What Telnyx sells — and what it buys
AI minutes. Complete voice-agent conversations at $0.08/min all-in — STT, LLM inference and TTS included — plus tokens: OpenAI-compatible APIs on open-weight models (Kimi K2.6, GLM-5.x, MiniMax, Qwen3-235B) from $0.21/1M, pay-as-you-go, no minimums, fine-tuning on the same key. Underneath sits the original carrier business: SIP trunking, numbers in 140+ countries, messaging, wireless.
GPUs outright — a 4,000+ fleet racked inside its own points of presence — plus carrier interconnects, numbering resources, and colocation and power at PoP sites on a private 400Gbps backbone. Notably absent: third-party cloud capacity. The call and the inference never leave Telnyx infrastructure.
Scale proof
The bet
That for real-time AI — voice above all — the decisive variable is distance, and the shortest path runs through a carrier's own facilities, not a cloud region. The edge that matters is the carrier edge: own the call path, the identity attestation and the GPUs in one operational domain, and you sell what a stitched stack structurally cannot — a sub-200ms turn with one SLA and one bill. The corollary: a fleet bought from cash flow, with no reseller markup and no venture clock, can hold $0.08/min and $0.21/1M as structural prices, not promotional ones.
Natural counterparty
Voice-AI and agent builders — contact centers, healthcare documentation, logistics — whose latency budget dies on inter-vendor hops; and regulated enterprises that need inference and communications inside one compliance boundary, in-region by architecture. On the buy side: GPU vendors and colocation and power providers at its PoP sites. Telnyx buys no capacity on the GPU-cloud market — which removes it from the deal flow that connects most of this ladder.
Buyer fit
You're building real-time voice AI and want the call, the models and the compliance boundary on one network — priced per minute, with no infrastructure to stitch.
You need proprietary frontier models, training infrastructure, or raw GPU capacity to operate yourself — Telnyx sells the output, never the hardware.
Five fields we track but don’t publish
Everything above is public — figures are Telnyx-published unless otherwise sourced, and this card is a draft pending direct provider verification. The fields below change weekly and are verified directly with each provider — they live in a private supply-demand ledger, not on this page.