The Probability Engine

We’ve spent a lot of time in this space together, looking at data centers from the outside in, how much power they draw, how much water they use, how much noise they generate, and how a hyperscale campus differs from the smaller, lower-density facilities that have served businesses for decades. That naturally raises a question I haven’t explored directly: what is all of that for? When a company spends billions of dollars on a hyperscale campus to train a “frontier” artificial intelligence model, what is actually being built inside? Once it’s built, what is the thing actually doing when you type a question into it?

The short version is this: a large language model is a probability engine. It is a machine for guessing what comes next. All that computing, all that data, all that power, noise, space, and cooling, is in service of making those guesses better. Understanding this one idea is essential to understanding the AI chatbots from OpenAI, Anthropic, xAI, and Google that are driving hyperscale data center buildouts.

Leave a comment

You already use a tiny probability engine every day. When you start typing a text message and your phone suggests the next word, that’s a prediction. Type “I’ll be there in five...” and your phone offers “minutes.” It doesn’t know what you mean. It has seen enough text to learn that, after those words, “minutes” is far more likely than “elephants.”

A large language model, the AI chatbots like Claude or ChatGPT, is the same idea taken to an almost unimaginable extreme. Instead of suggesting one of three words, it considers the entire vocabulary of human language and assigns a probability to every possible next chunk of text, then it picks one, usually one of the more likely options, adds it to the sentence, and repeats the whole process for the next chunk, and the next. One piece at a time, it writes an essay, a poem, or a block of computer code, never doing anything more exotic than answering the question: given everything so far, what is likely to come next?

That’s worth pausing on, because it’s the whole magic trick. The model is not looking up an answer in a database. It is purely computing odds.

A useful analogy: a weather forecaster doesn’t know whether it will rain tomorrow. What they have is an enormous record of past weather and a model of how conditions tend to evolve; from that, they produce a probability: “70% chance of rain.” The forecast is a very good, evidence-based guess, not a fact retrieved from the future.

A language model forecasts words the same way. Given the text in front of it, it produces something like a forecast: “the next word is 78% likely to be cooling, 9% likely to be power, 4% likely to be water...” and so on across thousands of options. Then it commits to one and moves on. String enough of these forecasts together and you get a coherent paragraph, but every single word along the way was a probability, not a certainty.

This is also why these models sometimes state false things with total confidence (the industry calls it “hallucination”). The model isn’t lying or confused because it was never checking its words against reality in the first place. It was always just producing the most probable continuation of what came before it. Usually, the most probable continuation is also true, because true statements are common in its training text, but not always. The machine optimizes for plausible, not for correct, and most of the time, plausible and correct happen to align.

So if the model is just guessing the next word, why does it take a hyperscale campus and a city’s worth of electricity to build one?

Because making good guesses is staggeringly hard, and the model has to learn how on its own.

Inside the model are billions, in frontier systems, often well over a trillion, adjustable numbers called parameters. Think of them as tiny dials; the specific settings of all those dials are what determine the model’s guesses. The right settings produce useful predictions; random settings produce gibberish. Nobody sets these dials by hand. The model has to find the right settings through practice, and that practice is called training.

Training works like this. The system is shown a mind-boggling amount of text, a large fraction of the public internet (Reddit is especially common), books, articles, code, and more. It plays a simple game, trillions upon trillions of times: a passage is shown with the next word hidden, the model guesses, and the guess is compared to the real word. Every time it guesses wrong, an automated process nudges the dials a tiny bit in whatever direction would have made the right answer more likely. Wrong again, nudge again. Repeat across this entire digital library, over and over. More data and more compute generally buy sharper, more capable probability estimates. That, in a sentence, is what the entire AI buildout is purchasing: better guesses.

This is the part that pushes toward hyperscale, but it’s worth being precise about why. Nudging a trillion dials, trillions of times, against an ocean of text is an arithmetic problem of absurd size. Here’s the point that often gets lost: an ordinary computer could, in principle, do every bit of that arithmetic. The math is the math. A single high-end machine would simply chip away at it one calculation after another, and at that pace, a frontier model would take centuries, plausibly millennia, to train. By the time it finished, the result would be a museum piece, and the company that built it would have gone bankrupt waiting.

