Voice AI that doesn't need the cloud.

Nemo-RT Pro is a bilingual (Spanish/English) voice AI platform that runs fully on-prem on your own NVIDIA hardware. Sub-second TTFA. Zero per-minute cloud fees. Full data residency.

Built by INFINITO CLOUD (US LLC). In production at a LATAM SIP trunk operator serving 200+ end-users across multiple tenants on a single NVIDIA DGX.

Book a 20-min discovery call See pricing

Why Nemo-RT Pro

Voice AI that works like software, not metered water. Benchmarked and tuned for production workloads — comparable to OpenAI Realtime, Gemini Live, Vapi, and Retell AI.

Zero per-minute fees

One deployment, one monthly support fee. 5,000 minutes or 500,000 — your cost doesn't change.

Full data residency

Runs on your NVIDIA hardware in your datacenter. Patient PHI and customer data never cross US clouds.

Sub-second latency

Sub-second TTFA at single-user benchmarks. Comparable to the best cloud voice APIs — without the cost curve.

Multi-tenant by default

One NVIDIA DGX serves multiple clients with isolated prompts, RAG, and MCP tools per tenant.

How it works

SIP call in. Voice AI agent that actually does things out. All on your hardware.

1
SIP / Asterisk

Your Asterisk PBX (or compatible SIP trunk) bridges the call via ARI WebSocket to Nemo-RT Pro.

2
VAD + STT

Silero VAD + NeMo Conformer CTC bilingual EN-ES. Acoustic and lexical language detection per turn.

3
LLM + MCP

Qwen3-14B-NVFP4 on TensorRT-LLM with full MCP tool-calling framework and per-tenant RAG.

4
TTS back out

NeMo FastPitch + HiFiGAN streaming audio back to the caller. 174 Latin Spanish voices plus English.

No data leaves your datacenter. No cloud API fees. Full stack ownership — ASR, LLM, TTS, SIP integration, and MCP tool calling delivered as one coherent product.

Built for production

Benchmarks-driven engineering. Every component chosen and tuned for real traffic.

<1s

TTFA at single-user benchmarks

20+

Concurrent conversations tested, zero errors

200+

End-users at 10:1 overbooking on one DGX

39%

TTFA improvement via TTS batch sizing

Model stack

  • LLM: Qwen3-14B-NVFP4 via TensorRT-LLM (4-bit quantized)
  • STT: NeMo Conformer CTC bilingual EN-ES
  • TTS: NeMo FastPitch + HiFiGAN (174 Latin Spanish + 10 English voices)
  • VAD: Silero with configurable thresholds
  • MCP framework: 5 tool-call parsing formats, multi-turn sequences, RAG per tenant

Hardware and deployment

  • Recommended: NVIDIA DGX Spark (GB10 Blackwell, 128GB unified memory)
  • Compatible: any NVIDIA GPU with 16GB+ VRAM (RTX 3090/4090, A100, H100, Jetson AGX Orin)
  • Install: Docker Compose, single-command script, Caddy auto-HTTPS
  • Stack: FastAPI + WebSocket, Python 3.11+, Node.js ARI bridge
  • Go-live: 4-week target from contract signature

Nemo-RT Pro vs cloud voice AI

Same conversational quality. Fundamentally different economics and compliance story.

Dimension Nemo-RT Pro OpenAI Realtime Vapi Retell AI
Deployment On-prem, your hardware Cloud only Cloud only Cloud only
Pricing model Deploy + fixed monthly ~$0.30/min $0.20 - $0.33/min ~$0.11/min
Data residency Your datacenter US-hosted US-hosted US-hosted
Multi-tenant native Yes (per-tenant RAG + MCP) No (build yourself) Limited Limited
LATAM Spanish quality 174 Latin voices Generic multilingual Castilian default Castilian default
Year 1 at 20K min/mo $28K early-access ~$72K $48K - $80K ~$30K

At sub-15K min/mo, pure-cloud providers may be cheaper. Nemo-RT Pro economics shine at 20K+ min/mo, multi-tenant operators, regulated verticals, and Spanish-first use cases.

In production at a LATAM SIP trunk operator

Regional telecom operator running automated medical appointment booking for a growing clinic network, 24/7, in native LATAM Spanish.

  • 5 clinic tenants on a single NVIDIA DGX, multi-tenant from day one
  • 200+ end-users serviced with 10:1 overbooking, zero errors in stress tests
  • Custom MCPs for medical scheduling API and civil registry lookup by national ID
  • Per-tenant RAG documents, system prompts, and API keys
  • Zero OpenAI API fees — they eliminated that P&L line item entirely
  • 100% data residency — patient PHI never leaves their datacenter
Read the full case study

"We evaluated every major voice AI platform and none of them fit: on-premise deployment, native LATAM Spanish, multi-tenant for our clinic customers, and no per-minute fees. Nemo-RT Pro was the only option that checked all those boxes."

— CTO, LATAM regional telecom operator (anonymized)
Tenant types running today
Basic system-prompt assistants RAG-powered medical appointment booking Custom call-transfer MCPs

Open-core by design

Start free on the community edition. Graduate to Pro when you need multi-tenant, support, and commercial licensing.

