The market’s obsession with silicon has run its predictable course. For the past three years, retail traders and institutional money alike have chased the same handful of hyperscalers and GPU designers, inflating multiples to eye-watering heights. Yet the most durable fortunes in the artificial intelligence boom are no longer being minted in Silicon Valley. They are being forged in the industrial parks of Ohio, the copper mines of Chile, and the nuclear cooling towers of Pennsylvania. The generative AI revolution has hit the physical limits of the material world. To scale further, the algorithms require a staggering mobilisation of concrete, electricity, and water.
The pivot from software to heavy industry is a mathematical certainty dictated by the sheer physics of compute. A standard Google search requires about 0.3 watt-hours of electricity. A single generative AI prompt consumes nearly 10 times that amount. When extrapolated across billions of daily interactions, the infrastructure deficit becomes glaring. The International Energy Agency estimates that global data centre electricity consumption will exceed 1,000 terawatt-hours by the end of 2026, an energy footprint roughly equivalent to the entire power demand of Japan.
This physical bottleneck has rewritten the entire tech investment playbook. Asset managers who spent 2024 obsessing over parameter counts and algorithmic efficiency are now scrutinising regional grid capacity and server thermal dynamics. The digital economy is fundamentally subsidising the industrial economy.
The Physical Price of Compute
Finding the unlikely AI stock winners requires looking past the blinking server racks and into the raw guts of the building itself. Liquid cooling systems, heavy-duty electrical switchgear, and backup generators are suddenly commanding premium valuations. Air cooling, the industry standard for 30 years, simply cannot dissipate the heat generated by high-density racks packed with H100 and B200 chips. The necessary transition to liquid-to-chip cooling has transformed obscure thermal management firms into critical choke points for the entire technology sector. Companies like Vertiv and Schneider Electric are no longer viewed as sleepy industrial suppliers; they are the absolute gatekeepers of the generative boom.
On March 14, 2025, hyperscale developers effectively admitted they were building faster than the supply chain could move. Lead times for specialised grid transformers and heavy-duty switchgear have quietly stretched from an average of 14 weeks to well beyond 80 weeks. Bloomberg’s tracking of electrical component supply chains reveals a persistent, structural shortage that forces data centre developers to warehouse equipment years before breaking ground. This hoarding behaviour has created unprecedented backlog visibility for legacy electrical manufacturers like Eaton and ABB, driving margin expansion that software companies would envy.
The materials required to build this infrastructure also present severe constraints. A 500-megawatt data centre—the new standard for frontier model training facilities—requires thousands of tons of high-grade copper for internal cabling alone, completely separate from the transmission lines needed to connect it to the grid. The mining sector, chronically underinvested for the past decade, cannot accelerate production fast enough to meet this concentrated spike in demand. Mining giants and commodities traders are quietly absorbing the capital flow that previously belonged exclusively to software-as-a-service startups.
Identifying the Second-Order AI Beneficiaries
What are the hidden stocks benefiting from AI? The most overlooked companies profiting from artificial intelligence are those controlling scarce physical and intellectual resources. These include utility companies holding nuclear generation assets, electrical equipment manufacturers producing grid transformers, and legacy media conglomerates licensing proprietary data archives to train large language models.
Power generation has become the ultimate proxy for AI expansion. You cannot deploy 100,000 GPUs if you cannot secure the baseline electricity to turn them on. Tech monopolies have realised that intermittent renewable energy sources—while politically popular—cannot provide the stable, 24/7 baseload power required by multi-billion-dollar compute clusters. This mathematical reality has triggered an unexpected renaissance for nuclear energy.
Utility companies that own existing nuclear fleets have suddenly found themselves holding the most valuable real estate in the technology sector. Constellation Energy and Vistra have seen their market capitalisations surge as hyperscalers aggressively bid for co-location rights. By building data centres directly adjacent to nuclear plants—known in the industry as “behind-the-meter” deals—tech companies bypass congested regional power grids entirely. These 20-year power purchase agreements guarantee utility providers a high, fixed price for their electricity while insulating the tech firms from grid instability. It is a massive wealth transfer from Silicon Valley to legacy power operators.
The Great Data Squeeze and Intellectual Property
Beyond the physical infrastructure, a different kind of shortage is minting another class of AI infrastructure companies. The largest models have effectively consumed the open internet. The low-hanging fruit of public data—Wikipedia, public forums, open-source code repositories—has already been scraped, tokenised, and ingested. As algorithmic progress increasingly depends on high-quality, domain-specific training data, the owners of that data hold enormous pricing power.
Legacy media empires and digital platform operators, long battered by the economics of the modern internet, are monetising their archives at staggering premiums. The Financial Times reported that AI developers are rapidly exhausting the supply of high-quality text, forcing them to sign multi-million-dollar licensing agreements with publishers just to feed their next-generation models. News Corp, Reddit, and Getty Images have transformed their historical content databases into high-margin revenue streams.
These deals aren’t merely about training data; they provide legal indemnification. As copyright infringement lawsuits wind their way through federal courts, holding a licensed data pipeline acts as regulatory insurance for the hyperscalers. For the publishers, it represents pure profit. The content was paid for and monetised years ago; licensing it to OpenAI or Google requires virtually zero marginal cost. This dynamic has turned old-guard publishers and niche platform owners into critical suppliers in the AI value chain.
The Bear Case: Overbuild and Capital Destruction
Still, the industrial mobilisation surrounding artificial intelligence carries echoes of historical manias. The bear case for these industrial and utility stocks rests on the assumption that tech monopolies are currently overbuilding in a panic, completely detached from actual end-user revenue.
During the late 1990s dot-com bubble, telecommunications companies spent hundreds of billions of dollars laying subterranean fibre-optic cables across the globe, anticipating exponential growth in internet traffic. When the bubble burst, the infrastructure companies collapsed under their own debt, leaving behind “dark fibre” that lay dormant for a decade. Sceptics argue we are witnessing the exact same misallocation of capital today. A highly circulated Goldman Sachs investment paper openly questioned whether the $1 trillion in planned AI capital expenditure will ever generate commensurate returns. If generative AI proves to be a highly useful but poorly monetised tool—rather than a completely transformative economic engine—the hyperscalers will inevitably slash their capital expenditure budgets.
If Amazon, Microsoft, and Google abruptly cancel their infrastructure build-outs, the multi-year backlogs at companies like Vertiv and Eaton will evaporate overnight. The power purchase agreements sustaining the nuclear utility stocks could face brutal renegotiations. The market is currently pricing these industrial suppliers as if the current rate of data centre construction will compound indefinitely. Physics dictates that growth curves eventually flatten, and when this one does, the cyclical nature of heavy industry will aggressively reassert itself.
That said, the technological arms race between sovereign nations provides an unnatural floor for this infrastructure spending. If corporate capital expenditure falters, government subsidies aimed at securing sovereign AI capabilities will likely fill the void. The Middle East, Europe, and Asia are all aggressively building domestic compute clusters, unwilling to rely entirely on American infrastructure.
The ultimate irony of the artificial intelligence boom is how intensely physical it has become. A technology designed to simulate human cognition and automate abstract thought is entirely dependent on men digging copper out of the Andes and engineers upgrading cooling valves in the American Rust Belt. Silicon Valley can write all the code it wants, but the future of artificial intelligence currently belongs to the companies that pour the concrete.
Discover more from Opinion Articles
Subscribe to get the latest posts sent to your email.
