My wife would probably call me a hoarder of artifacts, both physical and digital. We’re talking book reports from the fourth grade, 3.5-inch floppy disks filled with firmware I wrote in grad school, the spalled-out rear main bearing of a transmission I rebuilt between writing that same firmware, prototypes of products I developed 25 years ago, and half a dozen subtly tweaked Stanford manhole covers (yes, actual manhole covers…long story).

When I stumble on these old things, I feel a mix of nostalgia and the urge to pull out my red pen. The usual questions surface: What did I get wrong? Does it still hold up? How would I make it better today?

I had the opposite reaction when I rediscovered this slide from our 2016 annual meeting, titled “Deep Thoughts on Deep Learning.”

My speaker notes start with a prediction: “There won’t be a single AI company.” Instead, I argued, AI would be woven into the fabric of everything – how we interact with computers, diagnose diseases, secure networks, decide what to buy, and much more. My sketch depicts AI as a foundational layer cutting across every part of our economy, rather than a specialized tool for a narrow set of use cases.

Looking back, I’d pull out my red pen for one thing: I didn’t anticipate OpenAI emerging as the flag bearer for this technology. But my larger point holds – AI is becoming the omnipresent substrate of our world.

AI’s explosive trajectory explains why we seek out categories that have yet to fully form. At Foundation, we invest in “Zero Billion Dollar” markets – opportunities that don’t exist until a founder steps in and creates them. These markets initially seem speculative because, quite literally, they start from nothing. They’re spaces where significant latent demand exists, but there’s no clear way to serve it. If you read the McKinsey report, the size of the addressable market would be zero… 🙂

This thesis guided us to early investments in companies like Cerebras. Between 2012 and 2016, AI workloads grew by 300,000x, far outpacing Moore’s Law. To me, the opportunity for Cerebras wasn’t subtle or hidden – the workloads were growing like angry weeds. But the market had yet to recognize the need for a new computing paradigm purpose-built for AI’s insatiable demands for data and parallel processing.

Cerebras understood this challenge clearly but saw its solution differently from others. They stepped off the well-trodden path of GPUs and pursued wafer scale with total conviction. In doing so, they forged a new – and dramatically more performant – computing paradigm for both training and serving AI models. Shipping incremental improvements was never part of Andrew Feldman’s vision.

The same pattern played out with Seel, though in a very different corner of the economy. As online shopping became the norm, founder Zack Peng started to question why e-commerce had perfected the sale but neglected what came after. He realized that with enough data and the right AI models, returns could be predicted, priced, and managed like risk. What looked like a logistics nightmare was actually an insurance problem in disguise. 

Rather than compete in an existing market for return management, Seel built a new, universal post-purchase layer – one that turns refunds from a drag on margins into a predictable, and even now profitable, part of e-commerce infrastructure. Like Andrew, Zack internalized a growing problem and refused to accept the conventional solution. In just two years, Seel has grown 40x.

Or take Loft Orbital. Before Pierre-Damien Vaujour and Alex Greenberg started the company, launching a custom satellite into space required years of planning, specialized hardware, and upfront costs that priced out everyone except governments and big corporations. Pierre and Alex looked at the business model of cloud computing and asked: Why can’t space work the same way?

Loft lets its customers rent capacity on shared satellites – like booking compute time on AWS, but 300 miles up. Their satellites also serve as edge-computing nodes for space-based AI. Through Loft’s “virtual missions,” companies can run AI applications in orbit, right alongside the cameras, sensors, and radios that are tracking activities back on Earth.

More than simply lowering the cost of accessing space, Loft created a market for space-based software that barely existed before they built it. Soon, we’ll be able to ask critical questions about our planet – like “Where are all the active wildfires in Southern California right now?” or “What was the last commercial ship to pass through the Strait of Hormuz?” – and get instant answers from AI models that are analyzing data directly in space.

Reflecting on this 2016 slide, I see it as an expression of a belief I’ve long held: extraordinary founders don’t search for markets – they manifest them. Our strategy is to go early, go deep, and find these founders who are building the categories they’ll ultimately dominate.

And yes, it makes me think more deeply about what I’m sketching now that will seem inevitable in 2034…