Pantera Capital: 4 Major Opportunities in the Computing Power Market

Author: Jay Yu, Partner at Pantera Capital

Compiled by: Jiahua, ChainCatcher

Introduction

Spending on computing power and data center infrastructure has become a trillion-dollar market, comparable to U.S. consumer spending, and is gradually becoming an important pillar of U.S. economic growth.

However, as AI Agents gradually enter the real economy, the demand for computing power continues to grow, yet the procurement methods for GPUs remain quite traditional. Despite the emergence of computing power markets such as SF Compute, Vast AI, and Runpod, the buying and selling of a large number of GPU nodes still occur through group chats, over-the-counter brokers, and customized bilateral agreements.

Currently, GPU computing power is still in the early stages of financialization. Similar to electricity, computing power is not a completely homogeneous asset; it is constrained by factors such as chip models, delivery times, and geographical locations. However, in the next 5 to 10 years, computing power is likely to gradually become a commodity like electricity or oil, becoming a key resource supporting the operation of the global economy in the AI era.

This article will discuss how the computing power market can achieve this transformation, including insights from the electricity market, potential structures of the computing power market, products being developed in the industry, and the opportunities and challenges present in the commoditization of computing power.

1. Finding Answers from the Electricity Market

1.1 Structure of the Electricity Market

To understand how the computing power market may evolve in the future, we can first observe its underlying resource, which is the operation of the electricity market.

Like computing power, electricity is also an asset with heterogeneity, temporal attributes, and hub attributes, and it is an important foundation for global economic growth.

Since the 1990s, the U.S. electricity system has gradually been privatized. The previously highly vertically integrated electric utilities began to dismantle the segments of generation, transmission, and distribution.

Independent grid operators started to manage the electricity systems in different regions of the U.S. At the same time, large interconnected electricity systems such as PJM-West (East Coast), ERCOT North (Texas), and CAISO SP15 (Southern California) eventually became commodities that could be traded on exchanges like CME and ICE, forming important price benchmarks for the entire electricity asset class.

On a physical level, the electricity market can be roughly summarized as a structure of "grid, operators, nodes."

At the top level are mutually independent regional grids. Under each regional grid, there are a number of system operators, including Regional Transmission Organizations (RTOs) and Independent System Operators (ISOs). These operators are responsible for managing multiple substation nodes in their local areas.

For example, in San Francisco, a city may be powered by multiple substation nodes, such as Embarcadero, Larkin, and Mission.

The electricity price at each node is not exactly the same but is determined by an algorithm called "Locational Marginal Pricing" (LMP). This algorithm takes into account generation scheduling, user demand, transmission costs, and the laws of electricity flow to calculate the optimal price under constraints.

The LMP algorithm becomes a key link connecting the physical delivery of electricity and tradable electricity benchmarks such as PJM-West.

The hub prices and index prices in the electricity market are essentially weighted averages of the prices at multiple terminal LMP nodes. The largest and most liquid price benchmarks ultimately achieve stable trading depth on exchanges like CME and ICE, becoming reference prices for the entire electricity asset class.

1.2 Potential Structure of the Computing Power Market

Observing the development of the electricity market over the past 30 years allows us to make several speculations about the computing power market.

First, the computing power market may form a structure similar to that of the electricity market.

In terms of physical delivery, the closest computing power version to the electricity market's "grid, operators, nodes" structure may be "hardware, suppliers, clusters."

Just as there are multiple mutually independent regional grids in the electricity market, the computing power market may also form multiple markets based on different hardware categories. H100, H200, B200, B300, and other chip categories may each have their own independent but interrelated price benchmarks.

At the operator level, computing power suppliers such as AWS, Nebius, CoreWeave, SF Compute, and Ornn will provide different pricing mechanisms for specific clusters. These clusters will also be influenced by time, location, and SKU, similar to different nodes in the electricity market.

As for the counterpart of LMP in the computing power market, it may be a scheduling or routing algorithm. This algorithm dynamically generates computing power instance prices based on the availability of a certain SKU.

