The promise of AI has triggered a massive ripple effect. In pursuit of what feels like game changing technology, the world writ large has started tripping over itself to build out all the necessary components. What that revealed was just how interconnected all the pieces are.
On the one hand, you have talent. OpenAI’s original input was AI researchers. Now, you’re seeing the formation of “neolabs” at a rapid rate. Former OpenAI researchers who became former Anthropic researchers who became former Character AI spin outs and are now building their own lab. There are dozens of iterations of this. Many of these labs were started within the last 12 months. One of the most famous, Thinking Machines, founded by former OpenAI co-founder, Mira Murati, is rumored to be getting valued in at $50 billion.
We saw a similar supply and demand strain on AI talent this past summer as Mark Zuckerberg was hoovering up AI talent for Meta, either with buyouts or massive bonuses. The atomic unit of AI breakthroughs felt like they were coming from AI researchers, so having the best of the best was worth any price (potentially up to, and including, $100 million payouts.)
But while the talent end of the spectrum races towards what everyone hopes will be continued breakthroughs, whether in the form of higher performance models, higher quality agents, or more efficient architectures, there is the other end of the spectrum racing to enable those breakthroughs: compute.
The AI boom triggered a commensurate infrastructure buildout that shadows everything that has gone before it. One report indicated that, without data center investments, US GDP growth in the first half of 2025 would have only been ~0.1%.
What we have is a massive cyclone of dependencies. More AI talent leads to more AI breakthroughs. More breakthroughs leads to more AI deployments, more inference, more data centers, more compute, not to mention the requirement for more water, and more energy.
And that cyclone is, increasingly, being fueled by debt. One report indicated that OpenAI alone may spend $75 billion in 2026 capital expenditures. Other companies, like Meta, are weaving complex debt deals together with private equity firms that, apparently, are making Zuckerberg feel inclined to promise his team that the bet probably won’t bankrupt the company. Granted, an unnerving promise that a company with $30 billion in quarterly cash flow needs to make.
Unfortunately, that same cyclone of dependencies is starting to run smack into the physical realities of the world around us. In terms of energy to turn on that compute, the CEO of GE Vernova, one of the largest suppliers of hardware for turning on new energy, has said that “nearly all of the company’s output is booked through 2028.” Without the energy to turn on those data centers, you can’t get more compute. Without more compute, you can’t get more inference, more AI deployments, more breakthroughs.
As a result, we will increasingly have to wait and see. We have a finite amount of AI talent. We have a finite amount of physical infrastructure. What will give out first? Our ability to find AI applications worth deploying? Or our access to physical infrastructure to actually be able to turn that AI on? In a cyclone of dependencies, something will ultimately be the bottleneck. We’ll just have to wait and see what it ends up being.
