We started Ludwig because we've spent the last decade living inside a problem that the rest of the tech industry is only now running into: probability is fundamental to modern AI, but our computers still treat it like an afterthought.
The Inference Wall
The reality of generative AI inference is that performance and energy are no longer bottlenecked by arithmetic; they are dominated by data movement.
Generative AI is probabilistic by nature. Its core function is to map input probability distributions into output probability distributions. Yet, inference today runs on hardware optimized for deterministic, high-precision math. To get probabilistic behavior out of deterministic machines, we are forced to emulate it through massive memory traffic and repeated compute cycles.
This is why inference is expensive and power-hungry even on state-of-the-art GPUs. We are paying a massive “bits moved” tax. That architectural mismatch is the real wall the industry is hitting.
We didn't stumble into probabilistic computing - we built its foundation
We use pCompute to describe a simple idea with deep consequences: a full-stack computing paradigm where the algorithms, software, and hardware are all co-designed to natively compute with probability distributions. By generating stochasticity directly at the hardware level, we dramatically shrink the “bits moved tax” of traditional compute-memory shuffles.
For our team, pCompute isn't just a theoretical pitch. It is rooted in our decade-long research lineage pioneering the p-bit - a probabilistic building block designed to make stochasticity useful, controllable, and scalable.
Our trajectory has been a step-by-step mission to bring probabilistic computing out of the lab and into real systems:
- The Insight: The thesis that drives Ludwig today began over a decade ago. In our early device-level work (Nature Nanotechnology, 2010), we argued that new computing paradigms emerge only when you design the device, the memory, and the computation together, rather than stacking them in isolated layers.
- The Blueprint: By 2016, we formalized the solution. Instead of forcing Bayesian reasoning to be a software pattern layered on top of deterministic compute, we proposed building primitives that directly implement probabilistic inference (Nature Scientific Reports, 2016).
- The Physical Proof: We didn't let pcomputing stay at the concept level. We pushed through the hard engineering layers of device embodiments and circuit design, culminating in a flagship demonstration that proved stochastic hardware could solve complex problems in the real world (Nature, 2019). We also built large scale, fully digital FPGA designs to study real world scale implementations of probabilistic computing (IEEE, 2018) (IEEE, 2020).
- The AI Architecture: As generative AI began to scale, we tied our device, circuit, and architecture designs into a unified software-hardware framework - creating a complete, full-stack view of probabilistic computing with p-bits for generative AI (ICRC, 2025).
Today, our foundational work is widely recognized in low-power and unconventional computing roadmaps (APL Materials, 2024). We didn't pivot to this space when AI got hot; we've been building toward a probability-native computing stack for years.
Why pCompute matters for Generative AI right now
The generative AI stack has quietly become a massive sampling machine.
Sometimes that's explicit: temperature, top-p, speculative decoding, and uncertainty estimation. Sometimes it's implicit in model behavior. Either way, the industry is already paying the massive energy and latency costs of probability. Plenty of companies are trying to build faster matrix-multiply engines to brute-force this problem. That will certainly help, but it won't be enough. Ludwig exists because the next massive leap in AI efficiency won't come from squeezing another 5% out of dense, deterministic compute. It will come from changing the foundational primitive operations so that the hardware actually matches the true mathematical structure of inference.
We've already invented the primitive. Now, we are building the machine.