How Deepgram optimized voice AI at scale with AWS Trainium
Deepgram's Nova-3 on AWS Trainium2 delivers 11–70% more audio per dollar for batch workloads
“On Deepgram's offline, throughput-oriented batch workloads, that setup processed 11–70 percent more audio per dollar than the compute options Deepgram had been running.”
Deepgram ported its Nova-3 speech-to-text model to AWS Trainium2 using PyTorch and custom NKI kernels, achieving 11–70% better cost efficiency for offline batch workloads like contact-center analytics and medical transcription. The optimization runs eight model replicas on a single trn2.3xlarge instance and extends to customers self-hosting via Amazon SageMaker AI. While technically substantive, this is primarily a vendor co-marketing case study rather than a novel research or industry-shaping announcement.