AI made an expensive thing cheap, then sent the bill somewhere else.
A credible page of copy, a dozen product images, a software prototype, a personalized sales sequence: each can now appear faster than the people receiving it can decide whether it matters. The cost of production keeps falling. The number of hours in a day has been stubbornly unimpressed.
We can generate faster than we can absorb. The bottleneck moved.
Stephen Garner at Adobe put the problem cleanly: “faster production does not necessarily create a faster organisation.” His real question is whether a company can absorb, adopt, and act at the speed being asked of it. That is the fault line. It also points to a second question: what kind of compute belongs on the absorption side?
That is why I am interested in a strange computer being built outside Boston. Not because it promises to make more things. We have handled that problem, possibly too well. I am interested because the hardware is designed to make particular forms of inference under uncertainty much cheaper.
If it works at scale, the next big drop may be the cost of deciding what deserves attention.
Your spreadsheet is lying about the future.
Most business plans take uncertain numbers and make them behave. Demand is 4,000 units. A delivery takes six days. Twelve people will be available. Fraud will stay under one percent. Put the numbers into a spreadsheet, optimize the plan, and the result can be precise to two decimal places while being wrong before lunch.
Berend Markhorst's business primer on stochastic programming starts with a fashion company and a bad crystal ball. A deterministic plan acts as if uncertain demand were known. A stochastic plan represents several possible futures, assigns probabilities to them, and chooses a decision that can survive the distribution rather than win against one imaginary average.
One useful version has an ordinary rhythm: decide, observe, correct. Order inventory before demand is known. See what customers actually buy. Mark down, reorder, or move the remaining stock. The correction is called recourse, which is a wonderfully formal word for what every decent operator does after reality ignores the plan.
The framework can also ask sharper business questions. What is the cost of planning around the average instead of the full range? What would perfect information be worth if you could buy it? Those have formal names, the value of the stochastic solution and the expected value of perfect information, but the point is practical. Uncertainty has a price. Better information has a ceiling. Both can be estimated before somebody buys a larger dashboard.
This field has existed for decades. Extropic did not invent it, and it does not require a new chip. Stochastic programming in operations research and stochastic differentiable programming in Extropic's new Torx framework are adjacent ideas at different layers. One helps a business choose under uncertain inputs. The other makes random transitions into trainable building blocks for software and, eventually, hardware.
The shared idea is more important than the shared word: stop forcing an uncertain world through one pretend-certain answer.
A bit that will not sit still.
A normal bit is read as zero or one. A probabilistic bit, or pbit, is also zero or one at any instant, but it keeps moving between them. A parameter controls how much time it tends to spend in each state. Turn the control one way and the pbit behaves like a coin that lands heads nine times out of ten. Turn it the other way and the bias reverses.
One draw from that coin is a sample. It is not an answer. It is one possible state drawn from a distribution. Draw repeatedly and the shape of the distribution appears.
Digital hardware is engineered so ordinary thermal fluctuations do not change the answer. Extropic's Thermodynamic Sampling Units use transistor noise as part of the computation. The pbit does not ask software to manufacture randomness on top of deterministic machinery. The physical circuit produces a tunable random state directly.
That sounds like a small trick until many pbits are connected. Make one pbit influence another and correlation appears. Add local rules, repeat the transitions, and a network of little uncertain parts begins producing structured distributions.
Extropic's open-source Torx framework is the software layer for arranging those transitions. Instead of a graph made only of deterministic functions, Torx programs can contain stochastic kernels, small operations that move probability from one state to another. Their parameters can be trained with gradient-based methods. The accompanying paper demonstrates random walks, discrete diffusion, stochastic graph networks, jump diffusion, and Ising sampling, plus two estimators whose physical randomness came from Extropic's X0 test chip.
That is the bridge. Write a probabilistic program at a useful level, train it, then compile the right parts down toward hardware that performs sampling as a native operation.
Not every random program belongs on this hardware. The mapping still has overhead. The energy advantage can disappear if a workload needs too much decomposition, communication, or control around the sampler. The Torx paper says this explicitly. Workload fit is the whole game.
The valuable work looks a lot like filtering.
What happens after generation floods a business?
An inbox needs to separate a buying signal from polite noise. A support queue needs to estimate urgency, identify the likely issue, and route it to someone who can resolve it. A warehouse needs to revise its demand estimate when weather, promotions, and late freight disagree. A security system needs to decide whether a missing device is actually gone or simply failed to answer one network scan.
