Most AI invoices still land on a card. A sales lead who wants one AI agent on research and another on follow-up gets a seat license, a monthly minimum, and a ticket for finance. NEAR AI is pushing a stranger bill: keep the NEAR where it is, and let the yield it would have earned, or a stake-sized hosting allowance, pay for the model calls. You are not selling the principal to buy prompts.
The product rules come from NEAR AI’s own launch account. Staked NEAR stays with the user. Hosting and inference use different formulas. Privacy is a different path for open-source models than for frontier models. crypto.news is the source for the adoption headline, a July 30 launch it says covered 43 models, a one-dollar value on an inference credit, typical validator commissions of 1% to 10%, and an Oct. 21 pay-as-you-go price list. Those catalog and adoption details are not restated in the launch account. Whether credits already earned survive an unstake, and how long unbonding takes, are not documented here.
The line that outlet quotes from NEAR AI’s Oct. 21 post is the pitch:
Stake yours to pay for your AI. The NEAR stays yours.
Read that against the docs, not against the headline. Unstaking stops new allowance. The original NEAR stays intact. “Withdrawable” means withdrawable after any unbonding period this material does not document, not cash in the same minute. A separate one-off purchase locks NEAR for a fixed duration. That product is a lock. Withdrawability is claimed for the staking product only.
A deposit is not a workload
More than 1 million NEAR is a stake total. It is not dollars of inference, not a user count, and not a receipt from OpenAI, Anthropic, or Google. No tighter figure was published. An examination of AI token economics carried by the same outlet (analytical, not an official NEAR metric) split the story into three numbers that should not be mashed together: balances staked, actual model usage, and payments to providers.
Until the second and third show up, this is a parking story. Parked NEAR can earn. It can also sit there and answer nothing. A warehouse of prepaid stake is inventory, not a harvest. For AI for business, ask what ran.
The dollar value of that million is not in these materials either. Hosting credits ignore NEAR’s market price. Inference credits do not. Anyone turning the headline into a budget has to apply that day’s price and yield themselves.
Four ways the bill gets paid
Hosting allowance
Under current policy, staked NEAR divided by 100 is a monthly dollar credit for agent hosting. Five hundred NEAR produces $5 a month. The full budget is available at the start of the period and refreshes while the NEAR stays staked. A move in the token’s price does not resize it. A multi-year AI automation plan that assumes the ratio will hold is a bet, not a contract.
Tiers are a different lever. NEAR AI names them Starter, Basic, and Pro, with thresholds set in NEAR. Those ranges adjust if the price moves so the dollar-equivalent cost of each tier stays stable. One contrast covers both rules: 500 NEAR still throws off $5 of hosting if NEAR doubles, but the NEAR required to sit in Pro may fall.
Yield-funded inference
Confidential inference is the variable path. NEAR AI describes the same yield the stake would have earned anyway, redirected into compute under protocol-defined parameters. That is not “delegate to any validator and the invoice pays itself.” Neither the launch account nor the reporting lays out the contract path.
The budget moves with stake size, NEAR’s price, and staking annual percentage yield. Rewards are shared among stakers, so a larger share of supply participating can change the rate. The reported range for typical validator commissions is 1% to 10%, taken before rewards reach the user. That commission is not the price of a prompt.
Credits accrue per second. NEAR AI says they are usable on day one, which means you can spend them as they accrue. It does not mean a month of inference lands up front the way hosting does. Unused balances are reported to carry forward. Unstaking stops new credits. It is not clearly stated that credits already earned survive the unstake. If treasury cares about the leftover, read the docs before assuming it travels.
Reported docs value one inference credit at one US dollar of platform usage. Hosting is a monthly dollar budget on its own clock. These materials do not say the two draw from a single balance. Treat them as two allowances until the console shows otherwise.
A calculator can turn a monthly allowance into a required stake, and it lets you change price and yield assumptions. The page calls those figures illustrative and sends you to the cloud console for live model prices and to the wallet for the yield that actually applies. Sketch, then check.
Timed lock
One-off credits can be bought by locking NEAR for a fixed duration, without changing tiers. This is not the withdrawable stake. There is no honest unbonding number to quote.
Per-token rate
If yield variance is a bad way to run AI for sales, there is an ordinary price list. The Oct. 21 update, as reported, charges for input and output tokens, with no minimum spend, no subscription, no seat license, and no platform fee above the listed model rate. Secure execution and hardware-signed attestations are described there as included in those rates, not sold as an add-on.
On that reported list, GLM 5.3 Flash is $0.08 per million input tokens and $0.50 per million output tokens. Qwen 3.8 27B is $0.04 and $1.35. A million input tokens plus a million output tokens would be $0.58 on the first and $1.39 on the second, using equal counts and ignoring any separate service. Those version strings are as printed in the report. They were not matched here to vendor model cards, and the arithmetic says nothing about which model is fit for the job. Check the live list before anyone budgets against it.
