# Staking for Inference: Morpheus vs NEAR vs Venice

> Three networks let you stake a token and draw AI inference from it. Two pay the compute bill by printing; Venice pays it out of revenue. Which model survives the taper, and why the peg decides it.

- Published: 2026-08-13 · Updated: 2026-10-15
- Canonical: https://ownyourmind.ai/tokenomics/staking-for-inference-morpheus-near-venice/
- Cite as: Own Your Mind (ownyourmind.ai), licence CC BY 4.0. Link the canonical URL.

Charts and interactive widgets are omitted here; project data is at https://ownyourmind.ai/api/scores.json.

---

## The Rung Between Renting and Owning

The question that started this piece: is staking for inference a way of owning your compute, or just a nicer-looking rental? It sits somewhere between paying OpenAI every month and buying the hardware yourself, and that gap needs mapping before we get to whether the economics hold.

Four places you can put capital around AI inference. Only three of them get you any.

NVIDIA equity sits in that table as the control case, for a blunt reason. Owning a supplier's margin and buying a supplier's output are separate transactions. Shares in the company selling the shovels give you a claim on its profits and no claim at all on a GPU-hour: you still queue up and pay list price like every other customer, with no special rights of any kind. It's the purest bet on compute demand that delivers no compute.

That's what makes it the useful contrast. The staking column is offering usage rights attached to the capital, rather than exposure sitting next to it.

The staking column is also the only one where the capital comes back. That's the pitch, and it's a fair one. You're posting capital and drawing a yield in compute rather than in tokens, and the day you want out, you unstake. Compare that to a $200-a-month frontier subscription, where twelve months of payments leave you with nothing but a chat history the vendor also has a copy of. I priced my own habit in [what a heavy AI habit costs](https://ownyourmind.ai/journal/what-a-heavy-ai-habit-costs/), and the number that stings is that none of it accrues.

So the model is attractive. Now the harder question: who's paying for the electricity?

## Three Networks, One Primitive

Three protocols route a version of the same primitive, and they do it three different ways.

Venice publishes an exact conversion rate and holds it fixed, NEAR publishes a fixed ratio for agent hosting and a floating one for inference, and Morpheus puts its rate in the contract itself, readable any time from `getTodaysBudget` and `stakeToStipend`. The yield column shows the three positions on a spectrum: Morpheus never promised you yield, Venice lets you keep most of it, NEAR takes all of it.

## Morpheus: A Share of the Budget

Morpheus is the oldest of the three and the only one that never dressed the mechanism up as free money. Nothing is burned, nothing is spent, and the deposit comes back. I traced the full flow in [How Morpheus Pays for Inference](https://ownyourmind.ai/tokenomics/morpheus-compute-economics/).

It matters which of the two routes you take, because they share a name and almost nothing else. Run your own consumer node and you post MOR as returnable collateral on each session, and the contract sizes your allowance against **total MOR supply**. Other stakers arriving can't dilute you. What does shrink your slice is the network being used at all: every MOR paid to a provider leaves the compute pool and joins supply at the same moment, so the allowance ratio has fallen from 1.1526 to 1.0564 MOR a year per MOR held between March and August 2026.  It only moves one way.

Stake into the Marketplace API builder subnet instead, for credits at the gateway, and the denominator is **the total staked across every builder subnet**. There you are diluted by each new staker, exactly as you would be in any pro-rata pool. Morpheus's own calculator says so without hedging: yield falls as more MOR is staked.

The compute providers get paid from the 24% Compute pillar of MOR emissions, drawn from a protocol funding account on Base rather than from anything the user hands over. That's the subsidy, stated plainly in the contract's behaviour. And the scale of it isn't subtle.

Read the second card twice. Our on-chain SessionClosed index puts provider-attested input at 54.8 billion tokens and output at 4 billion, and the index's estimated-throughput basis at 81.9 billion. The direct-payment figure is computed against that second basis: 48,138 tokens paid for with a user's own MOR, with everything else drawn from the emission pool. (figures as of 2026-10-23)

The session basis is no kinder. Direct pay accounts for 4,829 of the 423,941 sessions ever closed, and it barely qualifies as a live alternative: almost all of them ran between 8 March and 27 April 2026, they paid providers 66 MOR in total, and only 6 have used it since May.  The contract explains why. Session length is computed from your stake without reading the direct-payment flag at all, so paying buys exactly the compute staking would have handed you free, and no interface offers the option anyway. (figures as of 2026-10-23)

