# dTelecom: Apple Silicon Infrastructure & AI-Powered STT

> How dTelecom leverages Apple Silicon's unified memory architecture to run WebRTC media servers and AI workloads on a single node — and why this changes the economics of real-time communication infrastructure.

## Overview

dTelecom is deploying **Mac Mini and Mac Studio** hardware as the physical nodes powering its decentralized real-time communication (dRTC) network. Each node runs two complementary workloads simultaneously:

- **WebRTC SFU on the CPU** — routing live voice and video traffic across the global mesh
- **AI inference on the GPU and Neural Engine** — speech-to-text, translation, voice processing, and content moderation

This is made possible by Apple Silicon's unified memory architecture, where CPU and GPU share the same memory pool but operate on independent compute units. The result: two revenue-generating workloads on a single device, with minimal resource contention and exceptional power efficiency.

dTelecom's first AI service built on this architecture is **real-time speech-to-text**, available today via the x402 micropayment protocol — no API keys, no accounts, pay-per-use with USDC.


## Why Apple Silicon

### The Dual-Workload Advantage

Traditional server hardware forces a choice: use the CPU for media routing, or add expensive discrete GPUs for AI. Apple Silicon eliminates this tradeoff.

| Resource | WebRTC SFU | AI / STT |
|---|---|---|
| CPU (Performance Cores) | **Primary** — packet forwarding, SRTP encryption, signaling | Minimal |
| GPU (Metal / MLX) | Not used | **Primary** — model inference |
| Neural Engine (16-core) | Not used | **Primary** — encoder acceleration via CoreML |
| Unified Memory | Moderate — packet buffers, connection state | Heavy — model weights, audio tensors |

The workloads are naturally complementary. The SFU saturates the CPU. The STT model saturates the GPU and Neural Engine. They coexist on the same memory bus with minimal interference because their access patterns differ: the SFU reads and writes many small packets; the AI model streams large tensor blocks.

An additional benefit of unified memory: **audio decoded by the SFU on the CPU is immediately accessible to the STT model on the GPU with zero-copy overhead**. No PCIe transfers, no DMA latency. This is uniquely efficient for the pipeline of receiving a WebRTC audio stream and transcribing it in real time.

### Hardware Lineup

| Device | CPU | GPU | Neural Engine | RAM | Max Power | Price |
|---|---|---|---|---|---|---|
| Mac Mini M4 | 10-core | 10-core | 16-core (38 TOPS) | Up to 32 GB | 65W | From $499 |
| Mac Mini M4 Pro | 14-core | 20-core | 16-core (38 TOPS) | Up to 64 GB | 140W | From $1,399 |
| Mac Studio M4 Max | 16-core | 40-core | 16-core (38 TOPS) | Up to 128 GB | 145W | From $1,999 |

For comparison, a typical 1U rack server draws 300–750W and costs $3,000–$15,000. An NVIDIA H100 GPU alone consumes 700W and costs $30,000+. A single Mac Mini M4 Pro at 140W max and $1,399 delivers both a competitive WebRTC SFU and a capable AI inference engine.

### Power Efficiency

Apple Silicon's efficiency is the foundation of dTelecom's cost model:

- Mac Mini M4 idles at **4W** — less than a night light
- Under full dual-workload load: **40–140W** depending on configuration
- A standard 5kW data center power budget can support **~35 Mac Mini M4 Pro nodes** vs. **7–10 traditional 1U servers**
- Annual electricity cost per node at 100W average: **~$130** (at $0.15/kWh)

This 5–10x power advantage directly translates to lower operating costs, enabling dTelecom's flat bandwidth pricing model.

### Data Center Deployment

Mac hardware in data centers is an established practice, not an experiment:

- **AWS** offers EC2 Mac M4 instances (mac-m4.metal, mac-m4pro.metal) — generally available since September 2025
- **MacStadium** operates tens of thousands of Macs across 5 data centers, with racks holding up to 460 Mac Minis each
- **Scaleway** offers Mac Mini M4 bare metal as a service from European data centers
- **MacWeb** deploys clusters of 50–500 Mac Minis for AI inference at scale

Colocation costs range from **$30–$200/month** per unit depending on provider and SLA.