The benchmarks that ultimately become industry reference prices in the electricity market, such as PJM-West, ERCOT-North, and CAISO SP15, are typically connected to the most liquid physical delivery infrastructures. This is similar to other commodity markets like oil.

As more price benchmarks emerge in the computing power field, the ones that ultimately prevail may be those connected to high liquidity physical delivery locations.

Additionally, in the electricity market, generators and electricity-consuming enterprises sometimes directly hedge in electricity exchanges, but most transactions are still completed by over-the-counter (OTC) trading departments operated by banks and trading companies. These OTC institutions exchange node settlement prices with hub settlement prices, earning the price difference and hedging their own risks.

A similar market structure may also appear in the computing power market. The actual supply and demand sides of the computing power market, namely new cloud service providers and AI companies, may prefer to procure specific SKUs through brokers rather than directly hedging in computing power exchanges.

Finally, the computing power market may face more severe basis risk than the electricity market.

According to the U.S. Federal Power Act, wholesale electricity prices must remain transparent under the supervision of the Federal Energy Regulatory Commission (FERC). However, the computing power market currently does not have similar transparency requirements.

This means that the dynamic prices and indices that can be established in the computing power market may only be based on its own order book and the order book data provided by new cloud service providers it partners with. These collaborations may be achieved through revenue sharing, data procurement, and other means.

2. Structure of the Computing Power Market

2.1 Three-Tier Structure of Inference Demand

The electricity market exists because the industrial system has a continuous demand for electricity. Similarly, the fundamental driving force of the computing power market is the demand for Tokens, especially inference Tokens.

The current inference industry chain can be roughly divided into three tiers, which together constitute the structural supply and demand sides of the computing power market.

New Cloud Service Provider Layer

Companies like Nebius and CoreWeave operate physical data centers, forming the sellers of GPUs.

On-Demand Layer

Developer platforms like Fireworks and Baseten transform bare-metal GPU environments into more complete GPU usage environments, allowing developers to run tasks directly or obtain inference Tokens at any time. This layer represents the buyers of GPUs.

Application Layer

Applications like Cursor, Perplexity, and Rime use inference platforms to provide final products for users and enterprises. This layer is the demand side for Tokens, and the demand for Tokens translates into demand for GPU computing power.

Although each company in the ecosystem has different GPU management strategies, overall, new cloud service providers are on the supply side of the GPU market, representing a structural short; the on-demand layer and application layer are on the demand side, representing a structural long.

On the other hand, large-scale cloud service providers like Amazon and Google operate their own products across all three tiers.

As for the flow of profits in the industry chain, a rough rule of thumb is that when top-tier applications spend $100 on Tokens, approximately $45 flows to the on-demand layer, $50 flows to new cloud service providers or the GPU layer, and the remaining $5 flows to routing layers like OpenRouter.

2.2 Structure of the Computing Power Capital Market

The Token economic structure of various participants in the AI industry chain will influence how the computing power capital market ultimately forms.

One possible structure is that the demand-side developer platforms and application layers, along with the supply-side new cloud service providers, will trade specific SKUs through computing power brokers and OTC trading platforms like SF Compute, Runpod, and Compute Exchange.

Subsequently, computing power brokers and OTC trading platforms will manage inventory and bear the basis risk between specific SKUs required by end Token consumers and the general H200.

They can hedge through computing power trading platforms like Architect or Pluto.

The prices on these trading platforms are established based on a computing power benchmark. This benchmark is calculated as a weighted average of the order book prices from new cloud service providers and OTC platforms that collaborate with them.

2.3 NVIDIA: The "Central Bank" of the Computing Power Market

In addition, NVIDIA recently announced that it will allow the computing power from its AI factories to become an "investable asset class."

In the financialization system of computing power, NVIDIA can be seen as the "central bank" of the computing power market, as it almost monopolizes the GPU technology stack.

Central banks typically have several core objectives:

1. Manage inflation in the economy;

2. Promote full employment;

3. Act as a lender of last resort in extreme situations.

To some extent, NVIDIA is playing all three of these roles.

