Investing in AI Startups: Panacea or Problem?

There has been a Cambrian explosion of AI startups. Since 2012, investment in AI/ML startups has risen exponentially. The first cohort of AI startups and academic projects were horizontal. Horizontal AI views AI as an elixir to solve problems across industries. Horizontal AI startups build general tools related to perception and computer vision. This is the MLaaS model (Machine Learning as a Service).

Established incumbents enjoy significant structural advantages in offering MLaaS. They have a monopoly on AI talent and other resources like data and capital. Additionally, the open source ethos in AI limits the proprietary nature of algorithms. Incrementally improving natural language processing and computer vision does not build enduring companies. This is a common problem with many “AI-first” startups. They don’t become enduring companies and lose focus on real customer needs. Low-level AI tasks will likely be commoditized.

2017 sparked the next generation of “AI-first” startups. These startups applied AI to specific verticals. Vertical AI startups solve a problem and build data defensibility and avoid disintermediation. Owning the end customer enables real data network effects. Building full stack solutions strengthen defensibility around the customer need.

As Bradford Cross notes, the goal is to build a “data flywheel” into your products. Applying AI to specific industries operate backward from a market instead of forward with technology. Constraining to a particular customer problem constrains the data to train those algorithms. Ownership of the data value chain creates a moat against new entrants and large incumbents.

There is some truth to “data is the new oil,” but only if data is proprietary in structure and content. As soon as that data becomes more openly available or new advancement within Deep Neural Networks are made (like GANs), there is no resource advantage. Corresponding technologies like the IoT ecosystem and Blockchains will likely make data access more accessible. Companies need to train models recursively with closed feedback loops. The only access point to this is delivering real, tangible value from the beginning and pull in data.

The most successful startups with the most valuable data network effects are not AI companies. Here’s Uber’s Data Network Effects, which will be prized in the race for autonomous cars:

Courtesy of Whitney Zimmerman.

Markets like financial services and healthcare have been active as it pertains to investments. There are many problems to solve and since the past few years, very little software penetration. These sectors feature favorable high-margin structures and desirable TAM sizes. But most of these opportunities will increasingly saturate. I’m excited about how AI/ML will enable true platform shifts in autonomous products, mixed reality, and new networked markets. This translates to untapped markets and new distribution channels.

Good ideas arise where everyone is looking. Great ideas happen where no one is.

Most of the activity around AI right now is noise on what AI can do and should do. There are many highfalutin ideas on what will happen like Greylock’s Systems of Intelligence (SOI) thesis. I think it’s wrong too. 1  SOI doesn’t have an accumulating advantage but an accumulating disadvantage. In a world of data portability and techniques with data reduction frameworks, the defensibility to an SOI strategy is questionable. Companies won’t have barriers to entry. Most enterprise and consumer companies are trending towards a world of technology inversion. Cloud/SaaS/AI technologies once created barriers to entry through network effects, but as they advance and democratize, they also reduce barriers to entry. Competitive advantages are becoming more transient, and deep tech is not the new moats in the Information Age.  I’m skeptical of predictions and the conventional wisdom around the potential of AI startups.

In 2017, out of the 120 AI companies that exited the market, 115 were acquihires. The M&A activity in AI is led by usual suspects like Google and Facebook.

I’m increasingly skeptical of founders who identify themselves as an AI startup. I think it’s an overused term. In a world where software is eating the world, every startup is a data company. Every $10B+ startup will need a fantastic data acquisition strategy. AI is an enabling technology to build great companies and products off of. But AI will not solve your Product-Market fit problem.

Leave a Reply