## Node Performance

### WebRTC Capacity

Estimated per-node capacity based on LiveKit SFU benchmarks, adjusted for Apple Silicon's per-core performance advantage over the Intel Xeon Cascade Lake used in reference tests:

| Config | HD Video Meeting | Livestream (720p) | Audio-Only |
|---|---|---|---|
| Mac Mini M4 Pro (14-core) | ~150–200 participants | ~2,500 viewers | ~3,000+ |
| Mac Studio M4 Max (16-core) | ~200–250 participants | ~3,000–3,500 viewers | ~3,500+ |

In livestream scenarios, the **network interface (10GbE) becomes the bottleneck before the CPU**. 3,000 viewers at 720p consume approximately 4.5 Gbps outbound. The 10GbE option on Mac Mini M4 Pro and Mac Studio is essential for production deployments.

### STT Throughput

Speech-to-text performance on Apple Silicon, running alongside the WebRTC SFU:

| Model | Implementation | M4 Pro Performance | Use Case |
|---|---|---|---|
| Whisper Large v3 Turbo | FluidAudio CoreML | ~0.19s per segment (~110x real-time) | Highest accuracy, multilingual |
| Parakeet TDT 0.6B | CoreML | ~0.5s per minute of audio (~110x real-time) | English, lowest latency |
| Whisper Medium | MLX-Whisper | ~1s per segment | Good accuracy/speed balance |
| Whisper Large v3 | whisper.cpp + CoreML | ~0.08x RTF (50s per 10 min audio) | Maximum accuracy |

A single Mac Mini M4 Pro can process approximately **8,000–12,000 minutes of audio per day** for STT while simultaneously serving WebRTC traffic on the CPU.

Memory requirements are modest — the Whisper Large model needs ~10 GB, Parakeet ~3–4 GB — leaving ample unified memory headroom for WebRTC connection state on a 24–64 GB device.


## Speech-to-Text via x402

### The Service

dTelecom offers real-time speech-to-text as a standalone API, powered by nodes running Whisper and Parakeet models on Apple Silicon GPUs. The service is accessible via the **x402 micropayment protocol** — no API keys, no accounts, no subscriptions.

**Pricing: $0.005 per minute of audio.**

| | dTelecom STT | OpenAI Whisper API | Deepgram Nova-3 | AWS Transcribe |
|---|---|---|---|---|
| Price per minute | $0.005 | $0.006 | $0.0043 | $0.024 |
| Streaming support | Yes | No | Yes | Yes |
| Account required | No | Yes | Yes | Yes |
| Payment method | USDC via x402 | Credit card | Credit card | Credit card |
| AI agent accessible | Yes — native | Requires API key provisioning | Requires API key provisioning | Requires IAM setup |

### Why x402

The [x402 protocol](https://www.x402.org/) — created by Coinbase and now governed by an independent foundation — enables HTTP-native micropayments using USDC stablecoins. It activates the long-dormant HTTP 402 status code for its intended purpose: payment required.

The payment flow is embedded in standard HTTP:

```
Client                          dTelecom STT Server
  |                                    |
  |--- POST /transcribe -------------->|
  |                                    |
  |<-- 402 Payment Required -----------|
  |    (price: $0.005/min, USDC, Base) |
  |                                    |
  |  [Client signs payment with wallet]|
  |                                    |
  |--- POST /transcribe + payment ---->|
  |                                    |
  |<-- 200 OK + transcription ---------|
```

No API keys. No OAuth tokens. No account creation. The client just needs a wallet with USDC. The SDK handles the entire payment flow automatically.

This matters for one specific reason: **AI agents**. An autonomous agent that needs to transcribe audio can discover the service, pay for it, and use it — all within a single HTTP request cycle. No human needs to sign up for an account or provision credentials. This is the native payment model for the agentic web.

x402 is production-ready: 100M+ payments processed, $24M+ in volume, backed by Coinbase, Cloudflare, Google, and Vercel, with V2 launched in January 2026.

