We buy voice capacity from regional operators in 100+ countries and sell it two ways: as raw carrier-grade minutes to enterprise buyers, or as turnkey private AI contact centers deployed on our own GPU infrastructure. Same plumbing. Two products. One operational team.
We are not an AI startup. We are not a pure wholesale broker. We are both — because in this market the two businesses reinforce each other.
The business that pays the bills. We aggregate voice capacity and DID numbers from regional telecom operators and resell them to Tier-1 carriers, CPaaS platforms, and large enterprise buyers at competitive wholesale rates.
Built on top of the same carrier-grade voice infrastructure. For enterprise clients who want more than minutes — a managed AI voice service deployed privately on our GPUs, not in anyone's public cloud.
The wholesale business is mature, high-volume, and grounded in regional carrier relationships we have spent years building. It funds the company and gives every downstream product carrier-grade voice lines.
Local minutes, DIDs, SMS termination from carriers across Africa, Middle East, South-East Asia, Latin America.
Aggregation, SIP interconnect, quality monitoring, billing, settlement. Carrier-grade routes built from the regional supply.
Tier-1 carriers, CPaaS platforms, large contact-center operators, fraud-prevention companies, DID aggregators.
Voice minutes, DID number pools, and SMS termination capacity. Primary coverage across Africa, the Middle East, South-East Asia, and Latin America — markets where local presence and regulatory relationships matter most.
Carrier-grade aggregated routes, E.164 number pools, and termination into Tier-1 networks. Buyers range from established carriers to CPaaS platforms to contact-center operators needing reliable volume at wholesale pricing.
Standardized Supply Agreements with regional suppliers and Service Agreements with enterprise buyers. Settlement exclusively in USD via bank wire. All commercial contracts governed by Delaware law.
The AI voice market is crowded. ElevenLabs Agents, Vapi, Retell, Bland, Voximplant — all strong products. All running on someone else's public cloud. Great for marketing, lead-gen, non-regulated support. Not an option for regulated industries.
Every well-known AI voice brand runs on a hyperscaler. That works beautifully for outbound sales, lead qualification, and casual customer support. It breaks the moment a hospital, bank, law firm, or government agency tries to adopt it.
Customer voice and transcripts leave the perimeter. Shared GPU pools. Vendor controls the model and the terms. Compliance requires heavy contractual work. For regulated buyers the conversation usually ends at "our data cannot leave our data center."
Same class of engines — Ultravox v0.7 for speech-to-speech, Whisper for transcription, Coqui and Piper for synthesis, Llama 3.3 70B as the reasoning backend. But they run on our NVIDIA A100 in our own data center. Nothing in the call path leaves our rack.
We started with one A100 (already enough for production latency at first-tenant volumes) and will scale to three or four cards as demand grows. For clients who want premium cloud voice quality, ElevenLabs is available as an opt-in API integration — never as a mandatory part of the stack.
Three products, transparent commercial terms. Wholesale is ready to trade today. AI is in MVP — shipping to first private-deployment clients this quarter.
All invoices issued by Goippro LLC (Delaware, USA). Card payments through Stripe for pilots and prepaid balances. Wire transfers in USD for production contracts.
Different regulations, different industries, different comfort with the cloud. We engineer all three paths with the same operational playbook.
Fully-managed voice AI on shared cloud infrastructure. Same class of engines the market already uses. Go live in days.
Voice inference runs on our A100 in our data center. Whisper, Ultravox, Coqui, Piper — all on hardware we own, audit, and can show you on a site visit.
Sensitive turns (identity, payment, medical) stay on our private GPU. Routine turns (FAQ, directions, hours) run on cost-efficient cloud inference.
Our wholesale side runs on FreeSWITCH, SIP interconnects with Telnyx and Twilio, and direct partnerships with regional carriers. Our AI side runs on top of that same telephony, adding Ultravox, Whisper, and GPU inference. The full picture:
Sub-500 ms speech-to-speech. End-to-end voice model running on our A100. We pair Ultravox with Llama 3.3 70B locally, so the reasoning backend is also in our infrastructure — full data sovereignty.
OpenAI's Whisper large-v3 deployed locally. 70+ languages with automatic detection. Used for real-time transcription, call analytics, post-call summaries, RAG retrieval — all without transcripts leaving the rack.
Coqui XTTS v2 and Piper run directly on our inference node. Natural barge-in: agents can be interrupted mid-sentence and recover gracefully. ElevenLabs is available as opt-in cloud API for clients who want premium voice.
Carrier-grade SIP infrastructure. Redundant interconnects via Telnyx (primary Tier-1) and Twilio (failover). Plus direct trunks to our regional wholesale suppliers. PSTN numbers in 100+ countries, full E.164.
Open-source CPaaS (jambonz) for the SIP/RTC layer. Our custom decision-gate handles handoff routing, session resume, multi-tenant isolation. Everything we run, we can modify — no black boxes.
