This report is meant to examine the emergence of compute as a new economic commodity and explain why its importance extends well beyond artificial intelligence itself. We will first define why commodities matter within economic systems, then examine how AI is turning intelligence into a production input, why the current buildout increasingly resembles an industrial CapEx cycle, and how memory is becoming one of the most important bottlenecks within the compute stack. We will then analyze the arrival of compute pricing, the credit loop forming around GPU-backed financing, the hidden inflation channel created by compute demand, and the broader implications for market structure, geopolitics, and capital allocation.
In our view, the market still largely frames AI through the wrong lens. Most discussion remains centered on applications, software revenue, productivity tools, model capabilities, and enterprise adoption. Those variables remain important, but they sit on top of a much larger physical infrastructure cycle that is increasingly determining the economics of the entire AI regime. The more important question is no longer simply which application captures value, but what happens when intelligence itself becomes a scarce economic input that must be produced, financed, priced, and distributed.
Understanding Commodities
Commodities occupy a unique position within every economy because they sit high in the production chain. They are rarely important only because of their raw form, but because they enable broader economic activity across many downstream sectors. Oil, copper, natural gas, electricity, steel, and agricultural inputs matter because businesses, consumers, governments, and financial markets all depend on their availability and price. Once an input reaches that level of importance, it stops behaving like an ordinary product and becomes part of the operating system of the economy.
Oil remains the clearest historical example. A barrel of crude rarely reaches the end consumer directly, yet modern transportation, manufacturing, agriculture, aviation, shipping, chemicals, construction, and consumer goods all depend on oil somewhere inside their cost structure. Because oil sits upstream of so much activity, changes in its price ripple outward through margins, inflation, consumer spending, central bank policy, currencies, and geopolitical strategy. Oil became systemically important not only because it powered machines, but because the entire industrial economy had to organize around it.
Compute is beginning to exhibit many of those same characteristics. Every AI training run requires compute, every inference request consumes compute, and every autonomous workflow compounds that demand as software begins to reason, search, execute, verify, and repeat. As AI becomes embedded into manufacturing, finance, healthcare, logistics, cybersecurity, retail, energy, robotics, and scientific research, more companies become direct or indirect consumers of computational capacity. The market may continue debating which model or application wins, but beneath that debate sits a more durable requirement for usable compute.
From Software to Physical Infrastructure
For most of the post-GFC period, markets rewarded asset-light business models. Software scaled globally with low marginal costs, recurring revenue expanded valuation multiples, and platform companies absorbed increasing shares of economic activity without requiring the same physical investment as traditional industrial sectors. That regime made software the organizing principle of the prior market cycle. Investors were rewarded for owning businesses that could compound revenue without proportionally expanding physical capital.
Artificial intelligence changes that structure because intelligence does not scale for free. Every incremental increase in model capability, inference volume, long-context reasoning, and agentic execution requires more semiconductors, memory, networking, storage, power, cooling, data center capacity, and financing. The application layer may still look digital, but the constraint has become physical. In that sense, the data center is becoming the transmission mechanism through which AI demand flows into the broader economy.
This is why the current AI cycle increasingly resembles a CapEx regime rather than a traditional software cycle. The previous decade rewarded companies that minimized capital intensity, while the emerging regime may increasingly reward the suppliers and owners of scarce infrastructure. Semiconductors, advanced packaging, electrical systems, optical networking, cooling equipment, utilities, industrial construction, and power infrastructure all sit directly in the path of AI deployment. As the buildout accelerates, the market may gradually shift from rewarding only the companies using AI toward also rewarding the companies supplying the physical inputs required to make AI work.
This creates a meaningful benchmark problem. Most major equity benchmarks still largely reflect the winners of the previous era, including software platforms, cloud incumbents, digital advertising businesses, and companies that benefited from asset-light scalability. Those companies remain important, but benchmarks are backward-looking by construction and may underrepresent the industries receiving the next marginal dollar of AI-driven investment. If intelligence becomes increasingly abundant at the application layer while the infrastructure required to deploy it remains scarce, value should continue migrating toward the bottlenecks.
Intelligence as a Production Input
Traditional economic frameworks describe production through combinations of land, labor, capital, and technology. AI does not eliminate that framework, but it changes the role intelligence plays within it. Historically, intelligence was embedded primarily in labor, management, software, and organizational processes. AI begins turning intelligence into a more measurable, scalable, and purchasable input.
