Peer-to-Pool economic design
Dolphin Network is designed as a peer-to-pool system to repurpose idle GPUs.
Each model runs across a pool of GPUs provided by the network.
This is different from most AI DePIN networks that link buyers of compute into a "session" where they rent a node directly from a provider.
On the supply side, nodes running the same model form a 'pool' that processes requests sent to that model. Nodes are randomly assigned jobs based on their availability and there is no direct link between the request sender and node provider.
Nodes are rewarded solely based on inference tokens processed with POD from the protocol treasury.
On the demand side, API users buy credit from the protocol directly.
Dolphin Network accepts payments in $POD, $ETH, $BTC, $USDC, $XMR & $ZEC
100% of all revenue sent to the protocol is used to buy back POD on the market - directly offsetting emissions
The decoupling of the buy and sell sides of the network means that more or less POD can be distributed to nodes than what is received from revenue.
A concrete example: Qwen 3.6 35B inference on Dolphin Network
- Current cost to run datagen.dphn.ai : $0.50 per 1M tokens
- Cheapest comparable price on OpenRouter : $1.00 per 1M tokens
- Dolphin charges users: $0.70
- Pays nodes: $0.50
- Net buyback pressure: $0.20 per 1M tokens generated
We undercut the cheapest centralized provider by 30% while still extracting $0.20 of net buy pressure per million tokens generated.
We believe this is the best application of DePIN as
1 - There is extremely high demand for AI inference
2 - The supply of idle gaming GPUs capable of running local AI models is very large. This type of network feels reminiscent of GPU proof-of-work mining, but with far greater potential revenue capture because the output has real commercial utility.
3 - Unlike many DePIN networks, location does not matter so we don’t have the coverage problem.
AI inference is largely location-agnostic: a few hundred milliseconds of latency make little difference to usability. This allows our network to connect consumers to compute resources across the globe, unlocking far greater scalability & utilization per node
4 - Liquid pooled compute is the only way to unlock the largest supply of GPUs we can target: gamers & enthusiasts.
It allows nodes to come online & offline at any time, without the fixed uptime requirements that P2P node rentals depend on.
Previous GPU DePIN projects required 1:1 consumer <-> node pairing, which does not work for idle GPUs like gaming PCs or datacenter cards where the operator may want to take back the machine at anytime. No one wants to rent a VM that disappears randomly when the GPU is taken back.
Token Alignment & Value Accrual
POD is the only valuable asset in the Dolphin ecosystem.
100% of revenue generated by the network is automatically used to buy back POD on the market.
We do not have any shareholder-based external equity structure & never will.
POD holders may stake their tokens in the xPOD vault to gain access to
- Direct auto-compounding dividends from network token buy backs
- Daily inference allocations that provide access to all models available on the network
- Subscription status in our Web chat, bots & other ecosystem apps
Our token design is inspired by the best parts of other projects that we found map well to our distributed inference & training network
- ETH style slashable bond deposits for node operators & validators
- CRV style reward boosting for node operators - up to 2x boost for bonding POD - above 1.5x to 2x is competitive based on the relative bond size to earnings ratios of other providers
- xSUSHI / yCRV style auto-compounding staking vault - not needing to claim rewards means that xPOD (staked Dolphin) can be deposited as collateral for the node operator bond
- stAAVE style cooldown period & withdraw window
- vlCVX / veCRV style bribe market for unused daily xPOD inference allocations - sell your unused inference allocations on the market to earn higher staking rewards
Bonding, Slashing & Reward Multiplier
Cheating is the largest problem facing a decentralized inference network.
A node operator can quietly swap to a smaller, quantized, or fake model and still earn rewards. Output quality collapses. Inference buyers leave. The flywheel never starts.
Dolphin Network uses slashable bonds to strongly align node operators with the value of the POD token while also making it economically irrational to cheat given that all node operators are liable to have a bond equivalent to 4 weeks worth of income slashed in any confirmed case of malicious activity.
Node operators earn bonded POD by default. Once a node has over 4 weeks worth of income held as bonded POD, they are able to choose between claiming bonded POD or liquid POD at the end of each weekly epoch.
Liquid claims attract a 20% fee that is routed directly to the xPOD staking vault - which grows the token balance of other stakers & bonded node operators.
Nodes will be able to further deposit xPOD into the bonding contract which will boost their rewards & allow them to validate other nodes on the network.
The POD reward multiplier determines how much additional yield an operator earns on top of their base node rewards. It is inspired by @CurveFinance's LP boost mechanism but adapted for a decentralized AI network with usage-based rewards, account-wide bonding and slashable operator bonds.
At a high level:
- Nodes earn base rewards from completed inference, validation, and related protocol work
- Accounts receive a multiplier on eligible node rewards earned by that account based on their bonded token balance relative to earnings
- The earnings denominator is based on a trailing multi-epoch average of the provider's base rewards, then updated with asymmetric smoothing: it rises quickly when their inference activity increases, but falls slowly when activity decreases
- Accounts that maintain over 3 months of earnings bonded and a minimum active bond of 50,000 POD can become validator-eligible
- Bonding the equivalent of 6 months of earnings (26 weeks) guarantees at least a 1.5× multiplier on rewards
- Bonding more than 6 months gives a competitive pathway up to 2×, based on both relative over-bonding versus other over-bonded operators and absolute over-bonding beyond the 6-month target
All calculations are done in POD units only, without any price oracle inside the reward engine. The bond is per account (wallet), and the resulting multiplier applies across all nodes owned by that account. Adding more nodes increases the account's base earnings and therefore requires proportionally more active bond to maintain the same weeks-bonded ratio.
Stay tuned for our paper 'Encrypted Live-Weight Proofs for Decentralized Inference' that we will be releasing tomorrow. This covers our lightweight verification system that lets us verify nodes are running the correct model across all hardware types - going beyond the standard TEE-based verification that can only be used on enterprise NVIDIA cards.
