Announcing Staking for NEAR AI: Put Your NEAR to Work Powering Confidential AI

Key Insights 

• NEAR AI has launched staking-based payments for confidential inference and agent hosting. Rather than paying through recurring fiat subscriptions, users stake NEAR to receive AI compute credits.

• Two products are supported. Agent hosting allocates credits based on a fixed stake-to-credit ratio, while confidential inference converts staking yield into compute credits using protocol-defined parameters.

• Staked NEAR remains withdrawable. Users retain ownership of the underlying stake while it generates recurring credits.

• Confidential inference leverages Trusted Execution Environments (TEEs). Hardware-signed attestations enable users to verify that inference executed inside a confidential enclave.

• Staking for NEAR AI gives you access to models from providers like Anthropic, OpenAI, and Google.

The most centralized thing in a crypto-native AI stack isn't the model. It's the billing. To call a model today, you open a cloud account, hand over a payment method, and let a provider meter your usage and hold your credentials. For most people that is just how software works. For crypto-native builders, it is a dependency in the wrong place, a centralized billing relationship wrapped around infrastructure that is supposed to be sovereign.

It is also a quiet form of extraction. The provider captures the spread between what your usage produces and what you keep, and the payment rail sits outside the wallet where the rest of your onchain life already lives.

Today, we are introducing Staking for NEAR AI, which removes the credit card from the loop. Lock NEAR, and it converts into monthly credits that can be used to pay for your confidential inference from NEAR AI Cloud, for hosting IronClaw, and for other NEAR AI services. So the trade isn't "spend your NEAR or keep it." You keep the NEAR, and it funds your NEAR AI usage at the same time. Stake more, get more. Unstake, and the allowance stops.

How the conversion works

Staking NEAR can now pay for two different NEAR AI services, confidential inference and agent hosting, and each converts your stake on its own formula.

Agent hosting (Starter, Basic and Pro Tiers). Hosting runs on a flat ratio: staked NEAR ÷ 100 = your monthly credit budget, denominated in dollars, at current policy. Stake 500 NEAR and you receive $5 in credits every month, refreshed each subscription period, for as long as the NEAR stays staked. The budget is pinned to your stake, so the dollar figure stays predictable regardless of where the market moves.

Confidential inference (staking farm). Inference converts the yield your stake earns rather than the stake itself. Here your budget is a function of how much you stake, the price of NEAR, and the staking APY, the same yield your stake would earn anyway, redirected into compute.

Whichever path you use, the credits are usable on day one. For agent hosting, you get the full monthly budget immediately, rather than accruing toward it, and it refreshes each subscription period for as long as the NEAR stays staked. For private inference, staking credits are received on a per second basis, accruing as a function of the size of the stake.

Two things follow from this that a metered card never gives you. First, the budget renews each period instead of draining your principal, so running AI is a recoverable position rather than a sunk bill. Second, when you are done, you unstake and walk away with your original NEAR intact.

Stake for confidential inference on NEAR AI Cloud

Inference requests for open source models (Qwen, DeepSeek, GLM) run inside a trusted execution environment (TEE), a hardware-isolated enclave on the GPU. Inside that enclave, your prompt and the model's output are sealed off from everything around them. The host operating system, GPU operator, and NEAR AI itself cannot read them. They are cryptographically locked out, not merely trusted to look away. And you don't have to take that on faith: each request returns a hardware-signed attestation in under 30 seconds, a proof signed by the chip that the code ran inside a genuine enclave and was not tampered with. Confidentiality here is a hardware guarantee you can check, not a contractual promise you have to trust.

For frontier models, we built a secure gateway that runs inside a TEE. It routes your traffic to OpenAI, Anthropic, and Google and users have the option of enabling a PII redaction function to strip out everything with certain identifiers . The provider sees a prompt arrive from NEAR AI's gateway and nothing else. It cannot tie the request to you, and because the gateway runs inside an enclave, NEAR AI cannot see what you're working on either. Think of it as incognito mode for AI: the model answers your question, but nobody, including us, knows it was your question.

Stake to deploy IronClaw agents

Now you can stake NEAR to activate your IronClaw, NEAR AI’s secure, open-source agent harness. Here your staked amount does two jobs at once: it sets your monthly credit budget and it determines how many agents you can run in parallel. Stake at least 50 NEAR and your first agent is live in about 30 seconds. 

IronClaw Benchmarks:

IronClaw takes the top spot on PinchBench, ClawBench, and OfficeQA, three benchmarks that stress different parts of what an agent actually does. All scores above pair each harness with the same base model, deepseek-v4-flash, so the comparison isolates the harness performance.

One onchain lifecycle for both

This entire lifecycle lives in your wallet. Stake more for more credits or more agents. Upgrade, downgrade, cancel, or resume a subscription with a signed transaction. Buy one-off credits by locking NEAR for a fixed duration when you need headroom without changing tiers.

One detail that keeps the pricing model predictable: hosting tier thresholds are denominated in NEAR, not dollars. If the market price moves significantly, the ranges adjust so the dollar-equivalent cost of each tier stays stable.

Principled design

NEAR AI already keeps your data confidential and your credentials out of the model. Payment should follow the same principle. A credit card routed through a centralized billing provider is a dependency that does not belong in sovereign infrastructure, so the staking model carries that logic through to how you pay: wallet-native, verifiable, and denominated in the same token as the rest of your onchain life. Staking closes the loop between what you hold and what you run, the capital and the compute finally live in the same place.

Start staking today

Staking for NEAR AI is a new way to put your NEAR to work, powering the inference you run and the agents you deploy. Deploy your first agent, or stake for confidential inference, by logging in with a NEAR wallet at near.ai.