Help grow World Market Atlas — a student project for individual investors. Every contribution counts.
WORLD MARKET ATLAS · PERSONAL COMMENTARY
Market commentary
My perspective on markets, the economy and events that matter to investors.
World Market AtlasEN
Compute as a potential new asset class — the financialisation of computing power in the AI economy
As of 23 September 2026. Announcements and forecasts do not imply completed investment or live trading.
NVIDIA logo. Source: NVIDIA Newsroom. NVIDIA owns the trademark; editorial use does not imply affiliation or endorsement. NVIDIA Newsroom ↗
The development of artificial intelligence is rapidly increasing the importance of computing infrastructure, particularly specialised GPU accelerators used for training and inference. Computing power, traditionally treated as an operating expense or part of IT infrastructure, is becoming the subject of standardised pricing, long-term contracts, infrastructure-backed financing and planned derivatives. NVIDIA describes this process as turning compute into an “investable asset class”. At the same time, the first academic studies on pricing compute as a financial asset and futures contracts based on GPU rental prices are emerging.
This paper assesses whether compute can genuinely develop into a separate asset class. The analysis suggests that computing power shares some characteristics with commodities, energy and infrastructure. However, it is also highly heterogeneous, subject to rapid technological change, impossible to store in the form of unused capacity, and concentrated among a small number of suppliers. The most likely outcome is therefore not one homogeneous “compute” asset class, but a complex of financial instruments spanning AI infrastructure, compute contracts, lending secured by equipment and revenues, and derivatives on GPU prices.
Keywords: artificial intelligence, compute, GPU, asset class, futures, AI infrastructure, data centres, financialisation, NVIDIA.
1. Introduction
Every technological revolution has produced not only new companies but also new assets to finance and invest in. Railways supported the development of infrastructure bonds; oil created extensive futures markets; telecommunications led to the financialisation of network infrastructure; and the internet contributed to an entire class of assets associated with data centres and digital infrastructure.
Generative AI may trigger a similar process. The fundamental resource that enables AI models to operate is computing power — compute. In this context, it encompasses not just an individual GPU but the whole system that converts electricity and data into computation: accelerators, processors, memory, networks, cooling, data centres and software.
In August 2026, NVIDIA CEO Jensen Huang argued that AI infrastructure was beginning to become a new “investable asset class”. NVIDIA also announced cooperation with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR on AI infrastructure financing platforms intended ultimately to mobilise more than $500 billion of third-party capital. However, NVIDIA’s announcement describes planned financing platforms and comes from a company with a direct economic interest in expanding the AI infrastructure market.
The research question should therefore not be whether NVIDIA calls compute an asset class, but whether computing power meets the economic and financial conditions required to qualify as a separate asset class.
2. What does “asset class” actually mean?
Not everything with a price is an asset class. In his classic article What Is an Asset Class, Anyway?, Robert Greer defined an asset class as a set of assets sharing fundamental economic characteristics that distinguish them from other assets.
A contemporary approach used by the CFA Institute also suggests that members of an asset class should be relatively homogeneous, represent common sources of risk, and offer sufficient scale and investability to form a meaningful part of an investment portfolio.
In the case of compute, at least four different economic objects can be distinguished:
Physical computing infrastructure — GPUs, servers and data centres;
A flow of computing services — for example, an hour of GPU usage;
Contracts and financial instruments based on compute — long-term agreements and futures;
Instruments that finance infrastructure — loans, bonds and project finance structures secured by GPUs or customer contracts.
This distinction matters. Owning a data centre resembles a conventional infrastructure asset generating cash flows, whereas an unused GPU-hour is more like a perishable economic good comparable to electricity.
3. Why is compute beginning to resemble a commodity?
A financial market requires an active market in the underlying good. Such a market is gradually emerging for computing power.
In (Early) AI Compute Asset Pricing, Bandi and Su argue that the economic object being traded should not be the GPU alone but access to a particular bundle of services: the accelerator, memory, CPU, network, location, reliability and operator infrastructure.
