Every Link Is a Node: The Mesh Network Thesis
The Inverted Economics of Compute
Every SaaS company in the world has the same problem: more users = more server cost. Growth is a liability. You scale by buying more AWS instances, and your margin gets thinner with every customer.
We asked: what if it worked the other way?
What if more users meant more infrastructure? What if every visitor to your website was simultaneously a consumer AND a provider of compute?
The Browser Is a Supercomputer
In 2024-2025, WebGPU landed in every major browser. This is the most underappreciated technology shift in a decade. Your browser can now access the GPU directly — the same GPU that runs your games, your video encoding, your ML inference.
Combined with WebAssembly (WASM) for CPU compute and WebRTC for peer-to-peer networking, every browser tab is now a full compute node.
The Math
- 1 browser tab: ~2-5 TOPS (modern laptop with WebGPU)
- 100,000 concurrent tabs: 200K-500K TOPS ($0 deployment cost)
- 1 million tabs: 2-5 PetaTOPS
- 10 million tabs: 20-50 PetaTOPS (competitive with GPU clouds)
A single viral post could spin up 10,000 compute nodes in an hour. At zero marginal cost.
How It Works
When you visit any BlackRoad site, a tiny script called mesh.js loads. It:
- Detects capabilities — WebGPU? How many cores? How much memory?
- Registers as a node — tells the signaling server "I'm here, here's what I can do"
- Connects to peers — WebRTC peer-to-peer connections, no central server needed
- Accepts tasks — when someone needs inference, your idle GPU cycles help
The Pi fleet (41.8 TOPS of Hailo-8 acceleration) serves as the permanent backbone. Browser nodes scale elastically on top. The mesh routes inference to the cheapest, fastest available node — whether that's a Pi in Minnesota or a laptop in Berlin.
The Precedents
Google (Larry Page): Every webpage did Google's ranking FOR them. The web was the infrastructure. The search engine was just the interface.
Bitcoin: Every wallet is a miner. Every participant makes the network stronger. Nobody can outspend the network because the network IS the participants.
Folding@Home: Hit 2.4 ExaFLOPS from voluntary compute donation. Proved that distributed volunteer compute works at planetary scale.
Coinhive: Proved browser-based distributed compute works at millions of devices. Died because mining is adversarial — but inference is cooperative.
The Revenue Model
Servers are a liability. Nodes are an asset. Every server is a bill. Every signup is free infrastructure that pays you.
- API: OpenAI-compatible at 50% of cloud pricing
- 70/30 split: Node operators earn 70%, BlackRoad takes 30%
- Hardware kits: $199 AI Home Server ($119 margin), buyers become permanent nodes
- Subscriptions: Premium models, priority routing, dedicated capacity
The Moat
Ollama is single-box. Petals is academic. Akash is GPU rental. Apple and Google are walled gardens.
BlackRoad is the only platform with a useful consumer product (chat) + distributed inference network (mesh) + permanent hardware backbone (Pi fleet) — all connected by links that ARE the network.
Every link you share is a new node. Every visitor is compute. That's the viral loop. That's the moat. — Alexa Amundson