Artificial intelligence data centres
Highlights:
- AI datacentres are currently estimated to consume around 1.5 percent of the world’s electricity.
- The recent dramatic increase in AI-related patent filings appears to have reached a plateau.
- Patent classification trends reveal shifting priorities within the AI sector: patent filings for applications of neural networks and combinations of neural networks are continuing to increase.
- Innovations in datacentre cooling systems and carbon-aware workload distribution offer opportunities to mitigate the environmental impact of AI.
A typical single generative AI prompt can use 10 to 100 times the energy of a simple internet (e.g. Google) search.
There is an increasing awareness that Artificial Intelligence (AI) is having an effect on almost every aspect of human endeavour. The cost of providing complex and comprehensive AI is not measured solely in the cost of purchasing and running computing equipment.
There is an increasing and potentially worrying environmental cost. This is due to the large amount of energy required to operate the infrastructure (e.g. data centres or server farms, comprising many computers).
Some of the estimates of the combined energy usage of all AI systems compare to the energy usage of entire countries. Energy usage at such a level has an inevitable environmental cost.
This area of technology is changing rapidly, and it is often difficult to give precise figures for the energy usage. With that caveat, we have researched the best available sources in researching this article.
The International Energy Authority (IEA) reports that the amount of energy consumed by datacentres worldwide in 2024 is estimated to have amounted to around 415 terawatt hours (TWh). This represents approximately 1.5 percent of the entire global electricity consumption, comparable to the annual consumption of large industrialised nations, such as Saudi Arabia, Iran or Mexico. By the end of 2025, this is projected to rise to between two and four percent of global electricity consumption. By 2030, the IEA estimate that global electricity usage could exceed 1,000 TWh by 2030. This is comparable to current annual electricity usage by China.
A typical single generative AI prompt can use 10 to 100 times the energy of a simple internet (e.g. Google) search. As more users worldwide make use of AI systems, it is easy to see how power consumption can only increase.
Many of the key players in this sector are keen to improve energy efficiency, not only to reduce Operational Expenditure, OpEx, but to meet environmental targets, whether self-imposed or put in place by governments and regulators.
Innovation in this area is driven by the companies designing and installing data centres. Key players in this regard include Intel, NVIDIA, Microsoft, Amazon Web Services (AWS), Samsung, Meta, Google, IBM.
Trends in AI patent filings
We can begin to get a feel for the magnitude of the recent surge in AI innovation by looking at global patent filings that are categorised under the International Patent Classification (IPC) system as relating to “Computing arrangements based on specific computational models”. This particular category (which goes by IPC subclass code G06N) includes patent filings directed to neural networks, machine learning, and other key AI technologies.
Since 2010, the rate of filings in this subclass has increased by over 20 times, plateauing at around 30,000 new patent families per year as of 2020. The biggest filers in this subclass include AI technology leaders IBM, NVIDIA, Microsoft, and Google.
Figure 1: new global patent family filings classified as computing arrangements based on specific computational models
While patent family filings within this category appear to have reached a relatively steady state over the last few years, looking at the categorisation of these patents in more detail reveals some interesting underlying trends.
The following graph shows the IPC “sub-group” classifications that have been assigned to the above patent applications. That is, it shows us the patent filing activity within the various niches that come under the umbrella of AI.
The data shows that, while filing activity within certain niches has levelled off in the same way as the overall AI-related data has (or even decreased), filings within other niches are actually continuing to increase year-on-year.
Figure 2: top 20 IPC sub-groups assigned to patent family filings classified as computing arrangements based on specific computational models
In particular, filings categorised in the following sub-groups appear to still be on the rise:
- Machine learning (G06N20/00)
- Using neural networks (G06V10/82)
- Combinations of networks (G06N3/045).
And, on the other hand, patent filing activity in the following sub-groups appears to have peaked or even started to decrease in recent years:
- Architecture (e.g. connection topology) of computing arrangements based on biological models (including neural networks) (G06N3/04)
- Inference or reasoning models (G06N5/04)
- Neural networks themselves (G06N3/02)
- Computing arrangements based on specific mathematical models (G06N7/00)
- Knowledge representation and symbolic representation (G06N5/02).
We can therefore see that the plateau in total filings within the G06N category masks trends across subcategories, with some areas continuing to attract increasing attention, while others see a decline. In particular, the trends suggest we may be seeing a flattening-out or decrease in patent filings directed toward “fundamental” technologies within AI such as neural networks themselves, different inference and reasoning models, and ways of representing data. In exchange, we’re seeing more filings directed to applications of machine learning, and the implementation of multi-agent networks.
This suggests that innovation in AI could be entering a more mature phase, characterised by specialisation and strategic focus.