So hyperscale is not a technical requirement; it is a business requirement. The whole game in frontier AI is finishing the training run before your competitors finish theirs, compressing what a lone computer would spend a thousand years on into a matter of weeks or months. The only way to do that is to split the work across tens of thousands of specialized chips running in lockstep, which is precisely what a hyperscale data center is built to do. The campus isn’t there because the problem is impossible without it; it’s there because speed wins markets, and speed at this scale costs a great deal. The power, water, noise, and space are the physical price of buying time, of forging a competitive model in months instead of millennia. The two distinctions we’ve drawn, hyperscale versus low-density, training versus everyday use, line up neatly: the hyperscale campus is the factory where the model is forged under a competitive deadline, and the lighter-weight facilities are more like where the finished model answers your questions afterward.

Seen this way, the stakes come into sharper focus. The same race that is being used to justify all this scale, finish before your rivals, is, at its core, a bet. Model builders are sprinting against one another toward a payoff no one can yet guarantee: the promise that these tools will transform the economy and society, and that promise is exactly what's in dispute right now. So far, many companies have struggled to turn these tools into clear, measurable economic returns, even as the firms building them pour in hundreds of billions of dollars, much of it spent well ahead of any profit, and much of it borrowed or financed through private equity and private credit markets in recent years. Put plainly, we are building factories at an enormous scale for a product no one is sure will ever justify the investment, and the communities that host those factories, across the country, and here in Southern Arizona, are the ones living with the consequences of the gamble.

Here’s the part that trips people up. If it’s “just” predicting words, why can it write a working program, walk through a math problem, or explain a complicated concept?

Because to predict human text really well, a model has to absorb the patterns buried inside that text, and a lot of those patterns are the structure of reasoning itself. Our writing is full of logic, cause-and-effect, step-by-step arguments, and factual relationships. A system that gets extraordinarily good at continuing such text ends up reproducing the shape of reasoning, because reasoning-shaped text is what it was trained to predict. When you ask it to “solve this step by step,” it generates the kind of words that tend to follow that instruction, and those words often add up to a correct solution.

So the capability is real and genuinely useful, but it emerges from prediction, not from a mind sitting behind the screen deciding what it believes.

This is where I want to be careful and honest, because it’s the question everyone actually has. Here is what I can say with confidence. A language model has no body and no senses; it has never seen or touched the world it writes about. By default, it has no memory of you from one conversation to the next; each session starts fresh. It has no goals, desires, or stake in the outcome of your question. It doesn't get bored, lonely, or curious between your messages because it isn’t doing anything. There is no evidence that it has feelings or any inner experience, and nothing about its design, a system tuning dials to predict text, gives a reason to expect that it would. On the question of sentience, genuine subjective experience, the lights being on inside, the honest answer is that there is no good evidence for it, and the burden of proof sits squarely on anyone claiming otherwise.

The word “think“ is where reasonable, serious people genuinely disagree. Some researchers argue that to predict text this well, a model must build real internal representations of the world, a loose kind of understanding. Others argue it remains a sophisticated pattern-matching with no comprehension behind it. That debate is live and unsettled, and pretending it’s closed would be incorrect. Notice that both camps agree on the mechanism described: the system computes probabilities for the next chunk of text. Whatever you want to call the result, the underlying engine is a probabilistic one.

For everyday purposes, that’s the mental model to hold onto. When a chatbot writes you something warm, or insightful, or eerily personal, it is not a being reaching out to you. It is a probability engine producing the words that, statistically, tend to follow your words, and it has gotten remarkably good at it.

You might wonder why a think tank focused on water, energy, and land use is writing about how chatbots work. The reason is that thinking about the physical buildout depends on thinking about the product.

Every claim made to justify a hyperscale campus, the jobs, the tax base, and “it’s the future” ultimately rests on what these machines are and what they’re worth. If we treat them as digital minds on the verge of consciousness, governments will make policy out of science fiction. Understanding what’s actually inside the box is the first step to developing responsible policy.

The Sonoran Think Tank is a civic research organization focused on data center development, energy, water, and land use policy in Southern Arizona. All of our work is free and publicly available.