Last week we explored why AI’s water crisis is fundamentally an architectural problem, not a cooling problem. The root cause is a learning algorithm that demands extraordinary computational precision at every step. This week we explore stochastic machine learning, a fundamentally different approach that trades that precision for probability, and in doing so, generates significantly less heat, consumes less energy, and demands far less water in the process.

The dominant paradigm in AI today is deterministic optimization. Feed a neural network enough data, run enough gradient calculations, adjust enough parameters, and eventually the model converges on something useful. It’s brute force mathematics applied at massive scale. And it works. But the energy cost of that precision is becoming untenable.

There is another way to learn. One that trades precision for probability, and in doing so, achieves results that are just as useful at a fraction of the computational cost.

That approach is stochastic machine learning. Stochastic simply means probability-based rather than exact. Instead of calculating a precise answer at every step, a stochastic system makes smart probabilistic observations that converge on the right answer over time, like a detective building a case from accumulating evidence rather than waiting for a confession.

And it’s the foundation of how Brain-CA’s architecture works.

Think of an experienced poker player reading an opponent across the table. They’re not calculating the mathematical probability that the opponent is bluffing. They’re observing: betting patterns, posture, the way they reached for their chips, the slight hesitation before they called. With each hand, the player builds a mental model of that opponent. Not a spreadsheet. Not a calculation. A pattern of states.

Aggressive. Cautious. Aggressive. Aggressive. Cautious.

When a new hand plays out and the opponent does something unexpected, the player doesn’t rebuild their model from scratch. They nudge it, shifting their read slightly toward “more aggressive than I thought” or “tighter than I expected.” One small adjustment. The model updates. No math. Just observation and a flip.
As the player tracks more dimensions of their opponent’s behavior, additional binary models layer on top: bluffing or honest, patient or impulsive, confident or rattled. Each one is its own simple two-state model, updated independently through observation. Together they compose a rich, accurate picture of the opponent without a single calculation.

This is exactly how Brain-CA’s Estimator works.

And here’s what makes that remarkable: the model the poker player is building in their head, without arithmetic or calculation, implies mathematical values. That pattern of states encodes a probability estimate just as accurately as a formula would. The math is implied by the mechanism. It doesn’t have to be performed.
A sharp reader might point out that flipping states is, in a formal sense, equivalent to binary arithmetic. That’s technically true. But the distinction matters for two reasons. First, the Estimator requires no concept of number, no numeric representation, and no arithmetic unit. It can operate entirely on symbols, colors, or states. Second, and more importantly, this is plausibly how biological intelligence actually emerged. Not from organisms that developed tiny calculators, but from organisms that developed the ability to store a state and flip it. The math is implied by the mechanism, not required by it.

To appreciate why this matters, it helps to understand what conventional neural network training actually does and why it’s so energy-intensive.

A modern neural network starts with a problem that stochastic learning avoids entirely: it begins large. Billions of parameters are randomly initialized at the start, essentially a system that begins by being wrong about everything. Training is the process of correcting that wrongness, over and over, across millions of examples.

Each correction requires a process called backpropagation, calculating exactly how wrong each of the billions of parameters was, and adjusting each one by a precise mathematical amount. This requires billions of high-precision floating-point operations per training step. Research from Stanford University finds that during training, up to 90% of that energy is spent simply moving data between the processor and memory, not performing the calculations themselves. That data movement generates heat. Heat requires cooling. Cooling requires water.

The waste isn’t incidental. It’s structural. It’s baked into the algorithm.

Brain-CA’s stochastic approach starts from a fundamentally different place: a null model with zero parameters.

Rather than beginning with a large, randomly initialized structure and spending enormous energy correcting it, Brain-CA’s model builds itself only as the data justifies it. Each new observation either confirms the existing model or prompts a minimal, targeted update. Nothing is pre-allocated. Nothing is corrected from a wrong starting point. The model grows in proportion to what it has actually observed.

Think of it like the poker player’s mental model of their opponent. On the first hand, they know nothing. By the tenth hand, they have a rough read. By the fiftieth hand, their model is detailed and reliable. The model grew with experience, not before it. They didn’t sit down at the table with a pre-filled spreadsheet of assumptions that they then spent the evening correcting. They started with nothing and built from observation.

