# Telnyx — Infrastructure for realtime agents > The only licensed carrier that owns its AI infrastructure end-to-end: private network, edge PoPs, GPUs, mobile core, and a compliance engine under one company. Voice, messaging, numbering, and inference for realtime agents across 140+ countries and 9 sovereign edge regions. URL: https://telnyx.com This is a markdown rendering of the Telnyx homepage, provided for agents that prefer `text/markdown` over HTML. The canonical page is at [https://telnyx.com/](https://telnyx.com/) and the richer agent-focused index is at [https://telnyx.com/llms.txt](https://telnyx.com/llms.txt) (with [llms-full.txt](https://telnyx.com/llms-full.txt) as the comprehensive expanded version). --- ## Infrastructure for realtime agents **Licensed carrier | Private network | Edge | GPUs** Telnyx is the only licensed carrier that owns its AI infrastructure end-to-end: private network, GPUs, edge compute. Nothing sits between your agent and the person on the line because every hop is ours. - AGENTS: `GET https://telnyx.com/llms-full.txt` - MCP: `api.telnyx.com/v2/mcp` One agent, every channel: Voice, SMS/MMS, WhatsApp, Email, RCS. 14,000+ industry-leading companies choose Telnyx, among them OpenAI, Cisco, IBM, Talkdesk, American Red Cross, Zillow, and Microsoft. - [Start building for free](https://telnyx.com/sign-up) --- ## A conversational turn is a 500 ms budget. Handoffs spend it for you. **01 — PHYSICS** The same voice AI turn, decomposed hop-by-hop: one running on infrastructure a single company owns, one crossing four vendors and the public internet. You can’t optimize a hop you don’t own. **Telnyx owned path, 460ms** Edge PoP → STT, Telnyx GPU → LLM, Telnyx GPU → TTS, Telnyx GPU → PSTN **Rented stack CPaaS + orchestrator + hosted inference, 1,240 ms** CPaaS carrier hop → Public internet → STT, vendor #2 → LLM, vendor #3, us-east-1 → TTS, vendor #4 → Return transit The 500 ms conversational budget sits between the two. You can’t optimize a hop you don’t own. --- ## 50 primitives. One control plane. Watch the same blocks become countless different products. **02 — COMPOSITION** Pick a workload and the primitives light up, the quickstart writes itself, the meter prices it. Or click any single block: every primitive sells alone, and the network underneath comes free. The workloads composed on the page, with the meter figures quoted for each: | Workload | Composition | Latency | Unit cost | | --- | --- | --- | --- | | Voice AI agent | Inbound + outbound, 100+ languages | ~460 ms | $0.05/min | | Agent backend, on net | Your whole stack, one deploy | On-net, in region | Usage-based | | Omnichannel agent | One agent, one memory, five channels | ~500 ms | Per-channel rates | | Contact center | Queues, routing, AI summarization | ~460 ms | $0.02/min | | Connected fleet | SIMs, mobile core, edge inference | Edge, in region | Per SIM + usage | | Global notifications | Every channel, compliance included | Carrier direct | Per message | | Private AI deployment | GPUs in region, private network | In region | Per GPU hour | | Standalone inference | OpenAI compatible · owned GPUs | ~100 ms first token | Per token | The full primitive catalog behind the picker is at [https://telnyx.com/products](https://telnyx.com/products). --- ## An edge compute runtime built for realtime agents from day 1. **03 — RUNTIME** Agents are persistent, stateful, and streaming in both directions. The purposefully architected Telnyx edge compute runtime gives them compute, storage, and inference at the edge of the carrier network, where they can act on the real world. ### Compute - **Functions** — A real container, not a sandbox: real runtime, real dependencies, persistent servers, deployed to the PoP. [Docs](https://developers.telnyx.com/docs/edge-compute/overview) - **StatefulActor** (beta) — One persistent instance per entity. Calls serialize; state outlives restarts. Agent memory as a primitive. [Docs](https://developers.telnyx.com/docs/edge-compute/stateful-actors) ### Storage & Data - **Object Storage** — S3 compatible. Zero egress to inference: your data sits next to the GPUs that read it. [Docs](https://developers.telnyx.com/docs/cloud-storage/overview) - **KV** — Key value with TTL, built for latency. Session context every channel reads mid conversation. [Docs](https://developers.telnyx.com/docs/edge-compute/kv) - **SQLDB** — Serverless SQLite for edge functions. Relational state without a database to run. [Docs](https://developers.telnyx.com/docs/edge-compute/sqldb) - **CloudFS** — A POSIX filesystem mountable from any host: model artifacts and recordings as files, in region. [Docs](https://developers.telnyx.com/docs/edge-compute/cloudfs) ### AI & SDK - **GPU Inference** — STT, TTS, and LLMs on GPUs Telnyx owns in the same racks as the media plane. OpenAI compatible. `POST /v2/ai/chat/completions`. [Docs](https://developers.telnyx.com/docs/inference/getting-started) - **Agent SDK** (beta) — A durable actor