Market Bubbles, Plateaus, and Gray Swans

Today, let's talk about market bubbles, plateaus, and gray swans.

For the last six months or so, I've been looking at AI in the context of the market, and it has felt pretty bubblicious—like we were hitting a bubble with a ton of companies jumping in. Some of them couldn’t even spell "contact center" a year ago. And some can’t spell "AI" most days. At this point, there are probably hundreds of companies doing pretty much the same thing.

Most of them are either focused on voice automation and chatbots—much easier to build today than even a year ago—or they're working on some flavor of call summarization, quality assurance automation, sentiment analysis, or CSAT. They take a call transcript, evaluate it against criteria through a prompt, generate output data, and put a front end on it—suddenly, it's a product.

We had already seen some of the air coming out of that, especially with the recent struggles of high-profile startups and the noise around deciphering real differentiation.

Then last week, DeepSeek released its R1 model, catching the market by surprise. It wasn’t a black swan event like GPT’s original release, which was a true paradigm shift, but it reset the market in many ways by unlocking new capabilities that didn’t exist a week ago.

Now, it's possible to run one of these high-performance models on commodity hardware, and cloud hosting providers are offering the full model at a fraction of previous costs. This shift opens up new possibilities for tackling complexity.

I no longer think the market was in a bubble. Instead, it had plateaued—everyone reached the same level, running in circles trying to figure out the next differentiation.

Some companies will recognize the opportunities that these new models and paradigms unlock, using them to create new products and categories. Others won’t, simply because these models introduce new capabilities that many don’t have the context to apply.

If a model touts PhD-level benchmarks, but you don’t have the expertise or the right questions to ask, you won’t get significantly different results from R1 than you would from GPT-4 or any other model. But those who understand the potential of models that don’t just generate code but can tell you what code to generate—or that don’t just create data but bring in vertical context to analyze it and surface new trends—will create an entirely new category of solutions.

We're going to see a divide: companies stuck at the plateau, having reached the limits of their competence and vision because they’ve only focused on automating what we already know how to do, and those who keep climbing, separating themselves by bringing entirely new capabilities to market.

Some of this will be driven by open-source models, making it possible to buy a Mac Mini and generate high-quality data 24/7 while leveraging more powerful models for reasoning-intensive tasks. The accessibility of these capabilities is shifting rapidly, and I expect we’ll see a wave of innovation until we hit the next summit.