NEAR AI Says Staking-for-Inference Deposits Have Exceeded 1 Million NEAR

How the staking model works

According to NEAR AI’s documentation, deposited NEAR remains under the user’s ownership while staking yield is converted into credits for platform usage. The size of the available computing budget depends on three variables: the amount staked, the market price of NEAR, and the staking annual percentage yield.

NEAR AI says credits accrue continuously, building up each second, and that unused balances roll forward. One credit is valued as one US dollar of usage on the platform. If a user unstakes, new credits stop accumulating, though the company says account changes are handled through signed wallet transactions.

Budget estimates depend on variable yields and prices

The service does not offer a fixed return, and NEAR AI notes that staking rewards vary because network rewards are shared among participants. That means the share of total supply being staked can affect the yield a user receives. Its documentation also says validator commissions typically range from 1% to 10%, with those fees deducted before rewards reach the staker.

To help users estimate spending power, NEAR AI provides a calculator that can work in both directions: it can estimate how much monthly inference a given stake may support, or how much NEAR would be needed for a target allowance. The company describes those figures as illustrative and advises users to check the cloud console for current model pricing and their wallets for the applicable staking yield.

Published inference prices vary by model

NEAR AI also lists pay-as-you-go pricing separately from staking-based credits. Its pricing page, updated Oct. 21, says there is no minimum spend, subscription or seat license under that plan, and no platform fee on top of the listed model rates.

The company’s dollar-denominated catalog includes GLM 5.3 Flash at $0.08 per million input tokens and $0.50 per million output tokens. Qwen 3.8 27B is listed at $0.04 per million input tokens and $1.35 per million output tokens. Based on those posted rates, a workload using 1 million input tokens and 1 million output tokens would cost $0.58 on GLM 5.3 Flash and $1.39 on Qwen 3.8 27B, excluding any separate services.

NEAR AI says secure execution and hardware-signed attestations are included in the inference rates it lists. Its enterprise tier is offered separately with features such as reserved GPU capacity, private models, volume pricing and service-level agreements.

Hosted agents use a separate formula

The company applies a different credit rule for hosted agents. In its July 30 product explanation, NEAR AI said the monthly hosting budget is calculated by dividing the amount of NEAR staked by 100, giving 500 NEAR as an example that would provide $5 in monthly credits.

Unlike inference credits, which accumulate from staking yield over time, the hosting allowance is made available at the start of a subscription period and refreshes while the tokens stay staked. NEAR AI describes this as current policy rather than a permanent conversion rate. It also says staking affects how many IronClaw agents can run at once, with a minimum of 50 NEAR required to activate a first agent.

Broader AI plans on NEAR

NEAR AI’s documentation also outlines private inference options based on trusted execution environments, or TEEs, which it describes as isolated processor areas designed to protect code and data during computation. For supported TEE-hosted models, the company says model providers and cloud operators cannot read prompts or use the data for training.

The staking milestone arrives as NEAR leadership continues to outline possible AI-native applications. On Oct. 3, NEAR co-founder Illia Polosukhin shared 13 application ideas and five distribution suggestions, including NFT-based digital pets with their own NEAR balances to pay for inference, and systems for testing AI models against private datasets without exposing the underlying information. For now, the latest confirmed development is the deposit total itself: NEAR AI says staking-for-inference has now passed 1 million NEAR.