A manufacturer using AI to optimize production schedules consumes compute. A hospital using AI-assisted diagnostics consumes compute. A financial institution using AI for underwriting, fraud detection, portfolio construction, and risk management consumes compute. A retailer using AI to forecast demand, automate inventory, and personalize customer engagement consumes compute.
This changes how corporate cost structures should be understood. If every company gradually becomes a consumer of intelligence, then every company eventually becomes a consumer of compute. Compute therefore begins behaving less like a narrow technology expense and more like an input into margins, productivity, pricing power, and capital allocation. The more AI becomes embedded into ordinary workflows, the more the price of compute begins to matter outside the technology sector itself.
The comparison to electricity is useful. Electricity allowed businesses to reorganize production by distributing power throughout factories, offices, and homes. Compute may allow businesses to reorganize cognition by distributing machine intelligence throughout workflows, software systems, industrial processes, and decision-making structures. The economic value comes not only from the model itself, but from the ability to produce intelligence reliably and cheaply enough for it to become embedded into production.
Memory as the Quality Differential
The most important recent development inside the compute stack is the rising importance of memory. Raw accelerator supply remains critical, but usable compute is increasingly determined by memory bandwidth, memory capacity, interconnect, power efficiency, and system-level integration. As context windows expand and inference becomes more interactive, the bottleneck shifts from simply owning GPUs to efficiently feeding those GPUs with data. This makes memory one of the most important quality differentials within the compute commodity.
In oil markets, different grades of crude trade at different prices because sulfur content, density, location, transportation costs, and refining economics all matter. Compute will likely develop similar quality distinctions because an hour of compute on one system is not economically identical to an hour of compute on another. Chip generation, memory bandwidth, interconnect, reliability, location, software stack, utilization, and power cost all affect the value of the underlying unit. The market often talks about GPUs as if they are the commodity, but the more accurate commodity is usable compute.
The memory cycle is becoming increasingly central because AI workloads are extraordinarily memory intensive. Training and inference both require accelerators to access and move massive quantities of data at extremely high speed. If memory bandwidth is insufficient, the system cannot fully utilize the accelerator capacity that has already been purchased. In that environment, the limiting factor is not simply the number of chips, but the efficiency of the entire compute system.
This has important market implications. If advanced memory supply becomes tight, usable compute becomes tight. If usable compute becomes tight, inference costs remain elevated and model economics become more difficult. That pressure can flow into cloud margins, enterprise AI pricing, contract structures, capital expenditure plans, and hardware collateral values. Memory is therefore not merely a semiconductor subcategory; it is becoming one of the main swing factors in the price and availability of intelligence.
The Missing Price
The largest contradiction in the compute economy has been that enormous amounts of capital are being committed to an input that historically lacked a transparent public market price. There have been cloud pricing schedules, private contracts, token pricing models, GPU rental quotes, and internal depreciation assumptions, but no deep public forward curve capable of organizing the market. Lenders underwriting GPU-backed credit, operators signing multi-year compute contracts, hyperscalers funding data center capacity, and AI labs modeling inference economics have therefore been forced to make assumptions around an underlying price that has not been widely observable.
This is a structural problem. A market that cannot price an input cannot allocate capital around it efficiently, and it cannot hedge that input properly. Operators have to make decisions about depreciation schedules, future utilization, resale values, contract pricing, and financing terms without a liquid reference point. The result is a system where large amounts of capital are being deployed against an asset whose economic value remains difficult to mark in real time.
That pricing layer is beginning to emerge through benchmarks, futures, and perpetual-style instruments. These products remain early, and liquidity will ultimately determine how important they become, but their existence marks a meaningful structural change. Compute is moving from an opaque operating cost toward a financial market with reference prices, hedging tools, and eventually a forward curve. Once that happens, compute becomes legible to capital markets in a way it has not been before.
A visible compute price changes the structure of the AI trade. It allows the market to determine whether compute is cheap or expensive relative to future demand, gives operators a way to compare spot pricing against forward pricing, allows cloud providers and AI labs to hedge exposure, and gives lenders a reference point for collateral valuation. Oil became a macro variable not simply because it powered the economy, but because liquid futures markets allowed energy risk to be financed, hedged, and transmitted through rates, currencies, credit, and inflation expectations. Compute may now be entering the early phase of that same transition.
