Tokenomics

Core principle: The token is not the product. Network usage is the product. The token captures value from that usage and coordinates the participants who supply it. A fixed supply serves as a backstop, not as the primary value engine.

---

1. Token Details#

• Name: Perspective AI Token
• Symbol: $POV
• Canonical chain: Ethereum mainnet (ChainID 1)
• Operating chain: an L2, reached through that rollup's own native bridge
• Standard: ERC-20
• Decimals: 18
• Total supply: 21,000,000 (fixed, pre-minted)

Rationale for the fixed cap#

The cap is a credible-commitment backstop, not the source of value. Against a fixed supply, every revenue-funded burn is permanently deflationary as real usage tightens supply irreversibly. Value derives from usage and the buyback-burn sink it drives, not from scarcity narrative.

Why the token is not native to the operating chain#

The network's high-frequency machinery — per-epoch settlement, node bonding, slashing, marketplace operations — is only economical on an L2, and that is where it runs. Supply is a separate question. A token whose canonical contract lives on a rollup inherits that rollup's sequencer, its upgrade keys, and whatever pressure can be applied to whoever holds them. For a network that sells uncensored access and resistance to deplatforming, that is the wrong dependency to accept at the base of the stack.

So the 21,000,000 are minted on Ethereum and projected onto the operating chain through the rollup's own native bridge — not a third-party bridge, and not a cross-chain token standard. Both of those introduce a party holding rights to mint $POV that nobody locked, which is an inflation switch by another name. Here no bridge holds mint rights, and the L2 balance is backed 1:1 by tokens escrowed on Ethereum.

The L2 representation is not a second token and not a child token (§5). It is the same $POV, and it cannot come into existence without a matching deposit on the canonical chain.

What lives where.

Two scheduled flows connect them, both batched monthly:

• Emission down. E_t (§3.4) leaves the vault as a single bridge deposit whatever its size. Providers then claim on the L2 at L2 gas.
• Burn up. Surplus burn (§4.3) and slashed stake accumulate in an L2 sink and are withdrawn once a month to be burned on Ethereum, so totalSupply on the canonical chain is always the true circulating figure with no cross-chain footnote attached.

At prevailing gas the pair costs single-digit dollars a month, against a mechanism that moves the network's entire margin.

---

2. The Two-State Token Model#

There is one token with two states — not two separate tokens.

• vePOV is minted by locking $POV and decays over the lock period.
• vePOV is non-transferable and is burned to unlock.
• vePOV cannot exist without a corresponding locked $POV position.

Design note: Non-transferability does not prevent vote-buying. Bribe markets can form around any valuable governance lever, especially emission direction. Emission-control votes should be designed assuming bribe markets exist. Untradability prevents position-flipping; it does not prevent renting the decision.

---

3. Supply, Emissions & Net Issuance#

Allocations#

The allocation splits into two groups, because only one of them can dilute.

Emitted for work — 40%. Released only against verified work or verified payment, under a ceiling that can only decrease. Nothing in this group has a calendar.

Allocated — 60%. Fixed at genesis. No issuance decisions remain inside it.

Total: 21,000,000 $POV, fixed and pre-minted. Float at TGE is approximately 19%: the public sale, the liquidity pools, and whatever genesis tranche has been earned. Thin float is cheap to push around, and an unvested public sale is what widens it.

How to read the cliffs. A cliff here delays the start of linear vesting. It does not release a retroactive tranche when it expires. At month 12 a team holder's unlocked balance is zero and begins accruing from that point — there is no 12/48 catch-up block, and no schedule above has a step in it. Every unlock in this table is either a single event at TGE or a continuous line.

Why there is no staking-rewards allocation. Emission paid for holding is not yield. It is a transfer from holders who do not lock to holders who do, denominated in the asset being diluted and at full participation it nets to exactly zero. Worse, in a design where the token is spent on credits, a nominal staking yield bids against the product's own usage sink. Locking $POV earns governance weight and nothing else (§4.2, §7). Every bucket above funds work, capital formation or market depth and none of it pays for holding.

Why team is 15%. Equity is folded into the token, so this allocation is the founding team's entire ownership rather than a supplement to it. The 12-month cliff is the condition on that size, not a detail.

Notation#

Symbols used throughout.

What providers require, and what the subsidy is#

Providers price capacity in dollars, because electricity, hardware and colocation are priced in dollars. T is the total reward providers require in a given epoch to serve the demand routed to them.

