Structured Infrastructure Commodities: Expanding the Compute Thesis into a Basket

Core Thesis

The next major asset class is not merely compute. It is structured infrastructure capacity.

Compute is the first frontier because GPU-hours are becoming benchmarked, scarce, financeable and hedgeable sooner than the surrounding infrastructure layers. But the durable basket is broader: compute, firm power, grid deliverability, data-centre capacity, cooling, water, fibre, silicon supply and environmental permissions.

This category can be framed as:

Structured Infrastructure Commodities, or SICs.

A SIC is not a raw material like oil or copper, and it is not simply an equity claim on a company. It is a tradeable claim on constrained infrastructure throughput over time: GPU-hours, megawatt-months, rack-kilowatt capacity, transformer-backed grid access, cooling capacity, water availability, bandwidth, storage and related delivery rights.

Compute is the best current example because futures-market infrastructure is already being formed around it. CME and Silicon Data have announced plans for compute futures based on daily GPU rental-rate benchmarks, while ICE and Ornn have announced GPU compute futures based on Ornn’s Compute Price Index, which tracks live-traded spot prices for GPU compute across hardware types.

That means compute has moved from a private procurement input to an emerging benchmarked commodity.

Benchmark infrastructure: CME/Silicon Data and ICE/Ornn have announced GPU compute futures tied to daily rental-rate and live-spot benchmarks.

1. Asset Class Definition

The five defining traits of a Structured Infrastructure Commodity.

Structured Infrastructure Commodities are standardised, financialised claims on critical infrastructure bottlenecks.

They have five defining traits:

Trait Meaning
Physical constraint Supply depends on real-world assets: chips, power, land, transformers, water, fibre, cooling and construction capacity.
Time-based utility The asset is valuable because it delivers capacity during a specific period: GPU-hour, MW-month, rack-year, Gbps-mile.
Benchmarkable price A reference price can be observed, indexed and eventually used for futures, swaps or structured notes.
Hedging demand Buyers need to lock future costs; sellers need to lock future revenue.
Financial wrapper potential Contracts can be securitised, tokenised, fractionalised or used as treasury reserves.

The key insight is that the commodity is not always the physical object. The commodity is often the right to use constrained infrastructure at a future date.

That is why compute belongs in the same conceptual family as electricity, grid capacity, shipping slots, spectrum, emissions credits, data-centre leases and cloud reservations.

2. Why Compute Is the First Frontier

Compute leads the SIC category because it has the cleanest direct connection between AI demand and a measurable unit of output:

GPU capacity over time.

AI compute is becoming economically critical for foundation models, inference infrastructure, robotics, autonomous systems, quant research, simulation, biotech, defence, gaming, media generation and enterprise automation.

The demand curve is no longer just tied to software companies. It is tied to nearly every sector that needs intelligence, automation or prediction at scale.

This makes compute different from a normal cloud-service expense. It is becoming a scarce productive input.

The stronger version of the thesis is:

Compute is not just a technology resource. It is becoming a priced claim on AI production capacity.

Compute is also easier to benchmark than many other infrastructure constraints. A contract can reference a specific GPU class, such as H100, H200, B200 or equivalent accelerator capacity, over a defined period.

This makes it possible to create:

That is why compute is the first frontier of the broader SIC asset class.

3. The Basket: AI Structured Infrastructure Commodity Basket

The AI Infrastructure Capacity Basket — eight sleeves from compute capacity through environmental permissions.

An expanded basket could be called the:

AI Infrastructure Capacity Basket, or AICB.

These are not recommended portfolio weights. They are research weights for defining the asset class.

Sleeve Illustrative Weight Commodity Unit Economic Role Near-Term Expression
1. Compute capacity 35% GPU-hour, accelerator-hour, inference-token capacity, reserved cluster capacity Core AI production unit. This is the most direct expression of AI infrastructure scarcity. Compute futures, OTC GPU swaps, reserved cloud-capacity contracts, DePIN compute markets.
2. Firm power and energy 20% MWh, MW-month, PPA strip, gas/nuclear/renewable generation capacity Compute cannot clear without power. The true scarce input may be firm power at the right node. Power futures, nodal power exposure, PPAs, capacity markets, gas/power spreads, nuclear/renewable firming contracts.
3. Grid deliverability and equipment 15% Transformer capacity, UPS capacity, substation access, interconnection rights Converts theoretical electricity into usable data-centre power. A major hidden bottleneck. Infrastructure funds, project finance, equipment-linked exposure, grid-access premiums, future interconnection-right markets.
4. Data-centre shell and rack density 10% Leased MW, rack-kW, data-centre availability-zone capacity, uptime-adjusted capacity Turns chips and power into usable compute environments. Location and density create scarcity. Data-centre leases, REITs, private infrastructure funds, capacity prepayments, regional rack-MW forwards.
5. Cooling, thermal management and water 7.5% kW cooled, gallons/MWh, water-use efficiency, liquid-cooling capacity Enables high-density AI clusters. Becomes more important as rack density rises. Cooling infrastructure, water rights, water-efficient facility exposure, thermal equipment, closed-loop cooling systems.
6. Connectivity and data movement 5% Gbps-mile, fibre route-mile, optical port capacity, latency-adjusted bandwidth AI training and inference both require fast data movement. Low-latency fibre becomes part of the compute supply chain. Fibre IRUs, bandwidth contracts, optical networking exposure, data-centre interconnects.
7. Silicon and memory bottlenecks 5% GPU units, HBM capacity, advanced packaging capacity, wafer starts Upstream supply constraint. Determines how quickly new compute supply can come online. Semiconductor supply contracts, advanced packaging exposure, HBM/memory exposure, equipment suppliers.
8. Environmental and regulatory permissions 2.5% RECs, carbon credits, water permits, heat-reuse credits, emissions intensity The licence-to-operate layer. Becomes more valuable where data-centre backlash grows. Renewable-energy credits, carbon markets, water permits, regulated site premiums.

