Self-Hosted Voice AI vs Pay-Per-Minute Platforms: Real Costs

Pay-per-minute voice AI platforms cost roughly $0.07 to $0.15 per minute once you add up the real components, so a business handling 20,000 minutes a month pays between $1,400 and $2,700. A self-hosted stack on LiveKit and Pipecat costs about $1,600 a month in infrastructure at that volume, plus engineering time. The crossover point, counted honestly, sits around 15,000 to 25,000 minutes a month. Below that, pay per minute. Above it, owning the stack starts to win.

This post breaks down what each option actually costs at real volumes, what the pricing calculators leave out, and when self-hosting is the wrong choice no matter what the spreadsheet says.

What you actually buy when you pay per minute

Every managed voice platform charges for the same four things: transcription, the language model, text-to-speech, and telephony. The differences are in how they bundle and mark them up.

Vapi charges a $0.05 per minute hosting fee and passes model costs through at cost. On their own pricing page calculator, a 1,000-minute month with Deepgram transcription, an OpenAI model, and ElevenLabs voices comes to $82 to $129. Telephony is extra if you bring Twilio: $0.008 per minute inbound, $0.014 outbound.

Retell AI advertises $0.07 to $0.31 per minute pay-as-you-go, with 20 concurrent calls included. The component breakdown shows their voice infrastructure at $0.055 per minute, standard TTS voices at $0.015 (ElevenLabs voices cost $0.040), and recommended LLMs at $0.064 per minute. A mid-range configuration lands near $0.134 per minute.

LiveKit Cloud sits lower. Their published worked example totals $0.0672 per minute: $0.01 agent session, $0.01 US telephony, $0.01 observability, a Gemma-class LLM at $0.0014, Deepgram Nova-3 transcription at $0.0058, and Cartesia TTS at $0.03. The free tier includes 1,000 agent session minutes a month.

Deepgram sells a bundled Voice Agent API at $0.075 per minute, or $0.059 if you bring your own LLM. Standalone Nova-3 transcription starts at $0.0043 per minute.

Three things follow from these numbers:

• Model choice swings the bill more than platform choice. The gap between a cheap LLM and a premium one on LiveKit's price list is over 40x per minute. Retell's $0.07 to $0.31 range is mostly model selection.
• The headline rate is never the whole rate. Telephony, recording, redaction, and concurrency limits all add cost or constraints.
• Managed platforms are priced for convenience, not scale. That is fine. Convenience is worth paying for, up to a volume.

The self-hosted stack, piece by piece

A self-hosted voice agent has the same four parts, but you run them yourself. The mature open-source options for each:

• Orchestration: Pipecat. An open-source Python framework, maintained by Daily, that wires STT, LLM, and TTS into streaming pipelines over WebRTC or WebSockets. LiveKit's own Agents framework is the main alternative. Both are production-grade.
• Media transport: LiveKit server. The WebRTC SFU that carries audio between the phone network and your agent. It is open source, and LiveKit publishes self-hosting documentation alongside its cloud offering.
• Speech-to-text: faster-whisper. A CTranslate2 reimplementation of Whisper that runs up to 4x faster than the original at the same accuracy. On a mid-range GPU it transcribes 13 minutes of audio in about a minute, so one card handles many concurrent calls.
• LLM: an open-weight model on vLLM or Ollama. A quantized 8B to 32B model (Llama, Qwen, Mistral, Gemma) handles scripted business conversations well. You can also keep the LLM as an API call and self-host only the audio path; that hybrid is common.
• TTS: Piper or Kokoro locally, or an API. Local voices are free but sound flatter than ElevenLabs. Many self-hosted stacks keep a cloud TTS for customer-facing calls and go fully local for internal lines.
• Telephony: Twilio SIP trunking at $0.004 per minute, or a local carrier in your country over SIP.

The hardware bill is smaller than people assume. Faster-whisper's large model needs under 3GB of VRAM in int8. An 8B LLM quantized to 4-bit fits in roughly 6GB. A single A100 40GB at $1.99 per hour runs the whole stack with room for ten or more concurrent calls. That GPU, on 24/7, is about $1,433 a month.

The worked example: 20,000 minutes a month

Take a business handling 20,000 voice minutes a month. That is roughly 2,000 ten-minute calls: a clinic appointment line, a SACCO member support line, a logistics dispatch desk.