The GPU-hour — an hour of access to a specified GPU type — could therefore become a unit analogous to a barrel of oil or a megawatt-hour.
The problem is that a GPU-hour is not homogeneous. An hour on an NVIDIA A100, H100 or B200, or an AMD MI300X, can deliver very different performance. Data centres also differ in network configuration, memory and infrastructure quality.
Bandi and Su describe the current market as heterogeneous and argue that financialisation requires indices that convert many different prices into a standardised benchmark.
This process has already begun. Indices publishing rental prices for A100, H100, B200 and other accelerators have emerged. Silicon Data publishes daily benchmarks expressed in dollars per GPU-hour, among other measures.
An index alone does not create an asset class, but it is a fundamental prerequisite for derivatives and a market for managing risk.
NVIDIA GB200 NVL72 — an example of AI computing infrastructure. Press image: NVIDIA. NVIDIA Newsroom ↗
4. The economic scale of the compute market
The enormous scale of investment provides a second argument for the financialisation of compute.
According to the Stanford AI Index, global AI infrastructure computing capacity had grown approximately 3.3-fold per year since 2022, reaching about 17.1 million NVIDIA H100 equivalents by the end of 2025. NVIDIA accounted for more than 60% of estimated global capacity, with Google, Amazon and other manufacturers also holding significant shares.
This infrastructure also requires substantial energy investment. Stanford estimated the power capacity of AI data centre infrastructure at approximately 29.6 GW in 2025.
The financial scale is even more striking. The International Energy Agency reports that capital expenditure by five major technology companies exceeded $400 billion in 2025 and was expected to increase by roughly another 75% in 2026. The IEA notes that capex by just five technology companies had become larger than global investment in upstream oil and gas.
This implies that AI infrastructure is reaching a scale at which financing exclusively through technology companies’ balance sheets is becoming less efficient. A natural consequence is the arrival of banks, private credit funds, infrastructure investors and bond markets.
5. Futures as a turning point
One of the most important stages in turning an economic good into a financial asset is the creation of a derivatives market.
CME Group and Silicon Data announced futures on NVIDIA H100 and B200 rental prices. According to CME’s announcement, trading is scheduled to begin on 5 October 2026, subject to the necessary regulatory approvals. As of 23 September 2026, this means that the market has been announced but does not yet have a long record of actual trading.
ICE and NATIVX announced similar plans, working on futures based on COIL, an index measuring the price of normalised computing power.
The emergence of futures matters for two reasons.
First, futures enable hedging. A company planning to train a model in 2027 could potentially hedge the future cost of compute today. A data centre operator could hedge the risk of declining rental prices.
Second, futures create financial exposure for investors who do not want to build a data centre or buy GPUs themselves.
At this point, compute begins to resemble markets for electricity, gas and other commodities.
6. Compute nevertheless differs from oil
One fundamental characteristic distinguishes compute from many traditional commodities: unused computing capacity cannot be stored.
If the owner of a storage facility does not sell a barrel of oil today, it can be sold later. If a GPU sits idle for an hour, however, that hour of productive capacity is lost permanently.
Bandi and Su note that the standard spot–futures arbitrage relationship familiar from storable commodities therefore does not apply directly to compute. In this respect, computing power is much closer to electricity.
The price of compute therefore depends on factors including:
Current infrastructure utilisation;
Energy availability;
The supply of new chips;
Demand for model training and inference;
The processor generation;
Memory availability;
Network infrastructure;
Data centre location;
Software efficiency.
As a result, infrastructure value can be highly sensitive to utilisation. GPUs with a nominal value of several billion dollars can generate substantial revenue near full capacity, but profitability may fall sharply if supply exceeds demand.
7. Technological obsolescence
Another difference from traditional infrastructure concerns the asset’s life cycle.
A power plant, gas pipeline or motorway can operate for decades. AI processors, by contrast, may lose competitiveness relatively quickly as new chip generations emerge.
NVIDIA argues that CUDA software can extend the economic usefulness of older hardware. The company cites the A100, introduced in 2020 and still in commercial use, as an example.