NEAR AI’s launch account does confirm access to models from providers like OpenAI, Anthropic, and Google, and names Qwen, DeepSeek, and GLM as open-source models. The count of 43 belongs to the news report, not to that account. An enterprise menu (reserved GPU capacity, private models, volume pricing, service-level agreements) is named in the reporting, with no prices and no contract terms. A line on a menu is not a term sheet.
What two agents actually clear
IronClaw is NEAR AI’s secure, open-source agent harness, and it is where this stops being a payment curiosity. A first agent needs at least 50 NEAR and, the company says, activates in about 30 seconds. Changes go through signed wallet transactions. Stake size sets both the monthly hosting budget and how many agents can run in parallel.
Apply the stated ratio to that minimum. Fifty NEAR clears the gate and produces $0.50 a month in hosting credits. That will not staff a research agent and a follow-up agent in any serious sense. Five hundred NEAR, the company’s own example, is $5 a month of hosting. Inference on the same stake is whatever yield is left after commissions, at whatever NEAR is worth that day. Activation and budget are not the same size.
The parallel cap is the constraint that matters if you are staffing AI agents the way you staff SDRs: one for research, one for follow-up, one that should never see the pricing sheet. These materials do not publish the NEAR required for a second or third agent, so there is no honest stake figure for that roster. If NEAR’s price doubles, the hosting dollars on a fixed stake stay put. Inference credits on that stake do not. They move with price and with yield, and yield can fall when more of the supply is staked.
Privacy is two doors, not one slogan
Illia Polosukhin, NEAR’s co-founder, has warned that ordinary inference can expose secrets and credentials to the provider. A CryptoBriefing write-up also identifies him as a co-author of the 2017 paper “Attention Is All You Need.” The worry is the right one if an agent is chewing on customer records, pricing, or pipeline notes. NEAR AI’s answer is hardware isolation. It is the company’s answer. Nothing here is an independent audit.
For open-source models (Qwen, DeepSeek, and GLM), NEAR AI says inference runs in a hardware-isolated enclave on the GPU. The host operating system, the GPU operator, and NEAR AI cannot read the prompt or the output. Each request returns a hardware-signed attestation in under 30 seconds.
Frontier models take a different path. For OpenAI, Anthropic, and Google, NEAR AI describes a gateway in a trusted execution environment that routes the request. Optional redaction can strip certain personal identifiers. The provider still sees a prompt arriving from that gateway. The stated protection is narrower: the provider supposedly cannot tie the request to you, and NEAR AI cannot see the contents because the gateway sits in an enclave. The company calls this “incognito mode for AI.”
Incognito is not invisible. A provider that receives the text can still have the text. Do not collapse those two doors into “nobody can read the prompt.” For a confidential workload, the open-source enclave is the claim that matches the fear. The frontier path is closer to an unlinked request than to a sealed one. Neither path is independently audited here.
Not on the purchase order
On Oct. 3, as reported, Polosukhin outlined application ideas rather than shipped products: NFT-style digital pets that hold NEAR balances and pay for their own inference, verifiable inference from prompts encoded onchain, and tests against private datasets that return a certified result naming the model and a hash of the test set without exposing the data. A pet with its own wallet is a decent picture of an agent that carries a compute budget. It is not something you can switch on this afternoon.
Questions worth asking first
- Do I spend the NEAR to pay for AI?
No. Staked NEAR stays yours, and yield or a hosting allowance pays for compute. Unstaking stops new allowance. A one-off credit purchase is a timed lock, and any unbonding delay on the staking product is not documented here.
- What does “more than 1 million NEAR” actually prove?
A stake total from a post the reporting dates to Oct. 21. Not a user count, not dollars of usage, and not proof that model providers have been paid.
- How do hosting credits differ from inference credits?
Hosting is a monthly dollar budget from stake divided by 100, granted up front and indifferent to NEAR’s price. Inference credits come from variable yield and accrue per second, and these materials do not show the two drawing from one balance.
- Can a frontier provider still see my prompt?
Yes, on the company’s own description. Open-source models are claimed to run in a GPU enclave outsiders cannot read. Frontier prompts still reach the provider, only supposedly unlinked from you. Neither path is independently audited here.
- Is the cheaper listed model the better one?
No. Reported rates put a million input tokens plus a million output tokens at $0.58 and $1.39 on two named models, and that says nothing about fitness for a sales or support workflow. Confirm the names on the live price list before you budget.
- Are the digital pets and private-dataset certificates live?
No. They were outlined as proposals, not described as shipped products. Useless as a procurement item.
- What share of that staked NEAR has been spent on inference?
Unknown. The milestone counts NEAR deposited, not yield actually consumed as model calls, and not dollars paid to providers.
If this lands on a shortlist, treat the calculator as a sketch, then read the wallet yield and the live price list. Ask which models sit in an enclave and which are merely unlinked from your name. The million-NEAR post does not say how much of that stake has been spent on thinking. That share is unknown.