The idle rich Gaming PCs sit unused nine tenths of their lives. Dolphin wants the other tenth, and will pay in its own currency.

When we wrote about Venice AI last week, we flagged three names sitting in its slipstream and promised to return to them. The first is Dolphin, and it is the most interesting of the three, partly because it is the one whose economics can actually be checked.

Start with the observation that everyone in decentralised computing makes and almost nobody has monetised. There are tens of millions of powerful graphics cards in the world sitting inside gaming PCs, and they are doing nothing for more than nine tenths of their lives. Meanwhile the market for artificial-intelligence inference, the business of running a trained model to answer a question rather than building the model in the first place, is forecast to more than double by 2030. One side has spare capacity. The other side is starved of it. The trade appears to write itself.

It has been attempted before, and the attempts have mostly disappointed. Akash and io.net built marketplaces where buyers rent a specific machine for a specific window. That works for a data centre with guaranteed uptime. It does not work for a person who wants their computer back when they sit down to play a game. The result, across most of the sector, has been a great deal of subsidised idle hardware and rather little paying demand: supply looking for customers, financed by token emissions.

Dolphin has approached the problem from the other end, and this is what makes it worth the attention.

 

A lab that grew a network

 

The team began in 2023 as a model laboratory, publishing what it calls uncensored versions of leading open-weight models: Meta’s Llama, Mistral, Alibaba’s Qwen, Google’s Gemma. In plain terms, they retrain other people’s models to strip out refusal behaviour. The output is downloaded more than five million times a month on Hugging Face, and it powers the uncensored chat product at Venice AI. Whatever one makes of the philosophy, the distribution is real and externally verifiable, which is more than most of the sector can say.

Only afterwards did the network arrive. Dolphin Network pools consumer and small data-centre GPUs to serve inference requests for those models. The architectural choice is what the team calls peer-to-pool. Machines running the same model form a shared pool; jobs are assigned at random; there is no link between the buyer and any particular provider. Operators can leave whenever they like. Nobody has rented a virtual machine that vanishes mid-session, because nobody has rented a machine at all.

That single design decision is what makes the gamer’s graphics card addressable at all. 

 

Borrowed machinery

 

The token layer is where things get familiar, and a reader who has followed decentralised finance for a few years will recognise almost every component.

Node operators post slashable bonds, in the manner of Ethereum validators, forfeiting roughly four weeks of income in any confirmed case of cheating. Bonding more POD boosts rewards, in the manner of Curve’s liquidity gauges: six months of earnings bonded guarantees a 1.5 times multiplier, with a competitive path to 2 times for those who over-bond further. Stakers deposit into an auto-compounding vault, in the manner of xSUSHI, which conveniently means the staked receipt can itself be posted as operator collateral. Withdrawals pass through a cooldown and window, in the manner of staked Aave.

The design borrows from Ethereum, Curve, Sushi and Aave. It is a DeFi tokenomics stack wearing an AI costume, and it is rather well tailored.

There is genuine craft here. Paying operators in bonded tokens by default, and charging a 20 per cent fee on liquid claims that is routed to the staking vault, is an elegant answer to the mercenary-supply problem that has hollowed out competing networks. The reward engine deliberately contains no price oracle, denominating everything in token units, which removes an entire category of manipulation. These are the decisions of people who have watched previous cycles carefully.

The question is whether mechanisms built to solve liquidity problems in finance transfer cleanly to quality problems in compute. A Curve boost aligns a depositor with a pool. It does not, by itself, verify that a machine on the other side of the world is running the model it claims to be running. Dolphin’s answer to that is a fingerprinting scheme based on sampled output probabilities, which is a sensible approach to a hard problem.

 

The arithmetic

 

To the team’s considerable credit, they publish unit economics, and they are worth working through.

For inference on a Qwen 3.6 35 billion-parameter model, Dolphin says its network cost is $0.50 per million tokens, the cheapest comparable price on OpenRouter is $1.00, and it charges users $0.70. Nodes receive the $0.50. The remaining $0.20 is used to buy POD on the open market. The network undercuts the cheapest centralised provider by 30 per cent while still extracting net buying pressure on its own token. Payment is accepted in POD, ether, bitcoin, dollar stablecoins, monero and zcash.

