Extropic is a US hardware company developing probabilistic computing accelerators for AI workloads, built on sub-threshold CMOS rather than on the digital arithmetic that GPUs use. Its first published chip is Z1, paired with an FPGA co-processor in a pipeline the company calls Z1T.
Z1 and the Z1T pipeline
Work on Z1T became public on September 7, 2026, reporting estimated energy figures for running transformer-like models across the pipeline (Source: extropic.ai).
| Measure | Figure as published |
|---|---|
| Z1T pipeline, per token | 294.52 nanojoules — 8.74 nJ of Z1 sampling and 285.78 nJ of FPGA work |
| Nvidia H100, per token | 40.9 microjoules at 10% model FLOPs utilization |
| Ratio at 10% utilization | About 139× |
| Ratio at 50% utilization | About 28× |
| Ratio at 100% utilization | About 14× |
Z1 itself carries 269,568 probabilistic bits across eight cores with 2,135,904 coupling edges at degree 16, drawing under 1 W.
Three qualifications are the company's own and bound the comparison. The projections are Extropic's best estimates rather than measurements of Z1 itself. The FPGA accounts for more than 95% of the pipeline's energy, so the headline ratio is dominated by conventional silicon rather than by the probabilistic part. And matching GPT-2-small's loss requires about an order of magnitude more training FLOPs with the sparser Z1T architecture — an efficiency gain at inference bought with a training-compute penalty. The ratio also falls by roughly an order of magnitude as the GPU baseline's utilization rises from 10% to 100%, so the comparison depends heavily on which baseline utilization is assumed.
Extropic released Z1T weights on Hugging Face and the training recipe on GitHub (Source: extropic.ai).
Relationships
- related: Inference Economics and Token Pricing
- related: Semiconductor Supply Chain
- related: AI Data Centers — energy per token is the unit in which data-centre power demand is ultimately denominated
- related: Nvidia & TSMC — AI Compute Infrastructure — the H100 supplies the comparison baseline