This argument does not eliminate technological risk. Bandi and Su point out that physical hardware ages while a financial index may gradually shift towards newer processor generations.
This creates an interesting situation: an infrastructure owner may hold an ageing physical asset, while an investor in a compute index may have exposure that dynamically migrates to newer technology.
The two instruments would therefore have very different risk profiles.
8. The emergence of compute-backed finance
The most advanced financialisation is probably occurring in credit markets rather than futures markets.
CoreWeave already uses structures in which financing is directly tied to GPU purchases and the fulfilment of customer contracts. In March 2026, one group company entered into a delayed-draw term loan agreement for up to $8.5 billion, intended in part to purchase GPU servers and infrastructure needed to fulfil a customer contract.
At the end of June 2026, CoreWeave reported approximately $35.6 billion in debt. Some structures were secured by special-purpose entities’ assets, including substantial infrastructure and receivables.
In May 2026, the company also announced publicly syndicated HPC financing secured by AI infrastructure.
A similar pattern is emerging more broadly. In September 2026, Reuters reported approximately $22 billion in financing for an infrastructure venture involving Blackstone and Alphabet, with collateral reportedly including TPU processors and customer contracts.
This resembles the historical development of other infrastructure asset classes: first a revenue-generating asset, then asset-backed lending, followed by securitisation, indices and derivatives.
9. The main risks of the emerging asset class
The greatest obstacle to the case for a separate asset class is the market’s current lack of maturity.
Technological risk stems from the rapid development of accelerators. A new generation can materially reduce the cost of a unit of computation and diminish the economic value of earlier GPUs.
Demand risk concerns the scale of future AI usage. Current investment assumes continued strong demand for model training and inference. If AI monetisation proceeds more slowly than expected, infrastructure utilisation could decline.
Technological concentration risk is especially important. The Stanford AI Index points to NVIDIA’s dominant position in global compute and the significant concentration of leading-edge semiconductor manufacturing.
Energy risk makes the price of compute partly dependent on electricity prices and availability. The IEA expects global data centre electricity consumption to rise rapidly over the coming years.
Basis risk arises when a company hedges costs using an H100 index but actually uses a different supplier, region or hardware configuration. Bandi and Su show that prices across market segments do not always move together.
Finally, financial risk follows from the sector’s high indebtedness. As AI investment grows, debt, leasing, project finance and residual-value guarantees become increasingly important. Credit markets consequently bear a growing share of the risk surrounding the future value of AI equipment and contracts.
10. Is compute already a separate asset class?
At this stage, the answer should be more cautious than the narrative advanced by infrastructure manufacturers.
Compute already has several features of an emerging asset class:
A very large and rapidly growing underlying market;
Measurable prices for units of computing power;
Price indices;
Long-term contracts;
The ability to finance assets against future cash flows;
Equipment-backed financing;
A developing derivatives market;
Natural market participants seeking to hedge price risk.
At the same time, compute does not yet meet all the criteria for a mature asset class. Price histories are short, indices are new, the market remains fragmented, products are not fully interchangeable, and futures lack a multi-year record of liquid trading. There is also insufficient historical data to assess long-run returns, volatility and correlations with equities, bonds, commodities and real estate reliably.
Interestingly, Bandi and Su’s preliminary results for synthetic futures suggest the possibility of a positive compute risk premium. The authors themselves emphasise the early nature of these findings.
It is therefore currently more precise to describe compute as an emerging market for assets and financial risk than as a fully established asset class.
11. Conclusions
AI development is changing the economic nature of computing power. Compute is no longer merely an IT expense. It is becoming a scarce productive resource with an observable price, a source of revenue, collateral for debt and a potential underlying asset for futures.
The key element of this transformation is therefore not simply growth in the GPU market’s value, but the financialisation of compute.
The process can be represented by the following sequence:
In this respect, 2026 may be a transitional year. NVIDIA and global financial institutions are creating infrastructure financing structures measured in hundreds of billions of dollars, while CME and ICE are developing instruments intended to transfer compute price risk to financial markets.
