Author: Jeremy Gorven
As the global race toward Artificial General Intelligence (AGI) accelerates, capital investment in AI infrastructure is reaching unprecedented levels. While this progress promises transformative gains in productivity, science, and innovation, it also introduces a significant — and growing — energy burden. For long-term investors and ESG-conscious stakeholders, the rapid expansion of compute-intensive AI systems poses a material challenge to near-term climate goals. This article explores the intersection of AGI development, big tech ambition, and the sustainability imperative, with a focus on the energy implications shaping the path ahead.
The global race for Artificial General Intelligence
Artificial General Intelligence (AGI) is the ability to efficiently acquire new skills and solve open-ended problems. The attainment of AGI will enable systems to invent and discover alongside humans, opening vast new possibilities. However, this advancement could also pose risks to near term climate-related goals, in particular the 2030 carbon neutral targets of many technology companies, given the substantial rise in energy consumption required to power AI data centres.
Advances in AGI are being achieved much faster than expected. For example, Open AI took 4 years to go from a 0% to 5% accuracy in a common AGI test (ARC prize – a benchmark test designed to evaluate a model’s ability to perform abstract reasoning and generalise pattern). In less than a year, they have improved to 88%. 100% represents the attainment of AGI. Sam Altman of Open AI believes that AGI could be achieved in 2025 and Elon Musk of xAI estimates that this could happen in 2025 or 2026. Given the potential economic value of AGI, the race to achieve it continues to escalate.
There is an AI race between the US and China driven by geopolitical reasons. For both countries, achieving superior AGI is a national imperative. This has resulted in supportive policies from both governments. The Trump administration seeks to further accelerate AI development in the US.
The US has thus far maintained a lead, but Chinese AI companies may be closing the gap despite fewer resources. Chinese AI startup DeepSeek has raised the stakes by achieving superior benchmark results than models from Meta, Open AI and Anthropic, with a model apparently built over 2 months for a fraction of the cost using NVIDIA’s reduced power H800 chips and using a Llama (Meta) open-source model.
Figure 1: ARC AGI score progression
Source: Arcprize.org
Scaling AI infrastructure and its energy demands
In AI, scaling laws refer to how greater compute is leading to significant improvements in AI. Open AI maintain that scaling laws remain intact and are expected to do so meaning that larger AI clusters are expected to continue to lead to better performance. Since the Chat GPT moment, AI capex has risen significantly. Estimates now reflect $234bn in capex spending in 2025 by the US Hyperscalers, more than double the capex spend in 2021. Further raising expectations is the recently announced $100bn – $500bn Stargate investment into AI infrastructure by Oracle, Open AI and Softbank.
The environmental impact of exponential compute growth
Previous annual improvements in computing power were far more gradual. NVIDIA, the leading supplier of AI systems, aims to improve computing power by approximately fourfold every year, leading to a million-time improvement over ten years. Greater computing power lowers the cost of compute, driving ever greater demand for data centre infrastructure to power AI systems. As such, data centre energy demand is expected to rise more than 5-fold over a decade, with the share of US energy production reaching 10% by 2028, from 2% in 2018.
- Despite the improving energy efficiency of chips, the sheer increase involved in compute makes AI enormously energy intensive.
- One study shows that each task on OpenAI’s o3 model is estimated to be equivalent to 684kg of CO2 emissions or 5 tanks of gas.
Figure 3: Total U.S. data center electricity use from 2014 through 2018
Source: Lawrence Berkley National Laboratory
Gas power: the energy backbone of AI infrastructure
Reliable baseload power is critical for AI infrastructure as a result of the unique and intensive demands AI places on data centres. Even minor interruptions in power can cause costly delays, loss of data, or the need to restart training processes. The primary sources of reliable baseload power are gas, coal, nuclear and hydro. Despite the recent hype around nuclear, a substantial portion of the new energy generation which will be required for data centres between now and 2030, is likely to come from gas generators.
Gas generators are far less complex and faster to deliver, making them ideal for short-term energy needs. For example, when xAI built its Colossus AI cluster in Tennessee, they immediately deployed mobile gas generators to provide immediate reliable base load power. The big downside of gas is of course the associated carbon emissions.