In technical terms, Brain-CA calls this Functional Growth. Storage requirements scale logarithmically: approximately 10 bits of storage after 1,000 observations, and approximately 20 bits after 1,000,000 observations. The model grows, but slowly and deliberately, in proportion to genuine new knowledge rather than pre-allocated guesswork.

This isn’t a marginal efficiency gain. It’s a fundamentally different relationship between how much data a system has seen and how much computational resources it needs to represent what it has learned.

At the heart of Brain-CA’s stochastic approach is the Estimator, a patented hardware component that implements this probabilistic learning process at the most fundamental level.

The Estimator maintains a single number, think of it as a compact summary of everything the device currently knows about the stream of data it is observing. When a new observation arrives, the Estimator compares it against its current state using simple bit-level logic, and either flips the relevant state, or leaves it unchanged.

That’s the entire update rule. Observe. Compare. Flip or stay. Repeat.

No multiplication. No division. No floating-point arithmetic. Yet despite this simplicity, the Estimator achieves what researchers call Bayesian-quality results, meaning it converges on the same probability estimate that a mathematician would calculate using far more complex methods. In plain terms: the Estimator figures out the right answer through observation and simple state adjustments, arriving at the same place a far more complex system would reach through calculation.

An 8-bit Estimator requires fewer transistors than an 8×8 multiplier, the kind of multiplier that conventional AI relies on billions of times per training step. At the scale of thousands or millions of Estimators running in parallel on a single chip, the energy difference becomes significant.

The Estimator models a single binary data stream, like a poker player tracking one dimension of their opponent’s behavior. But real-world problems involve relationships between many streams simultaneously.

Just as the poker player tracks multiple independent binary models at once (aggressive or cautious, bluffing or honest, patient or impulsive), multiple Estimators working in parallel can track relationships across many data streams simultaneously. Brain-CA’s Learning Engine scales by composing Estimators, building progressively more sophisticated models of the environment by tracking relationships across multiple streams in parallel.

Critically, each Estimator updates locally, without needing to communicate with a central processor. There is no data shuffling back and forth between memory and compute. No centralized bottleneck. Each component learns where it sits, from the data in front of it, in parallel with every other component doing the same thing.

This is how stochastic learning at the bit level becomes a viable architecture for complex, real-world AI tasks. Not by making individual operations more powerful, but by making millions of simple operations genuinely parallel and genuinely local.

The connection to AI’s energy and water footprint is direct.

Conventional neural network training demands precision: billions of high-precision floating-point operations per training step, each requiring data to be moved between compute and memory. That movement, as we explored in last week’s post on AI water consumption, accounts for up to 90% of training energy. All of that energy becomes heat. All of that heat requires cooling. Cooling requires water.

Brain-CA’s stochastic approach reduces the arithmetic. The local, in-memory update rule of the Estimator eliminates the data movement. Less arithmetic and less data movement means less heat, less cooling, and less water. Not as a side effect of better engineering, but as a structural consequence of a fundamentally different learning algorithm.

This is what it means to say AI’s resource problem is architectural rather than infrastructural. Better cooling towers don’t change the algorithm. A stochastic learning engine does.

Brain-CA’s stochastic architecture is protected by three issued U.S. patents. The Estimator and its underlying Cincinnati Algorithm were presented at the International Symposium on Computer Architecture (ISCA) 2025, where the work won Most Innovative Paper at the Energy Efficiency workshop. A proof-of-concept chip is currently being fabricated at TSMC.

The long-term vision is an AI hardware platform where learning happens at milliwatt power budgets. Not because of efficiency improvements layered on top of a conventional architecture, but because the learning algorithm itself requires significantly less computation to reach the same quality of result.
Probability, it turns out, is a more efficient path to intelligence than precision.

To explore the intuition behind the Cincinnati Algorithm further, read The Intelligence Evolution Casino. For the technical details of the Estimator’s convergence proof, see our Cincinnati Algorithm white paper.