Approaches to minimising AI datacentre emissions
As energy use becomes a central concern for companies seeking to deploy AI infrastructure, tech companies are exploring how to optimise computation at different levels, from chip to datacentre.
Not all of these innovations are apparent from reviewing patent literature – in part because chip-level patent filings often focus on low-level technical mechanisms without explicitly framing them in terms of energy efficiency or AI-specific workloads. In addition, leading companies like Google, NVIDIA, and Microsoft may choose to keep performance-related optimisations as trade secrets, rather than disclose them in patent filings, particularly when those innovations are difficult to reverse-engineer.
For example, Google’s patent application US20240403043A1 describes an AI accelerator architecture that reduces data movement by performing vector-matrix multiplications directly within memory modules. While this approach can substantially lower power consumption in deep learning inference tasks, the application broadly focuses on improvements to speed and computational output rather than energy efficiency.
Beyond chip-level innovations, many recent patent filings describe methods for reducing power consumption and/or emissions from datacentres by distributing computing workloads more effectively.
For example, IBM patent US2025037142A1 describes a “workload allocation engine” that assigns tasks to datacentre server clusters based on their predicted carbon emissions. The patent describes tracking various time-series variables for server clusters in order to forecast their carbon emissions. Based on this, the workload is then allocated to the cluster that is forecast to generate the lowest emissions.
Other recent IBM patent filings such as US2023418687A1 and US11803375B2 describe different “energy-aware” or “carbon-aware” approaches to distributing cloud computing tasks.
As workloads become more intensive, traditional air-cooling methods are increasingly seen as insufficient. In response, some datacentres are eschewing traditional air-cooling methods in favour of more direct approaches.
Datacentre cooling
The steep energy costs of AI datacentres don’t just arise from computation. Preventing hardware from overheating requires a significant amount of power: according to the International Energy Authority, cooling systems in “enterprise” datacentres account for over 30 percent of total energy consumption, and even the most efficient “hyperscale” datacentres devote around 7 percent of their total energy budget to cooling.
Developing more effective and efficient cooling systems for datacentres therefore represents a major opportunity for reducing the amount of energy consumed by AI datacentres.
The number of new patent families directed to datacentre-specific cooling technologies has increased in recent decades – from 87 new families in 2003 to 353 new families in 2023.
Figure 3: new global patent family filings related to server and datacentre cooling
This steady rise in patent filings reflects an increase in deployment of AI datacentres and a growing prioritisation of thermal management.
Leaders in filings in this area include the likes of IBM, Nvidia, and Google – however the space appears to be dominated by Chinese, Taiwanese and Japanese electronics manufacturers such as Foxconn (Hon Hai), Inventec, Hongfujin, and Fujitsu, as well as Chinese tech giant Baidu.
Figure 4: assignees for patents related to server and datacentre cooling
As workloads become more intensive, traditional air-cooling methods are increasingly seen as insufficient. In response, some datacentres are eschewing traditional air-cooling methods in favour of more direct approaches. One such approach being explored by the likes of NVIDIA and Microsoft is to use direct-to-chip cooling systems where liquid coolant is circulated in close contact with power-dense processing units.
NVIDIA patent number US20220256736 describes a liquid-cooled “cold plate” system which can be brought into direct thermal contact with semiconductor devices (such as the GPUs commonly used for AI computation). Internal channels in the cold plate precisely control the flow of coolant, enabling targeted cooling – particularly over hotter, harder-working components. Notably, the system features a replaceable intermediate layer, allowing the cooling configuration to be customised for different circuit board layouts and thermal profiles.
In patent US2025176095A1, Microsoft propose an approach in which microfluidic channels are routed directly into printed circuit boards (PCBs) themselves. These microfluidic PCBs integrate coolant channels in a manner similar to how electrical traces are laid out, allowing coolant to flow through the board via embedded pathways.
Pumps, which can be either external or embedded in the PCB, circulate coolant through these channels to provide localised, board-level cooling. This integrated approach reduces or eliminates the need for traditional thermal management components such as heat sinks or thermal pastes. Moreover, it offers application-specific adaptability: designers can route coolant around high-heat areas like power supplies or network switches, depending on the use case.
Other novel approaches to datacentre cooling include that of NVIDIA patent application US2024284640A1, which describes a combined cooling system for datacentre servers and fuel cells used to power the datacentre.
Implications for innovation and future patent filings
As AI adoption accelerates across industries, the demand for more energy-efficient AI technologies has become a critical priority, both to manage operational costs and to reduce environmental impact. We anticipate that the coming years will see a continued increase of innovation focused on minimising the energy consumption of AI hardware and datacentres, including advances in cooling technologies, workload optimisation, and low-power AI architectures.
Oliver Rigg Trainee Patent Attorney
Ean Davies Partner and Patent Attorney