First is managing "inflation." The "inflation" in the GPU economy can be likened to the depreciation cycle of GPUs, especially in relation to the depreciation speed of the latest generation of chips. By controlling the pace of product releases, such as launching Vera Rubin, NVIDIA can influence the rate at which GPUs depreciate over time to a certain extent.

Secondly, it is about keeping GPU utilization high. High GPU utilization usually indicates strong market demand for tokens, leading users to purchase more GPUs. Therefore, NVIDIA is motivated to promote the financialization of GPU computing power, making GPU hour transactions more liquid, thereby further enhancing GPU utilization.

Finally, acting as a lender of last resort. NVIDIA has also announced that it will provide support for GPUs up to 25% of their remaining value. This can be seen as a form of rescue or insurance mechanism for GPU value: when new cloud service providers and other GPU suppliers face liquidity crises, such support can safeguard the residual value of GPUs.

3. Opportunities and Challenges in the Computing Power Market

3.1 Product Forms in the Computing Power Market

As the financialization of the computing power market continues to increase, several main product types may emerge in the industry. Each product type has its own advantages and challenges.

Product One: Physical Delivery

Physical delivery refers to the actual delivery of GPU devices to end users who need computing power. Projects like SF Compute, Hyperbolic, Vast, Runpod, and Compute Exchange fall into this product layer. This layer is the closest to the hardware itself and directly connects both sides of computing power supply and demand.

Due to the high heterogeneity of computing power SKUs, many projects initially acted merely as brokers, earning commissions by matching GPU demand with new cloud service providers. In the long run, the goal of this layer may be to establish some form of "computing power spot exchange."

However, continuously ensuring the quality of physical GPU delivery is a very challenging issue, especially for projects whose supply comes from decentralized networks. The computing power market may need a rating agency similar to Moody's to certify the quality of deliverable GPUs to promote further maturity of this layer.

Despite the intense competition in the physical delivery layer, it may have the most enduring moat in the long run. A review of the oil and electricity markets reveals that financialized systems typically develop around the most liquid delivery locations of physical goods, gradually forming indices, exchanges, and lending financial products.

Therefore, whoever can first solve the problem of physical GPU delivery may gain significant market returns.

Product Two: Indices

This layer mainly refers to the "index curves" built on computing power, including computing price indices launched by projects like Ornn, Silicon Data, Compute Desk, and Semianalysis.

Establishing indices is an important step for further financialization and trading of assets like computing power. It can be said that many current discussions in the market about the "computing power market" have begun to heat up as computing power indices gradually mature. Large exchanges like CME and ICE have also announced collaborations with relevant indices.

However, there is still a significant price gap between many computing power indices and the individual GPU market.

The reasons for this price gap may mainly be twofold.

First, there are differences in the underlying products themselves. Different computing power platforms often vary greatly in product stability, interruptibility, and contract terms.

Second, the order books behind different computing power indices are not the same, and the computing power market lacks the transparency requirements seen in other asset classes. For example, FERC requires transparency in the electricity market.

Currently, many computing power indices simply bulk purchase order book data from new cloud service providers or establish initial datasets through cooperation and revenue-sharing agreements. Additionally, the monetization ability of independently operating index layers is weaker than that of exchanges and physical delivery layers.

Therefore, although the index layer can generate noise and plays an important role in the computing power market system, it is simultaneously squeezed from both upstream and downstream:

• The upstream physical delivery layer controls the source of prices;

• The downstream exchanges hope to take a share of the revenue from the indices.

Product Three: Derivatives Trading Platforms

The third layer is derivatives trading platforms, which typically offer cash-settled futures products. These platforms provide hedging tools for market participants based on computing power indices. Liquid Compute and Architect are developing related products.

In the long run, derivatives trading platforms may be the most anticipated and easiest to scale monetization products in the financialization system of computing power. However, this field is still in its early stages, and trading volumes have not yet reached significant scales.

Moreover, the tradable products on cash-settled trading platforms are usually generic H200 instances, rather than specific SKUs that can be directly used for inference. Therefore, those truly participating in these trading platforms may primarily be OTC trading platforms and computing power brokers, rather than end enterprises in the AI industry chain. The former seeks to hedge the computing power inventory held on their balance sheets, while the latter needs to manage their own computing power exposure.

