Last week, NVIDIA agreed to buy Hugging Face, the platform where developers share, download, and modify open-source AI models. The price paid by NVIDIA was very steep; the fact that it was willing to pay 86 times revenue points to just how important these open-source models are becoming.

Why did NVIDIA do this? One of the things holding back AI usage (and therefore demand for NVIDIA chips) is the cost of running a token through the leading models from OpenAI and Anthropic. The theory goes that if models become commoditised and effectively free, usage explodes. And someone still has to run all that cheap AI, on NVIDIA's chips.

But that’s not the end of the story. A few weeks ago, Stripe paid a similarly large sum for OpenRouter, a startup that helps businesses work out which model should handle which task. Both NVIDIA and Stripe are betting big money on the same vision of the future: real value will accrue to whoever makes it easy to switch between models depending on the job. For example, an expensive frontier model for creative or genuinely hard work, a cheap open-source one for the monotonous stuff, a specialist model for coding or legal…

Taken together, these acquisitions tell us about where the puck is moving next in the market, and where the tollbooths of the next stage of AI infrastructure are likely to sit. So here's a question for your next investment in the AI space: does this company get paid more when the number of models its customers use goes from one to dozens? If yes, it probably owns a tollbooth. If no, it's probably paying one.

Here is a company that fits the tollbooth profile described above. Here is another.

There are rumours that S&P Global is considering spinning off Capital IQ, a market intelligence and financial research platform used by finance professionals to analyse both public and private companies. What’s remarkable about this rumour is not that S&P Global would want to divest such a large portion of its business… After all, Capital IQ has been under some pressure lately, as it has grown slower than other portions of S&P Global’s business, and investors worry about the impact that AI can have on financial data & workflow platforms. What’s remarkable is that S&P Global acquired Capital IQ (through IHS Markit) in 2022 for a very large sum of money.

This case study is a perfect example of why large acquisitions usually fail. Major acquisitions are usually pitched with claims of massive synergies and cost savings. Management paints a rosy picture of a stronger company and rewarded shareholders. The reality is that larger does not mean better**.**

Acquiring IHS Markit gave S&P Global a large market intelligence offering, but the competitive edge of this segment was never going to be as strong as some of S&P’s other segments like the ratings or the indices. Capital IQ would have to compete with well-funded and well-established brands like Bloomberg, London Stock Exchange Group, and FactSet.

And then there’s all the money spent on these deals. S&P Global payed a big premium to acquire IHS Markit, paid again to integrate it, paid again to spin off its Mobility Unit (part of IHS Markit), and may now pay again to spin off Capital IQ. The winners in all of this are the investment bankers, not the shareholders.

Beware of mega-deals such as these: they rarely create value for the acquirers’ shareholders, and usually end up diworsifying the business.

Upgrade to unlock full access to the RatedA Portfolio (with weekly updates on every buy/sell and rationale), and the complete database of companies + ratings.