The Role of AI in Climate Tech

The world is currently abuzz with excitement about AI following the launch of ChatGPT and other generative AI models. At G2, we’ve been thinking a lot about the effect of advancing AI capabilities on climate tech.

First, we should note that AI has already made a significant impact on climate tech, and we have been investing in AI capabilities within traditional industries for quite some time. Many of our portfolio companies leverage AI in delivering new capabilities and insights to their customers. However, there has certainly been an “inflection point” in AI capabilities over the last few years, and we see this driving tremendous value in the years to come. So, let’s dive in!

The AI Value Chain

Before delving into the specific impacts of AI in climate and industrial tech, we wanted to share our framework for thinking about “AI”. Put simply, AI is an efficient way to estimate a function. To do this well, you need three key elements. First, you need data — both input and expected output. Second, you need algorithms — a design for your neutral network and your training process. Lastly, you need hardware to run the computations. With these three things — data, algorithms, and hardware — you can produce an application of AI.

This straightforward framework helps us think clearly about how AI has and will continue to improve. Simply put — better (and more) data, better algorithms, and better hardware. Improving these three inputs enables us to make better approximations more efficiently. Occasionally, these improvements will cross a threshold, resulting in a ‘tipping point’ in which a new problem area becomes addressable by AI. We witnessed this recently with text generation, and in the past with games like chess and Go.

With this baseline, let’s dive into our world at G2 Venture Partners — industrial and climate tech.

The Intersection of AI and ClimateTech

We have already seen the impact of AI in climate science in a variety of areas. Here are some examples:

1. New Product Development: Scientists at Stanford utilized AI/ML to simulate how different charging approaches affect new battery chemistries. These techniques showed the potential to cut down battery testing time from years to weeks.
2. Transparency: Climate Trace, spearheaded by Al Gore, aggregates and analyzes 60TB of data from thousands of sensors to identify and attribute emissions from nearly 80,000 specific sources globally.
3. Control: DeepMind demonstrated an impressive 17% energy savings when utilizing reinforcement learning agents to control HVAC systems in data centers, as opposed to a ‘rules-based’ policy.

Additionally, we have quite a few portfolio companies applying AI/ML to tackle industry challenges (although we will not exhaustively list them here, but check them out!). Looking ahead, we anticipate further disruptive innovations and are excited about the emerging companies that will shape the future in the years to come.

5 Archetypes

We foresee the emergence of five different “archetypes” of companies that will leverage AI to transform traditional industries.

1. Clarity from Chaos: The scientific community has already demonstrated the power of AI/ML in delivering superior models for complex and chaotic systems, particularly in climate modeling. This capability will extend to other complex systems (global logistics? agronomics?) enabling better planning and decision-making.
2. Intelligent Workflows: AI will unlock new levels of efficiency in executing a digital task. We envision ‘intelligent agents’ supplementing existing software to streamline workflows and provide additional capabilities, such as offering ‘expert-level’ support in solving a specialized problem.
3. Generative Industrial Design: There are many situations in the industrial economy in which an expert designs an asset or process. Generally, this expert will generate one design over a relatively long time period. Instead, an AI could generate 1000s of designs, quickly, and select the optimal one based on the most important parameters and expert input. This approach revolutionizes industrial design processes.
4. Learning Control Loops: We see significant opportunity in what we are calling “learning control loops”. Currently, most industrial control systems rely on schedules and setpoints, plus human oversight. But, what if the data from these controls could be brought into a digital environment, where reinforcement learning agents constantly test new operating paradigms, identify the optimal one, and push it back to the system?
5. Robotics: As algorithms and hardware continue to improve, robots are becoming increasingly capable of addressing new operating environments, driving efficiency and quality in the industrial sectors.

How will these products get built?

If this is the future we see, how will companies get there?

One challenge to AI’s ability to solve a problem, everywhere but particularly in the industrial and climate sectors, is the availability of data. Returning to our framework, even as algorithms and ‘foundation models’ become more capable out of the box, AI still requires examples of the task to be executed — and, potentially quite a few examples, depending on the complexity of the task.

Climate issues are real-world problems, and obtaining massive datasets for these problems is not as straightforward as accessing language or imagery on the internet. In reality, a substantial portion of the world’s data is privately held and not publicly available on the internet.

Consider the below:

The total amount of data in the world is a staggering 200,000,000 times greater than what is available on the public internet. This vast pool of data is protected by passwords, paywalls, and firewalls. Collecting this data and using it to build products is a massive opportunity. In fact, we believe this number undersells the opportunity, as we know that most industrial data goes uncollected today, suggesting the untapped dataset is orders of magnitude larger still.

So, while we at G2 are selectively investing in algorithms and hardware, prioritizing efficiency and sustainability (e.g. our investment in Crusoe), we are actively hunting downstream of this enabling tech, for companies with winning AI applications and products. These products provide access to proprietary data flows, either existing enterprise data or newly collected data, which will then be used to improve the performance of the product. This virtuous cycle (product -> data -> better product -> data…) will drive significant value creation for both companies and their customers.

We are interested in speaking with any company that is building this sort of product to solve problems in traditional industries!

Sources:

¹https://arxiv.org/pdf/2005.14165.pdf

²https://www.lesswrong.com/posts/6Fpvch8RR29qLEWNH/chinchilla-s-wild-implications

³https://commoncrawl.org/2023/04/mar-apr-2023-crawl-archive-now-available/