Managed platforms:

• Retell, mid-range config: $0.055 infrastructure + $0.015 TTS + $0.064 LLM = $0.134 per minute. Monthly: $2,680.
• Vapi, typical config: $0.05 hosting + $0.0099 Deepgram + ~$0.02 LLM + $0.0238 ElevenLabs + $0.008 Twilio inbound = ~$0.112 per minute. Monthly: ~$2,240.
• LiveKit Cloud, their published example: $0.0672 per minute. Monthly: $1,344.

Self-hosted:

• GPU: one A100 40GB, 24/7: $1,433
• Telephony: 20,000 minutes of Twilio SIP at $0.004: $80
• App server and storage: ~$100
• Infrastructure total: ~$1,613

Infrastructure alone, self-hosting saves about $1,000 a month against Retell's mid-range config and roughly breaks even with LiveKit Cloud's cheapest example. But infrastructure is not the whole cost.

Add engineering. Budget 10 to 20 hours a month for monitoring, updates, prompt and model tuning, and the occasional 2am incident. At $75 an hour, 15 hours is $1,125. Fully loaded, the self-hosted stack costs ~$2,700 a month, which is the same as Retell at this volume.

The break-even logic:

• Infra-only comparison: self-hosting beats $0.11 to $0.134 per minute platforms at roughly 12,000 to 15,000 minutes a month.
• Fully loaded with engineering: the crossover moves to roughly 20,000 to 25,000 minutes a month.
• At 50,000 minutes a month: Retell costs ~$6,700, Vapi ~$5,600, the self-hosted stack ~$2,900 fully loaded. Saving $3,000 to $4,000 a month justifies real engineering attention.
• If you already employ engineers who can absorb the stack, the crossover drops back toward 15,000 minutes, because the marginal engineering cost is near zero.

The costs the calculators leave out

Concurrency caps. Vapi's free tier allows 4 concurrent calls; extra lines cost $10 each per month. Retell includes 20. If your peak is 30 simultaneous calls, a managed platform charges for headroom. A self-hosted stack scales with GPU capacity you already paid for.

Compliance add-ons. Vapi charges $2,000 a month for HIPAA-eligible data handling with a BAA. Retell gates HIPAA and custom data retention behind enterprise pricing. If you handle health or financial calls, that add-on alone can exceed the entire cost of a self-hosted GPU.

Recording and transcription extras. On Twilio, call recording adds $0.0025 per minute and transcription $0.05 per minute. Deepgram charges $0.002 per minute extra for PII redaction. Self-hosted, recordings live on your disk and redaction is your code.

Data residency. Pay-per-minute platforms process audio in their cloud, mostly in US regions. If you operate under Kenya's Data Protection Act or a regulator that expects customer data to stay in a jurisdiction you control, self-hosting (or a region-pinned enterprise plan, which LiveKit offers from its Scale tier) may be the deciding factor regardless of cost. We covered the broader build-versus-buy economics in The Real Cost of Replacing SaaS With an Open Source Stack, and the same logic applies here.

Engineering reality. The number nobody can put in a calculator: your team's ability to run this. If nobody on staff has deployed a Python service with a GPU, the first month of self-hosting costs far more than any platform bill.

Latency and control: what changes when you own the pipeline

Humans notice conversational delays of a few hundred milliseconds. Voice pipelines hit that budget through streaming: transcription starts on partial audio, the LLM streams tokens, and TTS starts speaking the first sentence while the rest generates.

A managed platform makes these choices for you, and mostly makes them well. Self-hosting gives you three specific controls:

• Model placement. Colocate STT, LLM, and TTS on one machine and you remove two network round trips per turn.
• Model selection per call. Route simple IVR-style calls to a 8B model and complex ones to a larger model, paying for the big model only when needed.
• No shared-tenant jitter. Your GPU serves only your calls, so tail latency is yours to fix.

The honest counterpoint: a tuned managed platform with a good LLM will beat a badly configured self-hosted stack every time. The faster-whisper benchmarks show transcription is not the bottleneck on GPU; the LLM's time to first token usually is. If you self-host, measure that first.

When self-hosting is the wrong choice

• You are under 10,000 minutes a month. LiveKit Cloud's example rate costs $672 at that volume. No GPU rental plus engineering time beats that.
• You need to launch this month. A managed platform gets a competent agent live in days. A self-hosted stack needs provisioning, telephony setup, and prompt work before the first call.
• Nobody owns it. A voice agent that takes customer calls is production infrastructure. If there is no named engineer responsible for it, pay the platform.
• Your traffic spikes hard. Seasonal peaks of 10x are what per-minute pricing is good at. Self-hosted capacity sits idle in the off-season.
• You need premium voices. ElevenLabs' best voices, from $0.0146 per minute via Vapi or 5 cents per minute on their Business plan, are hard to match with local TTS today.

