Summary:
This proposal aims to pivot the protocol from pure DeFi to owned compute infrastructure, deploying 9x Mac Studio M5 Ultra operated by Cooper Labs to sell AI inference paid in USDp via x402, with net revenue flowing back to the DAO treasury.
Rationale:
The strategy we have pursued with USDp so far has not worked. Pure DeFi is a zero-sum game: the only thing that moves capital is governance token incentives, and those incentives buy temporary liquidity that leaves when the rewards stop.
We cannot incentivize like newer stablecoins. Parallel has been live since 2021 and 100% of the PRL supply is liquid: there is no token reserve to fund points or governance incentives. Newer stablecoins distribute unallocated tokens to bootstrap liquidity, but that only lasts a few months before it collapses and slowly dies. Very few stablecoins live since 2021 are still around today, precisely because they had no durable edge once paid liquidity left. Every incentive program we run is therefore paid in cash from the treasury. And what we spent to incentivize USDp on the secondary market cost us more than what the USDp backing returned to the protocol, even with co-incentive programs from partners. We were renting liquidity at a loss, with no durable revenue in return.
x402 story follows the same pattern. The main x402 payment volume today, both in transactions and in dollars, comes from AI inference via API, notably Blockrun and Cluster Protocol. The original plan was to go to the main merchants and have them integrate x402 with USDp, with exclusive discounts for their users. It would have taken months: due diligence, integrations, negotiations, before a single payment. The intermediate step we considered was launching our own merchant wrapping AI inference APIs, because that is where the volume is. But after 6 months and about $66K allocated, we would have made nothing. x402 is free, no transaction fees, and USDp balances of agents generate almost no revenue. Most agents would leave for other providers when incentives end: the same mercenary capital we see on stablecoin pools & DeFi. And to compete long term, we would have had to pivot the Parallel brand into a competitor of Blockrun and OpenRouter, building a B2C distribution brand from zero in a hyper-competitive market with zero margins.
So instead of controlling distribution, where competition is brutal and margins are zero, we control the layer below: compute. There is competition there too, but demand is such that there is room.
We are stopping the spending. No more liquidity incentives that do not stick, no more usage subsidies that do not convert. The DAO moves the unspent budget into real infrastructure: owned hardware that generates real revenue from real usage, settled onchain.
This is the first step of a product pivot: Parallel moves from a pure DeFi yield protocol to a payment and infrastructure platform, where USDp is used to pay for real goods and services (starting with AI inference via x402) instead of being parked in liquidity pools. The compute pilot is the first physical asset behind that thesis.
Budget: this proposal cancels the unspent budget lines below and approves a new budget of up to $125,000 for AI compute, funded from the DAO Treasury:
- x402 cashback program: ~$60,000 allocated, $0 spent, cancelled (the remaining ~$6,000 gas sponsoring for x402 transactions stays in place)
- DeFi liquidity incentives: $96,000 allocated, $48,000 already spent, the remaining $48,000 is stopped and cancelled
- New budget approved by this proposal: up to $125,000 one-time for the compute pilot. Hardware prices vary significantly by country, so the machines will be purchased across several countries where prices are best; the ceiling covers the price range, and any unspent remainder is not spent and stays in the treasury
Impact on USDp supply: if this proposal is approved, the USDp supply will decrease significantly as liquidity incentives end and secondary-market demand normalizes. This is expected and normal: part of the current supply is supported by cash incentives that cost more than they return. What matters is durable demand from real usage, not a supply number sustained by paid liquidity.
Why Apple Silicon:
The pilot uses 9x Mac Studio M5 Ultra (36-core CPU / 80-core GPU, 256GB unified memory), operated by Cooper Labs in Tbilisi, Georgia. Electricity is $0.05-0.087/kWh, roughly 5-10x cheaper than Western European rates, and the machines are insurable. Models served: Qwen3.8-27B (8-bit), Glimmer 30B, A3B.
Capacity estimates: the plan is based on Qwen3.8-27B 8-bit (mlx-community/Qwen3.8-27B-8bit) on vllm-metal, our production stack. Measured references: 38.3 tok/s on M3 Ultra (MLX 8-bit, 262K context) and 9.01 tok/s on M5 Pro (early vllm-metal build, pre-v0.2.0 which brought a 3.6x throughput improvement). The M5 Ultra has ~1.5x the memory bandwidth of the M3 Ultra (1.2 TB/s vs 819 GB/s), which would extrapolate to ~56 tok/s. We use a conservative 30 tok/s single-stream estimate, below the bandwidth extrapolation, until the pilot validates it with real telemetry. Aggregate capacity: 2,592 tok/s = 30 tok/s single-stream x 16 batch x 0.6 efficiency x 9 machines. These numbers are deliberately pessimistic and will be validated by the pilot; if real throughput is lower, we fail fast.
Economics (pessimistic scenario, locked):
- CAPEX: ~$105,800 at US reference prices (9x 256GB at $10,799 = $97,191 + $2,500 network equipment + 15 kWh battery backup at 16,000 GEL ≈ $6,100). Hardware prices vary by country, some markets list the same machine above $13,000, so the machines are bought across several countries at the best available prices, and the budget ceiling is set at $125,000. The economics below use the US reference prices; if procurement lands at the top of the machine price range, monthly operating costs rise by roughly $550 and break-even moves from ~29.8% to ~34% utilization, still leaving the pilot profitable at 50% utilization and loss-bounded at the bottom
- Operating costs: $3,605/month (36-month depreciation $2,939 including the battery + electricity $113 + cooling $11 + maintenance 5%/year of machine CAPEX $415 + internet $127)
- Sale price: $1.80/M blended, below current market
- Break-even: 29.8% utilization
- Net monthly profit to the DAO:
- At 30% utilization: +$23 (essentially breakeven)
- At 50% utilization: +$2,442 (accounting payback 43 months)
- At 100% utilization: +$8,489
- Bounded loss at 20% utilization: -$1,186/month (machines remain resellable)
- Cost price: $0.54/M at full load (75% discount vs the $2.125 blended market price)
- Real cash flow at 50%: +$5.4K/month (depreciation is non-cash), for a cash payback of ~20 months
To be explicit: revenue of $1.80/M at 50% utilization is approximately $6,047/month, minus operating costs of $3,605/month, leaves a net profit of $2,442/month flowing to the DAO.