To meet climate goals of lower or net zero carbon emissions, gas power plants would need to be paired with carbon capture solutions. Direct Air Capture (DAC) to extract CO2 directly from the atmosphere, is a technology that does exist and is being further developed. AI players are investing in carbon capture projects such as Climeworks DAC (Microsoft). This is a promising approach, however, there are significant hurdles to large scale commercial use. These include, high energy intensity, high cost per ton CO2 captured, lack of infrastructure to store or utilise captured CO2 , or lack of policy and or policy certainty around long term subsidies.
The future of nuclear energy: large-scale vs. Small Modular Reactors (SMRs)
Large scale Nuclear provides reliable, almost carbon free, base load power. In 2024, Microsoft agreed a deal to re-open the mothballed Three Mile Island nuclear plant by 2028. However, new projects take 10-15 years to deliver and encounter delays, technological challenges, cost over-runs and regulatory hurdles. In the last 30 years, only 3 new nuclear reactors were completed in the US. The most recent, Plant Vogtle took 18 years from planning in 2006 to being fully operational in 2024, including 10 years of construction with costs of $30bn, more than double the $14 billion budget.
Given the challenges of large scale nuclear, the hyperscalers are investing in small modular reactor (SMR) startups such as Kairos Power (Google), SMR technology (Amazon), Oklo (Sam Altman) and Terra Power (Bill Gates). Bill Gates views SMRs as critical to achieving global net-zero emissions by providing a steady, reliable source of carbon-free energy to complement intermittent renewables. SMR technology has a variety of potential advantages including lower costs, scalability, flexibility, faster deployment and on-site or off-grid use. SMR is however a new technology with significant challenges facing commercialisation such as limited operational history to validate performance, safety, and cost claims. In addition, regulation, supply chains and waste management are all yet to be established. As such, the timeline for SMR to enable the first SMR clean energy solutions for AI infrastructure is 2030 at the earliest.
2030 targets are at risk
Given that new energy mix for AI data centres is likely to be carbon intensive until at least 2030, there is a strong likelihood that the near-term ambitions of the Hyperscalers may be unachievable, including net zero targets for Google, Microsoft and Meta by 2030. For example, in 2024, Microsoft reported a 30% rise in total greenhouse gas emissions since 2020, with Google’s emissions rising 48% over the same period. And the outlook is for rising AI capex, meaning more energy requirements and likely more climate impact.
Even if the Hyperscalers reach net zero via carbon credits, the net result for the world is likely to be substantial net additions to carbon emissions from urgent new energy initiatives linked to AI infrastructure. This reality is not lost on the industry, and the investments currently being made into both carbon capture and SMR technology are essential to the long run sustainability of AI infrastructure.
We will closely monitor these companies’ carbon pathways and seek to engage with them to encourage a strategy which emphasises the importance of their net zero targets to our clients and to the world, in the context of climate change. For instance, we would make it know out support for investments into technologies that could help to offset these increasing carbon emissions.
Balancing AGI innovation with climate commitments
The rapid advancements in AGI are reshaping technological capabilities at an unprecedented pace, but they come with significant energy demands that threaten near-term carbon neutrality targets. While AI-driven companies remain committed to sustainability, the projected reliance on gas power for data centers indicates a widening gap between ambition and reality.
Long-term solutions such as carbon capture and SMR technology hold promise but are still in early development stages, making it unlikely that they will offset AI-related emissions before 2030. Given this, the industry must take proactive steps to mitigate its environmental impact. As investors, we will continue to monitor and engage with key players, advocating for a strategy that prioritises sustainable energy solutions and reinforces the importance of achieving net-zero targets in the face of AI’s accelerating energy footprint.
Conclusion: Investing in sustainable AI infrastructure solutions
Artificial intelligence stands at the frontier of innovation, but it is becoming increasingly clear that its energy demands are misaligned with the industry’s net-zero timelines. As investors, we must scrutinise the full cost of AI advancement – not only in terms of capital allocation but also in environmental impact. While the long-term potential of carbon capture and next-generation nuclear is encouraging, they will not arrive in time to resolve the pressures of this decade. Responsible capital must now engage meaningfully with technology leaders, advocating for credible, transparent energy strategies that align growth with global climate commitments.
This article was first published as part of our 2024 5th Annual Stewardship and Sustainability Report, released on 28 March 2025.