If several of these apply, start managed. Migrating to self-hosted later is much easier once you know your real call patterns, and the voice AI implementation cost breakdown we published covers the managed route in detail.

Common mistakes

Comparing GPU cost to platform cost and stopping there. Engineering time is 40 percent or more of the true self-hosted cost at moderate volumes. Include it or the comparison is fiction.

Self-hosting everything on day one. The pragmatic path is hybrid: self-host transcription (cheap, easy, keeps audio local) and keep the LLM and TTS as APIs. Fully local is a later optimization.

Ignoring telephony legwork. SIP trunking, number porting, and call recording laws in your country take longer than the AI parts. In Kenya, recording consent and data localization deserve a legal read before launch, not after.

Load-testing with one call. A stack that feels instant with one caller can fall apart at five concurrent calls if the LLM queue is not tuned. Test at 2x your expected peak.

Choosing the biggest LLM. For structured business calls, an 8B to 32B open-weight model with a good prompt performs close to frontier models at a fraction of the cost per minute. Our conversational AI development guide covers prompt and flow design that matter more than model size.

A decision checklist

Work through these in order:

1. Estimate your steady monthly minutes. Under 10,000: managed. Over 25,000: self-host is worth scoping. In between: keep reading.
2. Price your actual configuration on two platforms, not the headline rate. Include telephony and your real model choice.
3. Price the self-hosted stack fully loaded: GPU at 24/7 rates, telephony at SIP rates, plus 10 to 20 hours of monthly engineering at your real cost.
4. Check concurrency at your peak hour against platform caps and your GPU's realistic capacity.
5. Check compliance needs. HIPAA, financial regulation, or data residency requirements can settle the question before cost does.
6. Name the engineer. If no one owns the stack, the decision is made for you.
7. Plan the exit either way. Keep prompts, call flows, and recordings in portable formats so switching platforms or moving to self-hosted later is a migration, not a rebuild.

Frequently asked questions

How much does a voice AI agent cost per month? At 5,000 minutes a month, expect $350 to $700 on a managed platform depending on models. At 20,000 minutes, $1,300 to $2,700. A self-hosted stack runs about $1,600 in infrastructure plus engineering time regardless of volume, which is why it wins at high volume and loses at low volume.

Is Vapi or Retell cheaper? For a typical configuration they land close together: Vapi around $0.11 per minute all-in, Retell's mid-range setup around $0.134. Vapi charges a platform fee and passes models through at cost; Retell bundles infrastructure into its rate. LiveKit Cloud undercuts both at roughly $0.067 per minute in its published example. Model selection moves the number more than platform choice.

Can I self-host Vapi or Retell? No. Both are proprietary hosted platforms. Self-hosting means building on open-source components: LiveKit or Pipecat for orchestration, faster-whisper for transcription, an open-weight LLM, and a SIP trunk for telephony. Vapi does let you bring your own API keys and telephony, which reduces cost without self-hosting.

What GPU do I need to self-host a voice agent? One datacenter GPU handles the full stack. Faster-whisper's large model needs under 3GB of VRAM, and a quantized 8B LLM needs about 6GB, so even a 24GB card has headroom for several concurrent calls. An A100 40GB at $1.99 per hour supports ten or more concurrent calls comfortably.

Does self-hosting help with data protection compliance? It gives you control, which is the prerequisite for compliance but not compliance itself. Self-hosted means call audio and transcripts never leave infrastructure you choose, which simplifies data residency under laws like Kenya's Data Protection Act. You still need retention policies, access controls, and consent handling. Managed platforms sell this too: Vapi's HIPAA package costs $2,000 a month.

Should I start self-hosted and move to a platform later? Almost never. Start managed, learn your real volumes and call patterns, then self-host when the numbers justify it. The reverse migration happens rarely and usually expensively.

Where to go from here

If your volumes justify owning the stack, the build is a known quantity: LiveKit, Pipecat, faster-whisper, an open-weight LLM, and a SIP trunk, run properly with monitoring and backups. We deploy and manage exactly this kind of infrastructure through our managed open-source stack service, and for banks, insurers, and others with strict residency or governance requirements, our enterprise AI practice handles private deployment end to end. If you are still weighing managed platforms against self-hosting for your call volumes, talk to an engineer and we will run the numbers for your case.