Product Four: Financialization Tools

There may also emerge a broader category of financialized products in the computing power field, supporting financing for data centers and new cloud service providers. For example, lending agreements, Vault, and synthetic stablecoins like USD.AI may help data centers and new cloud service providers complete infrastructure construction.

At the same time, the market may also see risk transfer and insurance layers to smooth basis risk in the GPU economy. Observing the different levels mentioned above, it becomes relatively clear what the complete form of the computing power market may ultimately take.

The winners will be full-stack participants capable of achieving reliable physical GPU delivery. They can form price indices from their own physical order books and then establish computing power futures exchanges, making computing power a hedgable asset class.

However, the real divergence in the industry lies in the order of appearance of different products. Will the ultimate winner start with physical delivery, establish indices first, or launch trading platforms first?

3.2 The Frontier of Computing Power Token Economics

So far, we have mainly discussed the computing power market at the GPU level. However, we can also broaden our perspective to observe the more general computing power token economics, including its historical evolution, methods of value capture, and its game relationship with frontier model laboratories.

Historically, the computing power market is not a brand new concept. From 2023 to 2024, projects like SF Compute and Hyperbolic have already been discussing this model, while decentralized projects like Akash and IONet are also exploring similar directions.

Interestingly, many projects ultimately did not remain at the GPU market level. For example, Hyperbolic later became an inference service provider, directly selling tokens; SF Compute began directly signing contracts and operating computing power clusters.

From the value chain perspective, value flows from AI applications like Cursor, Harvey, Granola to routing layers like OpenRouter, then to inference service providers like Fireworks and Baseten, and finally enters new cloud service providers and computing power markets like Nebius, CoreWeave, and SF Compute, ultimately flowing to data center operators.

If a company remains in the middle of the industry chain, it may be squeezed from both upstream and downstream. To maintain a more stable profit margin, the company may ultimately need to extend upstream or expand downstream.

The token market also presents a very interesting game theory perspective, with three unavoidable forces at play.

The first force is NVIDIA, which controls the supply and depreciation speed of GPUs.

The second force is frontier model laboratories like Anthropic and OpenAI, which continuously release new models, potentially rapidly increasing token demand.

The third force is open-source models like Kimi, GLM, DeepSeek, and Qwen, which exert downward pressure on inference profit margins.

These three forces together provide continuous momentum and pressure for both supply and demand sides of the GPU market and token market. Any one of these parties could rewrite the entire market's pricing curve by launching significant products.

For example, the release of GPT 5.6 could become an event determined by enterprises but difficult for the market to predict in advance, triggering a chain reaction along the AI token value chain and causing fluctuations in token prices and GPU hardware prices.

Therefore, if a company can operate inference services and the GPU hardware market within the same system, it may be more commercially viable. This model can hedge against risks arising from changes in the profit structure of the industry chain's upstream and downstream.

Conclusion

Looking back at history, commodity markets like oil and electricity initially relied on bilateral transactions facilitated by intermediaries, with opaque trading processes and a lack of unified standards. However, over time, physical delivery gradually developed a series of agreements and financial infrastructures, including indices, trading standardization, and quality audits, ultimately evolving into a highly financialized asset class.

Today, the computing power market is undergoing a similar transformation. Numerous projects attempting to financialize foundational tools have emerged in the market, including computing power indices, cash-settled exchanges, lending products, and insurance products.

At the same time, the physical GPU market and cloud service providers are also continuously extending upstream and downstream in the industry chain, trying to capture the complete value in the token economy. As the underlying resource on which the entire AI economy operates, computing power is gradually transforming from a mere infrastructure into an independent financial asset.

In the future, this field may see multiple unicorn companies covering different levels of the market, forming various business forms, including physical computing power supply, brokerage services, lending, risk management, and combinations with DeFi and TradFi capital markets.

Computing power may be a rare new type of physical commodity in decades, and we may be witnessing its journey from privately facilitated early transactions to a complete asset class.