Deepgram's playbook for invisible, always-on voice AI
Key points
Key takeaways from theCUBE interview with Nick Mann, staff technical program manager at Deepgram, recorded at the Dell Technologies AI Leadership Symposium (September 2026):
Two very different customers. Mann says he serves two sides: an internal research team of around 45 unusually strong researchers that consumes massive compute and data, and tens of thousands of external API customers with unpredictable, scaling workloads — both depend on Deepgram's Dell partnership to keep infrastructure invisible and always on.
The AI-factory moment. He compares today's shift toward AI factories — semiconductor innovation, memory-first architectures and custom silicon, such as OpenAI building its own chip — to the IoT explosion he lived through working with Tony Fadell at Nest, arguing the tooling has compressed the path from idea to production even as latency, performance and scalability stay non-negotiable.
Enterprise-first discipline. Latency, performance and scalability are the bar Deepgram optimises for; the best ideas still need skill, experience and the ability to productionise properly, and hype doesn't change that a mediocre idea poorly executed is still mediocre.
Models that travel. Deepgram runs models both in its own cloud and on customer-hosted infrastructure — a large share of customers self-host on heterogeneous systems, and with Dell it spans small-business hardware, such as a GB10 powering an IVR system, through to enterprise-scale deployments, with Dell using Deepgram's technology inside its own innovation lab.
Continuously retrained models. Deepgram treats even its best models as temporary, expecting a month-old release won't remain state of the art.; it trains for hard, noisy domains like air-traffic control and emergency medical services, and sees synthetic data expanding what it can retrain on.
Invisible by design. Infrastructure's job is to stay out of the way of researchers and engineers with good ideas — Mann quotes C T O Adam's line that the team should never be blocked by hardware or scalability, and infrastructure must simply always be on.
From rack-scale to laptops. Beyond data-centre systems, Mann wants Deepgram's models to run natively, even air-gapped, on everyday AI-enabled laptops — framing personal AI and individual data sovereignty as the next frontier for voice AI.
Read more: theCUBE · SiliconANGLE
