Workload mix drives Gladia margins

The key margin question for Gladia is not how many audio hours it processes, but which mix of hours it processes and how efficiently it can serve them. A batch transcription job can be queued and packed onto compute more flexibly than a live stream that must stay responsive in real time, and optional layers like summarization, translation, and redaction add more work on top. That means discounted volume can still be attractive, but only when the delivery cost per hour falls faster than the selling price per hour.1 2 3 4 5

• Gladia prices async and live work differently because they stress the system differently. Current public pricing shows async from €0.16 to €0.27 per hour depending on model, versus €0.68 for live transcription, and Gladia describes live transcription as a low latency streaming service with sub second turnaround targets.1 2 4
• Add-ons and model choice change unit economics. Gladia sells extra analysis per audio hour, while its own product material separates Solaria-3 for async business audio and Solaria-1 for broad language coverage and streaming. That signals different compute profiles by workload, not one uniform cost base across all transcription hours.1 2 4 5
• This is a common pattern in infrastructure APIs. AssemblyAI also ties revenue to audio hours, streaming sessions, and add-ons, while noting that speech providers keep more only when usage fees stay ahead of inference and delivery cost. In adjacent video infrastructure, Mux and Wistia describe the same basic pressure from compute, storage, and bandwidth heavy workloads.3 6 7

Going forward, the winners in speech infrastructure will be the ones that turn accuracy gains and workflow features into better gross margin, not just more traffic. As transcription pricing keeps falling, Gladia's upside comes from steering customers toward higher value workflows, while using its own models and infrastructure choices to keep the cost of each additional hour under tighter control.1 3 6 7