effects.

GMI Cloud needs to be categorized separately. It is not a typical broker/dealer; it’s more like a neocloud: it builds data centers, holds assets, and sells GPU cloud capacity. At the same time, it is also the user of GPU debt/funding—most of its funding in the Series A was debt financing—so it is closer to a Layer 3 compute producer.

What the market lacks most right now is not an even more decentralized ideal cloud, but a trading layer that can deliver H100/H200 capacity today, is responsible for SLAs, and helps buyers reduce long-contract risk.

Is there already a compute price discovery market?

The mainstream form of compute trading today is still OTC / bilateral customized trading. Public quotes are increasing market transparency, but more often they are only the starting point of price discovery rather than the final unified trading price.

Taking H100 as an example, observable quote ranges are already appearing in the market: Andromeda’s price is about $1.83/hour; SF Compute’s average price is about $2.03/GPU-hour; GMI Cloud’s starting price is $2.00/GPU-hour; Mithril’s H100 SXM5 8-GPU instance spot price translated is about $2.92/GPU-hour.

This means that public-market H100 quotes generally fall in the $1.8–3.0/GPU-hour range. But these prices can’t be directly compared one-to-one because the underlying delivery conditions are not the same. The GPU form factor, region, network interconnect, cluster size, lease duration, SLAs, and workload types will significantly affect the final transaction price.

Therefore, what enterprises truly procure is usually not an abstract “H100 hour,” but a capacity contract designed around specific SKUs, regions, terms, cluster configurations, and delivery conditions. In other words, web quotes make compute prices more visible, but the core of actual trading in the current market is still highly customized OTC contracts.

Ornn: Trying to become the index layer of the compute market

Source: Ornn

Ornn’s core positioning is not simply selling compute, but building the price infrastructure for a compute financial market. Its Ornn Compute Price Index (OCPI) tracks real-time spot transaction prices for GPUs compute such as H100, H200, B200, and B300, and organizes these prices into indices that can be used for pricing, hedging, and settlement. Ornn’s official website states that OCPI is a compute reference price and is used for pricing, hedging, and settlement in the compute derivatives market.

This means Ornn aims to create “Platts/Argus/WTI-style” benchmarks for the compute market: first normalize scattered and non-standard GPU rental prices, and then allow the market to trade forwards, futures, or perpetual contracts around this benchmark.

Ornn’s roadmap can roughly be understood as three steps:

Step one: establish the spot price index, i.e., OCPI.

Step two: license OCPI to exchanges and derivatives platforms to become the contract settlement price.

Step three: develop futures, perps, hedging, and lending financial products around the index.

Architect: Bringing the perpetual contract structure into institutionalized compute trading

Architect is a type of player in the compute derivatives market that is more oriented toward a trading venue. It was founded by Brett Harrison, former president of FTX US. Its institutional trading platform AX collaborates with Ornn and plans to launch exchange contracts based on GPU rental prices and DRAM prices.

Mechanistically, Architect does not deliver real H100/H200 compute. Instead, it gives traders financial exposure to GPU rental prices and memory prices through contracts that track the Ornn compute index. Its product is closer to the perpetual contract structure in the crypto market: traders trade index-linked contracts using margin, and the contract price is then kept as close as possible to underlying GPU rental prices via index anchoring and funding-rate mechanisms.

Therefore, Architect’s significance is to introduce the crypto-native perpetual contract mechanism into more institutionalized, regulated compute trading scenarios. It is more like the derivatives trading layer in the compute market, while Ornn provides the index layer as a price benchmark.

Lighter: On-chain perpetual contracts provide early tradable price discovery

Lighter is more like an on-chain early compute perp venue. The platform has already launched $H100, allowing users to trade H100 compute price exposure with up to 10x leverage; this product tracks the Ornn H100 Compute Price Index.

The significance of this kind of product is that it lets the market form continuous on-chain price signals for GPU rental prices for the first time. It does not solve the real GPU delivery problem, nor is it the main channel for enterprise compute procurement, but it can serve as an early market for speculation, hedging, and price discovery.

