Voice AI's Hidden Margin Problem

Voice AI companies stack STT, LLM, and TTS costs per call. At scale, the per-minute COGS compresses margins faster than any other AI modality. Here is the math.

William Min

Technical Product Manager · Lovie

Published June 16, 2026

Creator of TokenOps. Technical PM at Lovie. 12+ years building payment infrastructure and fintech products across Asia and the US.

The Stack That Eats Margin

A text-based AI product has one inference cost: the LLM call. A voice AI product has three:

1. Speech-to-text (STT) — converting the caller's audio to text.
2. LLM reasoning — processing the transcript and generating a response.
3. Text-to-speech (TTS) — converting the response back to audio.

Each layer is a separate vendor bill. Each layer scales with call volume. And unlike text, voice has no batching opportunity — every second of audio must be processed in real time.

The Per-Minute Math

Consider a typical enterprise voice AI call:

A 5-minute call costs $0.15 to $0.30 in raw inference. A 15-minute call costs $0.44 to $0.90.

Now scale that to a company processing 3.5 million calls per week. At an average of 7 minutes per call and a blended $0.04 per minute:

3,500,000 × 7 × $0.04 = $980,000 per week.
That is $4.2 million per month in inference COGS before hosting, telephony, or support.

Why Voice Is Different From Text

Text-based AI products can cache, batch, and defer. A document summarization request can queue during off-peak hours. A RAG pipeline can cache embeddings.

Voice cannot defer. A caller on the phone expects a response in under 800 milliseconds. That latency requirement means:

• No batching. Every request is real-time.
• No caching of the primary inference path. The conversation is unique.
• Higher-tier models. Voice requires models with strong reasoning and low latency, which are the most expensive.

The result is that voice AI gross margins compress faster than any other AI modality as volume scales.

The Companies Facing This Now

Several voice AI companies have raised significant capital in recent months:

• Vapi closed a $50 million Series B in May 2026. They orchestrate STT, LLM, and TTS per call and won the Amazon Ring integration. Their COGS stacks per minute.
• Bland AI closed a $50 million Series C on June 16, 2026 — the freshest trigger in the space. They process 3.5 million calls per week.
• Synthflow raised $20 million from Accel for no-code enterprise voice agents.
• Retell AI hit approximately $50 million ARR on 30 people with only $5 million raised.
• PolyAI reached approximately $40 million ARR with a Series D focused on LLM and speech COGS.

These companies share a structural reality: their gross margin is the spread between what they charge per call and what they pay per minute of inference. That spread shrinks as they add enterprise customers with longer, more complex calls.

The Attribution Gap

Here is the problem that mirrors every other AI company, but amplified: most voice AI companies cannot tell you which customers are profitable.

They know their total vendor bill. They know their total revenue. But without per-customer, per-call cost attribution, they cannot see:

• Which customers generate the longest calls and burn the most tokens.
• Which use cases (customer support vs. sales vs. scheduling) have different margin profiles.
• Whether a pricing change would help or hurt at the customer level.

What Needs to Happen

Every voice AI call should carry:

1. Customer ID — who initiated or received this call.
2. Duration and components — STT minutes, LLM tokens, TTS minutes.
3. Dollar cost per component — computed from each vendor's pricing.
4. Use case or workflow — what the call was for.

With these data points, a voice AI company can answer the question that its board will eventually ask: "Which customers are we making money on, and which are we subsidizing?"

TokenOps handles this attribution. But the pattern — tag every request, compute every cost, join to revenue — applies to every AI company, regardless of modality. Voice just makes the math more visible because the costs are higher and harder to hide.

Sources: STT/TTS/LLM pricing from Deepgram, ElevenLabs, Cartesia, Anthropic, and OpenAI public rate cards. Funding data from TechCrunch, Crunchbase, and company press releases (Vapi Series B May 2026, Bland AI Series C Jun 2026).