At the same time, compute remains much more complex than a traditional commodity. It is perishable, technologically heterogeneous, energy-dependent, rapidly ageing and heavily concentrated around a handful of technology suppliers.
The most likely result is therefore not one simple “compute” asset class. A broader AI infrastructure finance ecosystem is more likely, encompassing infrastructure assets, private credit, bonds, compute contracts and derivatives.
If the coming years produce a liquid spot market, reliable benchmarks, a deep futures market and a sufficiently long price history to identify an independent risk premium, compute could genuinely begin to function as a separate asset class in portfolio theory.
For now, we are observing the beginnings of the process that could lead to its emergence.
Bibliography
Bandi, F. M., Su, Y. (2026), (Early) AI Compute Asset Pricing, arXiv:2607.12156, version dated 10 September 2026 ArXiv — full paper
CFA Institute (2026), Overview of Asset Allocation. CFA Institute
CME Group (2026), CME Group and Silicon Data to Launch Compute Futures on October 5 to Unlock New Way to Hedge AI Risks. CME Group
Greer, R. J. (1997), What Is an Asset Class, Anyway?, The Journal of Portfolio Management, 23(2), 86–91, DOI: 10.3905/JPM.23.2.86.
International Energy Agency (2025), Energy and AI. IEA – Energy and AI
International Energy Agency (2026), Key Questions on Energy and AI. IEA — 2026 report
Intercontinental Exchange, NATIVX (2026), ICE and NATIVX to Launch Energy-Normalized Compute Futures Contracts. ICE
NVIDIA (2026), Huang, J., NVIDIA AI Factory Compute Is Becoming an Investable Asset Class, 11 August 2026 NVIDIA Blog — source article
NVIDIA (2026), NVIDIA Partners With Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms to Mobilize Over $500 Billion of Third-Party Capital. NVIDIA Newsroom
Stanford Institute for Human-Centered Artificial Intelligence (2026), AI Index Report 2026 – Research and Development. Stanford AI Index 2026
CoreWeave Inc. (2026), SEC filings on GPU infrastructure financing and financial liabilities. SEC – CoreWeave 2026 filing
Silicon Data (2026), GPU Rental Price Indices oraz materiały dotyczące powstającego rynku compute futures. Silicon Data indices
Market commentary
My perspective on markets, the economy and events that matter to investors.
Compute as a potential new asset class — the financialisation of computing power in the AI economy
As of 23 September 2026. Announcements and forecasts do not imply completed investment or live trading.
The development of artificial intelligence is rapidly increasing the importance of computing infrastructure, particularly specialised GPU accelerators used for training and inference. Computing power, traditionally treated as an operating expense or part of IT infrastructure, is becoming the subject of standardised pricing, long-term contracts, infrastructure-backed financing and planned derivatives. NVIDIA describes this process as turning compute into an “investable asset class”. At the same time, the first academic studies on pricing compute as a financial asset and futures contracts based on GPU rental prices are emerging.
This paper assesses whether compute can genuinely develop into a separate asset class. The analysis suggests that computing power shares some characteristics with commodities, energy and infrastructure. However, it is also highly heterogeneous, subject to rapid technological change, impossible to store in the form of unused capacity, and concentrated among a small number of suppliers. The most likely outcome is therefore not one homogeneous “compute” asset class, but a complex of financial instruments spanning AI infrastructure, compute contracts, lending secured by equipment and revenues, and derivatives on GPU prices.
Keywords: artificial intelligence, compute, GPU, asset class, futures, AI infrastructure, data centres, financialisation, NVIDIA.
1. Introduction
Every technological revolution has produced not only new companies but also new assets to finance and invest in. Railways supported the development of infrastructure bonds; oil created extensive futures markets; telecommunications led to the financialisation of network infrastructure; and the internet contributed to an entire class of assets associated with data centres and digital infrastructure.
Generative AI may trigger a similar process. The fundamental resource that enables AI models to operate is computing power — compute. In this context, it encompasses not just an individual GPU but the whole system that converts electricity and data into computation: accelerators, processors, memory, networks, cooling, data centres and software.
