Virtual realities

Consensus estimates put aggregate capital expenditure by leading hyperscalers and AI infrastructure companies near USD 940 billion in 2026 and USD 1.3 trillion in 2027.¹ Private estimates point the same way, with UBS putting cumulative AI capex spending near USD 4.7 trillion between 2026 and 2030.²

Similarly, McKinsey’s split of future AI capex allocations suggests roughly 60% to chips and hardware, 25% to power and electrical equipment and about 15% to land and site development.² Put simply, the money is pouring into things you can drop on your foot and, importantly, the investment has grown too large to be funded from company balance sheets alone, making capital markets central to the build-out.⁴

Electricity is a key binding constraint at the moment. AI racks are far denser than traditional servers and the electric grid was not built for them. Interconnection queues stretch into years, transformers sit on long lead times and skilled labor is scarce. Although still modest, the aggregate figures are rising fast with US data centers consuming around 4% of national electricity in 2023 and projected to reach 7 to 12% by 2028.⁵ In Ireland, data centers already draw about a fifth of the country’s electricity, a share that could approach a third within a couple of years, which has pushed the grid operator to curb new connections around Dublin.⁶

Securing this vital power is becoming a discipline in itself. With long interconnection queues, developers are increasingly contracting firm, always-on supply directly – including long-term nuclear power-purchase agreements, on-site gas generation and batteries to firm up renewables⁷. The economics increasingly turn on tariffs, grid-connection timing and who ultimately bears the cost of the supporting infrastructure.⁸

The footprint also covers land and water. A single hyperscale campus can draw the power of a mid-sized city and the water of a small town to cool it.⁹ And although closed-loop and liquid cooling can cut a site’s own water use, they tend to raise its electricity demand, with much of that electricity coming from thermal power plants that themselves consume large volumes of water.

It is worth being precise here, because the environmental debate around AI-related activity is rife with misinformation.

The viral claim that a single AI prompt “drinks a bottle of water” is a misreading of a 2023 University of California, Riverside study that actually described roughly 500 milliliters per 10 to 50 responses – not per prompt.¹⁰ OpenAI’s Sam Altman has put a typical ChatGPT query at about 0.34 watt-hours and a fraction of a teaspoon of water, and an independent estimate from Epoch AI broadly agrees. But, as with any average, the tail matters: the same Epoch analysis finds energy use rising to 2.5-40 watt-hours for queries with very long or complex inputs. Newer “reasoning” models and agentic workflows can therefore draw far more than the simple chatbots these headline numbers describe.^11,12

So while there is a genuine footprint to be concerned about, we need to look at aggregate scale and indirect consumption – the water and power drawn by the plants feeding the racks (the Scope 3 if you will), rather than the mere per-prompt scare figures, which the IEA estimates account for the majority of data-center water.¹³

Step back far enough and the physical bottlenecks described above are a symptom of a much larger condition. In 1969 Buckminster Fuller described the planet as “Spaceship Earth”.¹⁴ By this he meant a vessel with finite supplies, no resupply and a crew that must manage its resources or perish.

AI is one stress test of that closed system – a new and fast-growing claim on the ship’s fixed stores of power, water, materials and the capacity to shed heat.

The most vivid response to this limit is the attempt to leave it behind. In early 2026 SpaceX filed to build a constellation of up to a million solar-powered satellites as orbital data centers, and folded in xAI ahead of a record-breaking listing.^15,16 Musk is not alone in this ambition, Google, Amazon, Nvidia-backed Starcloud (which flew a GPU in orbit in late 2025) and China’s space program are all chasing versions of the same idea.¹⁷

Serious skepticism exists. Independent estimates put the cost of power in orbit far above terrestrial solar or gas¹⁸, rejecting heat in a vacuum demands radiators at vast scale; the enabling rocket is still in testing; and orbital hardware is clearly challenging to maintain.

So the deeper point, in effect the one Fuller was making all along, is that you do not escape a closed system by bolting on a bigger one. Even the orbital dream is another physical-infrastructure bet – chips, launch, solar arrays and radiators. It merely widens and relocates the constraint rather than removing it.

But Fuller also understood the social and relational implications of a crew sharing a single, finite vessel. How the costs and gains from the system are distributed is just as important as the consumption volume. This is, arguably, where AI’s next collision lies.

