The AI Bandwidth Economy:
How AI Agents Buy Clean IPs From Device Fleets.

Cloudflare, DataDome, PerimeterX, Akamai, AWS WAF, Kasada, Imperva — the major anti-bot systems share one shortcut: they block datacenter IPs by default. AWS, GCP, Azure, Hetzner, OVH — entire ASN ranges flagged. Residential IPs work better but still get blocked when ISPs sell their DHCP pools. Mobile IPs are the last category that scales: carriers route thousands of real subscribers behind a single CGNAT IP, so blocking a mobile IP means blocking real paying customers. Anti-bot vendors are cautious about taking that risk.

That makes mobile IPs structurally scarce. There are only so many real phones in the world, and only some of them are willing to route external traffic. Every AI agent that wants to scrape Google SERPs, monitor Amazon prices, watch Reddit sentiment, or train on the public web is bidding for the same finite pool of clean IPs.

We ship four protocols that compose into the bandwidth economy. Each is independently usable; together they form the loop.

The peer network is a true two-sided marketplace. The same WebSocket relay, the same x402 payment rail, the same USDC settlement — but participants come from very different worlds. Here's how each side plugs in.

Rates are set per partner and may be a fixed rate or a share of collected revenue. There is no universal public percentage; the live rate is the earningsPerGB value returned when you register a device. IP type is classified server-side by ASN — device-reported types are ignored. Datacenter IPs are never listed as residential or mobile. Registration does not guarantee traffic or earnings.

Per-GB rates are set per partner and returned at registration as earningsPerGB. Datacenter IPs are never listed as residential or mobile. Server-side ASN classification — device-reported type is ignored.

Not every IP is worth the same, so rates are not flat — because the IPs aren't worth the same. A real mobile carrier IP is hard to block. A datacenter IP is often blocked at the edge, which is why datacenter IPs are never listed as residential or mobile.

Instead, every connecting device gets classified server-side via ASN lookup against a curated registry of 200+ mobile carrier ASNs, and 100+ datacenter providers. Device-reported metadata is ignored. You can't spoof your way into the mobile tier — the network sees what your IP actually is. That keeps the marketplace honest, customer success rates high, and scarcity verifiable.

All public-web work. No login walls, no private data, no identity scraping. Standard data intelligence — the same kind of work companies have done for 20+ years, now increasingly run by AI agents.

We document four ways to join — Android SDK 1.3.2, the Node reference client 1.6.6 (API_KEY=psx_... node reference-sdk.js, e.g. in a node:20-alpine container), the Go 1.6.8 Linux binary, or your own WebSocket client. But the only thing we actually require is: deliver an HTTP proxy endpoint from a real IP. How you get there is up to you.

Run a Raspberry Pi mesh on a residential network. Build a custom firmware for a router. Write a SOCKS5-to-HTTP bridge in Rust. Run a fleet of IoT devices behind a single relay client. Build something nobody has thought of. As long as the IP is real (server-classified), traffic flows through your device, and you connect via the relay protocol — we can route to you and track usage, and eligible traffic is paid under your partner terms.

Input:  A device with a real mobile or residential IP.
Output: An HTTP proxy endpoint that routes traffic through it.
Result: Eligible traffic is paid under your partner terms.

That's the entire requirement.

We're early to a real shift. The first generation of AI agents was just chatbots — text in, text out, no external state. The second generation is doing real work: filling forms, calling APIs, running scrapers, executing transactions. Every single one of those operations needs to actually reach the public internet, which means it needs an IP that won't get blocked.

Today the supply side of clean IPs is captured by a handful of proxy companies running their own infrastructure with billing departments and AML compliance. We think it becomes a marketplace where consented devices with real IPs can supply, and any agent with a wallet can pay. The hoped-for result: the cost of real-world data collection drops, the supply of clean IPs grows, and the AI agents doing useful work for humans get cheaper to operate.

The ultimate version: autonomous agent-to-agent commerce. An agent that earns USDC from its data work pays another agent for its bandwidth. That agent in turn pays for its compute, its hosting, its tools. It is a direction, not a finished product — and supplying bandwidth still means a signed Partner Agreement 2.0 that we have approved and an account-linked API key. This is what the four-layer stack (x402 + MCP + Peer Relay + Gateway) is designed for.

We've written extensively on the components of this thesis. Six places to go next:

The economy is live. Both sides are open. Pick your seat at the table.