Sustainable investing: power, compute and capital – inside the AI infrastructure supercycle

The rise of artificial intelligence (AI) is proving to be both the most disruptive force investors have faced and one of the most compelling structural opportunities. AI is not just changing individual business models; it is reshaping the economy itself and, in turn, the very framework through which we evaluate what constitutes a sustainable, durable business.

AI is unfolding as a multi-year, capital-intensive infrastructure build-out constrained by physical bottlenecks: land, grid connections, electricity, chips, cooling and capital. For sustainable investors, that shifts the debate from ‘software adoption’ to ‘industrial expansion’ – and it demands a more realistic view of near-term carbon trade-offs.

Tokens: The output of “intelligence factories”

Nvidia’s CEO Jensen Huang argues we should stop thinking about AI data centres as cost centres that
store files and start thinking of them as “intelligence factories”. Their output is tokens: the units of input and output AI models process to generate useful work. In that framing, the AI economy’s key throughput metric is token consumption – how much intelligence the system is being asked to produce.

The demand signal is already extreme. Weekly token consumption across the AI ecosystem has risen more than 3,800% over the past year. ¹ Looking ahead, forecasts suggest agentic AI (systems that execute multi-step tasks on a user’s behalf) could push token consumption 24x to around 120 quadrillion tokens per month by 2030 ², with enterprise agents a major multiplier (Figure 1).

Figure 1: Estimated token usage per month could be 24x today’s global capacity by 2030

Source: Strategas, as at 25 November 2025. There is no guarantee that past trends will continue or forecasts will be realised.

Power is the binding constraint – efficiency is the moat

Token growth runs into a hard ceiling: power. Operators can build “land, power and shell”, but once facilities exist, many are power limited. The implication is stark. In a world where every data centre operator is power-constrained, the operator who can generate 50x more tokens from the same power envelope generates 50x more revenue. Within a fixed electricity envelope, competitive advantage is determined by how many tokens can be produced per unit of energy.

This is why the arms race is shifting from “how much compute” to “how much compute per watt” and, increasingly, “tokens per watt”. Hardware, memory, networking and cooling are now evaluated as one production line converting electricity into tokens. Small gains in efficiency can translate into outsized gains in throughput (and economics) when power is scarce.

The scale of the build-out reflects both demand and constraint. Some estimates imply AI computing capacity is doubling roughly every seven months this decade, while broader industry analysis suggests global data centre spending could reach US$7 trillion by 2030.³ Physical constraints – grid capacity, manufacturing lead times and supply chain bottlenecks – are not dampening demand; they may extend the cycle.

The sustainability paradox: Higher near-term intensity, potential long-term benefit

Here is the uncomfortable reality: the AI build-out is carbon intensive. Semiconductor fabrication, heavy electrical equipment, grid reinforcement and cooling are energy hungry, asset heavy activities. Then, once data centres are built, they require continuous electricity, often from grids that still rely materially on fossil fuels.

Yet AI also has the potential to be a powerful productivity and efficiency multiplier for the broader economy: accelerating materials science for clean energy, optimising grid operations and industrial processes, and reducing the resource intensity of activities from drug discovery to infrastructure design. The sustainability question is not whether AI has a footprint – it does – but whether system-wide efficiency gains can outweigh transition costs over time.

What it means for sustainable investing: A systems-level lens

This is why we frame our approach as systems-level sustainability. Instead of judging companies solely on their standalone operational emissions, we assess whether they enable a more efficient, resilient and ultimately lower carbon economic system, including by improving the efficiency and carbon intensity of AI itself.

Practically, that points investors toward the bottlenecks where incremental efficiency gains can unlock disproportionate value: power and electrical infrastructure; thermal management and cooling; core computing and advanced manufacturing; memory/bandwidth and storage; and energy storage and
power management.

Stewardship matters. Engagement on credible transition plans, capital allocation discipline and transparent reporting becomes central, including acknowledging that portfolio carbon intensity may rise at times as exposure shifts towards the physical backbone of the AI economy.

Conclusion: Sustainability was always physical

For years, sustainability narratives often sat comfortably alongside capital light digital models. AI is a reminder that the modern economy runs on physical inputs: energy, metals, grids, factories and infrastructure. As token consumption surges, the scarcity value of power and efficiency rises with it.

For sustainable investors, the task is to invest in the AI infrastructure supercycle without ignoring its near-term costs – focusing on the technologies and assets that raise tokens per watt, reduce carbon intensity over time, and enable a more productive, resilient economy.