T is discovered, not decreed. A periodic reverse auction has providers bid the lowest rate they will accept per unit of served work, and T is that clearing rate applied to the work actually routed. Bidding a rate rather than a capacity commitment is what keeps T at zero when demand is zero, as a commitment to standing capacity would create an obligation the demand gate is designed to refuse to fund. Governance does not set it: a votable reward target would be the most valuable lever in the system, and bribe markets form around exactly that (§2). While the operator set is small enough for an auction to be thin or collusive, a published reserve price bounds it. That reserve is the one bootstrap parameter here, and it is disclosed rather than hidden.

The subsidy is the wedge between what users are willing to pay and what providers require to show up. Nothing more than that:

$$\text{subsidy}_t = \max\big(0,\; T_t - B_t\big)$$

where B is revenue-funded buyback (§4.3). Every dollar of real revenue mechanically displaces a dollar of subsidy, so decay is an identity, not a parameter — when B ≥ T the subsidy is zero that epoch, with no calendar, committee or vote involved.

The two sides are deliberately asymmetric. T counts only work the network serves. B draws on margin from all revenue — subscriptions, credits, agent operations, and inference routed to centralised frontier providers.

So frontier routing earns margin and owes no emission. Routing GPT, Claude and Gemini centrally is not a decentralisation apology but coverage-accretive: every frontier request raises B without raising T, pulling the end of the subsidy closer.

Emission — the residual, under three bounds#

The subsidy is paid in $POV, so it must be converted, and the conversion is where reflexivity lives. Emission is the residual, clamped by three independent bounds:

$$E_t = \min\Big(\underbrace{C_t}_{\text{token ceiling}},\; \underbrace{k \cdot \text{verified\_usage}_t}_{\text{demand gate}},\; \underbrace{\tfrac{\max(0,\; T_t - B_t)}{\bar{p}_t}}_{\text{residual}}\Big)$$

• C — the ceiling, ratcheting down. Bounds dilution in total, not just per epoch (see below). If the ceiling binds, providers are paid less and some leave: capacity shrinking is survivable, a mint spiral is not.
• k · verified_usage — the demand gate. No demand, no emission. verified_usage is credits burned against served, undisputed requests (§6). Without it, a residual would happily fund capacity nobody used — the failure that hollowed out a generation of supply-side networks.
• p̄ — trailing average, never spot. A dollar obligation divided by a token price mints more tokens as price falls, adding sell pressure exactly during a drawdown. The average stops a wick triggering a mint; the ceiling stops the loop running.

How the ceiling falls#

The ceiling is not a schedule anyone maintains. It is a fixed fraction of whatever remains in the provider-incentives allocation:

$$C_t = \lambda \cdot R_t, \qquad \lambda = 1\% \text{ per month}$$

where R is the unspent balance of the 7,350,000 $POV, and R falls only as emission is actually drawn.

Two properties fall out of the form rather than needing to be enforced. The ceiling can never rise, because R never rises: C next month is λR of a balance no larger than this month's. And cumulative emission can never exceed 7,350,000, because you cannot draw more than what is left. Roll-forward is therefore automatic: tokens not emitted stay in R and remain available later, without any epoch's ceiling ever exceeding an earlier one.

The allocation is therefore never exhausted in the strict sense. It asymptotes rather than terminates, in the same way a halving schedule does. Fitted to the same cap, a four-year halving would give a first-era ceiling of ~76,600 $POV per month, within a few percent of the figure above.

Because emission is a min, the ceiling binds hardest in bad conditions and is barely touched in good ones. A long stretch of weak coverage draws at the cap; a healthy network draws far less and leaves the balance intact.

When emission stops#

Two different things stop it, and they are not the same event.

Coverage reaching 1.0 stops it on the demand side. When B ≥ T the residual is zero, so emission is zero that epoch no matter how much ceiling is available. This is the intended path and it is per-epoch, not a switch: if coverage later falls below 1.0, emission resumes as a shortfall top-up, bounded as always by the ceiling and the demand gate. "Self-extinguishing" means zero whenever revenue covers what providers require — not permanently disabled.

The allocation running down stops it on the supply side. As R decays the ceiling decays with it, and emission becomes negligible regardless of coverage. Providers are then paid only what revenue funds. If that is less than T, capacity shrinks to the demand real revenue supports.

