Tensordyne

Tensordyne is a data center AI inference hardware system for hyperscalers, neocloud operators, and AI research organizations running large-scale generative AI workloads.

Key capabilities include:

• Logarithmic math number system that targets 16-bit precision at 4-bit power efficiency, favoring additions over multiplications to increase compute density while maintaining accuracy.
• High-throughput LLM inference with published performance examples including 3,000,000 tokens/sec for Llama 3.3 70B and 1.7M+ tokens/sec per rack for Deepseek R1, plus support for 10K+ concurrent users per rack.
• Scale-out interconnect and memory density using any-to-any connectivity (up to 460 GB/s unidirectional bandwidth per chip), tensor parallelism up to TP144, unconstrained MoE experts (up to 512), and HBM3e capacity (terabytes per instance).
• Operational tooling for deployment with Kubernetes-native deployment, push-button compilation via SDK (listed as coming soon on the website), and industry-standard observability for power, thermals, latency, and throughput.

Pricing: Usage-based, with an example rate stated as low as $0.05 per 1M tokens for Deepseek R1.

Notable for its full-stack approach (math, silicon, and system design), drop-in datacenter compatibility with air-cooled racks, and hardware specifications including , , and . It was formerly known as Recogni and was developed in collaboration with .