Are We In A Bubble?

We’ve been here before. In the 1990s, there was a massive overbuild of fiber and networking capacity associated with the arrival of the World Wide Web riding the Internet. Justifiable excitement led to the creation of uncounted numbers of companies having “something” to do with the Internet and the Web. The so-called “dot-boom” eventually led to the “dot-bust” in 2000. Is it possible we are seeing a similar scenario in 2025? There is massive expansion of subsea fiber networking, partly for capacity and partly for resilience in the face of increasing dependence on Internet-based applications. Much of this expansion is driven by new artificial intelligence (AI) developments that had their origins in the 2010s and 2020s with deep and reinforcement learning. (Computer-based AI research started in 1956 at a meeting at Dartmouth University before it went through three cycles—heuristic and symbolic processing, expert systems, and machine learning / large language models between the 1960s and the present.) Today, we have multi-modal large language models—sometimes called foundational models—that exhibit capabilities that seem closer than ever to artificial general intelligence (AGI). Inter alia, they draw, speak, write, program, sing, compose, and summarize. They are our notetakers, oracles, critics, and helpers.

Even though these huge generative models can produce incorrect output, they are improving rapidly in scope and reliability. Smaller, specially-trained models seem less likely to “hallucinate.” New kinds of models are emerging as “agents” that can engage in real-world actions like making reservations, writing articles, programming, controlling aspects of data center operation, and undertaking a multitude of other tasks. They are getting better at translating and transcribing scores and even hundreds of languages. Inter-agent interaction is being facilitated by the development of standard protocols. Just as TCP/IP and QUIC in the Internet facilitate interaction between programs / processes, new agent-to-agent (A2A) and Model Context Protocol (MCP) are being developed to facilitate and make reliable the exchanges between agents.

These systems consume significant amounts of computing power both during the training phase and the “inferencing” phase when they are put to work. Providers of AI applications are building data centers and networks at a ferocious pace and taking advantage of new hardware like graphical processing units (GPUs) and tensor processing units (TPUs). It is conceivable that this build-out represents another Internet bubble as companies vie for market share in a rapidly developing market. I am inclined not to think it is so obvious that this is a bubble. The application space of AI is enormous and expanding. Agentic AI will spawn many applications in the same way that smartphones supported the creation of millions of “apps” in the 2007+ period.

There are, of course, important economic considerations. The build-out is demanding significant capital expenditure. There are operating costs for running the networks and data centers, to say nothing of the programming and training of the models. The demand for electrical power is without precedent and is already leading to availability challenges. When such metrics begin to grow exponentially, we know that is not a sustainable path. More efficient hardware, smaller AI models engaged by agents driven by foundational models, incremental learning, and many other mechanisms may tame this demand curve. We are just beginning to see the landscape of AI applications emerging as business users and researchers start to build their own AI applications. Skill at using AI-based applications will grow, just as the use of spreadsheets made programmers out of everyone.

There is a hazard with the emergence of these powerfully enabling systems. They can be put to use generating misinformation, disinformation, and other socially harmful practices. That’s a topic for another essay, but it deserves a great deal of attention. Agentic AI will tie systems together in increasingly complex operations. We are really just at the beginning of that transformation. So, are we in a bubble? Perhaps it is too early to tell, but I am persuaded that a lot of the growth of AI applications is rooted in practical uses that will become part of our daily fabric of activity.  AI will become another norm. People will say, “What did we do before we had microwave ovens, the Internet, the Web, smartphones, and AI?”