Where does your AI really come from?
Why the source of your AI access matters: model substitution, proxy data risks, and how Xava Inference approaches capacity, pricing and rewards.
AI access sold through an unvetted proxy can introduce model-substitution, account-reliability and data-handling risks. Xava Inference uses purchased model capacity, provider list prices and the model you select. Check each model’s retention policy and current availability before sending sensitive data.
Inference is the meta right now, and tokens that come with AI credits are getting a lot of attention. But how often do we stop and ask where those credits actually come from?
Some are bought through other platforms and passed on, which adds a layer between you and the provider. Others are sold cheap on AI marketplaces, made from leftover startup credits, shared accounts or token subsidies. They work, until a provider shuts the accounts down or the subsidies run out, and users are left with nothing.
In some cases, it goes further. Oxford China Policy Lab researcher Zilan Qian describes grey-market services advertising Claude access at as little as 10% of the official price, along with account abuse, model substitution and privacy risks. The original investigation cautions that the scale of stolen-card inputs and claims of systematic log resale remain unverified; it does not establish that every proxy sells user data.
Cheap AI usually has a reason for being cheap.
How we built Xava Inference
From day one, Xava Inference was built on one simple rule: buy model capacity directly from the providers, at scale, and run it as a real inference business.
That rule shapes everything you experience. Requests use the model you select; we do not silently replace it with a cheaper model. Model availability can change, and an unavailable model can return an error. We do not store prompt or reply text; each model provider’s own retention policy still applies unless the model is marked ZDR supported. And because we buy it directly, your access doesn’t depend on someone else’s spare account staying open.
It’s also why we think of Xava as an aggregator rather than a marketplace. Our aggregator model brings supported AI models behind a single key at provider list prices. The model catalog lists current availability and capabilities.
And then, it pays you back
Doing it properly is only half the story. The other half is what happens every time you use it.
Eligible AI usage can earn cash back, capped by the available margin on each request. Holder credits use daily distributions for wallets with a time-weighted average balance above 1,337 $INFX, without staking or claiming. Production holder-credit issuance is not enabled yet. And when you top up in USDC, there’s no fee, so every dollar you add goes straight into AI.
None of this leans on borrowed capacity or subsidies. Cash back comes out of our margin on real AI usage, and holder credits come from real trading. Both depend on the activity and funding available; neither is a guaranteed return.
That’s the kind of AI we wanted to use ourselves: one you can count on, that gives something back.
New to Xava? Start here, and see how holder credits work.
$INFX is a utility token, not an investment. Free AI credits depend on trading and can be small or zero. Cash back is paid from our margin and depends on your usage. Nothing here is financial advice.
Frequently asked questions
Does Xava Inference replace the model you choose?
No. Requests use the model you select. We do not silently replace it with a cheaper model. If a model is unavailable, the request can return an error.
Does Xava store prompts and replies?
No. We do not store prompt or reply text. Each model provider’s own retention policy still applies unless the model is marked ZDR supported. Check that policy and current availability before sending sensitive data.
Are cash back and holder credits guaranteed?
No. Cash back is paid from our margin on real AI usage and is capped by the margin available on each request. Holder credits depend on trading and funding, and production issuance is not enabled yet. Neither is a guaranteed return.
