Is OpenAI the Next Webvan?

December 17, 2024 / Steffen Bartschat

If you could invest in OpenAI right now (valuation: $157B), would you do it? After all, we are still supposed to be in the early stages of the AI revolution. When Facebook was the same age as OpenAI is now (about 9 years old), it had a similar valuation ($140B market capitalization). Today, Facebook (now known as Meta) is worth around $1.6T!

Webvan reached the height of its valuation at $10B in November 1999, only three years after the company was founded. I still remember receiving our first Webvan order around that time, from a cheerful delivery employee, with every piece of meat and produce carefully wrapped with a sticker of my name on it – all for the same price as buying from the local Safeway. Webvan was going to disrupt grocery shopping, and users loved it. There was just a slight problem here – groceries are a low-margin business (1-5% profit on average). With these slim margins, how was Webvan ever going to earn returns on its massive capital investment including regional distribution hubs, delivery vans, and technology portals? And it was still cheaper to shop at Costco – nothing was going to beat driving my minivan to Novato to fill it full of pallets of foodstuffs for the little ones. Webvan soon disappeared from view – a bloodbath for regular investors and institutionals like Benchmark Capital and Softbank.

Is OpenAI on the same path? It is making a decent amount of revenue ($3.7B expected this year) but it is far from cash-flow positive, investing huge amounts on training its latest LLM models. A lot has been written about the immense sums of money to buy the computing time needed for training – but soon that cost will be outweighed by the operational costs, mostly made up of the processing power required to provide the answers to all of our queries. And… it turns out that this is a surprisingly low-margin business.

According to some independent research, Google’s search business has an operating margin of only 34%. When we type a query into Google’s search box, it costs the company about 1.06 cents to service it, and it generates 1.61 cents in revenue (from advertising and “suggested” links). A ChatGPT query consumes about 10x more electric power than a simple Google search, so we can assume that genAI query costs are substantially higher. This explains why Google is not making its search results much smarter with the latest genAI: the added costs would destroy its unit economics.

The unit economics for OpenAI are better, since the company charges a $20 monthly subscription to access its latest version of ChatGPT, as compared to Google’s main search business, which is free of charge to its users. The key question is, in which direction is this monthly subscription price headed? There is good reason to think that genAI services will soon become a commodity, given the number of competing LLMs coming online, many free to use. I had an OpenAI subscription for a few months, but have since switched to Perplexity Pro – it’s included in my Zoom subscription at no extra charge.

So there will be revenue pressures and the cost pressures won't ease up – the latest gen AI techniques are adding “inference time compute” steps during the query to improve accuracy – adding even more processing costs per query.

The bottom-line? While I don’t expect OpenAI to implode like Webvan did, I still wouldn’t invest in OpenAI right now. But then, I also thought Facebook was grossly overvalued back in 2013…