The collaboration claims a 3.6× reduction in infrastructure cost and higher throughputs with faster response times compared to leading GPU alternatives for enterprise-scale voice AI deployments.

Lightning V2 is rebuilt for Tenstorrent’s architecture, with over 95% of the model in reduced-precision arithmetic (LoFi) and more than 80% in BlockFloat8, achieving zero audible degradation at production quality.

The system supports 550 simultaneous voice calls with zero idle time between requests, underscoring its suitability for high-volume, on-prem or regulated deployments that must stay within data-security requirements.

Tenstorrent’s NoC architecture and on-chip data movement between cores, coupled with large SRAM and avoidance of DRAM round-trips, enable faster real-time inference and improved memory efficiency, addressing bottlenecks in single-stream TTS deployments.

Lightning V3 is in development and expected to surpass Lightning V2, but Lightning V3 is not part of this release; Lightning V2 is available now with usage-based pricing and no upfront commitments.

The partnership focuses on enabling on-prem, private-cloud deployments for regulated industries and growth-stage companies, reducing economic barriers to large-scale voice AI adoption.

Smallest.ai and Tenstorrent collaborate to run Lightning V2, its real-time TTS model, on Tenstorrent hardware, marking the first production-grade TTS system to achieve production-quality audio under aggressive low-precision compute without quality loss.

Smallest.ai notes the collaboration signals a structural shift in inference economics for voice AI, making high-quality voice capabilities economically feasible at scale across industries.

Compared to NVIDIA L40S GPUs, 11 L40S GPUs cost about $100,000, while Tenstorrent’s P100 accelerators cost around $27,000 for the same capacity, demonstrating a 3.6× lower total cost.