Deepgram moved its Nova‑3 speech‑to‑text model to AWS Trainium2 chips. The port runs eight model replicas on a single trn2.3xlarge instance. In offline batch jobs like contact‑center analytics, Deepgram measured 11‑70 % more audio processed per dollar compared with its previous compute options. The work involved using TorchNeuron for PyTorch, swapping layers with custom kernels, and building a Rust runtime to avoid Python overhead.
Why it matters
Customers using Deepgram’s API can expect lower transcription costs for large‑scale batch workloads.