The capital side runs the same way. Morpheus captures the staking yield on deposited stETH at roughly 3% and pays capital providers a blended pool rate of 5.1% in MOR. The buyback engine that's meant to close the loop is running at 8.3% of emissions on a 30-day average. That gap is the whole sustainability question in one ratio. (figures as of 2026-10-23)

> **Fact:** Morpheus has served 81.9 billion inference tokens on our index's estimated basis, of which 48,138 were paid for by a customer. (figures as of 2026-10-23)
>
> **Take:** the marketplace works, the attestation works, and the providers show up. What hasn't arrived is a paying customer at any scale the chain can see, and the emission schedule covering for that runs down to tail levels after roughly year 16. This is a well-built network on a 16-year runway, not a business yet.

## NEAR: Your Yield Is the Invoice

NEAR shipped the cleanest articulation of the model at the end of July 2026, and co-founder Illia Polosukhin described it on [Bankless](https://www.youtube.com/watch?v=lRuaC0JkWhU) without any spin at all: "you are effectively trading off your yield," with that yield "being paid for the AI inference and capacity under the hood."

The [terms of service](https://near.ai/terms-of-service) are blunter. NEAR AI's own document (the version dated 24 September 2026) defines staking rewards as those "routed, protocol-direct, to NEAR AI as consideration for the Services,"  and states that they "are not paid, credited, or otherwise made available to Customer as income, yield, or return." You keep the principal. The yield is the invoice.

The contract is plainer than the legal wording. "Protocol-direct" describes intent; the deployed mechanism is an ordinary NEAR staking pool at `nearai.pool.near` whose `get_reward_fee_fraction` returns 100.0% to the operator. It is a validator commission set to the whole reward. Nothing routes at the protocol layer, and nothing needs to: a 100% fee is the simplest possible expression of "your yield is the invoice", and it is readable by anyone in a single call. (figures as of 2026-10-19)

At the top of each tier the formula is simply staked NEAR divided by 100, paid monthly in dollars. Annualise that and the network is promising 12% of your staked token count, denominated in US dollars, against a nominal staking yield of about 5.3% before validator commission. Work the break-even and the yield only covers the credit when NEAR is worth somewhere between $2.30 and $2.60, depending on which commission you pay.

At the Starter floor it isn't close. Stake the 50 NEAR minimum and you get the same $5 a month as someone who staked ten times more, which works out to a credit worth roughly $1.20 per NEAR per year against a yield of about 0.053 NEAR. For that to fund itself, NEAR would need to trade above $22, and its all-time high was $20.44 in January 2022.

That tier is a customer-acquisition subsidy, and it's priced like one.

Inference runs on a different formula, and the documentation for it arrived late. When I went through the NEAR AI Cloud developer docs on 13 August 2026 there was no staking or credit-conversion page at all, though the [launch post](https://near.ai/blog/staking-for-near-ai) said inference credits convert "the yield your stake earns rather than the stake itself". By 8 October the docs carry a [Staking for Inference](https://docs.near.ai/cloud/staking-for-inference/overview) page and a calculator whose code sets the monthly credit at stake × NEAR price × staking APY ÷ 12, accruing per second, with unused credits carrying forward. The calculator calls its figures illustrative, and the parameters that convert reward units into credits sit behind an authenticated API, so a staker can model coverage but can't yet read the binding rate.

That formula moves NEAR's inference product into a different column. It still prices credits in dollars, but the dollars float with the token and the staking rate, so a falling NEAR shrinks the user's allowance rather than NEAR AI's margin. That's the trade the Morpheus gateway makes, with the staker carrying the gap. The fixed-dollar exposure worked through above now applies to agent hosting only.

### The Part of NEAR That Does Earn

Set the AI arm aside and NEAR has the strongest fee business of the three by a distance. NEAR Intents generated $53M in gross fees over the trailing year per [DeFiLlama](https://defillama.com/protocol/near-intents), and the protocol turned on the fee switch in February 2026. That's a working revenue engine attached to a chain that has run since 2020. (figures as of 2026-10-19)

The catch is what reaches the token. NEAR's own dashboard at [revenue.near.org](https://revenue.near.org) traces 2.47 million NEAR of all-time protocol revenue across its three collection wallets, of which 446,604 NEAR sits in the buybacks wallet. DeFiLlama puts trailing-year protocol revenue at $8.5M against those gross fees, so well under a fifth of what the swaps generate is retained. The network issues about 88,356 NEAR a day under the 2.5% inflation rate set by the October 2025 halving. (figures as of 2026-10-19)

Every NEAR the protocol has ever captured amounts to roughly four weeks of issuance, and the buyback slice runs to about five days. And NEAR's buybacks are held rather than burned, per DeFiLlama's methodology note, so the offset is a transfer to holders rather than a reduction in supply. Morpheus and Venice both destroy what they buy.

