INN™ intuitive neural networks, BPU brain-inspired silicon, two-phase cooling, and liquid metal — a full stack built for the next generation of intelligence.
Better intelligence comes from better first principles — rethinking every layer, from neurons to silicon to the way heat leaves the room.
INN™ isn't a bigger Transformer — it's a new architecture closer to biological intuition. Explainable, low-power, inference on CPU alone.
BPU brain-inspired silicon and wafer-scale systems, designed from first principles for sparse, event-driven computation — not as a GPU afterthought.
Two-phase liquid cooling and liquid-metal interfaces carry heat out of 45KW cabinets silently — no high-power CDU, PUE convergent at 1.08.
99.9% less energy isn't a marketing number — it means intelligence can live anywhere there's power, not just in hyperscale data centers.
When inference runs on CPU and cooling needs no chiller loop, operational costs fall by 50% — not through compromise, but through better design.
Research led by Chinese Academy of Engineering academicians, building INN™ from foundational theory — not engineering optimization, but scientific breakthrough.
Neogenint · Zhuhai, China
2024 — 2026
For a decade, the story of intelligence has been a story about scale — bigger models, more GPUs, hotter rooms. We're not going to keep telling it.
We're starting over in three places: the algorithm, the compute, the cooling. An algorithm that behaves like a brain, not a fatter Transformer. Compute built for sparse, event-driven thought, not matrix multiplication. Heat carried away quietly, not pinned down by a building full of chillers.
These three have to happen together. Any one alone is an improvement. All three together is a generation change.
That's why we exist. Our name points to the Neogene: the period when much of life began to look recognizably modern.
Algorithm, compute, cooling. Only all three together make a generation change. We ship all three.
Drop energy by two orders of magnitude and the question changes — intelligence lives wherever power does, not only in hyperscale halls. Here's how we compare to the traditional approach across every dimension that matters.
An order of magnitude isn't optimized — it's redesigned from first principles.
INN™ accuracy on public scientific classification datasets, measured against traditional systems.
Tests run independently on public datasets. Accuracy figures represent INN™ vs. equivalent traditional classifiers on the same data.
One small machine, one wall of power, one decision that actually matters — that's what intelligence should look like.
Explainable assistive diagnosis on imaging and records — where the doctor can ask the model back: why did you decide that.
Inference that keeps pace with wet-lab throughput, compressing discovery cycles from months to weeks.
Sparse models resist overfitting in non-stationary markets and return auditable decisions with calibrated confidence.
Thermal as a first-class concern — the cabinet carries its own heat away quietly, no building-scale chiller loop required.
On-line, on-device quality inspection — CPU-only, no accelerator card per production line.
Models that run at the intersection, the substation, the service hall — not every decision round-tripped to the cloud.
Real-time recommendations and fraud detection decided in milliseconds — not every transaction round-tripped to a remote data center.
Explainable learning-path assessment — the teacher knows why the model flagged a student for extra support, not just that it did.