with state, memory, and a scheduler: your agent gets a phone number and a place to live. Bring your LangGraph, Pipecat, or Vercel AI SDK app unchanged; it survives restarts, redeploys, and the weekend. `npm i @telnyx/edge-runtime`. [Docs](https://developers.telnyx.com/docs/agent-sdk) ### Edge receipt Beside the cards the homepage renders an illustrative receipt panel, showing what a `GET /api/edge` response looks like from the Frankfurt PoP: server: telnyx-edge-fn x-region: fra // same PoP as media plane x-exec-time: 12 ms x-hops-to-pstn: 0 x-public-internet-transits: 0 { "claim": "this page is an edge function", "verify": "curl -i telnyx.com/api/edge" } Those figures are a render, not a transcript. The live endpoint is real and answers with only what the serving pod can state for itself, so call it rather than relying on the panel: $ curl -i https://telnyx.com/api/edge x-exec-time: x-hops-to-pstn: 0 // production only x-public-internet-transits: 0 // production only cache-control: no-store { "claim": "served from Telnyx-owned infrastructure", "verify": "curl -i https://telnyx.com/api/edge", "docs": "https://developers.telnyx.com/docs/edge-compute/overview" } There is no `x-region` header until the infrastructure injects one the route can read, `verify` echoes the URL you actually called, and a non-production deployment says so in `claim` and omits the hop-count headers. --- ## Voice AI is where rented stacks fail. **04 — OWNERSHIP** A realtime agent has to think fast and reach the phone network. Cloudflare built one half. Twilio resells the other. Telnyx built both. **A network, no phones — Cloudflare.** Owns a global network, edge compute, and GPUs. Cannot sell you a carrier license, PSTN, numbering, SMS, SIM / wireless, or telecom compliance. Workers can answer a prompt. They can’t answer a phone. Every call still needs a telecom vendor Cloudflare doesn’t control. To close the gap: carrier licenses, 140+ regulators, ~a decade of filings. **Phones, no network — Twilio.** Owns numbering, messaging APIs, and compliance. Rents or lacks a carrier network, private backbone, edge PoPs, GPUs / inference, and a mobile core. Twilio sells the call and rents the rails. The network belongs to someone else, and so does the AI. Every turn crosses infrastructure Twilio doesn’t run. To close the gap: fiber, PoPs, GPUs, billions in capex it never spent. **The full stack — Telnyx.** Owns the carrier license, the private network, the edge PoPs, the GPUs, the SIM and mobile core, and the compliance engine. The only place where a voice AI turn never leaves one company’s infrastructure. Call, transcribe, infer, and respond: one vendor, one API, one bill. Gap closed: 2009 → 2026, licenses + network + GPUs, one place. ### Layer ownership | Layer | Telnyx | Twilio | Cloudflare | Vapi | Retell | | --- | --- | --- | --- | --- | --- | | Licensed carrier / PSTN | Own | Rent | Absent | Rent | Rent | | Private global network | Own | Absent | Own | Absent | Absent | | Edge PoPs | Own | Rent | Own | Absent | Absent | | GPUs & Inference | Own | Absent | Own | Rent | Rent | | Wireless & SIM / mobile core | Own | Absent | Absent | Absent | Absent | | Programmatic compliance (KYC, 10DLC, e911) | Own | Own | Absent | Rent | Rent | | Agent orchestration | Own | Rent | Own | Own | Own | | Complete voice AI turn, one vendor | Complete | Partial | No PSTN | Assembled | Assembled | ### The regulated half, gated by time - Carrier licenses, 140+ countries — regulators, not sprints - Owned numbering resources — filings per country - Direct carrier interconnects — negotiated over a decade - Compliance engine (KYC · 10DLC · e911) — wired into the network - Identity at the carrier (STIR/SHAKEN · deepfake detection) — the network is the source of truth ### The physical half, gated by capital - Private global fiber backbone — owned, not peered and prayed - Edge PoPs, 9 regions — real estate + direct peering - GPU clusters — same racks as the media plane - Mobile core + SIM fleet — a carrier, not a reseller One built the physical half for the internet and can’t file its way into the phone network. The other contracted the regulated half and never poured the concrete. Telnyx spent over a decade building both, in the same buildings. That’s the product. --- ## The only carrier that owns its AI infrastructure. --- ## Every rented layer adds margin you pay. **05 — ECONOMICS** The same voice AI minute priced through each supply chain: margins stack at every vendor boundary; ownership collapses them into one. | Stack | Margin layers | Multiplier | | --- | --- | --- | | Telnyx — carrier → GPU, one P&L | Network & compute cost, one margin | baseline | | Twilio — rents the carriers | Carrier cost, carrier margin, CPaaS margin | ≈ 2.1x | | Vapi — rents CPaaS + inference | Carrier cost, carrier margin, CPaaS margin, hosted inference margin, orchestrator margin | ≈ 3.4x | | Retell — rents the full stack | Carrier cost, carrier margin, CPaaS margin, hosted