So NEAR has revenue, the revenue is growing, and it's nowhere near the size of the issuance it would need to offset. The AI product draws on a separate pocket anyway: your staking rewards, which are themselves newly issued NEAR.

## Venice: A Perpetual Dollar Claim

Venice went furthest. Lock staked VVV in escrow, mint DIEM along an exponential cost curve, and each DIEM you stake pays you $1 a day of API credit for as long as you hold it. A fixed daily budget against the full model catalogue, resetting whether you spent yesterday's or not. The mechanics are pulled apart in [the DIEM deep dive](https://ownyourmind.ai/tokenomics/diem-venice-perpetual-ai-credit/).

Two design choices deserve credit. Venice caps the total obligation: the DIEM supply target reached 40,000 on 14 September 2026, which puts a ceiling of about $14.6M a year on the credit it owes at list price until Venice raises the target again. And the escrow is reversible at any time by burning the DIEM, so the lock-up is a decision you can undo rather than a cliff you fall off. Neither NEAR nor Morpheus has an equivalent cap on aggregate exposure.

Venice also runs the strongest offset of the three. It has cut emissions eight times since August 2025, down to 2M VVV a year from 1 October 2026 (an on-chain `setEmissionRate` call), and it buys VVV back off subscription revenue through two separate burners that have removed about 287,000 VVV between them since December 2025. Add the monthly discretionary buyback to the programmatic per-event burns and July 2026 came to roughly 38,100 VVV against 254,795 VVV issued, about 15% coverage.  That's the best ratio in this comparison and it's still one dollar of burn for every seven of issuance.

The exposure is the perpetuity itself. A $1-a-day claim that never expires is a fixed liability in dollars sitting on a company whose funding asset is a volatile token, and Venice still runs a roughly 20-person operation with no published board and no independent privacy audit. The escrow mechanism is elegant. What stands behind that escrow is the question, and it's the one the next section takes apart.

## The Rest of the Balance Sheet

Erik Voorhees got there before the critics did. Explaining the DIEM mint curve [on X](https://x.com/ErikVoorhees/status/1954293990146428966) on 10 August 2025, he wrote that it "sets a natural asymptote on DIEM supply (required since each DIEM is a liability of Venice)."

That's the founder using the accounting word, in public, as the design rationale for the cap. It also exposes what I've been doing wrong for most of this piece, and what most coverage of these networks does wrong: judging the staking product as though it were a standalone business. At NEAR and Venice it's a division inside a company that earns money elsewhere, and an obligation is only half a balance sheet.

### Venice: A Sized Liability Against a Growing Burn

Take the liability seriously first. At the 40,000 DIEM supply target, Venice owes about $14.6M a year of inference at its own list price, in perpetuity, and it can't inflate that away because the claim is fixed in dollars.

Now the other side of the ledger:

- **Revenue-funded burns.** July 2026 ran roughly $445,000 across the discretionary monthly buyback and the programmatic per-event burner, an annualised rate near $5.3M and rising. That's about 40% of the retail obligation, paid for out of subscriptions rather than issuance.
- **The obligation costs COGS, not retail.** Venice serves that credit at its own cost of inference. The burn covers the entire DIEM float's cash cost at any inference gross margin above roughly 60%, and falls short below it. Venice doesn't publish the margin, which makes it the single most useful number a reader could ask them for.
- **Equity capital.** Venice closed a $65M Series A at a $1B valuation in July 2026, led by Dragonfly with Coinbase Ventures and North Island Ventures. That's company equity rather than a token sale, so it adds no VVV supply overhang.
- **A cost-side attack.** Management says the Series A funds owned GPUs and data centres to replace leased compute. Every dollar off the cost of serving inference is a dollar off the perpetuity's running cost, permanently.
- **Self-reported profitability.** Voorhees stated at the raise that Venice is profitable on annualised run-rate revenue above $70M. That's a company claim with no independent confirmation, and it should be read as one.

