6 February 2025
Hugh Falcon, Senior Sustainability & Energy Consultant
Two years on from ChatGPT’s release, the global AI frenzy continues and the role of data centres in powering AI infrastructure undoubtedly remains front and centre in the minds of investors.
This theme has continued to dominate news flow in 2025, with the Trump administration announcing a further ramp in US AI infrastructure investments through the OpenAI-led Stargate project, which plans to invest $500 billion by 2029[1]. Two days later, the state-owned Bank of China’s announced that it will provide one trillion yuan ($138 billion) in support for the AI industry supply chain in the country over the next five years[2].
Appetite for AI infrastructure from investment firms is no different: globally there were 50 real estate funds participating in capital raising actively seeking exposure to data centres in October 2024, with $20 billion earmarked for this specific asset class[3].
The combination of AI becoming an even greater national security interest, coupled with insatiable private sector investment demand, leads us to the following question: what could possibly cause a slowdown in this Dot Com Bubble-style buildout?
Is DeepSeek a gamechanger for the AI investment landscape?
In a move which has caused investors to question their unwavering demand for all things AI, DeepSeek, a Chinese AI company, recently publicly released its latest large language and reasoning models, DeepSeek V3 and R1. The models match the performance of leaders such as ChatGPT 4o and o1 on many benchmarks, but crucially, they were claimed to be significantly cheaper to train and can be operated with lower costs[4].
This can be explained by the advanced architecture of both models, which allow them to ‘think’ far more efficiently and economically – a major breakthrough in the industry. Specifically, Deepseek’s novel use of Multi-Token Prediction for its V3 foundation model, and use of Chain-of-Thought architecture for distilled reasoning in its R1 model, has enabled the company to significantly drive down the cost of developing and using AI.
While there is ongoing debate regarding the true cost of training the model and the hardware it requires (i.e., the type and quantity of graphics processing units (GPUs) used)[5], as well as the fact that V3 and R1 seem to have used data from other previously released closed-source LLMs in their training[6], this marks a change in character for future investment plans. The focus has changed from (pre) training-dominated improvements to improvements gained through reasoning advancements, which may impact the global demand for further investment in large scale AI data centres globally, at least in the near term.
Despite this, there is nuance to this pivot: many have applied ‘the Jevons Paradox’[7] to gauge the future implications of Deepseek’s innovative approach to AI development. In this context, the greatly reduced cost to develop and operate Deepseek V3 and future LLMs could result in far higher usage of AI technologies, ultimately leading to the need for more power, computing, and data centres to accommodate this demand.
The question of whether investment in AI infrastructure will slow therefore likely needs to be reframed. We need to move from simply looking at the future steepness of the demand curve, to a different question: what will contribute to this expected exponential growth in data centre demand for the foreseeable future?
As breakthroughs in the efficiency and reasoning capabilities of LLMs such as Deepseek’s continue, AI agents should proliferate. These essentially enable automation of entire project tasks or workflows with minimal human oversight. AI agents will rely on inferencing to carry out tasks, which will require a great deal of computing power from data centres. This gives us the confidence to predict continued high growth for the asset class over the years to at least 2030.
Data Centres and AI in the Context of the Climate Crisis
In a world already grappling with the climate crisis and associated obstacles of the energy transition, this ‘full-steam-ahead’ AI data centre build-out puts further pressure on global climate change mitigation efforts.
Data centre power consumption already made up 4.4% of the US’ total energy demand in 2023 and is projected to increase to as much as 12% by 2028[10], and much of this incremental demand will likely be met with fossil fuel combustion[11].
To taper this projected growth in power demand, innovative and efficient approaches to AI training and inference, such as those used in DeepSeek V3, will need to be developed quickly to ensure any hope of reaching the goals outlined in the Paris Agreement.
Alongside relying on hardware and software efficiency gains, data centre owners, operators and developers should be framing investment decision-making in the context of both physical and transition climate risks. For example, regarding physical climate risks present in regions experiencing water stress, closed-loop evaporative or immersion cooling technology for data centres is more appropriate to deploy[12]. Data centres also need to be future-proofed against transition risks by being situated in areas with sufficient renewable energy supply to minimise their carbon footprint.
While the AI revolution presents many opportunities for tackling dangerous climate change, reducing the climate impact of data centres must be prioritised as the AI infrastructure build out continues. Decarbonisation and efficiency measures including cooling optimisation, server management and virtualisation, demand response and on-site renewable energy can be deployed throughout design and operation to ensure that data centres align with net-zero pathways. Data centres have the opportunity to deliver huge technological benefits while limiting their impact on the climate: it is imperative that these two objectives align as we move forwards into the age of AI. This way we can ensure that we see the envisaged benefits of AI in solving problems associated with the climate crisis before it is too late.
As a specialised ESG consulting firm, Longevity Partners can assist you in improving the sustainability of your data centres through services such as life cycle assessments (LCAs), sustainability due diligence and data centre sustainability certification services. To find out more, please contact info@longevity.co.uk
[1] OpenAI, SoftBank, Oracle to invest $500 bln in AI, Trump says | Reuters
[2] Bank of China to provide comprehensive financial support to help the development of the artificial intelligence industry chain
[3] Global real estate funds target more than $20bn for data centre investment amid growing investor interest | News | About Us | Linklaters
[4] (PDF) DeepSeek-V3: A High-Performance Mixture-of-Experts Language Model
[5] DeepSeek Debates: Chinese Leadership On Cost, True Training Cost, Closed Model Margin Impacts – SemiAnalysis
[6] DeepSeek Debates: Chinese Leadership On Cost, True Training Cost, Closed Model Margin Impacts – SemiAnalysis
[7] https://x.com/satyanadella/status/1883753899255046301?mx=2
[8] Blackstone holds firm on data center investments despite DeepSeek turmoil | Reuters
[9] Microsoft, Meta back big AI spending despite DeepSeek’s low costs | Reuters
[10] lbnl-2024-united-states-data-center-energy-usage-report.pdf
[11] Chevron to build gas plants to power data centers amid AI boom | Reuters
[12] https://www.microsoft.com/en-us/microsoft-cloud/blog/2024/12/09/sustainable-by-design-next-generation-datacenters-consume-zero-water-for-cooling/