The strongest version of the basket is not “long every AI stock.” It is a multi-layer scarcity basket around the infrastructure stack required to produce AI compute.

4. Why Each Sleeve Belongs in the Same Asset Class

4.1 Compute Capacity: The Benchmark Layer

How a commodity financialises: benchmarks, forward markets, listed futures, liquidity, structured products.

Compute is the cleanest first contract because the unit is intuitive:

a specific GPU or accelerator available for a specific period.

This is exactly how commodity markets begin to financialise:

  1. Benchmarks emerge.
  2. Forward markets develop.
  3. Futures are listed.
  4. Liquidity grows.
  5. Structured products follow.

Compute is moving into this sequence now.

4.2 Power: The Shadow Commodity Beneath Compute

Firm power is the embedded cost curve under compute — the shadow commodity every GPU-hour depends on.

Power is the embedded cost curve beneath compute. As AI workloads scale, compute pricing becomes connected to electricity availability and the cost of firm energy.

A GPU-hour is only useful if it has power.

Power is the embedded cost curve under compute. The more AI workloads scale, the more compute pricing becomes connected to electricity availability, grid access and the cost of firm energy.

This means the basket should not be purely digital. Compute is the visible commodity, but firm power is the underlying constraint.

4.3 Grid Deliverability: The Hidden Bottleneck

Grid bottlenecks — transformers, substations, interconnection queues, UPS and switchgear — convert theoretical electricity into usable data-centre power.

Power supply alone does not solve the problem. The power must be delivered, conditioned and made reliable enough for hyperscale compute.

The critical bottlenecks include:

This is why the basket needs a grid-equipment and interconnection sleeve.

The highest-value asset may become not electricity in the abstract, but deliverable electricity at a specific node with sufficient transformers, UPS, substations and permits.

4.4 Data-Centre Capacity: The Real-Estate Wrapper

The data-centre floor — the physical factory where chips, power and cooling become usable compute.

Data-centre capacity: the commodity unit is available, powered, cooled, networked rack capacity — not square footage.

Data centres are the physical factory floor for compute.

But unlike generic real estate, the commodity-like unit is not square footage. It is:

available, powered, cooled, networked rack capacity.

A future market could evolve around leased MW capacity by region, rack density, uptime class and cooling type.

This is where the asset class starts to look like a hybrid of power, logistics, cloud infrastructure and industrial real estate.

4.5 Cooling and Water: The Density Enablers

Cooling and water as density enablers — liquid cooling, closed-loop systems and water-efficient facility design.

AI servers are denser, hotter and more power-intensive than legacy workloads.

Cooling and water are no longer side issues. They are density enablers.

The more AI clusters move toward high-density rack architectures, the more valuable liquid cooling, closed-loop cooling and water-efficient facility design become.

This makes cooling capacity a valid part of the SIC basket.

4.6 Fibre and Connectivity: The Transportation Layer

Connectivity as the transportation layer — route-specific, time-specific and latency-adjusted data movement.

If compute is the refinery, fibre is part of the pipeline network.

AI workloads require fast data movement across:

Bandwidth can therefore become a structured infrastructure commodity: route-specific, time-specific and latency-adjusted data movement.

4.7 Silicon and Memory: The Upstream Supply Constraint

Upstream semiconductor constraints — advanced nodes, HBM, packaging and foundry capacity.

GPU supply depends on deeper semiconductor constraints:

This sleeve captures the upstream bottleneck that determines how quickly new compute supply can come online.

4.8 Environmental and Regulatory Permissions: The Licence-to-Operate Layer

The licence-to-operate layer — RECs, carbon credits, water permits, heat-reuse credits and emissions intensity.

The more data centres stress local grids, water systems and land-use policy, the more valuable permissions become.

This includes:

In constrained regions, the permission to operate may become a scarce asset in itself.

5. The Compute Treasury Concept

A compute treasury could hold prepaid GPU capacity, forward contracts, GPU lease receivables, tokenised cluster shares, power PPAs, grid rights, water rights and carbon credits.

The treasury point becomes stronger when expanded beyond server ownership.