So the honest version of the Venice sustainability question is narrower than "can a startup afford a perpetuity". It's whether burn growth plus falling unit costs outrun a hard-capped obligation. Framed that way, Venice is the only one of the three with all three levers moving in the right direction at once.

The cost is a new one. The Series A puts an equity preference stack ahead of VVV holders for the first time, so economics can now accrue to preferred shareholders while the token sits behind them. Our [Venice review](https://ownyourmind.ai/projects/venice/) marks that down in Value Accrual.

### NEAR: Three Legs, One Missing Linkage

NEAR makes the consolidated argument more explicitly than Venice does, and it deserves to be put properly rather than waved at. Asked on Bankless how the token captures value, Polosukhin answered "this is AI money" and named three legs that are meant to work together:

- **Sovereign security.** The chain itself, which is what lets the token act as a store of value.
- **Access.** Staking, which converts holdings into inference and agent capacity.
- **Transaction capture.** NEAR Intents, which takes a fee on the volume flowing through the stack.

He then closed the loop out loud, describing a buyback mechanism funded from the revenue the protocol generates on those transactions. So NEAR's fuller answer to "who pays for the compute" runs wider than the staker's yield: the Intents engine buys the token back while the staking layer hands out the utility, and the two legs are meant to support each other.

The engine is real. NEAR Intents turns over $53M of gross fees a year and retains $8.5M of protocol revenue per DeFiLlama, with the fee switch live since February 2026. (figures as of 2026-10-19)

There's a second revenue line that gets missed entirely, and it's the one most relevant to this article. NEAR AI Cloud is sold wholesale to consumer brands: Polosukhin named Venice and Brave as customers, plus the government of Bermuda and the remittance firm Abound, with Brave's user base put at over 100 million on air. All of that is self-reported and unaudited, but the shape of it matters. Venice buys its most private inference from NEAR, which means two of the three networks in this comparison sit in a supplier relationship, and some of the same GPU hours stand behind both companies' promises.

The demand-side argument is sharper than the retail framing suggests too. Polosukhin's case is that a fully autonomous business holds NEAR on its balance sheet so it never runs out of inference and stops being able to operate. That converts an operating expense into a treasury asset nobody can switch off, which is a structurally different reason to hold a token than yield-chasing, and it's the most interesting demand thesis any of the three has articulated.

Now the tests, because a narrative of integration isn't the same as a mechanism.

- **The buyback is held, not burned.** Per DeFiLlama's methodology note, captured Intents revenue buys NEAR on the open market and returns it to holders rather than destroying it. That's a transfer, and it leaves supply where it was.
- **Nothing routes Intents revenue to the AI obligation.** The two legs are argued as one system and wired as two. Polosukhin conceded the accounting gap himself in the same interview, calling it an open question how to equate revenue with the staking approach, since the staking approach generates revenue "in this yield way" rather than by direct payment.
- **The obligation is small, and the documents are the wrong place to look for it.** No figure appears in NEAR AI's launch post, its terms of service, the developer documentation or revenue.near.org. The pool answers it directly: `nearai.pool.near` holds 1 million NEAR across 7 delegator accounts, with a single account holding 98.3% of it and the pool's own operator treasury holding most of the rest. At the tier-top formula of staked divided by 100 per month, that bounds the aggregate obligation at roughly seven thousand dollars a month. The product is real and the exposure is currently trivial. One caveat: the dominant depositor may itself front many end users, which the pool cannot show. (figures as of 2026-10-19)

One number from that interview went unchallenged. The host read off revenue.near.org that 20% to 50% of NEAR emissions were being "captured and burned" by the protocol, and nobody corrected the burned half. Buybacks are held. On the capture side, all-time protocol revenue across the three collection wallets comes to 2.47 million NEAR against issuance of about 88,356 NEAR a day, so the cumulative total is roughly four weeks of issuance even if a recent month's ratio reads far higher.

### Morpheus: No Second Business

This is where the consolidated view is unkind. Morpheus has no equity arm, no subscription line and no separate fee engine. Its only material revenue is the yield captured on capital-provider deposits, and that collapsed along with TVL.