Pricing model:
The pilot makes USDp a real payment rail for a real service, instead of a token parked in liquidity pools:
- Standard payment: market price
- x402 payment in USDp: cost price ($0.54/M at full load), a permanent structural discount instead of a temporary subsidized program
This replaces the planned cashback program with a structural discount that is sustainable because it is backed by owned hardware at Georgian electricity costs. Every inference request paid in USDp is a real transaction: real usage, real settlement, verifiable onchain.
Pilot, not commitment:
This is a pilot, not a commitment. If utilization stays below break-even, the machines are resold and the loss is bounded (approximately $1,186/month worst case, plus the resale value of the hardware). If the pilot validates demand, the model scales: the hardware stack can grow, and larger models (Kimi K3 class) become addressable in a later phase.
Structure:
The DAO has no legal personality and cannot hold hardware. Cooper Labs (FZCO) owns and operates the machines under a contractual revenue-share arrangement. Onchain inference payments settle via x402 in USDp: what is processed is what is billed, verifiable end to end. A portion of the revenue will be offchain (direct B2B clients billed in fiat); that revenue is onramped and sent to the DAO treasury, and reported on the same dashboard. Revenue accrues to a dedicated Safe multisig, operating costs are deducted, and the net is transferred to the DAO treasury monthly.
How the DAO verifies the machines are working:
- Machines are served through Darkbloom, an existing AI inference provider specialized in Mac inference (an Eigen Labs-ecosystem coordinator with a model registry): real-time utilization and per-machine connectivity are visible to the DAO at any time through its dashboard.
- Later, when we run our own OpenRouter provider, the same telemetry is exposed: utilization, connectivity, and model availability per machine.
- Onchain revenue settles via x402 in USDp: what is processed is what is billed, so any diversion of machine usage shows up as a revenue mismatch onchain. Offchain revenue is reconciled through the public dashboard.
- Physical access is restricted to Cooper Labs operations in a dedicated commercial space in Tbilisi; the machines run protocol workloads only.
Machine environment:
- Dedicated commercial space in Tbilisi with restricted access, not a personal residence, to avoid residential-to-commercial reclassification and to match the “controlled facility” standard.
- 15 kWh battery backup: ~6 hours of autonomy at full load (2.5 kW), covering power cuts; smart PDU cycles outlets and auto-restarts machines on recovery. Cost: 16,000 GEL (~$6,100), included in the CAPEX above.
- Triple connectivity: fiber + Starlink + 5G modem, so the cluster stays reachable if one link fails.
- Machines insured with a Georgian provider (Aldagi or comparable): equipment against fire, theft, natural disasters; business interruption coverage discussed for revenue loss during downtime.
- A public dashboard will show, in simple terms: costs, revenue, net profit, utilization per machine, connectivity, and onchain x402 revenue. Same format as the transparency reports Cooper Labs already publishes.
Means:
- Human Resources: Cooper Labs handles hardware setup, operations, uptime, and telemetry (Darkbloom shows real-time utilization and per-machine connectivity; our own OpenRouter provider later exposes the same telemetry). Multisigners sign and execute the initial transfer and the monthly net revenue transfers.
- Treasury Resources: One-time budget of up to $125,000 in USDC and/or USDp from the DAO Treasury, allocated to Cooper Labs for hardware and network equipment purchase (CAPEX ~$105,800 at US reference prices including the battery; hardware prices vary by country and the ceiling covers the expected range). This is an investment in resellable assets, not a burned expense. The cancelled budget lines (~$60,000 cashback, ~$48,000 remaining liquidity incentives) are removed from the protocol spend, and this proposal approves a new budget of up to $125,000 for AI compute. Operating costs (electricity, internet, maintenance, ~$666/month cash) are covered by pilot revenue before the net is transferred to the DAO. No additional recurring treasury outflow.
Technical Implementation:
- Initial transfer: Transfer of up to $125,000 in USDC and/or USDp from the DAO Treasury multisig to the Cooper Labs entity for the purchase of 9x Mac Studio M5 Ultra 256GB and network equipment. The machines are bought across several countries where prices are best, and the final amount reflects actual invoices, capped at $125,000. The exact recipient address will be published in this thread before the Snapshot vote opens.
- Setup: Cooper Labs configures the machines (Qwen3.8-27B 8-bit, Glimmer 30B, A3B), connects them to the inference routing layer (Darkbloom provider first, own OpenRouter provider later).
- Billing: Inference is served and billed via x402 in USDp. Payments settle onchain.
- Revenue: Accrues to a dedicated Safe multisig. Monthly, operating costs are deducted and the net is transferred to the DAO Treasury. Offchain revenue is onramped and sent to the DAO Treasury.
- Telemetry: Real-time utilization per machine, connectivity, and onchain x402 revenue are public.
Voting Options:
- For the Compute Pilot
- Against / Rework the Proposal
- Abstain
Author(s): Cooper Labs
Community Poll:
- For the Compute Pilot
- Against / Rework the Proposal
- Abstain