Many of the most critical data sets are the ones that we have long been unable to access or interpret.
Value on the internet has largely compounded through data loops that continuously strengthen with scale. A product or platform collects data, that data makes that product better, and the better product earns the right to, in turn, collect more data. That self improving loop sits beneath most durable software businesses and has been a significant pillar of the USV network effects thesis since Andy wrote about how it sits beneath application layer network effects in 2015.
Today, this is truer than ever. In the AI era, data is the ultimate currency. Labs are spending accordingly and companies like Mercor are racing to billions of dollars of revenue.
The constraint to the data network effect has always been scope and reach–what data is in range and what is out. Data that software could capture was most accessible because the vast amount of data outside of software (the environment around us, the physical world, the human body) was too expensive and difficult to capture, too hard to process, and thereby inaccessible.
Now, a convergence of powerful forces is turning this on its head. Intelligence is abundant and the cost is declining. Models can quickly process even the messiest, unstructured inputs that were too difficult for software. The cost and timeline to build hardware has rapidly gone down. And we’re experiencing a proliferation of observability–through increasingly inexpensive and ubiquitous sensors, satellites, cameras, etc–making capturing the data from the world around us more doable than ever before. In aggregate, the ability to collect, process immediately and intelligently, and build on top of these inputs in unprecedented ways allows data loops to form in places that were totally dark a few years ago. This isn’t AI bringing efficiency to existing markets; it’s a completely new set of opportunities all together.
There are lots of examples of where this is coming into play. Ambient conversation is one. We have been able to record speech for a century but now our ability to transcribe, structure, and act on it has turned it into a utility dataset. That births vertical opportunities like Abridge to take that dataset and build applications that transform how particular markets operate or Granola to create horizontal infrastructure and tooling. The recording was never the hard part but the processing and productization was previously not possible.
The human body is another. The cost of testing is collapsing, the ability to interpret results is improving, and it is increasingly possible to personalize a program on top of the data. The body becomes both reachable and useful as a data source.
But perhaps the largest opportunity here is the physical world.
The physical world holds enormous amounts of data that has long been either out of reach to collect or too messy to process but is essential to automation, optimization, and understanding. Now, sensors are proliferating, robots are becoming more capable and cheaper, and processing messy data quickly is achievable. Models to train robots in increasingly challenging tasks are improving at rapid speeds and sucking in more data to do so than ever before. We’re seeing what’s possible in the physical world transition from experiment to commercialization. That data loop is especially strong here. More deployments produce more real world data, better data makes the models better, and better models make the next deployment faster and cheaper than the last.
Building into these opportunities in the physical world is both very early and very hard, with the data flywheel only just emerging. Within the software flywheel, we are just starting to see massive progress in going from learning (using data to train a model) to reinforcement learning (defining reward functions so the system can learn which actions lead to better outcomes through interaction) to continuous learning (allowing models to keep improving as new data arrives.) In the physical world, we are just starting to scratch the surface of reinforcement learning that can happen when robots interact with the physical world.
But the opportunities that await from this physical world data flywheel are massive, market-shifting, and previously unreachable. The most interesting use cases aren’t making hard things easier but allowing insight and action that we’ve never been able to achieve.
Sensors on every utility pole, for example, will allow for infrastructure observability that was previously impossible because batteries had to be swapped every 6 months making it cost prohibitive to embark on. Now, with batteries that can last 10 years, deep knowledge of our infrastructure will not just be affordable, it’ll be achievable. Models that can both take sensor input from hugely varied sources in the world around us and also synthesize them together and make sense of the noise will give us understanding of our weather patterns in a granularity and accuracy we’ve never had, the most critical step to eventually being able to change them. Autonomous transportation (made possible through sensors) is quickly en route to create unprecedented ability to move people and goods around in a completely new cost structure. We can now understand our oceans and thereby acquire new knowledge of how to defend our land, navigate our ships, and maintain our planet.
There is a massive opportunity to reinvent every layer of the physical world stack. We’ve been investing significantly across it at each level and will continue to do so (with several unannounced investments we’re excited to share more on soon.) Generalist is building foundation models that give robots general dexterity, the ability to perform the tasks we actually want done. Tutor Intelligence runs the full loop of robot deployment through data collection and model improvement, enabling robots to quickly become productive in days rather than a six month integration (and then feeds that data back into its own model to continually improve.) Sofar Ocean leverages the massive sensor proliferation to provide infrastructure for both owned and third party software networks on top. Viam sits in between, as the operating layer for data, AI, and automation across fleets of devices. Efficient Computer is at the base, building silicon efficient enough to make new use cases economical out at the edge.
The second order effects are also significant. Once you can capture and act on physical world data at scale, you can run far more efficient factories through automation and agentic operating systems, for example, which is what Isembard is doing. And, of course, there is the dramatic need for increased energy and power underneath this layer enabling this level of compute and all the inputs required there (more efficient data centers, abundant and safer batteries, new formats of scaled and clean generation, etc.)
The below market map is illustrative of the stack exploring and acting on our physical world.
We are very early. Most of these datasets are barely tapped, and most of the products that will run on them do not exist yet. Finding them, reaching them, and putting them to work will change how we interact with the physical world. We want to explore the entirety of that edge with the founders running toward it.
A special thank you to Josh Gruenstien, Brandon Lucia, Alex Iskold, Kanyi Maqubela, Tina Haibodi, Nikhil Trivedi, and Chad Byers for sharpening our thinking on this post.