Which changes the comparison. On the standalone product all three look subsidised. On a consolidated view Venice has the clearest path, NEAR has the biggest engine wired to the wrong wheel, and Morpheus is the one where the critique lands in full, because for Morpheus the staking product is the whole business.

> **Fact:** Voorhees designed the DIEM mint curve as a supply asymptote specifically because each DIEM is a liability of Venice.
>
> **Take:** a founder who books his own token as a liability and caps it accordingly is doing something the other two haven't. NEAR's obligation is uncapped and unquantified in public, and Morpheus has no second balance sheet to lean on. The obligation was never the problem. An obligation left unbounded and uncosted is.

## The Peg Decides Everything

Here's the distinction I keep coming back to, and it's the one the marketing on all three sides skates past.

A staking entitlement denominated in dollars is a short position on your own token. One denominated in capacity isn't.

Morpheus promises you a slice of whatever the network can serve. If MOR halves, your slice is unchanged and the capital behind that slice is worth less, which is a loss you took as a token holder, not a promise the protocol broke. If inference gets cheaper to serve, the same emission buys more capacity and every staker's slice quietly grows. The mechanism self-adjusts because it was never denominated in anything external.

That holds for consumer-node sessions. The gateway is a third category, and it's the one most API customers actually touch. Its daily credits are USD figures computed as your share of builder emissions multiplied by the MOR spot price, and the documentation is explicit that price moves change the dollar value of your allowance. So a gateway staker's entitlement does fall with the token. The difference from NEAR and Venice is who absorbs that: they promise fixed dollars and the issuer eats the gap, while Morpheus hands the shortfall to the staker and carries no fixed liability at all. Cleaner for the protocol, worse for the user, and a distinct position from either column above.

NEAR and Venice both promise dollars. If the token halves, the obligation is unchanged and the yield funding it drops by half in value, so the gap the issuer has to cover widens exactly when the issuer can least afford it. NEAR hedged this with the phrase "at current policy" in its own announcement, which is an honest way of saying the divisor can move. Venice hedged it with the supply cap.

Neither hedge is free. A policy that can be rewritten is a subscription, not a property right, and the reason people locked capital was to stop being a subscriber.

## What the Subsidy Actually Costs

Before we go bearish on all of it, the headline subsidy is smaller than the numbers suggest, and this is the part critics get wrong.

A dollar of credit costs the network its wholesale price for serving that inference, and wholesale sits a long way below retail. NEAR AI Cloud's own catalogue on 13 August 2026 lists DeepSeek V4 Flash at $0.17 per million input tokens and Kimi K2.6 at $0.81, against the $15 per million that our [Render, Akash and io.net comparison](https://ownyourmind.ai/tokenomics/render-vs-akash-vs-ionet/) recorded for a closed frontier model. So a $5 monthly credit is perhaps a couple of dollars of actual cost, and the accounting subsidy overstates the cash subsidy by a wide margin.

### What a Burn Does and Doesn't Pay For

Separate two things that get run together, because the difference decides how much credit the burn deserves.

A buyback-and-burn doesn't pay an electricity bill. Somebody settles that in fiat, and on Morpheus and NEAR that somebody is funded by newly issued tokens. What the burn does is compensate holders for the dilution those tokens caused. Paying the cost and offsetting who bears it are different transactions.

They converge at one specific point. If a network burned exactly as many tokens as it issued, bought with customer revenue, holders would end the year undiluted and the customer's cash would have cancelled the tokens the providers were paid in. At that point the revenue has funded the compute, with the token acting as the medium. That's burn-mint equilibrium, and it's why the coverage ratio above is the number that matters rather than the burn's headline size.

Nobody is close. Venice runs the best of the three at roughly 15%, so about a seventh of its issuance is offset. Morpheus sits at 8.3%, and its buyback runs on yield from capital-provider deposits rather than on customer money. NEAR's 2.6% is customer revenue, and it's held rather than burned, so it doesn't reduce supply. (figures as of 2026-10-23)

That cuts the criticism down but doesn't kill it. Two dollars of real cost multiplied across a user base, funded by printing, is still funded by printing. The [Chutes analysis](https://ownyourmind.ai/tokenomics/chutes-bittensor-revenue-machine/) put a number on where this ends: strip the emission subsidy out and the unsubsidised break-even price lands above what centralised competitors charge. Every emission-funded compute network eventually has to answer that same question.

## The Bull Case

I want to give this its proper weight, because the sceptical read is easy and the constructive one is harder to argue.