A future AI-native treasury could hold a reserve of:

Treasury Reserve Asset Purpose
Prepaid GPU capacity Hedge future training or inference costs.
Forward compute contracts Lock supply before market tightness.
GPU lease receivables Turn compute revenue into financeable cash flows.
Tokenised server or cluster shares Fractionalise ownership or revenue participation.
Power PPAs Secure energy cost and availability.
Grid/interconnection rights Preserve future expansion optionality.
Water/cooling rights Protect thermal capacity in constrained regions.
Carbon/renewable credits Preserve licence to operate and ESG compliance.

For an AI company, a compute treasury could become as important as a fuel hedge is for an airline.

For a sovereign AI programme, compute reserves could become strategic infrastructure.

For a DePIN network, tokenised compute reserves could become the balance-sheet base layer.

The institutional version is likely to start with traditional contracts: PPAs, leases, GPU financing, project finance and OTC hedges.

The crypto-native version is more likely to develop around tokenised claims on server revenue, GPU-hour delivery rights or compute-backed treasury tokens.

6. The Structured Product Form

A mature version of this market could support several products.

6.1 Compute Futures

The first listed contracts are likely to be cash-settled GPU-hour futures by chip class:

Cash settlement is likely because physical delivery of compute is complex. Delivery quality depends on:

6.2 Compute-Power Spreads

Once compute futures trade, the natural spread is:

GPU-hour revenue minus electricity cost.

This creates an AI version of spark spreads in power markets.

A neocloud or data-centre operator could hedge GPU rental revenue while also hedging the power cost needed to deliver that compute.

6.3 Regional Compute Basis

Compute will not be perfectly fungible.

A GPU-hour in a power-constrained region with expensive electricity and grid queues is not the same as a GPU-hour near cheap power, fibre and permissive regulation.

Regional basis markets could develop around locations such as:

6.4 Rack-MW Forwards

Data-centre operators and AI labs could trade future claims on powered, cooled rack capacity.

These would look like real-estate leases, power contracts and cloud-reservation agreements blended into one capacity instrument.

6.5 Tokenised Compute Notes

A tokenised compute note could represent:

This is plausible, but it will need:

7. Key Trade Expressions

The thesis creates multiple trade expressions rather than one single long.

Trade Expression Thesis
Long compute scarcity GPU rental prices and capacity premia rise when AI demand outpaces supply.
Long firm power vs generic power AI loads reward electricity that is reliable, local and immediately deliverable.
Long grid bottlenecks Transformers, UPS systems and interconnection rights become gating assets.
Long liquid cooling / short legacy cooling Higher rack density favours facilities that can cool next-generation accelerators.
Long regional basis volatility Compute prices diverge by location as power, regulation and fibre constraints differ.
Long new accelerator capacity / short obsolete GPU capacity Hardware cycles create relative-value trades across chip generations.
Long tokenised capacity optionality Fractional ownership and programmable settlement may expand access, but this is higher risk and earlier stage.

The best formulation is not simply “compute prices go up.”

The better thesis is:

AI demand will create persistent bottleneck premia across the physical infrastructure stack, and compute will be the most visible, liquid and benchmarked expression of that broader scarcity.

8. Risks and Challenges

The asset class is real, but the thesis should still be challenged.

8.1 Compute Is Not Fully Fungible

Chip type, interconnect, memory, uptime, region, security, software stack and latency all matter.

This makes compute more like electricity, freight or bandwidth than crude oil.

8.2 Benchmarks Are Fragile

If GPU pricing data is too thin, private, self-reported or dominated by a few platforms, futures markets may struggle to build trust.

The success of compute futures depends on trusted reference prices.

8.3 Efficiency Can Cut Both Ways

Better models and cheaper inference can increase demand through Jevons paradox, but they can also reduce the need for some capacity or cause overbuild in older GPU generations.

8.4 Regulation Is Becoming a Real Constraint

Data-centre expansion is becoming a political and environmental issue in regions where power use, water use, utility bills and grid reliability are under pressure.

Regulatory permission may become one of the most important scarcity layers.

8.5 Tokenisation May Arrive Later Than Futures

Fractionalised server ownership is plausible, but institutional capital will demand:

Tokenisation can be powerful, but it may follow the benchmark and futures layer rather than lead it.

9. Final Thesis Formulation

Compute is the first major frontier of a broader Structured Infrastructure Commodity asset class. AI is turning infrastructure capacity into a financial asset: GPU-hours, firm power, grid access, data-centre rack density, cooling, water, fibre and silicon supply are becoming measurable, scarce and financeable. Compute will likely be the first liquid expression because benchmarks and futures are already emerging, but the superior basket is a multi-layer AI infrastructure capacity basket that captures the entire bottleneck stack.

Even more simply:

Compute is the oil. Power is the fuel. Grid access is the pipeline. Cooling is the refinery constraint. Fibre is the transport network. Tokenisation is the ownership wrapper. The asset class is structured infrastructure capacity.

10. Strategic Implication

This creates a frontier asset class for investors, infrastructure operators, protocols, AI companies and treasuries.

The opportunity is not only to trade compute. The opportunity is to define, benchmark and structure the entire AI infrastructure bottleneck stack.

The winning market participants will not only own GPUs. They will control or hedge the scarce inputs that make compute deliverable:

Compute is the entry point. Structured infrastructure capacity is the full asset class.