The Credit Loop
The financialization of compute matters because the AI buildout is increasingly connected to credit. A GPU is no longer just equipment sitting inside a data center; it is becoming a financeable asset that can support debt, contracts, and infrastructure investment. When demand is strong, the loop is powerful because capital flows into AI infrastructure, hyperscalers and neoclouds purchase hardware, suppliers receive the spending, lenders finance the assets, and collateral values remain supported by expectations of high utilization. This is the upside of financialization because it allows the buildout to scale faster than it could through retained earnings alone.
The same structure creates vulnerability if pricing or financing conditions change. If compute prices decline sharply, new chip generations compress residual values, utilization disappoints, or power costs rise faster than expected, collateral assumptions can change quickly. GPU-backed debt may then become more sensitive to marks, while infrastructure operators face pressure from higher funding costs and weaker resale values. The more credit becomes tied to compute collateral, the more important the compute price becomes for the broader financing environment.
This is why the compute commodity thesis cannot be separated from rates. Data centers, semiconductor facilities, grid expansion, power generation, networking infrastructure, and industrial automation all require substantial upfront capital investment. The economics supporting those projects become increasingly sensitive as rates rise and liquidity tightens. Even structurally strong themes can experience severe volatility when the market’s willingness to finance long-duration cash flows begins to compress.
The credit loop can therefore work in both directions. Strong compute demand can support collateral values, encourage lending, expand infrastructure investment, and reinforce the AI buildout. Weak compute prices or tighter financing conditions can pressure collateral values, reduce lending appetite, slow capex, and force investors to reassess valuations across the AI complex. Once compute has a visible curve, that curve becomes one of the key signals for the health of the entire regime.
Compute as a Hidden Inflation Channel
Compute also creates a hidden inflation channel. If compute prices rise, the pressure may not immediately appear in CPI, but it can still show up through corporate margins, cloud costs, software pricing, AI subscription fees, enterprise contracts, electricity bills, and capital expenditure budgets. The most visible official channel is likely power, because data centers require grid connections, substations, transformers, turbines, and long-duration electricity commitments. Monetary policy can influence the price of money, but it cannot directly create advanced memory, accelerators, transformers, gas turbines, or new grid capacity.
This is where compute begins to resemble energy from a macro perspective. A supply-side bottleneck in a strategic input can influence margins, inflation expectations, growth, and policy even when official inflation data captures it only partially. If companies become increasingly dependent on AI workflows while compute remains scarce, the cost pressure may first appear in equity margins and investment plans rather than headline consumer prices. That makes compute difficult for traditional macro frameworks because its economic significance can grow faster than its visibility inside standard inflation baskets.
The power constraint also explains why the AI cycle can support parts of the industrial economy even when other areas weaken. Data centers require steel, electrical equipment, cooling systems, construction labor, industrial machinery, turbines, transformers, and transmission infrastructure. As a result, AI demand does not remain contained within software or semiconductors. It leaks into the physical economy through the infrastructure required to turn electricity into intelligence.
This dynamic creates a more complicated macro environment. On one side, AI infrastructure spending can support growth, industrial activity, earnings for suppliers, and long-duration investment. On the other side, the same demand can create pressure across power markets, equipment supply chains, local grids, and capital budgets. The result is a cycle that can be growth-supportive and inflation-sensitive at the same time.
Market Structure and Speculative Flows
The financialization of compute is likely to change how investors express the AI trade. Today, most participants still express AI exposure through semiconductor equities, hyperscalers, power companies, data center beneficiaries, software platforms, or private infrastructure vehicles. Those proxies remain important, but they compress several different risks into a single trade. Compute pricing would allow the market to separate the input cost from the equities currently used as substitutes for that input.
Once compute markets become more liquid, investors could express more precise relative-value views. They could trade compute scarcity against semiconductor margins, memory pricing against model economics, power costs against cloud profitability, or infrastructure suppliers against the companies funding the buildout. The AI trade would become less of one broad narrative and more of a cross-asset market structure. This should increase dispersion across the ecosystem as the market begins separating the price of compute from the price of the companies exposed to compute.
This transition is unlikely to occur in a slow or purely institutional manner. Modern speculation is increasingly frictionless, continuous, social, and integrated across brokerages, crypto venues, prediction markets, options platforms, and mobile applications. Compute has nearly every ingredient required to become a major speculative narrative: technological importance, scarcity, unclear valuation, geopolitical relevance, large capital expenditure, and direct connection to the dominant equity theme in markets. If liquid compute instruments emerge, they are unlikely to remain quiet hedging tools for long.