In August 2026, NVIDIA CEO Jensen Huang argued that AI infrastructure was beginning to become a new “investable asset class”. NVIDIA also announced cooperation with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR on AI infrastructure financing platforms intended ultimately to mobilise more than $500 billion of third-party capital. However, NVIDIA’s announcement describes planned financing platforms and comes from a company with a direct economic interest in expanding the AI infrastructure market.
The research question should therefore not be whether NVIDIA calls compute an asset class, but whether computing power meets the economic and financial conditions required to qualify as a separate asset class.
2. What does “asset class” actually mean?
Not everything with a price is an asset class. In his classic article What Is an Asset Class, Anyway?, Robert Greer defined an asset class as a set of assets sharing fundamental economic characteristics that distinguish them from other assets.
A contemporary approach used by the CFA Institute also suggests that members of an asset class should be relatively homogeneous, represent common sources of risk, and offer sufficient scale and investability to form a meaningful part of an investment portfolio.
In the case of compute, at least four different economic objects can be distinguished:
Physical computing infrastructure — GPUs, servers and data centres;
A flow of computing services — for example, an hour of GPU usage;
Contracts and financial instruments based on compute — long-term agreements and futures;
Instruments that finance infrastructure — loans, bonds and project finance structures secured by GPUs or customer contracts.
This distinction matters. Owning a data centre resembles a conventional infrastructure asset generating cash flows, whereas an unused GPU-hour is more like a perishable economic good comparable to electricity.
3. Why is compute beginning to resemble a commodity?
A financial market requires an active market in the underlying good. Such a market is gradually emerging for computing power.
In (Early) AI Compute Asset Pricing, Bandi and Su argue that the economic object being traded should not be the GPU alone but access to a particular bundle of services: the accelerator, memory, CPU, network, location, reliability and operator infrastructure.
The GPU-hour — an hour of access to a specified GPU type — could therefore become a unit analogous to a barrel of oil or a megawatt-hour.
The problem is that a GPU-hour is not homogeneous. An hour on an NVIDIA A100, H100 or B200, or an AMD MI300X, can deliver very different performance. Data centres also differ in network configuration, memory and infrastructure quality.
Bandi and Su describe the current market as heterogeneous and argue that financialisation requires indices that convert many different prices into a standardised benchmark.
This process has already begun. Indices publishing rental prices for A100, H100, B200 and other accelerators have emerged. Silicon Data publishes daily benchmarks expressed in dollars per GPU-hour, among other measures.
An index alone does not create an asset class, but it is a fundamental prerequisite for derivatives and a market for managing risk.
4. The economic scale of the compute market
The enormous scale of investment provides a second argument for the financialisation of compute.
According to the Stanford AI Index, global AI infrastructure computing capacity had grown approximately 3.3-fold per year since 2022, reaching about 17.1 million NVIDIA H100 equivalents by the end of 2025. NVIDIA accounted for more than 60% of estimated global capacity, with Google, Amazon and other manufacturers also holding significant shares.
This infrastructure also requires substantial energy investment. Stanford estimated the power capacity of AI data centre infrastructure at approximately 29.6 GW in 2025.
The financial scale is even more striking. The International Energy Agency reports that capital expenditure by five major technology companies exceeded $400 billion in 2025 and was expected to increase by roughly another 75% in 2026. The IEA notes that capex by just five technology companies had become larger than global investment in upstream oil and gas.
This implies that AI infrastructure is reaching a scale at which financing exclusively through technology companies’ balance sheets is becoming less efficient. A natural consequence is the arrival of banks, private credit funds, infrastructure investors and bond markets.
5. Futures as a turning point
One of the most important stages in turning an economic good into a financial asset is the creation of a derivatives market.
CME Group and Silicon Data announced futures on NVIDIA H100 and B200 rental prices. According to CME’s announcement, trading is scheduled to begin on 5 October 2026, subject to the necessary regulatory approvals. As of 23 September 2026, this means that the market has been announced but does not yet have a long record of actual trading.
ICE and NATIVX announced similar plans, working on futures based on COIL, an index measuring the price of normalised computing power.