- **The capital comes back.** This is the structural difference from a subscription and it isn't a small one. Unstake and you hold the same token count you started with, which is a claim you can't make about twelve months of frontier API payments.
- **The economic cost is smaller than the headline.** A chunk of any proof-of-stake yield is compensation for dilution you'd suffer anyway. On a chain issuing 2.5% and paying about 5% nominal, the yield you hand over is worth nearer half the headline rate in purchasing power.
- **You get something the frontier labs won't sell you.** NEAR AI Cloud returns an Intel-signed attestation on every request and added [independent verification through Intel Trust Authority](https://near.ai/blog/near-ai-cloud-integrates-intel-trust-authority) on 12 August 2026. Venice runs four privacy tiers including hardware-attested modes. Morpheus is permissionless at the provider layer. None of that is on the menu at OpenAI.
- **The entitlement is a call option on the network.** If the network works, the token appreciates and your inference habit gets cheaper at the same time. The two payoffs are correlated in your favour, which is unusual.
- **The cost of servicing the promise falls every year.** Open-weight models are closing on frontier quality at a fraction of the price, and that trend does more for these balance sheets than any tokenomics tweak could.

That last point is the strongest one available to any of these networks, and it's the one I'd build a bull case on. A perpetual dollar-denominated claim written in 2026 gets cheaper to honour every year that open-weight inference deflates. Venice's DIEM, viewed coldly, is a geared wager on inference costs falling faster than the promise does.

## The Bear Case

- **Nobody is paying.** Morpheus's on-chain index shows 48,138 tokens paid for with user MOR against 81.9 billion served on the same basis. NEAR's AI arm doesn't touch the Intents revenue. Venice doesn't publish revenue at all, and the on-chain buybacks are the only figure anyone can verify. (figures as of 2026-10-23)
- **The peg is a liability.** Two of the three promise dollars out of a token-denominated funding source, which is the wrong way round when the token is the volatile leg.
- **The rules are unilateral, with one exception.** "At current policy" is doing heavy work in NEAR's announcement, and Venice moved the DIEM supply target four times this year. Morpheus is the exception: `getTodaysBudget` and `stakeToStipend` are live functions on the deployed Diamond, so the entitlement formula is contractual even though the pool parameters feeding it stay under multisig control. On the other two you are a customer of a policy rather than the holder of a claim.
- **Emission tapers are scheduled, demand isn't.** Morpheus's emission declines by about 2.5 MOR a day and reaches tail levels around 2040. Venice has cut emissions seven times. NEAR halved inflation in October 2025. Every one of those cuts shrinks the pot that pays for the inference, and none of them are matched by a committed transition to fee funding.
- **The governance record isn't spotless.** NEAR pushed the inflation halving through after the community vote failed to reach threshold, implemented by validator adoption instead. Chorus One called it "a dangerous precedent". Good outcome, bad process, and it tells you what happens when a policy and a vote disagree.

## The Scissors: Cheap Inference Cuts Both Ways

Now the forward-looking part, flagged as speculative up front.

The bull case above rests on inference getting cheaper. So does the bear case, and that's the trap. If the price of a million tokens of frontier-class open-weight inference falls far enough, the cost of honouring a $1-a-day claim collapses, which is good for the issuer. The demand for locking capital to obtain that claim collapses with it, which is fatal.

Nobody escrows 500 VVV to secure a daily allowance of something that's nearly free. Nobody stakes 2,000 NEAR to avoid a bill that rounds to nothing. The entitlement only holds value while inference has a price people want to avoid, and the entire open-weight trend is pushing that price toward zero.

Which suggests the survivors are the networks selling something that stays scarce when tokens per dollar stop being scarce.

That's privacy, verifiability, censorship resistance, and the ability to run a model nobody can take away from you. Attested confidential inference has a durable premium because the alternative is handing your prompts to a company that monetises them. Commodity token throughput has no premium at all. On that test NEAR AI and Venice are positioned better than a pure marketplace, because both sell a verifiable privacy property rather than tokens per dollar, and the [private inference ladder](https://ownyourmind.ai/why-deai/private-ai-inference/) is where that value actually sits.

**How firm this is:** the direction of inference deflation is well evidenced by open-weight pricing, but its timing and the size of the demand effect are our judgement.