The first is success. The second is the backstop, and it is deliberately slow: the ceiling exists precisely so that a sustained shortfall shrinks the network gradually instead of draining the allocation in months.

The invariant that keeps the subsidy un-farmable#

An operator who is also a user can buy credits, route the request to its own node, and collect the subsidy. With P paid per request, m the platform's gross margin, e the operator's true compute cost and s the per-request subsidy, that round trip nets s − m − e. Wash-farming is profitable exactly when s > m + e, so the binding design constraint is:

$$s_t < m_t$$

Deliberately tighter than the s < m + e breakeven, because e is unverifiable and falls as hardware improves — a bound that depends on it drifts the wrong way over time. Any discount d for paying in $POV enters directly as s < m − d, and a large discount can breach the invariant on its own. This is continuously monitored, not calculated only once.

Rewards are additionally weighted by distinct paying counterparties rather than by raw credits burned, so demand routed from a single source cannot dominate the distribution.

Net issuance#

$$\Delta\text{supply}_t = \underbrace{E_t}_{\text{residual subsidy}} - \underbrace{\text{burn}_t}_{\text{grows with usage}}$$

• Early phase: net-positive. This is dilution, and it is buying supply-side capacity. Holders are funding provider acquisition, which is a legitimate way to build a two-sided market.
• Mature phase: as the unit subsidy rate falls and coverage rises, E reaches zero and burn dominates.

The two numbers we publish#

Coverage — where we are:

$$\kappa_t = \frac{B_t}{T_t}$$

Bounded, readable as "revenue covers 82% of what providers require," and crossing 1.0 at exactly the moment emission reaches zero. Published monthly with B and T in absolute dollars, because a ratio is scale-free and a four-provider network satisfies one perfectly.

Unit subsidy rate — whether we are moving:

$$\sigma_t = \frac{r_t - c_t}{r_t}$$

Growth alone does not end the subsidy. If providers require 20% more per GPU-hour than revenue delivers, that gap exists at ten nodes and at ten thousand — T and B both scale with the network, so the wedge scales with it. Only four things close it:

1. Rising utilisation. The largest lever by far, and a routing and demand problem rather than a tokenomics one. A node paid per request but idle most of the time needs a high rate; fill the duty cycle and r collapses.
2. A falling risk premium, as track record accumulates and providers stop pricing in the chance we disappear.
3. Users paying more per unit through higher-value work — agent operations, verified classes — not through raising the base rate.
4. Hardware cost per token falling. Exogenous, real, and works in our favour over years.

Alongside both, we publish cumulative draw against the cap, ΣE as a percentage of the 7,350,000 ceiling. Bounded and monotone, so it cannot be flattered.

If coverage never reaches 1.0, the ceiling decays with the remaining allocation until emission is negligible, and the network shrinks to the demand that real revenue supports. That is a designed outcome rather than a failure mode.

---

4. Token Utility & Value Capture#

Credits — the usage sink#

Every billable action consumes credits: inference, preview-agent calls, agent operations, command execution.

• Credits are denominated in fiat-stable units (e.g. "$5 of credits") and are the sole unit of account for every billable action.
• Spending a credit meters usage. The $POV consequence happens downstream, in the revenue waterfall (§4.3), and is the same whichever way the credit was bought.
• End users need never see, hold, or understand $POV. Crypto-native users may interact with the raw layer.

Transparency rule: Cost legibility (showing what an action costs) is mandatory and a core trust feature. Token-volatility exposure is abstracted away from non-technical users. These are independent concerns.

What locking earns#

Locking $POV mints vePOV, which carries governance weight (§7) and nothing else. There is no yield of any kind: no emission for locking (§3.1), no share of revenue (§4.3), and no allowance of platform credits.

Why locking pays no credits. A credit is a claim on inference, so granting credits for locking is a real provider cost, and that cost has only two possible sources. Margin, which makes it a distribution to capital under a different name. Or emission, which is the staking bucket §3.1 rejects. Either way it is a payment for holding, in product rather than in tokens — and it would enter the anti-farm invariant as a 100% discount on every request it funded (§3.7).

What holders hold instead is the whole of the argument in §3.1 and §4.3: a fixed supply, a burn that grows with usage, and no dilution paid to anyone for the act of holding. Yield that is not funded by surplus is dilution wearing a costume.

