AI's Trough of Disillusionment

One of my favorite conceptual tools when evaluating new technologies is the Gartner Hype Cycle, which outlines five key phases of technology evolution. It starts with a “technology trigger,” such as the rise of generative AI, and then proceeds into the “peak of inflated expectations” and then into the “trough of disillusionment,” where expectations and reality diverge while a technology matures.

As we continue to ride the hype train (toot toot!) of GenAI, it's natural to start considering where it fits into the hype cycle. Earlier this week, I was reading an analyst report outlining IT predictions for 2025, and it included a bold assertion that GenAI would enter the trough of disillusionment this year. This got me thinking – what historical tech disruptions are the most analogous to GenAI, and what does that tell us about its progress across the hype cycle?

What Does History Tell Us?

There’s no question that GenAI is transformative, and it is so disruptive because it has the power to change the way we work, measure and understand productivity, and how we deploy people and capital to achieve business goals. Even at this relatively early stage, GenAI has completely upended software development, content marketing, and knowledge management to the point where people are asking some pretty big questions about how humans fit into the equation. Will Software Engineers be replaced by AI? How about Product Marketers? Research Analysts? What should modern organizations even look like given the broad applicability of AI?

The last time I heard these sorts of questions being asked was during the explosion of cloud infrastructure services across compute, storage, and networking. As early adopters began deploying workloads on the cloud, their technical teams redefined their roles and organizations. Were data center “rack and stackers” replaced? In businesses that went all-in on cloud, the need for these roles was reduced dramatically. Instead, we started to see titles like “Cloud Engineer,” “Cloud Architect,” and “DevOps Lead.” Those who embraced the change found themselves in highly impactful and lucrative roles. Similarly, software defined networking tossed a grenade into a space with entrenched giants and armies of “Certification Collectors.” What need do you have for a CCIE cert if you’re programmatically defining your networks in the cloud?

The Speed of Innovation is Unprecedented

It's clear to me that Cloud is a reasonable analog for GenAI’s disruption. Cloud came bursting onto the scene in the early to mid 2000’s, starting with the foundational elements of compute, storage, and networking, and before long, ran into obstacles to adoption – security concerns, the CapEx to OpEx transition, talent shortages, etc. I vividly remember 2010, when analyst reports and industry pundits posed the same question we’re considering today… was Cloud entering the trough of disillusionment, and what does that mean for the industry?

When you look at transformative technologies, including cloud computing, there's usually a point when innovation gives way to optimization and refinement. New tools emerge, best practices crystallize, and the tech becomes “boring” in the best possible way as it enters the “plateau of productivity.” Generally speaking, the rate of change at the core gives way to innovation in the periphery.

Now, how is GenAI different from Cloud? A significant chunk of IaaS is changing the way businesses procure and utilize pre-existing technologies. Hypervisors and virtualization existed before EC2. As did scalable block storage and object storage. In the early days, cloud infrastructure services were frequently “Things as a Service,” where pre-existing open source software like PostgreSQL, Memcached, Redis, and a fleet of Apache projects were turned into fully managed, API-driven cloud services.

GenAI seems a bit different to me. The disruptive technological leap was the introduction of powerful generative foundation models, trained on massive corpuses of data. Since then, the number of models and the capability of those models has skyrocketed, with no end in sight. OpenAI and other players are claiming that AGI is within reach this decade. Even if AGI isn’t fully realized, pushing toward that goal is driving an unprecedented rate of change.

Take context windows, for example. In 2024 alone, we went from having models that could handle ~8,000 tokens to hundreds of thousands, and even millions of tokens. That's a massive leap in capability that opens up entirely new possibilities. Huggingface now lists over a million models across a variety of modalities. The core of GenAI isn’t settling into a steady plateau, it's accelerating.

What About Economics?

One of the key drivers of Cloud’s entry into the trough of disillusionment was financially driven. Moving from CapEx to OpEx was challenging. The very nature of IaaS makes it so accessible compared to traditional infrastructure that it's easy to get into a position where a lack of governance leads to explosive costs. AWS and the other hyperscalers became more efficient over time, and introduced tools like reservations, storage classes, etc., to address these challenges. S3, for example, started at $0.15/GB at launch in 2006, and now, decades later, the price has dropped by over 80%.

One of the biggest early objections to GenAI was the high cost of running models. The token price for GPT-4, the state of the art in 2023, was $30/million. In less than two years, the token price for models with similar or better performance is now between 10 and 30 times lower. That’s an absolutely stunning change that has been driven by intense competition and incredible innovation, especially when compared to the historical price of cloud services. While training these models is very expensive and has a high environmental impact, running them is now quite efficient.

What Can We Conclude?

Because the definition of the trough of disillusionment is subjective, I think it will be hard to pin down exactly when GenAI will have entered the trough. In my opinion, the pace of innovation is simply too high, and until we see it slowing, the hype train isn’t pulling back into the station. We're still a long way from the “boring” plateau of productivity, and while there will inevitably be cooling of the sky high hype, I believe that the dramatic pace and rapidly improving economics indicate that in 2025, GenAI will remain a passenger on the hype train. Choo-choo, baby!