## What Would Actually Fix It

The diagnosis fits in one sentence: in all three networks the consumer stakes and the provider gets paid in printed tokens. Kyle Samani's 2018 work-token essay for Multicoin argued the opposite arrangement, where service providers stake for the right to earn and "end users don't ever need to purchase the token". Our three have inverted it, and the inversion is precisely why more usage doesn't move the token.

That's a design problem, and design problems have known fixes. Five of them are already running in production somewhere in this sector.

### Burn on the Paid Path

The most consequential change is also the most proven. Keep the dollar-denominated experience the user actually wants, then fund it by burning tokens at an oracle rate instead of by diverting yield.

Helium runs it today: a Data Credit is fixed at $0.00001, non-transferable, and created only by burning HNT at the Pyth-reported price. Akash shipped its own version in March 2026, where AKT is taken in at the oracle price to mint soulbound ACT, a compute escrow denominated in dollars but collateralised by the token. Its AKT moves into a vault as backing rather than being destroyed, so it's a circulating-supply reduction, and the on-chain permanent-burn counter still reads zero. The user gets price stability. The token absorbs the volatility on the way in, and usage removes supply instead of diluting it.

Applied here, a dollar of inference should be a dollar of token burned rather than a dollar of yield redirected. That single change flips consumption from supply-neutral to supply-reducing, which is the mechanism that connects adoption to price.

Our [Akash BME analysis](https://ownyourmind.ai/tokenomics/burn-mint-equilibrium-what-changes-for-akt/) is also where the honest caveat lives. Burn-mint equilibrium is inflationary in a falling market, and Akash's burns still offset only a fraction of issuance, even after Proposal 322 halved the inflation ceiling to 4% in May 2026. It aligns the product with the token. It doesn't manufacture demand that wasn't there.

### Charge the Supply Side for the Right to Earn

Only Morpheus asks its providers for anything, and it asks softly: pay is capped at the provider's own stake net of what it has already earned in the period, which forces earnings to be held rather than capital to be committed up front. Require a stake against the capacity a provider wants to serve, slashable on a missed service level, and two problems close at once.

Token demand starts scaling with expected future revenue rather than with the size of the subsidy, which is the whole point of the work-token model. And it puts collateral behind quality, which is the exact blocker Polosukhin named on Bankless when he explained why NEAR still runs its own GPUs: the SLA and quality problems for third-party providers are unsolved. A slashable bond is how every other infrastructure market solves that. Bittensor already charges on this side, burning TAO to register a subnet at a six-figure dollar commitment that doubles with each new registration.

### Clear the Capacity Instead of Administering It

Every fragility in these three designs carries the same signature. NEAR's divisor holds only "at current policy", Venice has moved the DIEM supply target four times this year, and Morpheus writes its budget rule into the contract while leaving the inputs to that rule under multisig control. A formula is a firmer promise than a policy document, and it still sets the rate by administration rather than by anyone bidding for capacity.

That's administered pricing, and administered capacity pricing has a long record of being gamed and then repriced. Electricity markets worked through this decades ago and landed on periodic capacity auctions against a published demand curve. The equivalent here is straightforward: the network offers a defined quantity of daily capacity, stakers bid their stake for it, and the clearing ratio is published every period.

Do that and the entitlement stops being a policy the issuer can rewrite and becomes a property right with a discoverable price. Which is what people thought they were buying when they locked the capital.

NEAR is the interesting case here, because it's already building this machinery one layer down. Polosukhin describes compute as a market with oil-like heterogeneity and electricity-like non-storability, calls the current wholesale market "a complete disaster" of opaque pricing and multi-year contracts, and frames Intents as the clearing abstraction: an underspecified request, solvers sourcing the GPUs, prices negotiated and settled on-chain. That is a capacity auction. It just clears the wholesale GPU layer rather than the staking entitlement sitting directly above it, where the divisor is still set by policy.

### Publish the Glide Path Off the Subsidy

None of the three has published a rule for stepping emissions down as fee coverage rises. Not a date, not a trigger, not a threshold.

A commitment as simple as "when fee-funded capacity passes 30% of served capacity, the emission share steps down by a fifth" would make the taper priceable rather than a cliff nobody can date. It would also force the number that matters onto a public dashboard.

### Make the Lock a Claim on Fees

The last one is the most familiar from DeFi. Vote-escrow designs pay a share of protocol fees that scales with lock duration, so the lock is durable because it's a claim on something the protocol actually earns.