The emergence of futures matters for two reasons.
First, futures enable hedging. A company planning to train a model in 2027 could potentially hedge the future cost of compute today. A data centre operator could hedge the risk of declining rental prices.
Second, futures create financial exposure for investors who do not want to build a data centre or buy GPUs themselves.
At this point, compute begins to resemble markets for electricity, gas and other commodities.
6. Compute nevertheless differs from oil
One fundamental characteristic distinguishes compute from many traditional commodities: unused computing capacity cannot be stored.
If the owner of a storage facility does not sell a barrel of oil today, it can be sold later. If a GPU sits idle for an hour, however, that hour of productive capacity is lost permanently.
Bandi and Su note that the standard spot–futures arbitrage relationship familiar from storable commodities therefore does not apply directly to compute. In this respect, computing power is much closer to electricity.
The price of compute therefore depends on factors including:
Current infrastructure utilisation;
Energy availability;
The supply of new chips;
Demand for model training and inference;
The processor generation;
Memory availability;
Network infrastructure;
Data centre location;
Software efficiency.
As a result, infrastructure value can be highly sensitive to utilisation. GPUs with a nominal value of several billion dollars can generate substantial revenue near full capacity, but profitability may fall sharply if supply exceeds demand.
7. Technological obsolescence
Another difference from traditional infrastructure concerns the asset’s life cycle.
A power plant, gas pipeline or motorway can operate for decades. AI processors, by contrast, may lose competitiveness relatively quickly as new chip generations emerge.
NVIDIA argues that CUDA software can extend the economic usefulness of older hardware. The company cites the A100, introduced in 2020 and still in commercial use, as an example.
This argument does not eliminate technological risk. Bandi and Su point out that physical hardware ages while a financial index may gradually shift towards newer processor generations.
This creates an interesting situation: an infrastructure owner may hold an ageing physical asset, while an investor in a compute index may have exposure that dynamically migrates to newer technology.
The two instruments would therefore have very different risk profiles.
8. The emergence of compute-backed finance
The most advanced financialisation is probably occurring in credit markets rather than futures markets.
CoreWeave already uses structures in which financing is directly tied to GPU purchases and the fulfilment of customer contracts. In March 2026, one group company entered into a delayed-draw term loan agreement for up to $8.5 billion, intended in part to purchase GPU servers and infrastructure needed to fulfil a customer contract.
At the end of June 2026, CoreWeave reported approximately $35.6 billion in debt. Some structures were secured by special-purpose entities’ assets, including substantial infrastructure and receivables.
In May 2026, the company also announced publicly syndicated HPC financing secured by AI infrastructure.
A similar pattern is emerging more broadly. In September 2026, Reuters reported approximately $22 billion in financing for an infrastructure venture involving Blackstone and Alphabet, with collateral reportedly including TPU processors and customer contracts.
This resembles the historical development of other infrastructure asset classes: first a revenue-generating asset, then asset-backed lending, followed by securitisation, indices and derivatives.
9. The main risks of the emerging asset class
The greatest obstacle to the case for a separate asset class is the market’s current lack of maturity.
Technological risk stems from the rapid development of accelerators. A new generation can materially reduce the cost of a unit of computation and diminish the economic value of earlier GPUs.
Demand risk concerns the scale of future AI usage. Current investment assumes continued strong demand for model training and inference. If AI monetisation proceeds more slowly than expected, infrastructure utilisation could decline.
Technological concentration risk is especially important. The Stanford AI Index points to NVIDIA’s dominant position in global compute and the significant concentration of leading-edge semiconductor manufacturing.
Energy risk makes the price of compute partly dependent on electricity prices and availability. The IEA expects global data centre electricity consumption to rise rapidly over the coming years.
Basis risk arises when a company hedges costs using an H100 index but actually uses a different supplier, region or hardware configuration. Bandi and Su show that prices across market segments do not always move together.
Finally, financial risk follows from the sector’s high indebtedness. As AI investment grows, debt, leasing, project finance and residual-value guarantees become increasingly important. Credit markets consequently bear a growing share of the risk surrounding the future value of AI equipment and contracts.