That reflexivity can be productive and destabilizing at the same time. Speculation improves liquidity and price discovery, but it also introduces leverage, crowded positioning, volatility, and narrative feedback loops. Oil financialized over decades, while compute may financialize much faster because the market infrastructure for continuous speculation already exists. The result should be more opportunity across the AI complex, but also more volatility as the market learns how to price a new commodity in real time.
Geopolitics of Compute
Compute is also becoming a geopolitical resource. For much of the twentieth century, geopolitical power depended heavily on energy reserves, industrial capacity, shipping routes, military infrastructure, and access to dollar funding. Those variables remain important, but advanced compute is now joining the list. Nations increasingly compete for semiconductor fabrication, advanced packaging, memory supply, accelerator access, data center capacity, power availability, engineering talent, and sovereign AI capability.
The AI race should therefore not be understood simply as a race to build better consumer applications. It is increasingly a race to control the infrastructure that produces intelligence. Export controls around advanced chips, national semiconductor policies, state-backed investments into AI capacity, and strategic power development all reflect the same underlying reality. Compute is not only a commercial input, but an industrial, military, intelligence, and geopolitical input.
If compute becomes a liquid commodity, geopolitical risk may eventually be priced through compute curves in the same way oil markets price sanctions, wars, shipping disruptions, production decisions, and strategic reserves. Chip export controls, Taiwan risk, memory shortages, power constraints, cyber risk, data sovereignty rules, and national AI policies could all become variables in the market price of compute. The strategic reserve of the next era may therefore include not only energy inventories and critical minerals, but also access to secure compute capacity. That would represent a meaningful shift in how markets price national power.
Market Implications and Risks
The main investment implication is that AI leadership should continue evolving. The first phase of the trade rewarded the obvious beneficiaries, including semiconductor leaders, hyperscalers, and large technology platforms. That was logical because they controlled early bottlenecks, had the balance sheets to invest, and sat closest to the public narrative. As the cycle matures, however, the economics increasingly shift toward the suppliers of scarce infrastructure, especially memory, power equipment, electrical systems, cooling, networking, data center capacity, and grid-related assets.
This does not mean the hyperscalers lose strategic relevance. Many of them will remain central to AI deployment and may ultimately capture significant downstream value through cloud services, enterprise integration, platform distribution, and model access. The issue is that their cash flow profiles are changing as capital expenditures rise, depreciation grows, competition intensifies, and investors begin scrutinizing returns on hundreds of billions of dollars in infrastructure spending. In the near term, the companies receiving those expenditures may continue to have cleaner earnings leverage than the companies funding them.
The broader thesis remains constructive, but not unconditional. Commodity cycles rarely move in straight lines, particularly when they are tied to leverage, technological change, public-market narratives, and long-duration infrastructure financing. Compute can become one of the most important commodities of the next decade while still producing drawdowns, overinvestment cycles, credit scares, positioning unwinds, and valuation compression along the way. The structural view is that compute becomes increasingly unavoidable, not that every asset exposed to compute rises indefinitely.
The forward curve will become one of the most important signals in this cycle once liquidity develops. If long-dated compute prices trade below spot prices, the market may be assuming future supply growth overwhelms demand. If long-dated prices rise, the market may be pricing persistent scarcity across chips, memory, power, or data center capacity. Either outcome will provide information that the current AI equity complex does not fully reveal.
Conclusion
The market understands that artificial intelligence matters, but it still underestimates what AI is becoming. AI is not simply another software cycle; it is a physical infrastructure cycle built around the production of machine intelligence. The scarce input in that cycle is compute, while the quality differential inside compute is increasingly determined by memory bandwidth, interconnect, power availability, location, and system-level integration. Once compute develops benchmarks, futures, perpetual-style instruments, collateral structures, and forward curves, it becomes legible to capital markets in a way it has not been before.
Oil organized the twentieth-century industrial economy because it powered production, shaped inflation, influenced geopolitics, and became embedded inside credit and financial markets. Compute may organize the twenty-first-century intelligence economy for similar reasons. It powers training, inference, agents, robotics, automation, scientific discovery, and digital labor, while directing capital into semiconductors, memory, power, cooling, data centers, and grid infrastructure. The market still largely views AI as a digital story, but the structure developing underneath it increasingly looks like a commodity story.
The recent developments across memory, infrastructure financing, and compute pricing all point toward the same structural direction. Compute is becoming more than an internal technology cost or a private infrastructure contract. It is becoming an economic input, a financial asset, a geopolitical resource, and increasingly, a commodity. As intelligence becomes embedded throughout production itself, understanding compute may become as important to the next macro regime as understanding energy was to the last one.