Client knowledge bases ingested into Qdrant on our infrastructure. Retrieval-augmented generation runs entirely on-premise. Post-call analytics, sentiment, topic clustering — all private, all auditable.
Every serious voice deployment in 2026 is a combination of carrier, AI engine, orchestration, and contact-center front end. Here is how the market actually lays out, and what we compose for clients.
| Provider | Role in our stack | Latency | Voice quality | Cost profile | When we use it |
|---|---|---|---|---|---|
| Ultravox self-hosted | Speech-to-speech core on our A100 | ★★★★★ <500 ms | ★★★★ | GPU capex + ops | On-prem and hybrid deployments. Client data never leaves the rack. |
| Whisper self-hosted | Speech-to-text, 70+ languages | ★★★★★ | — | GPU share | Always, for analytics, transcripts, RAG, post-call summaries. |
| Coqui XTTS / Piper self-hosted | Open-source text-to-speech | ★★★★ | ★★★★ | GPU share | Default TTS for private deployments. |
| ElevenLabs cloud API only | Premium cloud TTS, conversational agents API | ★★★★ ~1 s cascade | ★★★★★ | $0.08–0.10/min | Opt-in for clients who want premium voice and can accept cloud. Production elsewhere at Revolut, Klarna, Deliveroo, Deutsche Telekom. |
| Telnyx primary carrier | PSTN/SIP Tier-1 — our main voice wholesale path | ★★★★ | — | Low wholesale | Primary telephony both for wholesale resale and AI call termination. Tier-1 global DIDs. |
| Twilio failover carrier | Redundant PSTN/SIP, Flex UI | ★★★★ | — | 20–30% above Telnyx | Automatic failover when Telnyx degrades. Flex for clients already on Twilio's platform. |
| Voximplant Kit optional | Turnkey cloud contact-center | ★★★ | — | SaaS per-seat | For fully cloud-grade turnkey setups without data-sovereignty requirements. Native Ultravox integration. |
| jambonz self-hosted | Open-source CPaaS, SIP/RTC | ★★★★★ | — | OSS + ops | Our orchestration layer. Hosts the Ultravox app, manages transfers, bridges to human operators. |
| Vapi / n8n as needed | Workflow orchestration and rapid prototyping | ★★★★ | — | Low | Multi-agent orchestration, proof-of-concept builds, CRM and ticketing integrations. |
Model-agnostic and carrier-redundant by design. The only non-negotiable: on-premise deployments run core inference on hardware we own.
We are not an Ultravox reseller. We are engineers who have read every page of their documentation and every GitHub issue. We chose it for two reasons: sub-500ms speech-to-speech latency, and the ability to swap the underlying LLM for a model we fully control.
Traditional voice AI cascades three separate models: STT, LLM, TTS. Every stage adds latency, and the conversation feels robotic. Ultravox v0.7 (GLM 4.6 base, late December 2025) compresses this into a single end-to-end model — the agent perceives tone, interrupts gracefully, responds in under half a second.
Ultravox's strength is the speech layer, but the reasoning backend is swappable. We re-bind the audio encoder to Llama 3.3 70B on our own A100 — an open-weight, industry-grade model we own, with no per-token cloud billing and no third-party exposure of conversation content.
Whisper large-v3 runs locally on the same A100 — 70+ languages, automatic detection, real-time factor under 0.3×. Speech synthesis via Coqui XTTS v2 and Piper on our own hardware. ElevenLabs is available as an optional API integration for clients who want its premium voice, but it is never part of our self-hosted path.
Ultravox has real limitations every serious engineer reads about first: documentation gaps, memory accumulation under long sessions, inconsistent tool-calling, hardware-specific audio issues. We addressed each: session windowing, GPU memory watchdogs, tool-call schema validation, pinned kernel-level audio pipeline. The result runs unattended in production.
Our wholesale business is already live. Tell us what you have and what you need, and we will come back within one business day with a routing plan and an offer.
This form takes 60 seconds. Every field helps us match you with the right route or buyer faster.
The hardest problem in voice AI is not the AI — it is the graceful handover when the AI reaches its limits. We built our handover architecture from the principles published in Ultravox's seamless-hybrid design and refined it in live production.
80%+ of routine inquiries: order status, appointment booking, FAQ, lead qualification, payment reminders.
Operator receives conversation ID, full transcript, reason for handoff, concise summary. Customer never repeats themselves.
Because the business has two products, it has two sets of customers. Both are enterprise-grade B2B. Both sign USD contracts under Delaware law.
The engine class we deploy is already in production at the world's most demanding voice operations. Our differentiator: we install the same technology privately, without any of this data leaving your infrastructure.
Whether you need a million voice minutes a month or a private AI agent for a regulated industry — talk to us. One operational team, one contract, one compliance posture.
sales@goippro.com →