These are not content-generation problems. They are inference problems. The useful output is not another artifact. It is a better estimate of a hidden state, a better ranking of possible explanations, or a decision made against several futures instead of one.
I am starting with the boring one because boring problems are honest. SquirrelOps Home currently sees a device answer an ARP scan and marks it online. If the device misses one reply, it can be marked offline immediately. Phones sleep. Wi-Fi gets weird. One missed observation is weak evidence, but a Boolean field has no room for weak evidence.
The experiment I have specified replaces that instant flip with a small hidden Markov model. It asks for the probability that a device is present given the current scan, recent scans, time of day, and the fact that a false reply while absent is much less likely than a missed reply while present. There are 107,000 historical events available for a backtest. The gate is concrete: cut false offline flips by at least half without adding more than about one scan cycle of detection delay, or do not put it into the product.
Torx may be useful as a modeling bench for that work. It may not. The eventual SquirrelOps runtime is planned as roughly one hundred lines of plain Python with no new dependencies, because a probabilistic chip company does not need to appear in a product for the underlying decision model to be useful.
The larger opportunity is the same across the examples. A flooded business does not need every system to talk faster. It needs systems that preserve context while filtering, update beliefs when new evidence arrives, and spend human attention where the expected consequence is highest.
Probability is not judgment.
There is a tempting mistake here. If a machine can produce a distribution, people will call the distribution calibrated. If it can sample quickly, they will call the result judgment. Neither follows.
Language models are already probabilistic. They produce distributions over tokens, even when the interface gives us one confident sentence. What they do not reliably provide is calibrated confidence about a business decision. A pbit does not fix that. An energy-based sampler does not know whether the model of the business is any good. Calibration still comes from comparing predictions with reality, correcting the model, and knowing when the available evidence is thin.
Probabilistic hardware changes the cost of running the model. It does not accept responsibility for choosing the model.
The Ising ring is not a diagram of an organization. People are not spins, and consensus is not automatically truth. It demonstrates something narrower and more useful: local uncertain states can interact until global structure becomes visible. That family of math appears in denoising, state estimation, anomaly detection, and energy-based models.
The number to watch is not 10,000 yet.
Extropic has already manufactured dozens of XTR-0 desktop systems around its X0 test chip and says it has sent them to early adopters. The next chip, Z1, is planned with 269,568 pbits, a sampling rate above 50 MHz, and power consumption below one watt. The planned M.2 compute stick would carry two chips. The planned PCIe card would carry sixteen. Early access for the Z1 systems is targeted for 2027.
The company's headline claim is up to 10,000 times greater energy efficiency than GPUs for the right workloads. “Up to” and “the right workloads” are carrying real weight. Z1 has not shipped. The public Torx examples are research examples, and the X0 hardware result in the paper shows physical samples producing estimates consistent with a software pseudorandom baseline. That is meaningful evidence that the primitive works. It is not evidence that a production inference stack is 10,000 times cheaper.
On July 29, Extropic also announced a nonbinding letter of intent for up to $75 million from the US Department of Commerce. The money remains subject to definitive agreements and project milestones. It would support Z1 clusters, generative AI benchmarks, and a domestically fabricated Z1.5. Serious interest, serious caveat.
So I would not build a 2027 budget around thermodynamic hardware. I would build the ability to recognize a sampling-shaped problem now.
That is why I am working from several directions at once: a browser playground that makes pbits visible, a local Torx lab bench that lets an agent construct and test small circuits without writing JAX by hand, an evidence-gated presence filter for SquirrelOps, and small open-source contributions where the framework proves incomplete. The hardware can miss its date and each of those still teaches something useful.
Not everything should move at the same velocity.
Modern computers are already mixed machines. CPUs handle control. GPUs handle dense parallel math. Specialized chips encode video, process signals, and move packets. If Thermodynamic Sampling Units earn a place, it will be beside those processors, doing the probability-shaped work that fits them.
A business should be mixed the same way.
Let deterministic software enforce the rule that must never bend. Let probabilistic systems maintain beliefs where the evidence is incomplete. Let people set the utility, inspect exceptions, and own decisions whose consequences belong to them. Faster generation can live beside slower judgment. The goal is not to make every part of the organization move at model speed. That would be exhausting and stupid.
The next shortage is not words, images, code, or options. It is warranted attention.
If the cost of a sample collapses, judgment does not become free. We get more chances to ask what might be true before a person decides what to do. That is less spectacular than generating another world in six seconds.
It may be the first part of AI that makes the rest bearable.