Now apply the throughput. Dolphin reported 1.2 trillion tokens generated in a recent two-week window, across roughly a thousand GPUs. Annualise that and the network is running at something above 30 trillion tokens a year. At $0.20 of net buyback per million tokens, that is roughly $6m of annual purchases against a market capitalisation of $22m. At $0.50 per million paid to operators, it is roughly $16m of annual token issuance, which at the current price is around 6 per cent of maximum supply and something close to 70 per cent of the tokens presently in free circulation.

Both figures rest on the same assumption, and it is the assumption that matters.

Throughput is not revenue. Tokens generated for nothing are a cost, not a customer.

Much of that measured volume is synthetic data generation, which is to say Dolphin’s own workload, run to produce training datasets. It demonstrates that the machinery functions at scale, which is a real and underrated achievement. It does not demonstrate that anyone has paid. The buyback figure above is a ceiling that assumes every token processed is billed at full price to a third party; the emissions figure is not a ceiling, because the protocol pays operators from the treasury whether or not revenue arrives. The team is explicit that the two sides are decoupled by design, and that the network is intended to run at a loss while it bootstraps supply. That is honest. It is also, for now, a one-way flow.

 

The Apollo Crypto View

 

Dolphin is the rare decentralised-compute project that arrives with distribution rather than hoping to buy it, and the rare token design whose author has clearly read the last five years of DeFi post-mortems. The peer-to-pool architecture solves a real problem that its competitors have not solved, and the bonded-by-default reward structure is the most thoughtful answer we have seen to mercenary node supply.

None of that changes the fact that the flywheel described in the tokenomics has one component still on the bench. The network has proved it can generate tokens at industrial scale. It has not yet proved that anyone will buy them. Every mechanism downstream of that, the buybacks, the staking dividends, the reward multipliers, is contingent on paid demand arriving on roughly the schedule the treasury is spending against. That being said, if anyone should succeed in generating such paid demand it should be Dolphin, the structural advantage from their peer to pool model allows them to offer compute at genuine discounts due to its unique advantage of being able to capture idle compute without restraining suppliers. Dolphin has an extensive roadmap ahead of them where they are eventually aiming to shard model inference across many machines in parallel in future, if successful this would cut the cost of inference down even more drastically. 

There is also the small matter of size. Across every venue, roughly $130,000 sits within 2 per cent of the mid-price. That is a binding constraint for any professional allocation. POD is soon to be listed on Coinbase so this might assist liquidity here.

For those watching rather than acting, the tells are unusually clear. Watch for the public API opening and for buyback transactions appearing on Base. Watch whether paid throughput, as distinct from internal data generation, begins to show. And watch the treasury wallet. The mechanism design will be judged by none of these things, and the token by all of them.

 

 

Disclaimer: Apollo Crypto has exposure to POD 

This report (‘Report’) has been prepared for informational purposes only and does not constitute an offer to sell or a solicitation of an offer to purchase any security of financial product or service. This Report does not constitute a part of any Offer Document issued by Apollo Crypto Management Pty Ltd (ACN 623 059 227, AFSL 525760) or Non Correlated Capital (ACN 143 882 562, AFSL 499882), the Trustee of the Apollo Crypto Fund. Past performance is not necessarily indicative of future results and no person guarantees the performance of any Apollo Crypto financial product or service or the amount or timing of any return from it. This material has been provided for general information purposes and must not be construed as investment advice. Neither this Report nor any Offer Document issued by Apollo Crypto or Non Correlated Capital takes into account your investment objectives, financial situation and particular needs. The information contained in this Report may not be reproduced, used or disclosed, in whole or in part, without prior written consent of Apollo Crypto. This Report has been prepared by Apollo Crypto. Apollo Crypto nor any of its related parties, employees or directors, provides and warrants accuracy or reliability in relation to such information or accepts any liability to any person who relies on it. You should obtain a copy of the Information Memorandum, issued by Non Correlated Capital before making a decision about whether to invest in the Apollo Crypto Fund.