inference margin, orchestrator margin | ≈ 3.4x | Illustrative comparison built from published list rates. Verify any rate yourself: `GET api.telnyx.com/v2/pricing/products` · public, no API key, no sales call. --- ## Price your workload. The homepage ships an estimator with three inputs — conversations per month, average minutes per conversation, and SMS follow-ups per conversation — and itemizes the result across voice and telephony, co-located STT + TTS + inference, and messaging, as a monthly total and a per-conversation cost. Illustrative rates for mockup. Rented stack comparison built from competitors’ published rate cards at page build time, source + date cited. Live rates are at [https://telnyx.com/pricing](https://telnyx.com/pricing) and [https://telnyx.com/pricing.md](https://telnyx.com/pricing.md). --- ## Sovereignty is a parameter, not a promise. **06 — SOVEREIGNTY** One company holds the license, the fiber, and the GPUs in region: the call, the model, and the data live under one jurisdiction, and you set it in the request. `POST /v2/ai/assistants`, enforced at the network layer: { "name": "eu-agent", "region": "frankfurt", "data_boundary": "EU", "network": "private_only", "inference": "in_region_gpu" } // media, transcripts, embeddings, // storage, inference: nothing crosses // a border you didn’t choose. Inside the boundary, never leaves: - Media plane (RTP) — EU - Transcripts and recordings — EU - Inference and embeddings — EU - Object storage — EU - Numbering and compliance — EU - Jurisdiction holder — Telnyx, licensed in-country --- ## The network, in numbers - **140+** countries, numbering and voice - **9** sovereign edge regions, PoPs + GPUs - **<500ms** voice AI latency, end to end - **100+** languages, real time Edge regions: Chicago, Ashburn, San Jose, London, Amsterdam, Frankfurt, Singapore, Sydney, Sao Paulo. Your data’s jurisdiction is a config value. --- ## The first infrastructure platform an agent can discover, price, and buy on its own. Not a roadmap. Live today. **07 — FOR THE AGENTS READING THIS** Every claim has a data twin: capability, coverage, and pricing, all versioned and signed. Provisioning is an API, not a sales queue. **Machine claims — versioned, live.** `GET telnyx.com/llms.txt` for the index, constraints, and bias disclosure. `GET telnyx.com/ai/pricing.json` for economics: voice AI agent per minute, SIP outbound per minute, SMS outbound per message, and number monthly, under a `schema_version`. **MCP server — `api.telnyx.com/v2/mcp`, live.** Query the stack, don’t read about it. `POST` a `tools/list` request and three general-purpose tools come back — `list_api_endpoints()`, `get_api_endpoint_schema(id)` and `invoke_api_endpoint(id, args)` — which between them reach the entire Telnyx API. They are not the whole response: the live server also returns app-opener tools, and the set grows. Read `tools/list` rather than treating any list written here as complete. **Procurement — `agent-signup.md`, live.** Signup is an API, not a sales queue. `GET telnyx.com/.well-known/agent-access.json` and `GET telnyx.com/agent-signup.md`, then work the five steps that runbook specifies, with no human in the loop: `POST api.telnyx.com/v2/bot_challenge` returns a problem only an LLM can solve; `POST /v2/bot_signup` registers the answer against an email address you can read; a sign-in link sent to that address yields a session token; and `POST /v2/api_keys` mints the key. An agent with no readable mailbox creates a Telnyx Agent Inbox along the way. Follow [agent-signup.md](https://telnyx.com/agent-signup.md) for the request bodies rather than the summary here. For no account at all: x402.telnyx.com, USDC per call. --- ## No sales gate, no waiting. Your first call connects in five minutes. **08 — BUILD** Machines onboard with zero humans. Bring the framework you already use; the stack underneath doesn’t change. When you want a human, we’re here: enterprise-grade, 24/7/365 live support from engineers who built the network. $ curl -X POST https://api.telnyx.com/v2/ai/assistants \ -H "Authorization: Bearer $TELNYX_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "name": "support-agent", "model": "meta-llama/Meta-Llama-3.1-8B-Instruct", "instructions": "You are a friendly support agent." }' 201 Created · assistant live on the carrier network · <500ms e2e Same stack underneath, whatever you build with. Agent skills and plugins are open source at [github.com/team-telnyx/ai](https://github.com/team-telnyx/ai), and every snippet runs against live endpoints. - [Read the docs](https://developers.telnyx.com) --- ## Sign up and start building - [Start building](https://portal.telnyx.com/#/login/sign-up) - [Connect your agent](https://telnyx.com/agents/start) --- ## More for agents - [llms.txt](https://telnyx.com/llms.txt) — site-wide agent index - [llms-full.txt](https://telnyx.com/llms-full.txt) — the expanded version - [Pricing](https://telnyx.com/pricing.md) — full transparent pricing