Locking is not maximised on purpose. A lock sink and a deep market compete for the same tokens, and this design pays for depth twice — an unvested public sale to widen float, and 8% into protocol-owned liquidity. Thin float is cheap to push around (§3.1), and tokens locked to farm an incentive are tokens unavailable to absorb a sell.

Lock-gated product features may still exist: router priority, rate-limit headroom, early access to new capacity. They allocate capacity rather than purchasing it, they carry no claim on revenue or emission, and they are product terms rather than tokenomics. They belong in the product documentation, not on this page.

Node operators posting slashable stake are a separate case entirely (§4.3). They are doing work and carrying risk, and they are paid from the provider allocation.

Where revenue goes#

Revenue arrives as fiat or as $POV, and either way it runs through the same waterfall:

$$\text{revenue} \rightarrow \underbrace{\text{provider cost} + \text{OpEx}}_{\text{held in stables}} \rightarrow \underbrace{\text{open-market \$POV buy } (B_t)}_{\text{the margin}} \rightarrow \underbrace{\text{provider payment, up to } T_t}_{\text{displaces emission}} \rightarrow \underbrace{\text{burn}}_{\text{surplus only}}$$

Revenue must perform real open-market buying. Rewards must never be funded by relabelling emissions as "revenue."

Providers first, always. B pays providers up to T before anything is burned. This ordering is what makes the subsidy terminate. Route B anywhere else and E = T in perpetuity: the ratchet never engages, and the claim of the token being "self-extinguishing, with no calendar and no vote" becomes false.

Only the surplus burns:

$$\text{burn}_t = \max\big(0,\; B_t - T_t\big)$$

Which means that below coverage there is little or no revenue-funded burn. Burn exists from day one, but from sinks that are independent of provider coverage: the burned share of every agent clone (§5) and the tax on curation surfaces.

No revenue distribution to holders or lockers. Locking $POV earns governance weight (§7) and no yield of any kind (§4.2) — never a share of revenue, and never an allowance of product. Paying capital for the act of holding produces no work and funds no capacity, and it is no more legitimate when drawn from margin than when drawn from emission (§3.1). Node operators posting slashable stake are a different case: they are doing work and carrying risk, and they are paid from the provider allocation.

How OpEx is bounded. "Whatever operations require" is a quantity the company would size for itself. Three rules bound it instead:

• A fiat buffer, senior to everything below it. Six months of trailing-average OpEx, held in stables. Runway is never denominated in $POV.
• A contractual floor: at least 30% of gross margin routes to B every epoch, whether or not the buffer is full, and regardless of what operations would prefer. Reducible only by timelocked vePOV vote.
• A stated policy that exceeds the floor. Once the buffer is full, everything above OpEx routes to B. There is no equity class with a claim on retained earnings, so margin the company does not need has no shareholder to serve.

The floor rises to 75% of gross margin once coverage holds at or above 1.0 for two consecutive quarters — the point at which the subsidy is finished and the business is carrying itself.

Published monthly: the floor, the actual percentage routed, and the buffer's coverage in months. The gap between the promise and the practice is meant to be visible.

Enforced, not promised. The waterfall is an on-chain splitter with the floor and the timelock written into it. Governance may adjust the surplus band above provider coverage. It may not reorder the waterfall.

What providers are paid in#

Provider rewards are denominated in USD and settled in $POV, converted at the trailing-average price p̄.

Both halves carry weight. Denominating in dollars insulates a provider's economics from token volatility, which is what keeps their required return — and therefore the subsidy — as low as it can be. Settling in $POV is what makes the buyback mean anything: revenue → buy $POV → pay providers is the sequence by which revenue becomes token demand. Pay providers in stables and the buy never needs to happen, revenue converts into nothing, and $POV stops being a work token at all.

Providers may elect immediate on-market conversion at settlement. That reduces their exposure without changing the economics: the buy still occurs, and the sale is the same one they would have made themselves. A convenience, not a change in value capture.

Treasury rule. Neither the protocol nor the company holds $POV as operating reserve. Provider-cost and OpEx portions convert to stables on receipt, with no discretionary "we'll sell later" — runway denominated in your own token falls hardest exactly when you most need to spend it.

Margin model#

The platform runs on cost-plus-margin pricing — a low-margin utility, not a high-margin product. Holder value comes from volume × sink, not from per-call margins.

