Deepgram's cost floor and Mistral's price floor landed within 24 hours
AWS's engineers built Deepgram's cost floor, then open-sourced the tool that gives its rivals the same floor, while Mistral published a price list that undercuts Deepgram. Amazon wins either way; Deepgram's investors wait.
Vincent Jiang · 2 min read
Deepgram transcribes speech for more than 200,000 developers and 1,400 organizations, processing quadrillions of audio tokens a year 1. This week, two documents reset what a minute of that speech costs from both directions, and the two have not been read together until now.
The cushion Amazon just built
On October 19, AWS published a case study of Deepgram's Nova-3 speech-to-text model ported onto Trainium2, Amazon's−0.72% — Amazon, down 0.72 percent today in-house AI chip 1. Eight model replicas run on a single trn2.3xlarge instance, and on Deepgram's own offline batch workloads the setup processed 11 to 70 percent more audio per dollar than the compute Deepgram had been running 1. Deepgram delivered Nova-3 as standard PyTorch and ported it over several months together with AWS's Forward Deployed Engineering team 1.
One day later, at the PyTorch Conference, AWS open-sourced TorchNeuron, the tool that made the port possible, now a one-line change for any PyTorch app, with more than 30 early customers including Deepgram's voice rival Cartesia 3. Colin Brace, VP of Technology Software at AWS Annapurna Labs, the unit that makes Amazon's custom chips, told TechCrunch to look at Deepgram, "processing 70% more audio per dollar" 2. The reference customer is also the sales pitch. Note the range is company-reported, never independently reproduced 2.
The price floor arrives from Paris
On October 20, Mistral launched Voxtral 3 and named Deepgram as a vendor it undercuts: batch transcription at $0.0035 a minute on a 4-billion-parameter model, speech generation at $13 per million characters, a full voice agent API at $0.04 a minute 4. That is eleven times the transcript price; the margin in voice AI is migrating up the stack 5.
The voice agent API costs eleven times what the transcript costs
Data
| Value | |
|---|---|
| Batch transcribe | $0 |
| Flagship transcribe | $0 |
| Realtime transcribe | $0.01 |
| Agent (Small 4) | $0.04 |
| Agent (larger model) | $0.08 |
Whose money moves
Deepgram is private, so the exposure runs through Amazon and the chip race. TorchNeuron attacks Nvidia's−0.54% — Nvidia, down 0.54 percent today CUDA lock-in, where AWS says roughly 85% of AI apps live 2. Amazon wins either way: the case study sells Trn2 capacity, and the tool makes that capacity a one-liner for Deepgram's competitors too 5. The cost floor Deepgram earned is now a cushion against a price cut Mistral$14B — Mistral, private, latest valuation $14B has already made, not a moat. Watch the first independent benchmark of Voxtral's accuracy claims; the price list is real, the quality comparisons are Mistral's own 5.
Get Model by email, free
The AI labs, their models, and the business behind them. In our daily email, every weekday.
