Over the last 18 months, we’ve experienced an Artificial Intelligence boom, triggered by advancements in Generative AI.

During this period, we’ve seen total AI compute capacity grow 50-60% each quarter.

We’ve also seen demand for data centers skyrocket - in the US alone demand is expected to increase 10% per year until 2030, driven primarily by AI as well as bitcoin, work from home, and autonomous technologies.

Because of this growth, electricity demanded from all data centers will nearly double over the next 2 years.

By 2030, data centers are projected to take up a staggering 10% of the global power supply and consume more electricity than the entire nation of Brazil.

Many believe the answer is to speed up renewable energy development. And while this is of utmost importance for the energy transition broadly, energy is a bottleneck to AI growth.

‍Energy as a Bottleneck

Today, renewable energy demand is outpacing supply, which is further exacerbated by lengthy interconnection queues. Some utilities are using load growth as an excuse to delay retiring coal plants. Others are adding natural gas generation capacity for more than just peak demand, falling back on core competencies during this period of unclear growth. This is not the case universally - some utilities, such as Xcel Energy, are retiring coal plants and using those existing interconnection points for renewable energy deployment.

Further, AI-specific data centers require significantly more power and chips than traditional data centers. Because of this, data center development is limited to where sufficient power and interconnection is available. Land availability is becoming the next bottleneck, if policy doesn't restrict development first.

Big Tech will keep plugging in chips as long as they have the power—if forced to choose, they will choose dirty energy over their climate goals to prioritize the AI arms race.

While co-locating renewables with data centers is an option, this is best suited for flexible workloads that do not require low latency given renewable intermittency. For example, autonomous applications are not flexible in nature. MIT’s Lincoln Laboratory has proven methods to increase energy efficiency when training models and running inference, saving money and emissions. However, this requires behavior change to implement in practice - incentives are required to make more workloads flexible.

Additionally, nuclear is receiving a lot of attention as a stable clean energy source for data centers. Nuclear is not a viable option in the near term due to costs, safety, and lack of US engineering expertise. Nuclear should be viewed as a longer term solution, which has a different time horizon than satiating data center power demand today.

Given the unrelenting demand for data centers, rising cloud costs, and energy bottlenecks, we need to reduce the resources required to run AI. We need to focus on data center efficiency and energy efficient compute and storage as solutions.

The answer is not to consume unlimited amounts of renewable energy, especially considering current constraints.

Energy Efficiency Solutions are Key

‍Buoyant pulled together a market map of non-exhaustive digital solutions for decarbonizing data, with a focus on efficiency:

‍‍

‍Clean Energy Procurement‍

Startups helping companies procure clean energy through PPAs and RECs, some focused on the infrastructure and tracking of 24/7 renewable energy matching. For example, CFEX is focused on streamlining processes across the REC supply chain.‍

Renewable Energy Development‍

Software that accelerates the deployment of renewable energy, from financing to project management. For example, CapeZero is focused on optimizing the financing for renewable energy deployment, while Blumen is focused on streamlining project planning and management.‍

Grid Enhancing Technologies‍

Software that unlocks available grid capacity, such as ThinkLabs which is creating a co-pilot for grid operations and Enline which focuses on dynamic line rating.

‍Data Center Energy Efficiency Software‍

Startups using AI to create autonomously controlled data center operations, such as Phaidra.‍

GreenOps & DevOps‍

Startups in this category help customers track and reduce cloud emissions from a variety of activities, such as software deployment, writing code, and platform engineering. For example, Chicory AI is supporting data engineers using co-pilots to rewrite code in the most efficient way, while Cycloid is optimizing platform engineering for software delivery.‍

Green Computing, Green Data Processing and Storage‍

This category of startups are focused on reducing the cost of training models, running inference, storing and processing data, and data analytics. These companies may be cost driven, but have significant energy efficiency benefits. For example, Ocient, a Buoyant portfolio company, is a database software company that helps customers analyze hyperscale datasets at a fraction of the cost and significant energy savings.

Additionally, Radium is focused on developing integrated cloud computing services for Gen AI training, inference and deployment, which is purpose built to run efficiently. We need software written, delivered, trained, run, and stored to be as efficient as possible.‍

Path Forward

AI holds a lot of potential for societal benefit: from new material discovery, to early disease identification, and everything in between. But we must stay aligned with our climate goals - we cannot grow for the sake of growth. AI/ML practitioners need to measure and disclose impacts of their work for the industry to learn together.

We are entering an extreme period of energy load growth thanks to AI and broad-scale digital transformation.

We need to fundamentally change the way we build and manage computing resources if we want energy supply to keep pace with this growth, which is creating opportunities for innovation.

‍
To hear more on this topic, please find a highlight reel from our research and panel discussion here.