- Co-inventor of the p-bit (Nature 2019)
- Landmark Nature paper with 2025 Nobel Laureate John Martinis
- $25M+ federal research as PI
A hardware lab for probabilistic computing
We are buildinga new class of computer
Flucta invents probabilistic algorithms and designs the hardware around them. Our algorithms change what a machine needs to compute, which data it needs to read, and how often its processors need to communicate. They run on NVIDIA GPUs today, and our own FPGA accelerator runs models end to end.
250:1
Energy to read a byte vs. compute on it
0
Retraining or fine-tuning needed to deploy
Near-dense
accuracy in our benchmarks, vs. the dense baseline
2027
SPU tapeout, 7 nm custom silicon
Why a new class of computer
Reading a byte costs
250× more than computing on it.
On a modern accelerator, reading one byte from HBM costs roughly 25 picojoules; computing on it costs about 0.1. We found that models tolerate substantial sparsity: in attention, in experts, in long contexts, a small fraction of the bytes carries the answer. Our algorithms already use that tolerance on GPUs, and they run into the machine itself: a GPU is architected for dense execution and cannot exploit sparsity in full. That is why Flucta is building its own hardware, as the only hardware company that designs the algorithm and the processor as one.
Memory
HBM · hundreds of GB
KV cache, weights
bytes the model needs
Memory interface
Bottleneck25 pJ / byte
Compute die
0.1 pJ / byte
Waits on memory most of every token
Bytes the model needs
a small fraction
Bytes dense hardware reads
all, every token
250×
move ≫ compute
Widening the bus and stacking more memory does not remove the overhead: dense hardware still reads bytes the model never needed.
Flucta is redesigning compute, data movement and connectivity around that tolerance: the machine performs selected work, reads selected data, and communicates less often.
The same answerfrom a fraction of the reads.
Paris
Reads / tokenthe whole context
Legacy · Dense Read
Reads everything, every token
VS
Paris
Reads / tokena small fraction
Flucta · Sampled Read
Reads only what carries signal
From algorithms to FPGAs to purpose-built silicon.
The speedups run as software on today's GPUs, with no retraining. The same kernels are moving onto FPGA prototypes now, and measurements from that hardware drive the Stochastic Processing Unit: purpose-built ASICs first, then 3D-stacked memory-on-logic. Every stage runs the same probabilistic abstraction on hardware with less overhead than the stage before.
Where we’re going
Today
End-to-End Probabilistic Algorithms
Running on GPUs today · FPGA validated · ICML 2026
Early 2027
SPU tapeout
7 nm · custom memory hierarchy
2028
3D-stacked SPU
Near-memory compute · datacenter
Physicists and chip architects, building from first principles.
Flucta's founders started the field of probabilistic computing with p-bits, and many of the field's Nature-level firsts are theirs: the first demonstration of 1,000,000 p-bits, the first carbon nanotube computer, and firsts in 3D chips and in hardware/software co-design.
Founding team
* ConsultingSilicon
Kartik Prabhu
Ph.D., Stanford
Algorithm / Kernel
Shuvro Chowdhury
Ph.D., Purdue
Algorithm / Kernel
Corentin Delacour*
Ph.D., Univ. of Montpellier
Algorithm / Kernel
Kemal Selçuk
Ph.D., UC Santa Barbara
Algorithm / Kernel
Nihal Sanjay Singh
Ph.D., UC Santa Barbara
Silicon
Kai Bartolone
M.S., Stanford
Silicon
Kevin Callahan-Coray
Ph.D. Candidate, UC Santa Barbara
Silicon
Eric Cheng
M.S. Candidate, Stanford
Inference Serving
Zengxiao He
M.S., Stanford
Research Consultant
Osman Yağan
Professor of ECE, Carnegie Mellon
Research Consultant
Nikhil Shukla
Associate Professor of ECE, UVA
Advisors
Funded by investors
who back foundational technology.
Frequently Asked Questions
Flucta builds probabilistic computing across the stack. Our algorithms and runtime speed up inference end to end on today's GPUs, with accuracy close to the dense baseline in our benchmarks, and the same principles carry into FPGA prototypes and purpose-built silicon.
Chip companies optimize for algorithms as they find them, and algorithm companies optimize within the silicon they are given. Flucta designs the algorithm, the runtime and the processor as one system. The orders-of-magnitude gains live between those layers, and a company that owns only one layer cannot reach them.
Not initially. Though our principles will naturally lead us there in terms of custom ASICs, 3D-CMOS, and near-memory computation, our initial focus operates at the system and architectural level. Instead of waiting on novel physical devices, we re-architect memory hierarchies, routing logic, and execution pipelines on standard silicon processes to natively support sampling and probabilistic operations today.
Enormous efficiency is lost between the design abstractions of the modern stack. Algorithms are written as if moving memory were free, runtimes assume dense and deterministic execution, and silicon is built for worst-case precision. Each boundary hides what the layer below actually needs, and the waste compounds. Flucta removes these boundaries with a top-down design philosophy, carrying one probabilistic abstraction from the algorithm through the runtime into the silicon, so the gains at each layer multiply.
Email founders@flucta.ai to request the technical brief. We are hiring first-principles thinkers across physics, computer science, computer engineering, kernel development, and inference optimization.