Displacement of labor will likely be concentrated rather than universal. Language- and code-intensive white-collar tasks appear to be the most exposed.¹⁹

Most corporate AI projects still aim to augment rather than fully automate – barely a fifth target a hands-off end state.²⁰ So the near-term reality is roles reshaped more than eliminated, with the first pressure showing up in hiring, especially at entry level, rather than in wages.²¹ But this is cold comfort if it hollows out the bottom rungs of career ladders and concentrates the gains at the top.

Nor is exposure limited to white-collar work. Transport and logistics show how quickly the boundary can move once technology, capital and regulation align. Take John Smith, CEO of FedEx Freight, for example, who has alluded that autonomous trucks are close to operational readiness, with deployment now depending less on whether the technology can function than on when regulators allow it to scale.²²

Who actually captures the benefits and bears the costs is a vital question. The same AI dollar is at once a capital-expenditure boom, a potential labor substitute and a margin expander. And firms are increasingly restructuring around AI and citing efficiency, even where the productivity payoff is unproven.²³

For equity holders, that equates to potential margin expansion; for the wider economy, however, it raises an uncomfortable issue with circularity. A consumer-driven economy needs consumers with incomes and if AI lifts profits by displacing the workers who are also the customers, who sustains demand?²⁴

In an orderly transition, new AI-complementary work absorbs displaced labor and the productivity premium lifts incomes. In a disorderly one, however, the anticipation of job losses drives precautionary saving, softens growth and forces governments to step in with support, retraining and redistribution.²⁵

AI systems now sit inside decisions that are both consequential and legally protected: i.e., who gets hired, gets credit, gets insured, gets housing and so on. Where those systems learn from historic patterns, they tend to reproduce them. In Mobley v. Workday, an applicant alleged that AI-driven recruitment screening disproportionately rejected older, Black and disabled candidates; the court has allowed age-discrimination claims to proceed as a collective action.²⁶ Arguably the most relevant detail for investors is that plaintiffs are pursuing the vendor, not only the employer.²⁷

The stakes are not lost on the field’s own leaders. In 2023 the most-cited AI researchers – Geoffrey Hinton and Yoshua Bengio among them – joined the heads of the major labs in a one-line statement declaring that “mitigating the risk of extinction from AI should be a global priority alongside… pandemics and nuclear war.”²⁸

Although the regulatory architecture is firming up, the timetable is shifting. Under the EU AI Act, the high-risk obligations have been pushed to December 2027.²⁸ Existing anti-discrimination law already applies, litigation is not waiting for the statute, and a widening set of duties around data provenance, privacy and human-rights due diligence is arriving regardless.

In terms of stewardship and engagement norms, a striking gap is opening up. Research from the Oxford AI Governance Initiative finds that asset owner demand for AI governance is currently low, and that where asset managers do engage, it often reflects the manager’s own materiality judgment rather than client pressure (for now).³⁰

Governance quality should be thought of as a proxy for control in AI, as it is and has been for years in many other areas of corporate activity. A company that cannot explain how its models reach consequential decisions seems unlikely to be able to effectively govern them either, and could well be carrying contingent liabilities that no one has yet priced.

Taken together, the physical, finite and human elements of the AI roll-out are imbued with both opportunities and risks.

As investors, we often focus on limiting constraints when searching for untapped value. In this case, the value may sit in contracted and reliable power; transmission, transformers and grid equipment on multi-year lead times; the commodities that electrify it; and digital infrastructure with assured power and permits. The implications are not hypothetical as AI-linked names already drove outsized market swings in 2026, repricing sharply on each shift in sentiment about the build-out’s durability.

Thinking about the issue in Spaceship Earth terms has the effect of turning efficiency and circularity into investment themes. Liquid and closed-loop cooling, water recycling and heat reuse; grid efficiency and long-duration storage; and the hardware-longevity and e-waste chain. Orbital compute, meanwhile, is perhaps best held as long-dated optionality – a call option on launch costs collapsing – rather than an allocation.

There is also a reflexive angle in which AI is not just the subject of sustainable investing but increasingly a tool for it, used to read ESG signals from filings and to analyze satellite imagery for deforestation, methane and Scope 3 emissions, testing corporate claims against physical evidence.³¹ Google’s DeepMind, for example, cut the energy used to cool its data centers by about 40% – a 15% reduction in overall power-usage effectiveness – by letting a model, rather than static setpoints, run the cooling plant.³² That early result drew later questions over reproducibility, and production systems tend to report more modest gains of 14 to 21%, but it served as a proof-of-concept for others to follow.³³ In effect, AI can be both hero and villain.