Mechanistically, it is closer to perpetual contracts in the crypto market: traders do not actually settle for H100 compute delivery; they trade a contract that tracks the H100 index. The contract price is anchored via the index and funding-rate mechanisms.

Its advantages are fast go-live, low participation barriers, and support for around-the-clock trading. Its disadvantages are that liquidity may be thinner, and there is still basis risk between it and real enterprise-grade compute capacity contracts.

ICE × Ornn: A roadmap for a regulated futures market

ICE is a more traditional, regulated exchange route. In May 2026, ICE announced plans to partner with Ornn to launch a set of GPU compute futures contracts, with the underlying benchmark being the Ornn Compute Price Index. In its announcement, ICE explicitly mentioned that OCPI tracks live-traded spot prices of major hardware types such as H100, H200, B200, and B300. The related contracts are planned to be USD-denominated, cash-settled, and pending regulatory approval.

ICE’s mechanism is different from Lighter. Lighter is on-chain perpetual, suitable for quickly forming trading prices and speculative liquidity; ICE is a regulated futures market, more suitable for institutional participation, clearing, risk control, and compliant hedging.

But ICE’s contracts are not for physical delivery; they are cash-settled. That means traders won’t truly deliver or receive H100 capacity. Instead, they settle profit and loss based on indices such as OCPI. This reduces delivery complexity, but also means whether the contracts succeed depends on whether the index is credible enough and resistant to manipulation, and whether it can represent real market prices.

Market Outlook

Three directions worth tracking

Institutionalization at the OTC desk
The end state of the compute market may not be that industrial players directly trade futures on exchanges. It may more likely be that dealers absorb customized demands from industry players, and then manage risk through indices, futures, or perpetual contracts. In the next 12–24 months, what’s most worth watching is whether players like Andromeda and SF Compute can upgrade from “compute procurement platform” to a truly “compute trading desk”: on one hand handling spot and reservation demand at the SKU level; on the other hedging inventory and basis risk in the index market. Whoever completes this step first has the opportunity to become the core intermediary in the compute market.

A closed loop of credit and derivatives
If “GPU-collateralized financing + futures hedging” can work, lenders can manage GPU price volatility and residual value risk more effectively, thereby reducing haircuts and financing costs. This would directly improve the capital efficiency of AI infrastructure, and is one of the most important implications of financializing compute for the real AI industry.

Formation of price benchmarks and a clearing system
For compute to become a truly tradable and financeable asset, it must first form credible price benchmarks and clearing mechanisms. Index providers such as Ornn, Silicon Data, and NATIVX, as well as trading venues such as ICE, CME, Architect, and Lighter, are not only competing for a single product opportunity, but rather for the entry point to pricing power in the future compute market.

Open questions

Regulatory approval

Products from CME, ICE, Architect, and others still need to pass regulatory approvals. How compute will be defined—whether as a commodity, a service, or a new kind of tradable resource—still lacks clear precedents.

The underlying spot market is still relatively thin

The credibility of an index depends on the depth of real spot transactions. Today’s public spot and secondary circulation markets are still in early stages, and the vast majority of compute trading is locked into long-term contracts between hyperscalers, neoclouds, and AI companies. Insufficient underlying spot activity may affect the index’s representativeness and resistance to manipulation.

Cycle risk
If AI capital expenditure slows down, spot liquidity could shrink before derivatives markets mature. Meanwhile, GPU rental rates have already fallen meaningfully from their peak, and there is a lack of sufficient historical data for GPU residual value and depreciation curves, which will further amplify uncertainty in credit assessment and derivatives pricing.

Reference

About Gate Ventures

Gate Ventures is Gate’s venture capital arm, focusing on investments in decentralized infrastructure, ecosystems, and applications. It aims to reshape the world in the Web 3.0 era. Gate Ventures partners with global industry leaders to empower teams and startups with innovative thinking and capabilities, redefining how society and finance interact.
For more information, visit: Official website | X | Telegram | LinkedIn | Medium

Disclaimer :

This content does not constitute any offer, solicitation, or advice. You should always seek independent professional advice before making any investment decisions. Please note that GateVentures may restrict or prohibit all or part of the services from restricted regions. Read the user agreement for more information; link: *.