Why AI Commodity Demand Survives an AI Stock Correction

Imagine every AI stock fell 50% tomorrow morning. The NASDAQ panics, model providers get repriced, and the headlines write themselves. And yet, at hundreds of construction sites around the world, the cement mixers would keep turning.

The concrete already poured does not care about equity valuations. The copper cable already ordered ships regardless. This is the tension at the centre of how markets misread AI commodity demand: investors bundle it with software multiples and treat it as one speculative bet, when the physical resource chain behaves nothing like a share price.

The construction pipeline, the grid interconnection queue, and the long-lead equipment orders are already placed. Those commitments are what this piece unpacks. After reading, you will be able to separate the demand for physical resources from the price of AI shares, and see clearly which part of the AI thesis survives a speculative correction and which part does not.

Why physical infrastructure spending is categorically different from software valuation

Here is the assumption most investors carry: if an AI bubble bursts, the AI-linked demand story unwinds with it. Copper, gas, and uranium exposure all get repriced alongside the software names, because they were part of the same trade.

That assumption confuses two entirely separate layers of the AI economy.

• The speculative valuation layer: software companies, AI startups, and model providers. Value here is a multiple on expected future monetisation. It can compress in weeks. Nothing physical is destroyed when it does.
• The physical infrastructure layer: data centres, power networks, cooling systems. Value here is committed capital that has been spent or contracted. It exists as concrete, copper, and transformers, and it takes years to build.

A correction in the first layer does not automatically propagate to the second, because the mechanism of demand destruction is different. A software multiple falls when sentiment shifts. A half-built data centre with signed grid contracts does not stop consuming steel because a share price dropped.

The scale of committed capital in the second layer is the point. Global data centre capital expenditure reached USD 455 billion in 2024, a 51% year-on-year rise, according to Dell’Oro Group. That is not positioning ahead of a trade. That is money already deployed into physical assets.

The pace has since accelerated. Combined capex for the top five hyperscalers exceeded USD 140 billion in Q4 2025, having roughly quadrupled since GPT-4’s release, according to Epoch AI.

Infrastructure investment cycles of this magnitude, where hyperscaler capex quadruples within three years, have historically preceded multi-year commodity demand curves that persist well beyond the initial technology cycle that triggered them, because the physical assets require ongoing fuel, maintenance, and grid capacity additions throughout their operational life.

Individual firms tell the same story. Amazon SVP and CFO Brian Olsavsky put the number plainly.

AWS capex in 2025 was roughly USD 125 billion, covering data centres and AI infrastructure.

Once permits are secured, civil works begin, equipment is ordered, and grid interconnections are contracted, that capital is committed. Cancellation becomes economically and contractually costly. That tells you the commodity demand embedded in this build-out is a near-term reality, not a future forecast waiting on AI shares to hold their value. The top 10 hyperscalers account for more than half of all global data centre capex, which concentrates the demand signal but does not make it speculative.

What data centres actually consume, and why electricity demand locks in commodity needs

Capex figures in the hundreds of billions are abstract. The energy appetite behind them is not.

A single hyperscale data centre drawing 100 MW or more consumes electricity comparable to roughly 350,000-400,000 electric cars per year, according to the IEA. That is one facility. The global pipeline contains a great many of them, already announced or under construction.

Data centre energy demand at the facility level translates directly into generation investment obligations, because utilities that accept a hyperscale campus onto their grid must plan sufficient firm capacity to serve it under peak and contingency conditions, not just average load.

This is where AI commodity demand stops being a bet on software and becomes a question about the grid. Electricity has to be generated, transmitted, and managed. Each of those steps has a physical resource attached.

The demand chain runs in sequence:

1. Electricity demand growth. Data centre consumption is rising steeply and is already in motion, not theoretical.
2. Generation capacity additions. Utilities must build new supply to meet that load.
3. Copper for grid infrastructure. Substations, transmission lines, and high-capacity cabling are copper-intensive.
4. Natural gas for dispatchable and backup power. Firm, on-demand capacity to keep facilities running.
5. Uranium for baseload nuclear generation. Continuous power for facilities that cannot tolerate interruption.

