These invisible factors are limiting the future of AI

AI is no longer just a cascade of algorithms trained on massive amounts of data. It has become a physical and infrastructural phenomenon, one whose future will be determined not by breakthroughs in benchmarks, but by the hard realities of power, geography, regulation, and the very nature of intelligence. Businesses that fail to see this will be blindsided. Data centers were once the sterile backrooms of the internet: important, but invisible. Today, they are the beating heart of generative AI , the physical engines that make large language models (LLMs) possible. But what if these engines, and the models they power, are hitting limitations that can’t be solved with more capital, more data centers, or more powerful chips? In 2025 and into 2026, communities around the U.S. have been pushing back against new data center construction . In Springfield, Ohio; Loudoun County, Virginia and elsewhere, local residents and officials have balked at the idea of massive facilities drawing enormous amounts of electricity, disrupting neighborhoods, and straining already stretched electrical grids. These conflicts are not isolated. They are a signal, a structural friction point in the expansion of the AI economy. At the same time, utilities are warning of a looming collision between AI’s energy appetite and the cost of power infrastructure. Several states are considering higher utility rates for data-intensive operations, arguing that the massive energy consumption of AI data centers is reshaping the economics of electricity distribution, often at the expense of everyday consumers . This friction between local resistance to data centers, the energy grid’s physical limits, and the political pressures on utilities is more than a planning dispute. It reveals a deeper truth: AI’s most serious constraint is not algorithmic ingenuity, but physical reality . When reality intrudes on the AI dream For years, the dominant narrative in technology has been that more data and bigger models equal better intelligence. The logic has been seductive: scale up the training data, scale up compute power, and intelligence will emerge. But this logic assumes that three things are true: Data can always be collected and processed at scale . Data centers can be built wherever they are needed . Language-based models can serve as proxies for understanding the world . The first assumption is faltering. The second is meeting political and physical resistance. The third, that language alone can model reality, is quietly unraveling . Large language models are trained on massive corpora of human text. But that text is not a transparent reflection of reality: It is a distillation of perceptions, biases, omissions, and misinterpretations filtered through the human use of language. Some of that is useful. Much of it is partial, anecdotal, or flat-out wrong. As these models grow, their training data becomes the lens through which they interpret the world . But that lens is inherently flawed. This matters because language is not reality: It is a representation of individual and collective narratives. A language model learns the distribution of language , not the causal structure...

Preview: ~500 words