Ultimately, turning models into deployed capacity requires power that can be delivered, water that can be justified, materials that can be sourced, a workforce that can adapt and governance credible enough to withstand scrutiny. Much of that cost sits outside a model provider’s reported unit economics but inside society’s balance sheet.

It makes sense to treat large data-center projects less like ordinary technology capex and more like airports, ports or utilities – assets whose returns depend on physical capacity, planetary accounting and social legitimacy being underwritten together. Investors need to figure out whether the system supporting AI can scale responsibly, profitably and with public consent. It is here that the physical, finite and human dimensions converge.

¹ Bloomberg Consensus Estimates, August 2026
² Invest in transformational innovation, UBS CIO GWM, November 2025.
³The cost of compute: A USD 7 trillion race to scale data centers, McKinsey & Company, April 2025.
⁴Key Questions on Energy and AI, International Energy Agency, 2026.
⁵2024 United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory / US DOE, December 2024.
⁶AI: Five charts that put data-centre energy use – and emissions – into context, Carbon Brief, September 2025.
⁷Beyond the Hype: Assessing Hyperscaler Nuclear Commitments Against U.S. Energy Realities, Carnegie Endowment for International Peace, June 2026.
⁸Key Questions on Energy and AI, International Energy Agency, 2026.
⁹Data Drain: The Land and Water Impacts of the AI Boom, Lincoln Institute of Land Policy, February 2026.
¹⁰Google games numbers to make AI look less thirsty, The Register, August 2025.
¹¹ Sam Altman: ChatGPT queries consume 0.34 watt-hours of electricity and 0.000085 gallons of water, Data Center Dynamics, June 2025.
¹² OpenAI's Sam Altman superintelligence blog (Epoch AI GPT-4o estimate), The Register, June 2025.
¹³ Energy and AI, International Energy Agency, 2025.
¹⁴ R. Buckminster Fuller, Operating Manual for Spaceship Earth, 1969
¹⁵ SpaceX files plans for million-satellite orbital data center constellation, SpaceNews, Jeff Foust, January 2026.
¹⁶ Will data centers in space work? Elon Musk says yes, NPR, April 2026.
¹⁷ SpaceX files for million satellite orbital AI data center megaconstellation, Data Center Dynamics, June 2026.
¹⁸ SpaceX Vow To Loft 1 Million AI Satellites Could Spark Doomsday Dive, Forbes, Kevin Holden Platt, May 2026.
¹⁹ Generative AI and Jobs: A Refined Global Index of Occupational Exposure (ILO Working Paper 140), International Labour Organization & NASK, May 2025.
²⁰ The AI and labor landscape 2026, S&P Global, July 2026.
²¹ Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Brynjolfsson, Chandar & Chen, Stanford Digital Economy Lab, November 2025.
²² FedEx Freight CEO says self-driving trucks are ready for prime time, The Wall Street Journal, 1 June 2026.
²³ Every major tech layoff in 2026 that has name-checked AI, TechCrunch, July 2026.
²⁴ The simple macroeconomics of AI, Daron Acemoglu, Economic Policy 40(121), January 2025.
²⁵ Gen-AI: Artificial Intelligence and the Future of Work, IMF Staff Discussion Note SDN/2024/001, January 2024.
²⁶ Workday AI Lawsuit Explained: Implications for HR, OutSolve, April 2026.
²⁷ Why AI Hiring Discrimination Claims Are More Dangerous to Employers Than They Look, ABA Business Law Today, June 2026.
²⁸ Statement on AI Risk, Center for AI Safety, May 2023.
²⁹ EU Approves Delays and Other Amendments to Certain EU AI Act Obligations, Morgan Lewis, June 2026
³⁰ The Role of Investors in AI Governance, Oxford AI Governance Initiative, April 2026.
³¹ Addressing the “AI” in sustainability, Bank of America Institute, September 2025.
³² Evans, R. & Gao, J., DeepMind AI Reduces Google Data Centre Cooling Bill by 40%. Google DeepMind, 2016.
³³ Zhan et al., Data Center Cooling System Optimization Using Offline Reinforcement Learning, 2025.