The demand curve is already moving. The IEA’s Energy and AI base case, published in August 2026, projects data centre electricity consumption more than doubling to around 945 TWh by 2030, approaching 3% of total global electricity demand.

AI is the driver, not a side story. AI servers already represented roughly 24% of server electricity demand and about 15% of total data centre energy demand in 2024, according to Nature and IEA figures, with that share expected to rise.

From a single facility to a global demand curve

Scale the 100 MW facility up to the full pipeline and the aggregate arrives as arithmetic rather than assertion. Dedicated electricity generation for data centres is projected to rise from about 460 TWh in 2024 to over 1,000 TWh by 2030, and to approximately 1,300 TWh by 2035, per the IEA’s Energy and AI report.

That doubling of dedicated generation tells you utilities, grid operators, and fuel suppliers are already planning for this load. The investment cycles for copper, gas, and nuclear fuel are being activated now, not when AI proves a commercial case.

Geography sharpens the signal. The United States accounts for roughly 45% of global data centre electricity load, according to IEA and DataCenter Frontier data, which is where grid and fuel investment pressure is most acute. Your read here is straightforward: AI commodity demand is not a wager on whether any single AI product succeeds commercially. It is a wager on whether the physical grid built to power these facilities will need fuel, and grids always need fuel.

The Dot-Com parallel, what the fibre-optic lesson actually teaches commodity investors

When the Dot-Com bubble peaked in March 2000, most of the capital destroyed was concentrated in conceptual, early-stage software ventures. The valuations collapsed. The fibre-optic networks and long-haul telecom systems built during that boom did not.

That physical infrastructure persisted and became the backbone of the internet expansion that followed. The companies that funded it lost most of their market value; the cables in the ground kept carrying traffic.

The same structural mechanism operates today. AI-ready data centres and the power infrastructure serving them are long-lived physical assets. They will continue operating and consuming fuel regardless of what happens to AI company share prices, because the cost of decommissioning recently built generation and network assets typically exceeds the cost of continued operation.

The regulatory framework reinforces this. Once a large facility is energised and grid-connected, system operators treat it as a must-serve load.

A must-serve load is a large consumer, such as an industrial facility or data centre, that grid operators plan to supply under almost all conditions. Once a data centre reaches this status, its electricity demand becomes a planning obligation for the utility, not a discretionary variable that moves with technology sentiment.

The IEA frames current regulatory debates around AI data centres as questions of managing and decarbonising their energy use, not preventing their deployment. The infrastructure is treated as a permanent fixture.

The UK critical national infrastructure designation for data centres, confirmed in 2024, placed these facilities alongside water and energy networks in the regulatory hierarchy, making their continued operation a government planning obligation rather than a commercially discretionary outcome.

Where the analogy holds and where it breaks down matters for how much you trust the conclusion:

• Funding model: Dot-Com build-outs were often funded on speculative future revenue. Today’s are funded largely from existing hyperscaler cash flows.
• Asset longevity: Both cycles produce durable physical assets that outlast the valuations that financed them.
• Revenue linkage: Dot-Com business models were frequently pre-revenue. Modern hyperscalers integrate AI into profitable cloud, advertising, and productivity businesses, tying compute more closely to current revenue.
• Regulatory treatment: The 2000 telecom boom was largely unregulated. Today’s build-out sits within climate policy and grid decarbonisation frameworks that channel demand toward particular fuels.

That closer revenue linkage strengthens rather than weakens the structural demand case. The mental model to take from this: when physical assets are built, grid-connected, and designated must-serve, the commodity demand they represent becomes structurally anchored in a way software valuations never are.

The honest counter-argument: where AI commodity demand could actually weaken

The structural case is strong, but it is not bulletproof. A reader who has followed it this far deserves a genuine stress-test, not reassurance.