10. Is compute already a separate asset class?
At this stage, the answer should be more cautious than the narrative advanced by infrastructure manufacturers.
Compute already has several features of an emerging asset class:
A very large and rapidly growing underlying market;
Measurable prices for units of computing power;
Price indices;
Long-term contracts;
The ability to finance assets against future cash flows;
Equipment-backed financing;
A developing derivatives market;
Natural market participants seeking to hedge price risk.
At the same time, compute does not yet meet all the criteria for a mature asset class. Price histories are short, indices are new, the market remains fragmented, products are not fully interchangeable, and futures lack a multi-year record of liquid trading. There is also insufficient historical data to assess long-run returns, volatility and correlations with equities, bonds, commodities and real estate reliably.
Interestingly, Bandi and Su’s preliminary results for synthetic futures suggest the possibility of a positive compute risk premium. The authors themselves emphasise the early nature of these findings.
It is therefore currently more precise to describe compute as an emerging market for assets and financial risk than as a fully established asset class.
11. Conclusions
AI development is changing the economic nature of computing power. Compute is no longer merely an IT expense. It is becoming a scarce productive resource with an observable price, a source of revenue, collateral for debt and a potential underlying asset for futures.
The key element of this transformation is therefore not simply growth in the GPU market’s value, but the financialisation of compute.
The process can be represented by the following sequence:
GPU → AI infrastructure → GPU-hour rental → compute price index → long-term contract → compute-backed financing → futures → potential asset class.
In this respect, 2026 may be a transitional year. NVIDIA and global financial institutions are creating infrastructure financing structures measured in hundreds of billions of dollars, while CME and ICE are developing instruments intended to transfer compute price risk to financial markets.
At the same time, compute remains much more complex than a traditional commodity. It is perishable, technologically heterogeneous, energy-dependent, rapidly ageing and heavily concentrated around a handful of technology suppliers.
The most likely result is therefore not one simple “compute” asset class. A broader AI infrastructure finance ecosystem is more likely, encompassing infrastructure assets, private credit, bonds, compute contracts and derivatives.
If the coming years produce a liquid spot market, reliable benchmarks, a deep futures market and a sufficiently long price history to identify an independent risk premium, compute could genuinely begin to function as a separate asset class in portfolio theory.
For now, we are observing the beginnings of the process that could lead to its emergence.
Bibliography
Bandi, F. M., Su, Y. (2026), (Early) AI Compute Asset Pricing, arXiv:2607.12156, version dated 10 September 2026 ArXiv — full paper
CFA Institute (2026), Overview of Asset Allocation. CFA Institute
CME Group (2026), CME Group and Silicon Data to Launch Compute Futures on October 5 to Unlock New Way to Hedge AI Risks. CME Group
Greer, R. J. (1997), What Is an Asset Class, Anyway?, The Journal of Portfolio Management, 23(2), 86–91, DOI: 10.3905/JPM.23.2.86.
International Energy Agency (2025), Energy and AI. IEA – Energy and AI
International Energy Agency (2026), Key Questions on Energy and AI. IEA — 2026 report
Intercontinental Exchange, NATIVX (2026), ICE and NATIVX to Launch Energy-Normalized Compute Futures Contracts. ICE
NVIDIA (2026), Huang, J., NVIDIA AI Factory Compute Is Becoming an Investable Asset Class, 11 August 2026 NVIDIA Blog — source article
NVIDIA (2026), NVIDIA Partners With Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms to Mobilize Over $500 Billion of Third-Party Capital. NVIDIA Newsroom
Stanford Institute for Human-Centered Artificial Intelligence (2026), AI Index Report 2026 – Research and Development. Stanford AI Index 2026
CoreWeave Inc. (2026), SEC filings on GPU infrastructure financing and financial liabilities. SEC – CoreWeave 2026 filing
Silicon Data (2026), GPU Rental Price Indices oraz materiały dotyczące powstającego rynku compute futures. Silicon Data indices
Posts reflect the author’s personal views as of publication. They are not investment recommendations.