Nemo-RT Community

MIT license, free forever.

  • Same ASR, LLM, TTS, and MCP stack as Pro
  • Single tenant self-host
  • Any NVIDIA GPU with 16GB+ VRAM
  • Community GitHub support
View on GitHub
Most popular

Nemo-RT Pro

Commercial license + support + SLA.

  • Everything in Community, plus:
  • Multi-tenant architecture (unlimited tenants)
  • Admin panel with per-tenant config
  • Transfer Call MCP included, custom MCPs available
  • Monthly updates + engineering support
  • 4-week go-live target
See Pro pricing

Built for these operators

Where the economics and compliance story of on-prem voice AI lands hardest.

Medical appointment booking

Clinics and health networks automating 24/7 scheduling in native LATAM Spanish with RAG and scheduling MCPs.

Call center automation

Contact centers deflecting routine calls from human agents, with call-transfer MCPs to escalate when needed.

AI-powered SIP trunks

SIP operators reselling multi-tenant voice AI to their client base without per-minute margin pressure.

Government voice services

Municipalities and public entities needing Spanish-native voice access with strict data residency requirements.

Transparent pricing. No per-minute fees.

You pay the same whether you run 5,000 minutes or 500,000 per month. 4-week go-live target from contract signature.

Early-access

Pro — Early-access

First 3-5 clients this quarter.

$10,000 deployment
+ $1,500 / month support

Year 1 total: $28,000

  • Multi-tenant, unlimited tenants
  • Full Spanish voice library (174 voices)
  • Qwen3-14B-NVFP4 on TensorRT-LLM
  • RAG + MCP tool calling framework
  • Transfer Call MCP included
  • 4-week go-live target
  • 20h / month engineering support

In exchange for reference case study rights + roadmap input.

Book discovery call

Pro — Standard

Q3 2026 onwards.

$12,000 deployment
+ $2,000 / month support

Year 1 total: $36,000 · Year 2+: $24,000/year

  • Multi-tenant, unlimited tenants
  • Full Spanish voice library (174 voices)
  • Qwen3-14B-NVFP4 on TensorRT-LLM
  • RAG + MCP tool calling framework
  • Transfer Call MCP included
  • 4-week go-live target
  • 20h / month engineering support

White-label, SLA 99.5%, and custom MCP bundles available in Enterprise tier.

Contact sales

All prices in USD. LATAM companies can invoice in local currency via our US LLC. Hardware (NVIDIA DGX Spark or compatible GPU) procured directly from NVIDIA or your reseller — we don't mark up hardware.

Credentials and open-source track record

Founder credentials and defendible production signal across 10+ countries.

36

GitHub stars on asterisk_to_openai_rt

68

Forks (+ 23★ / 27⑂ on community variant)

10+

Countries with production deployments

NVIDIA-Certified Associate
AI Infrastructure & Operations (2026-2028)
AWS Partner
Generative AI Essentials
Google Cloud
Speech API certified
Aviatrix ACE
Multicloud Network Associate
Y Combinator
Startup School 2020 alumni
NVIDIA Inception
Member portfolio company
View asterisk_to_openai_rt View Nemo-RT Community

Frequently asked questions

Any NVIDIA GPU with 16GB+ VRAM. Recommended: NVIDIA DGX Spark (GB10 Blackwell, 128GB unified memory). Compatible: RTX 3090/4090, A100, H100, Jetson AGX Orin. We don't sell hardware — you procure from NVIDIA or your preferred reseller. We provide the software, deployment, and support.

4-week go-live target from contract signature. Core install is typically 2-4 hours on a ready DGX. The remaining time covers tenant setup, custom MCP development if needed, pilot testing with real traffic, and team training.

Yes. Nemo-RT Pro ships with a pre-built Node.js ARI bridge. Works with Asterisk 16+, FreePBX, and 3CX via SIP trunk. Other PBX systems are possible via SIP-to-WebSocket bridge. Custom integration is available as a services engagement.

20 hours per month of engineering time (Pro tier), monthly security patches, at least one feature release per month, access to new MCPs as released, LLM model updates, bug fixes (priority for production issues), monthly health reports, and email/Slack incident support.

Your deployment keeps running. You own the hardware and the software license stays installed. What stops: monthly patches, new features, engineering support, access to new MCPs. What keeps working: everything installed today runs indefinitely. Very different from cloud where "stop paying" = "service off."

Yes. Annual prepay: 10% off. 2-year prepay: 15%. 3-year prepay: 20%. Partner program pays 33% margin on sourced deals. For Enterprise (white-label + SLA 99.5%) or large multi-region deployments, we build a custom quote during the discovery call.

Ready to see if Nemo-RT Pro fits your operation?

20-minute discovery call. No slides. We look at your volume, your compliance needs, your hardware, and tell you honestly whether this is a fit — or what would be better.

Book a 20-min discovery call Contact us

Questions? Email yan.frank@infinitocloud.com