Changelogs: Week of May 25, 2026

Last week, we focused on expanding PolarGrid's West Coast footprint, validating Blackwell GPU performance, and improving the developer experience around model selection and billing transparency.

Streaming STT is live

Transcripts now arrive word-by-word as users speak. Rather than waiting for an utterance to end before processing, PolarGrid's edge nodes stream partial transcripts in real time — enabling voice AI pipelines to begin reasoning and generating responses before the speaker has finished their sentence.

This is a foundational capability for sub-300ms voice experiences. Streaming STT is available now on supported transcription models via the standard streaming API.

San Francisco is live

PolarGrid's edge inference network has expanded to San Francisco, bringing the fleet to six availability zones across North America. The fleet now spans California, New York, Texas, British Columbia, Ontario, and Quebec.

This expansion puts PolarGrid closer to West Coast developers and AI teams building in Silicon Valley. Every node offers a full model library, with intelligent routing to ensure your users receive the lowest TTFT possible. Updated SDK region configs and documentation reflect the expanded footprint.

NVIDIA Blackwell GPU benchmarks published

Inference benchmarks on NVIDIA Blackwell hardware are now published in the documentation, with server-side timing instrumentation for precise latency measurement.

This gives developers a clear view into Blackwell GPU performance on PolarGrid and supports more informed evaluation of latency-sensitive workloads.

View benchmarks →

Run multiple STT/TTS models on a single node

A single PolarGrid edge node can now serve multiple speech-to-text and text-to-speech models simultaneously. Models load and warm up automatically — no manual provisioning required.

This means you can choose the best model for your use case without needing dedicated hardware per model.

Per-model pricing now visible in the billing portal

The billing page in the management console now displays per-model pricing, updated for May 2026. You can see exactly what each model costs before running inference — no surprises at the end of the month.

This is part of our ongoing work to make PolarGrid fully self-serve for developers evaluating and scaling production workloads.