Deepgram targets air-gapped laptops for native voice AI model execution

Deepgram Inc. is targeting air-gapped laptops for native execution of its voice artificial intelligence models. According to a report by SiliconANGLE, the company aims to move its technology from cloud-based infrastructure to personal devices that operate without network connectivity.

Nick Mann, a staff technical program manager at Deepgram, stated that the goal is to run these models directly on hardware. This shift addresses the growing demand for voice AI as users replace traditional screen interactions with conversational interfaces, requiring new compute and storage strategies for listening and talking models.

Infrastructure and Partnership Strategy

Deepgram requires substantial computing resources for its internal research team and tens of thousands of external customers. Mann told John Furrier of theCUBE Research that researchers consume large amounts of compute and data. The discussion took place during an exclusive broadcast at the Dell AI Leadership Symposium.

Dell Technologies supplies infrastructure for Deepgram and hosts its models in the Customer Solution Center Innovation Lab. Mann noted that Dell’s customer base ranges from small businesses to large enterprises, providing broad reach for these voice AI solutions.

On-Premises and Local Deployment

A significant portion of Deepgram’s customers host models themselves rather than relying on the company’s cloud. A recent partnership with Fortanix Inc. enables private, on-premises voice AI for regulated industries. This approach supports customers who need to keep data within their own secure environments.

Deployments vary in scale, from small businesses using compact Nvidia GB10 systems for interactive voice response to larger enterprise setups. However, Mann identified AI-enabled laptops carried by students as a key target for future native model execution.

Model Development and Optimization

Deepgram has spent thirteen years tuning models for noisy settings like air traffic control and emergency medical services. The company assumes every released model will soon be surpassed and treats each version as a candidate for retraining on better data.

The Flux model, expanded to ten languages earlier this year, exemplifies this approach. Mann explained that continuous retraining allows the company to improve performance, ensuring that voice AI remains effective even in challenging acoustic environments.

What should you do?

Organizations planning to deploy voice AI on local devices should assess their hardware capabilities for native model execution. Security teams must verify that air-gapped systems maintain necessary data privacy controls without network connectivity.

What to do and how to stay safe: Deepgram

• Evaluate whether your current laptop hardware supports the computational requirements for native voice AI model execution.
• Review data privacy policies to ensure air-gapped deployments meet regulatory requirements for sensitive voice data.
• Monitor vendor announcements regarding updates to models like Deepgram’s Flux for improved local performance.
• Consult with infrastructure partners to determine if on-premises hosting solutions fit your organizational security needs.

Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What is Deepgram’s target for voice AI deployment?

Deepgram aims to run its voice AI models natively on air-gapped laptops and personal devices.

Which partners help Deepgram with infrastructure?

Dell Technologies supplies infrastructure, and Fortanix Inc. enables private on-premises voice AI for regulated industries.

How many languages does Deepgram’s Flux model support?

Deepgram expanded its Flux model to support ten languages earlier this year.