NEAR AI has reported more than 1 million NEAR staked with its service, where users can put staking rewards toward AI inference while retaining their underlying tokens. Summary Users retain their staked NEAR while rewards pay for AI inference. Unused inference credits carry forward, and unstaking stops new credits. Computing budgets vary with NEAR’s price, staking yield, and the amount deposited. Agent hosting follows a separate formula, with 500 NEAR generating $5 in monthly credits. In an Oct. 21 X post, NEAR AI said that the amount staked had passed 1 million tokens, with NEAR Protocol repeating the figure minutes later and describing the deposits as staking for inference. “Stake yours to pay for your AI. The NEAR stays yours,” NEAR AI wrote. The protocol introduced the payment option on July 30, as crypto.news previously reported when NEAR launched staking-based AI payments. At launch, NEAR said the mechanism covered 43 models from providers including OpenAI, Anthropic and Google, alongside confidential inference and hosted agents. NEAR AI uses staking rewards to fund inference NEAR AI’s inference documentation says the service converts staking yield into computing credits, leaving the deposited tokens in the user’s ownership. The available budget depends on the amount staked, NEAR’s price, and the staking annual percentage yield, or APY. Credits build up each second, according to the documentation, and unused balances carry forward. NEAR AI values one credit at one US dollar of platform usage, while unstaking stops new credits from accruing. The company says users can manage changes through signed wallet transactions. The reward rate is variable. NEAR AI explains that network rewards are divided among stakers, so the proportion of supply participating in staking affects the yield. Its documentation puts typical validator commissions at 1% to 10%, deducted before rewards reach the user. Those figures describe validator charges, not the price of an inference request. For budget planning, NEAR AI provides a calculator that translates a monthly inference allowance into the required stake, or estimates what a chosen token balance can fund. Users can change its price and APY assumptions to compare different budgets. The calculator page describes its figures as illustrative and directs users to the cloud console for current model prices and their wallets for the applicable staking yield. Pay-as-you-go prices depend on the model and token usage NEAR AI publishes a separate pay-as-you-go price list, updated Oct. 21, that charges for model input and output tokens. Its pricing page says the service has no minimum spending requirement, subscription or seat license on that plan, and no platform fee above the listed model rate. The dollar-denominated catalog lists GLM 5.3 Flash at $0.08 per million input tokens and $0.50 per million output tokens. Qwen 3.8 27B costs $0.04 per million input tokens and $1.35 per million output tokens. Using those published rates, a workload with 1 million input tokens and 1 million output tokens would cost $0.58 on GLM 5.3 Flash or $1.39 on Qwen 3.8 27B. These calculations use equal token counts and exclude any separate services; they do not measure output quality or compare the models’ performance. NEAR AI says secure execution and hardware-signed attestations are included in the listed inference rates. Its enterprise offering separately includes reserved GPU capacity, private models, volume pricing and service-level agreements. Agent hosting follows a different credit formula For hosted agents, NEAR AI’s July 30 product explanation sets a different rule. Its stated policy divides the amount of NEAR staked by 100 to determine a monthly dollar credit budget, giving the example of 500 NEAR generating $5 in monthly credits. The hosting allowance becomes available at the start of the subscription period and refreshes while the tokens remain staked, the company says. Inference credits instead accumulate continuously from yield. NEAR AI describes the hosting formula as current policy, not a permanent conversion rate. The company says staking determines both the hosting budget and how many IronClaw agents can run in parallel. Its explanation lists a minimum stake of 50 NEAR to activate a first agent, with subscription changes handled through wallet transactions. NEAR AI’s developer documentation describes trusted execution environments as isolated areas inside processors that protect code and data during computing. For supported TEE-hosted models, the company says model providers and cloud operators cannot read prompts or use the data for training. The documentation provides separate routes for model access, private inference and verification, alongside top-up and pay-as-you-go payment options. NEAR’s application proposals include agents with their own balances On Oct. 3, NEAR co-founder Illia Polosukhin outlined AI-powered application ideas including NFT-based digital pets carrying NEAR balances to pay for inference. His proposed design would use NEAR AI to run verifiable inference from prompts encoded onchain, with feeding, dressing, and possible competition features. Polosukhin’s list contained 13 application ideas and five distribution suggestions. His proposals included testing AI models against private datasets without exposing the underlying information, with certified results identifying the model and the test-set hash. An Oct. 6 examination of AI token economics identified balances staked, actual model usage, and payments to providers as separate measures for evaluating NEAR’s AI service. It described the network’s roadmap as combining private AI services with cross-chain execution infrastructure.
