What Aile is, and what it isn't yet

Most AI inference capacity in the world is idle right now. A 4090 in someone's closet, a Claude Code subscription that gets used two hours a day, an API key with a monthly cap nobody hits. Meanwhile the person who needs a few thousand tokens at 2am pays list price, if they can get a key at all.

Aile (say it like isle) is a marketplace that puts those two groups in the same room. Lenders list capacity. Buyers reach it through one OpenAI- and Anthropic-compatible endpoint and pay per call in USDC on Solana. This note is the short version of how it actually works, including the parts we haven't finished.

Three kinds of capacity, one listing surface

Every listing on Aile is one of three things, and the type is never hidden from the buyer.

• Self-hosted. The lender's own machine runs the model. Weights and prompts never leave that machine except as the inference output. Nothing is uploaded to us.

• Provider subscription. The lender links a provider account, an API key or a consumer subscription like Claude Code, Codex, or Gemini CLI. Aile holds that credential encrypted at rest and injects it at serve time. The lender's node only relays TLS ciphertext; it can't read the prompt or the key.

• Nodeless. API-key providers only, opt-in. No node in the request path at all: Aile dials the provider directly. Lowest friction, nothing to install, at the cost of the node-side kill switch.

The response headers tell you which one served you. x-aile-capacity names the type and x-aile-blind says whether the node in the path could see your prompt. We'd rather be honest about the trade than pretend every path is equally private.

What a buyer does

Point your existing SDK at Aile. POST /v1/chat/completions works with the OpenAI client and POST /v1/messages works with the Anthropic client, and either can serve a model from any provider. You pick a provider by prefixing the model, cc/claude-sonnet-5 or codex/gpt-5.5, and Aile routes strictly to that provider. There is no silent cross-provider fallback; if nobody is serving what you asked for, you get a 503 that says so, not a different model dressed up as the one you wanted.

Payment is per call. If your client can sign x402, USDC goes straight from you to the lender who served the request and Aile takes nothing. If it can't, you pre-fund your own embedded wallet and Aile debits it as you go, taking a 10% fee on that path.

What a lender does

Run the node on your hardware, or link a provider account, and list what you can serve at the price you want. Settlement lands in your own wallet; the private key lives in Privy's enclave, not on our server, and you can export it and leave whenever you like.

Lenders who elect to be paid in $AILE, our SPL token, pay 0% platform fee on the balance path. Same for buyers who fund in $AILE. That's the token's only job: it's a fee election, not a credit, a voucher, or a promise. Where a deployment has no mint configured, everything runs USDC-only.

What we can and can't verify

We try not to overstate this. For a handful of providers we can prove the account link: a provider-signed token with our nonce in it, which says "this lender authenticated with the provider just now." We can prove a custodied credential recently worked. For self-hosted capacity we run a known-answer probe and show the result on the listing card.

What we cannot prove yet is what a node actually serves: model identity, quality, or a quiet downgrade to something cheaper. The probe is evidence, not proof. Evals and reputation are designed and not built. If you need a hard guarantee of model identity today, Aile isn't the place to get one, and we'd rather you know that before your first call.

Where it stands

The relay is real code with tests: listings, discovery, the in-path proxy, x402 and the balance ledger, wallet custody, metering, and fee accounting. The launch is the point where other people's traffic tells us which of our assumptions were wrong. If you have idle capacity or an unmet need for some, come find out with us.