### SDKs and Integration

The STT service is available through multiple integration paths:

**MCP Server** — for AI assistants (Claude, Cursor, etc.):
```bash
npm install -g @dtelecom/stt-mcp
```

**TypeScript SDK:**
```bash
npm install @dtelecom/stt
```

**Python SDK:**
```bash
pip install dtelecom-stt
```

All SDKs handle x402 payment negotiation, session management, and auto-renewal transparently. The developer provides a wallet private key; the SDK does the rest.

**Open source:** [github.com/dTelecom/stt-mcp](https://github.com/dTelecom/stt-mcp) | [github.com/dTelecom/stt-client-ts](https://github.com/dTelecom/stt-client-ts) | [github.com/dTelecom/stt-client-python](https://github.com/dTelecom/stt-client-python)


## Economics

### Why This Architecture Enables Flat Pricing

dTelecom's flat bandwidth pricing (from $0/month for 50 Mbps to $3,600/month for 2.5 Gbps, unlimited minutes) is viable because the per-node operating cost is fundamentally lower than competitors running traditional infrastructure:

| Cost Component | dTelecom (Mac Mini M4 Pro) | Traditional Cloud Server |
|---|---|---|
| Hardware (amortized over 3 years) | ~$39/month | ~$80–$200/month |
| Colocation / hosting | $30–$200/month | $200–$500/month |
| Power (100W avg) | ~$11/month | ~$50–$150/month |
| Total per node | **~$80–$250/month** | **~$330–$850/month** |

At the same time, each node generates revenue from two streams:
1. **Bandwidth fees** from WebRTC media traffic (the core B2B product)
2. **STT micropayments** from AI inference via x402

This dual-revenue, low-cost model is what makes flat bandwidth pricing sustainable at scale.

### Node Operator Economics

As dTelecom's network decentralizes, clients and third-party operators can run their own Apple Silicon nodes:

- **Hardware investment:** $499–$3,000 (one-time)
- **Staking requirement:** ~$10,000 in $DTEL tokens
- **Revenue share:** 75% of client payments flow to node operators
- **AI workload bonus:** Nodes capable of running STT/AI receive additional revenue from inference tasks
- **Geographic incentives:** Higher rewards for nodes in underserved, high-demand regions

The low hardware cost and power consumption make Apple Silicon nodes accessible to a wider range of operators than traditional server hardware, supporting dTelecom's goal of genuine infrastructure decentralization.


## Roadmap

### Available Now
- dTelecom WebRTC infrastructure on Apple Silicon nodes
- Real-time STT via x402 ($0.005/min) with Whisper and Parakeet models
- MCP server for AI assistant integration
- TypeScript and Python client SDKs
- Open-source: full stack on GitHub

### Coming Next
- Expanded AI services on Apple Silicon GPU: real-time translation, text-to-speech, noise suppression, voice cloning
- Additional x402-powered AI APIs for the agentic web
- Client-hosted Apple Silicon node program — run your own node on a Mac Mini
- Multi-model inference: LLM-powered voice agents running on Mac Studio M4 Max with 128 GB unified memory
- Network expansion to additional data center regions


## Technical References

- [dTelecom Documentation](https://docs.dtelecom.org)
- [dTelecom Whitepaper](https://dtelecom.org/whitepaper.md)
- [x402 Protocol](https://www.x402.org/)
- [LiveKit SFU Benchmarks](https://docs.livekit.io/home/self-hosting/benchmark/)
- [Apple Silicon in Data Centers — AWS EC2 Mac Instances](https://aws.amazon.com/ec2/instance-types/mac/)
- [MLX Framework](https://ml-explore.github.io/mlx/)
- [Whisper — OpenAI](https://github.com/openai/whisper)
- [Parakeet TDT — NVIDIA](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2)

---

*dTelecom — Build, Own, Earn. [dtelecom.org](https://dtelecom.org)*