Every revenue line contributes to the buyback: subscriptions, credit purchases, agent operations, and the margin on inference routed to centralised frontier providers. There is no equity class holding a competing claim on retained earnings, so margin above what operations require has no shareholder to serve — it routes to B (§3.2).

Across every payment rail, the burn is bounded by margin above provider coverage. Providers are paid in dollars-equivalent, so the portion of revenue covering their cost cannot be burned without being re-minted somewhere else. The payment rail changes timing, processing cost and censorship-resistance; it does not increase value capture.

The platform does not compete with subsidized incumbent pricing on price-per-token. It competes on privacy, uncensored access, user ownership, and resistance to deplatforming.

---

5. Model & Agent Marketplace#

No tradable ownership shares. No child tokens. Models and agents are curated by measured demand, not by votes or ratings.

Permissionless listing#

Anyone may publish a model or agent. The market funds what it uses; governance does not vote model-by-model.

Curation by revealed demand#

No ratings or review system is operated (sybil-attackable; requires a human arbiter). Instead:

A paid, completed, undisputed request is the quality signal. Faking it costs real money.

Provider incentives flow to models/agents by measured metrics — credits burned, uptime, completion rate — never by popularity vote. Governance sets the formula; the data selects recipients.

Clone / fork economics#

A two-layer structure ensures recurring rather than one-shot demand:

1. Clone = acquisition. Pay $POV once to instantiate a private copy of a verified or community agent. Split: portion burned, portion to the original creator.
2. Run = recurring. Every operation the cloned agent performs burns credits — the sink that scales with usage.
3. Creator royalty. Original creators earn an ongoing cut of credits burned by every clone of their agent.

$$\text{preview} \rightarrow \text{clone (burn + creator cut)} \rightarrow \text{run (recurring burn)} \rightarrow \text{creator royalty on all clones' usage}$$

What the user buys: not open weights (free), but configuration, hosting, uptime, and zero-terminal convenience. Defensibility is convenience for non-technical users, not DRM. Lock-in on technical users is not attempted, as it would undermine open-source credibility.

---

6. Trustless Verification#

What the network is asked to do#

Operators run two classes of work. Both draw on the same provider-incentives allocation and the same T discovery, so the class of work changes what the network can sell, never how much it can print.

They are listed in the order verification reaches them, which is also the order they arrive:

The reward formula below applies cleanly to inference and partially to agent execution. The network therefore opens on inference and adds agent execution as those workloads containerise.

How work is verified#

Reward eligibility for nodes and providers is computed from on-chain-verifiable signals only.

$$\text{node reward} \propto \underbrace{\text{credits burned (paid demand)}}_{\text{on-chain, sybil-costly}} \times \underbrace{\text{probe pass rate}}_{\text{uptime}} \times \underbrace{\text{stake at risk}}_{\text{slashable}}$$

Engineering note: LLM inference is non-deterministic (sampling, floating-point, batching). Redundant verification cannot byte-compare outputs. Pin deterministic settings (temp 0, fixed seed, pinned kernels) for verification runs, or compare semantic similarity above a threshold.

---

7. Governance#

First principle#

Token holders govern the rules of the game and where the money goes. The team and contributors govern day-to-day execution. The market, via measured demand, decides which models win.

Three tiers#

Tier 1 — Binding on-chain votes (vePOV): the ceiling schedule C (downward only), treasury spend above threshold, incentive split between categories, the surplus band above provider coverage, fee parameters, contract upgrades.

Three things are deliberately not votable, because making them votable would create the levers bribe markets price: the provider-first ordering of the revenue waterfall (§4.3), the direction of emission between providers — which is settled by revealed demand rather than by preference — and the reward target T, which is discovered by auction (§3.2).

Tier 2 — Signaling votes (off-chain, non-binding): roadmap priorities, integration prioritization, model-category direction. The team retains how and when.

Tier 3 — Pure execution (no vote): hiring, security response, bug fixes, daily operations, emergency pause.

Model additions#

Not a vote. Permissionless listing plus measured-performance rewards. Governance sets only the guardrails a model must meet (safety policy, genuine open-weight requirement, minimum performance bar, node-hardware compatibility) — never individual choices.

Progressive decentralization#

Launch begins with greater foundation/team control (thin initial float is cheap to attack), accompanied by a published, committed schedule of which powers migrate to vePOV at which milestones. The schedule itself functions as a trust signal.