Today we all get a free and decent weather forecast thanks to government institutions such as NOAA. But climate change has increased weather volatility: airplane turbulence is getting worse and the pace at which storms morph into monster hurricanes is unprecedented. As a result, if we want to keep businesses and governments running on time, we need better weather prediction capabilities. Fortunately, we are seeing a set of emerging technologies that could lead to a 10x improvement in forecasting. 

Given the convergence of increased need and novel tech, we wanted to ask ourselves whether there is an opportunity to build a category-defining company in this space. At Innovation Endeavors, we’ve spent the last few months diving deep into this space, meeting with experts and founders. We are sharing our learnings to get feedback from the community. If you are deep in this space, reach out!

Status quo

Today, government agencies forecast weather using primarily Numerical Weather Prediction (NWP), an attempt to mathematically solve complex fluid dynamics equations at scale. These models are run by government agencies in the US and Europe, which use input data from both governments and publicly available sensors (namely from satellites, radars, weather stations, etc.).

There are several important limitations to the NWP approach. First, it is extremely computationally intensive, requiring ~6 hours of processing time to update a forecast. But in a world where tropical storms can become hurricanes in a matter of hours, as Hurricane Helene did, each hour matters. Second, NWP models can only use a fraction of the observational data that exists today, limiting their accuracy in a world where sensors are increasingly commoditized. Third, NWP has now been shown to underperform relative to foundational AI models when it comes to predicting extreme weather events which, arguably, are the most important ones to forecast accurately. Last, NWP tends to forecast more accurately in areas with more sensor readings; unfortunately that means that developing countries with fewer sensor readings are more likely to have, in addition to a higher level of vulnerability to natural disasters, a disadvantage in terms of knowing when they will hit.

Breakthroughs in weather prediction: Sensing and foundational models

Better weather forecasting should be hugely value accretive. That is why we are excited about some of the technical advances we are observing, including:

Novel sensor platforms: We are seeing both new and improved sensors alongside a crop of innovative companies building constellations of low-cost sensors across unique domains: Tomorrow.io in space, Windborne in the atmosphere, and Sofar in the ocean (among others). These sensors enable us to collect novel forms of data that can make existing models more accurate – albeit we do not yet know whether any of these approaches individually will be sufficient for a 10x improvement. Hence, one of the main challenges with this approach is that it can take significant capital and time to determine how much better these forecasts are than existing alternatives.

‍Foundation models for weather: Spurred by advances in LLMs, we’ve seen a number of projects building foundational weather models that are already better than NWP ones. Google’s GraphCast initiative is at the cutting edge and recently announced another major breakthrough. We are intrigued by teams attempting to build models that create forecasts purely by learning from observational data, which would radically cut down the time it takes to get the results into customers’ hands. Furthermore, we see a world where customers can augment AI weather models with their proprietary data. We’ve seen teams taking different approaches here, including Jua and Brightband. Ultimately, we are excited by AI-first teams that are laser-focused on commercialization while also being able to run a lean organization until product-market-fit is clear.

Barriers and opportunities for weather companies

Despite our belief that better technology could save lives and generate enormous value, we have also seen a common set of challenges for companies in this space. Those include:

• Competing against a public good: Weather as a ‘product’ today is ~free. As a result, we often hear customers question why they would pay for what is otherwise a public good. Thus, founders need to find customers for whom pain point translates into their bottom line AND to whom they can provide a 10x better product. For example, auto insurance companies need days or weeks to be able to turn a hail warning into actual savings because moving vehicles takes time. We tend to hear customers say that the 3-52 week time frame is where they would see the most value but don’t see great solutions today. Furthermore, as future models are likely to be open-sourced, including GraphCast, teams need to understand where long-term value will accrue.

• High fixed costs to build a differentiated product: Weather is a complex global phenomena that requires gathering and processing massive amounts of data – a reality for both companies building constellations and/or training foundation models – all of which costs lots of money. We’re curious to see if there are ways to prove PMF and market size before spending significant time and capital.
  
  
• Customers generally don’t value weather, they value its derivatives: Weather is almost always an input into an end product. For energy traders, weather is an input into their power models. For logistics companies, shipping time and cost is what they value; they need waypoint based navigation, not merely a weather forecast. Companies that only build an input (weather) limit the amount of value they can capture relative to selling a final good (power, coordinates for supply chain assets, etc.). 

• Customers don’t have in-house expertise: Most customers don’t have in-house meteorology teams, making it hard for them to use weather data outputs directly. Only very large organizations where weather plays a critical role can do this: Amazon, SpaceX, the military, etc. As a result, we often find that teams need to build deep into the workflows of customers, which is more expensive and time consuming than founders expect. This can also pigeonhole a founder into a small TAM and/or low ACVs (e.g. an ‘ag weather company for Europe’) versus building a platform that can both be deep and, eventually, broad.

Path ahead

Given these challenges, we find ourselves filled with a healthy dose of skepticism that a category-defining company will be built in weather forecasting alone. That said, we are eager to be proven wrong and keen to meet teams that can break through the aforementioned barriers.

More importantly, we see the speed with which AI has radically improved weather forecasting as a harbinger of AI’s potential to improve a plethora of processes in the physical economy, from hardware engineering and material science discovery to backfilling labor shortages and accelerating the deployment of renewable energy. If you are building AI tools for the physical economy, let's chat!