Thoughts On The AI Reset Wall

Typically I’d see an article like Groundbreaker‘s The Teaser Period and I would link to it in a note and say a blurb about it. Something I find insightful that you might find insightful as well. This time is different though. This time I feel like writing more than just a blurb.

For those that have the time and energy, I encourage you to read the whole thing. It’s fascinating and worth it. Go read it and come back. I’ll wait.

The article does a great job of comparing the 2007/2008 home loan teaser rate resets with the upcoming AI compute commitments coming online. While the comparison is fun for a nerd like me, the main point here is that over the next 18 months data centers built for this AI boom will start coming online. Data centers that OpenAI and Anthropic are committed to funding, whether they still want to or not.

Now, the numbers are a little hazy. In one chart, Groundbreaker shows OpenAI’s 2027 compute bill as being $132 billion. Since OpenAI isn’t public yet, this is just a guess and probably a minimum since we do know what they have announced. Either way, the point of the article is that OpenAI will go from $37B in compute spending this year to $132B in compute spending next year. Yet, OpenAI’s revenue as of today is nowhere near $132B. The latest report from Bloomberg was an annualized revenue run rate of $40B.

That obvious gap between $40B ARR and $132B in commitments next year sets up a very interesting 2027. The $90B gap will have to be closed somehow. Maybe OpenAI can raise the needed cash through an IPO. Maybe the revenue does show up. Maybe construction slips on enough of these projects that the gap ends up being $50B instead of $90B. Maybe Amazon/Google/Microsoft negotiate equity for waiving some of the compute costs.

Lots of maybes out there. But some of them are brutal. Let’s walk through the one worth thinking about. Anthropic has the same compute wall coming. It too will have a jump in compute costs next year. We could see a situation where OpenAI can’t actually pay for this compute and Anthropic doesn’t want it, and no one else is willing to step into the void at a price that makes cloud providers whole. The hyperscalers then start slowing down or stopping construction on 2028/2029 data centers, which means the memory chip makers and Nvidia start getting orders canceled. TSMC finds itself with less to do. And all the companies I’ve just named make up nearly HALF of the S&P 500.

It doesn’t have to get ugly. We might find buyers for every single cycle of the compute and the ROIC might be amazing and the party continues! But what we do know is that starting next year OpenAI in particular will be an interesting situation to watch. The bills are coming due for commitments made two years ago. And if they don’t find a way to pay these bills in 2027, 2028, 2029, etc, things will get spicy.