Three risk vectors could weaken AI commodity demand relative to the central projections:

• Capex reversal. Hyperscaler spending nearly tripled year on year to USD 142 billion in Q3 2025, according to Synergy Research Group. With more than 50% of global data centre capex concentrated in 10 firms, a coordinated slowdown in response to disappointing AI monetisation could moderate power and material demand growth far faster than the base case implies.
• Efficiency gains. The IEA explicitly models energy efficiency improvements and hardware performance gains that partially offset higher compute demand. The 945 TWh 2030 figure is a central case, not a floor. Better chips, improved cooling, and model optimisation could reduce absolute consumption growth.
• Policy and grid constraints. Regulatory scrutiny of AI’s energy footprint is growing. Efficiency mandates, low-carbon sourcing requirements, and grid bottlenecks in dense data centre clusters could cap demand or shift the fuel mix away from gas and toward renewables, moderating commodity intensity.

CIO Dive, citing Synergy Research Group, characterised hyperscaler capex as having “ballooned” and “fuelled an AI bubble,” naming capex retrenchment and project delays as the key unwinding mechanisms. This is not a fringe worry. It is a documented risk at the demand origination layer.

AI capacity expansion risks are most acute at the growth layer above the structural floor, where hyperscaler spending decisions remain discretionary and concentration among a handful of firms means a coordinated pause could slow demand trajectories faster than grid planning cycles can adjust.

The concentration point deserves emphasis. With demand originating from so few firms, AI commodity demand carries a genuine single-point-of-failure risk: a pause by a small group of hyperscalers could slow growth meaningfully, even where the structural logic holds for facilities already under construction.

What is already locked in versus what remains contingent

The honest read separates two categories.

Locked in: facilities under construction, grid connections already contracted, and long-lead equipment already ordered. Cancelling these is economically and contractually difficult, which is what creates the demand floor. That copper, gas, and uranium demand is coming regardless of where AI shares trade.

Contingent: everything above that floor. Continued growth depends on discretionary hyperscaler investment decisions, which can be paused, deferred, or redirected. The practical takeaway is that the structural floor is real, but the trajectory above it is not guaranteed.

What the structural floor actually means for commodity exposure over the next decade

Pull the threads together and a clear framework emerges. Dedicated generation for data centres is projected to rise from 460 TWh in 2024 to over 1,000 TWh by 2030, and to around 1,300 TWh by 2035, per the IEA. That trajectory is embedded in utility planning cycles, grid investment decisions, and long-lead equipment orders already underway.

Those multi-year construction timelines and must-serve load designations are the mechanisms that create the floor. It persists even under a significant AI valuation correction, because the physical assets and their grid obligations do not disappear when a share price does.

This is why commodity exposure sits in a different category from AI software exposure. Software valuations track monetisation expectations. Producers supplying copper, gas, and uranium into the data centre chain are exposed to the physical build-out cycle instead. The pace of that cycle is real: hyperscaler quarterly capex has quadrupled since GPT-4’s release, according to Epoch AI.

For readers wanting to understand how tightening mine supply interacts with infrastructure-driven demand, our full explainer on copper supply constraints examines the structural gap between projected data centre copper consumption and current global mine output through 2030.

The 2030-to-2035 generation trajectory tells you the outer edge of the committed planning horizon is already priced into infrastructure decisions. AI commodity demand has an operational runway extending well beyond the current speculative cycle in software valuations.

Three variables are worth watching to judge where demand heads next:

• Hyperscaler capex decisions. The demand origination signal. A slowdown here is the first warning that growth above the floor is weakening.
• Grid interconnection queue data. The construction commitment signal. A lengthening queue confirms committed load moving toward energisation.
• AI hardware efficiency improvement rates. The energy intensity signal. Faster efficiency gains could soften absolute consumption growth even as compute rises.

That leaves you with a two-layer view: a structural demand floor anchored in committed construction, and a growth trajectory above it that depends on continued hyperscaler spending. It is a more precise lens than the binary “AI boom or AI bust” framing that dominates market commentary.

This article is for informational purposes only and should not be considered financial advice. Investors should conduct their own research and consult with financial professionals before making investment decisions. Past performance does not guarantee future results, and financial projections are subject to market conditions and various risk factors. Forward-looking statements referenced here are speculative and subject to change based on market developments.