Chip giant Nvidia announced on Tuesday, 9 June 2026 that voice AI platform Deepgram now supports fully on-premises deployment with end-to-end encryption of audio data and model weights, powered by a joint stack combining Fortanix Confidential AI and NVIDIA Confidential Computing.

Context

The announcement addresses a long-standing concern in enterprise voice AI: sensitive audio — spanning call-centre recordings, medical transcriptions, and financial conversations — has historically required transmission to public-cloud inference endpoints, creating data-residency and compliance risks. Deepgram's new fully on-premises mode resolves this by keeping both the audio stream and the AI model weights encrypted inside hardware-protected enclaves, never exposing them to the host system or hypervisor.

Nvidia's post stated plainly: 'Voice AI calls just got a privacy upgrade.' The integration means organisations can run Deepgram's speech-to-text and voice AI workloads entirely within their own infrastructure, with cryptographic guarantees enforced at the silicon level.

Policy Backdrop

Nvidia introduced Confidential Computing support for its Hopper GPU architecture in 2022, enabling encrypted model execution without exposing data to the host operating system. The technology relies on trusted execution environments (TEEs) that isolate computation inside hardware enclaves — a design that prevents even privileged system administrators from accessing model inputs or outputs in plaintext.

Fortanix, a specialist in runtime encryption, supplies the confidential-computing platform layer that bridges Nvidia's hardware capabilities with application-level workflows. Together, the two companies provide a stack that encrypts not just data at rest or in transit but data in use — the hardest category to protect and the one most relevant to live voice inference.

This move mirrors a broader industry pattern: hardware vendors and AI application companies are jointly packaging confidential-computing stacks to help regulated industries — banking, healthcare, defence, and government — meet data-residency mandates without abandoning modern AI capabilities. Similar enclave-based approaches have previously been adopted in database and analytics sectors.

Stakeholders and Impact

The primary beneficiaries are enterprise AI teams operating under strict data-governance frameworks, including compliance officers in sectors governed by regulations such as India's Digital Personal Data Protection Act, the EU's GDPR, and US healthcare privacy rules. For these organisations, the inability to send voice data off-premises has been a hard barrier to adopting cloud-native voice AI; fully on-prem encrypted deployment removes that barrier.

Deepgram, which provides real-time transcription, call analytics, and conversational AI APIs, expands its addressable market significantly by reaching customers who previously could not use its platform. Fortanix gains a high-visibility reference deployment that validates its Confidential AI platform for inference workloads at scale.

For Indian enterprises in particular — where data localisation requirements are tightening and regulated sectors such as banking and telecom generate large volumes of voice data — the availability of a certified on-premises voice AI stack with hardware-level encryption is commercially significant.

What's Next

The Confidential Computing Consortium, an industry body that standardises TEE interoperability, is expected to see increased participation from GPU vendors and AI application companies as this model of jointly packaged confidential-computing stacks gains traction. Analysts will watch for similar announcements from other voice and large-language-model vendors seeking to address the same compliance gap.

For Nvidia, the Deepgram-Fortanix integration adds another enterprise reference case to its Confidential Computing portfolio, reinforcing the company's positioning as infrastructure for regulated AI — a segment that carries higher margins and longer sales cycles than commodity cloud GPU supply.