Applied to inference, longer locks would earn a larger share of the fee-funded capacity pool rather than a larger grant of subsidised capacity. The distinction sounds academic and isn't. A lock backed by fees survives the taper. A lock backed by emissions unwinds the moment the emissions stop.

### What Any of This Does for the Price

Locking supply is necessary and nowhere near sufficient, and this is where most token-design writing stops one step early.

If the lock is attractive only because of a subsidy, the supply reduction is procyclical in the wrong direction. Emissions taper, the lock stops paying, holders unstake, and supply returns to the float exactly when the network can least absorb it. A supply sink that depends on the subsidy it's meant to replace isn't a sink at all.

The durable version prices the token as a multiple of something the network earns. Burn on usage and the float shrinks with adoption. Charge providers for the right to serve and the required stake rises with expected revenue. Pay locked holders out of fees and the lock has a yield that survives a taper. Each of those turns a number the protocol controls into a number the market sets.

> **Fact:** none of the three currently routes a customer payment into a permanent supply reduction on the inference path.
>
> **Take:** that's the one sentence separating these tokens from an equity-like claim, and it's fixable with mechanisms already running elsewhere in this sector. Until one of them ships it, the honest metric to judge all three on is fee-funded share of served capacity. Everything else is a proxy for how generous the emission schedule happens to be.

One counter deserves airtime, because the sceptical case can be run too hard. Subsidised acquisition is a rational strategy when the subsidy buys durable users below what the equivalent Web2 customer-acquisition cost would be, and by that standard emissions may be cheap money well spent. The test is what share of subsidised users stay once the subsidy tapers. None of the three publishes cohort retention, so nobody can run that test from outside, which is itself telling.

## Verdict

Staking for inference is a better deal than a frontier subscription and a worse deal than the marketing implies, and both halves of that matter.

As a distribution mechanism it's excellent. It removes the credit card, it turns holders into users, it gives a token a reason to be held that isn't speculation, and it returns your capital when you're done. If you're already long one of these networks and you use AI daily, staking for access is close to free money in the narrow sense that you were holding the token anyway.

As a funding mechanism it's half-built. Morpheus and NEAR still pay the compute bill by printing, the best burn coverage in the group offsets about 15% of issuance, and none of the three has published a rule for stepping emissions down as fees grow into them. Venice is the exception that shows what finished looks like, and even there the staking yield is still emission-funded. That's a runway, and runways end.

The repairs aren't exotic. Burn on the paid path, charge providers for the right to serve, clear the capacity rather than setting it by policy, and publish the glide path off the subsidy. Every one of those is already running somewhere in this sector. What's missing is a network that ships them on the inference path.

Which one wins depends entirely on the lens, and three lenses apply. On peg durability Morpheus takes it, because a capacity-denominated claim can't break a promise it never made in dollars. On fee-engine scale NEAR takes it, by an order of magnitude, and then wires none of it to the AI arm. On the consolidated balance sheet Venice takes it, and not narrowly: a capped obligation, a burn growing out of subscription revenue, and equity capital pointed at the cost of serving the claim.

Venice is the only one with all three levers moving the right way at once, which is why I'd rank it first overall today despite running the peg risk this article spends most of its length warning about. That's the trade the sector hasn't resolved: the most durable mechanism sits on the weakest business, and the strongest business runs the riskier mechanism.

### What I'm Watching

Three things, in order:

- **Fee-funded share of served capacity.** On Morpheus this reads directly off the chain as direct payment against total tokens served, so it can't be spun. NEAR and Venice would both have to publish it. It's the single number that separates a runway from a business.
- **NEAR's binding inference rate, readable in public.** The docs now publish the formula and a calculator, but the live conversion parameters sit behind an authenticated endpoint. Until anyone can read them, a staker is trusting the illustration.
- **A stress test on the Venice perpetuity.** The supply target has moved four times this year. What happens to the $1 a day when revenue and obligation diverge is the test that matters. 

Anyone selling you staking-for-inference as free AI is describing an emission schedule, not a business model. That still has value while it lasts. Just size the position for what it actually is.

---

*I hold MOR as a capital provider with a six-year Power Factor lock, and stake MOR for inference access through a builder subnet. I hold NEAR. I hold staked VVV and a small DIEM position which I stake for daily inference credit. So I'm exposed to all three models described here, and none of this is investment advice.*