Why Decentralized AI Compute Needs Two Assets, Not One

The Bittensor network emits approximately $8 in TAO tokens for every $1 of real AI revenue it generates, a 7.6-to-1 ratio that reveals a structural flaw in decentralized AI compute networks. A developer argues that this "subsidy treadmill" stems from forcing a single token to serve as both a utility settlement mechanism and a speculative equity instrument, creating a destructive feedback loop where token price drops drive away contributors, reduce service quality, and trigger further emissions. The proposed fix is a two-asset model: a stable utility credit for billing compute work and a separate tradeable governance token for value capture.

Bittensor pays roughly eight dollars in TAO token emissions for every dollar of real AI revenue that flows through the network. The exact ratio fluctuates by quarter, but the shape is durable. Q1 2026: about $328 million in annual emissions against $43 million in real AI revenue. That is 7.6 to 1. It is what the crypto-skeptical press has called "extractive by default." It is also what the crypto-friendly analysts call "the subsidy treadmill." The Bittensor engineering team is sophisticated. The subnet validators run real ML evaluation. The miners serve real inference. The revenue is real. The emissions are also real. The cause is the token model itself. One asset is asked to do two jobs that do not belong together. I want to be specific about this part, because every other decentralized AI compute network I have looked at has the same problem, and the fix is well-known. A token in a decentralized AI compute network does two structurally distinct things. The first job is utility settlement . Contributors run inference, and someone has to pay them for the compute work they did. The payment medium has to scale with usage, has to be denominated in something the contributor can spend on the network or convert to fiat, and has to remain stable enough that contributors can plan around it. This is a billing system. The second job is value capture . Early supporters, investors, and contributors take risk to bootstrap a network that does not yet exist. They have to be paid back for that risk in a way that scales with the eventual success of the network. The payment medium has to be a speculative asset that appreciates as the network grows. This is an equity instrument. A billing system and an equity instrument want opposite things. A billing system that is also a speculative asset means that contributors who get paid in it cannot help but hold a speculative position. An equity instrument that is also a billing system means that token-price volatility shows up in the unit economics of the network itself. The single-token model conflates these two jobs. Bittensor's TAO is both. Akash's AKT is both. Render's RNDR is both. The conflation is what produces the subsidy treadmill. Walk through what happens in a single-token network at steady state. Real AI revenue flows in at rate R. Token emissions flow out at rate E. Contributors decide whether to keep running nodes based on whether R + E /N where N is the number of active nodes clears their economic threshold. In a healthy network, R rises as the network finds product-market fit, and E declines on a published glide-path. The two converge somewhere around year 5 or 7. At that convergence point, the network operates on real revenue and the equity holders capture the value that has accrued in the token. The problem is what happens before that convergence. Contributors are sensitive to the dollar value of their pay, and the token-denominated component E is large compared to the real revenue component R . When the token price falls for any reason macro selloff, competing subnet, ETF rejection, founder error , contributors leave. When contributors leave, the network's quality of service falls. When quality of service falls, real revenue falls. When real revenue falls, the token price falls further. The feedback loop is right there in the mechanism. The way out of the loop is more emissions, faster. That is the subsidy treadmill: emit faster to retain contributors, which dilutes the token, which weakens the contributor payoff, which requires more emissions. Bittensor's 7.6 to 1 ratio is the ratio of "what we have to emit to keep contributors here" to "what the network actually does." Most defenders call it a bootstrap phase. The math says it's the equilibrium that one-asset mechanism design produces. Separate the assets. Use a stable utility credit for the billing job. Use a tradeable governance token for the equity job. Let each asset do the job it is good at. The utility credit is denominated in compute hours. One credit equals one normalized A100-minute equivalent. Contributors earn credits by serving inference. Users earn credits by purchasing them with fiat at a rate set by the network. Users spend credits by consuming inference. The credit is non-transferable in v1, has no exchange listing, and has no speculative premium. It is a billing system and only a billing system. The governance token is tradeable on a public market and exists for value capture. It is allocated to the team, early investors, the foundation treasury, the ecosystem fund, and a contributor airdrop based on cumulative credit earnings. It carries governance rights over protocol parameters, treasury allocation, and slashing policy. A percentage of public-pool fees buys back and burns the token, so token holders capture the upside of network adoption. The two assets are kept apart on purpose: utility on one side, value capture on the other. A contributor who wants to participate in the value-capture side can earn credits and then convert their cumulative credit history into a governance-token airdrop at a Phase 2 milestone. A contributor who does not want speculation exposure can earn credits, redeem them for inference services, and never touch a tradeable asset. The credit's value floor is the compute it represents. One credit can be redeemed for one normalized A100-minute of inference at any time. That redemption right is what makes the credit stable. The redemption right is a use right, not an FX peg. A user who holds 1000 credits can run 1000 A100-minutes of inference. The credits do not need to trade against the dollar or against other crypto assets; they just need to clear inference requests at the rate the network publishes. This is the same shape as data credits in Helium's redesign one credit = one IoT data packet . Helium v1 used a single-token model and produced the same subsidy treadmill we see in Bittensor today. The v5 redesign split utility Data Credits from value capture HNT , and the unit economics stabilized. It was the standard fix that mechanism design produces when the original mechanism fails. Nothing clever about it. A governance token wants the opposite of stability. It wants to be a tradeable instrument that captures the value of the network as the network grows. A token whose only economic function is governance plus fee accrual is what the post-2024 DeFi consensus has converged on. MKR is the canonical example: it pays no one for operations; it just captures value through buybacks funded by DAI fees and votes on protocol parameters. The market prices it on expected future fee accrual. When the network grows, fees grow, buybacks grow, the token appreciates. That mechanism does not work if the same token is also paying contributors. Contributor payments dilute the token. Buybacks concentrate it. The two operations cancel out, and the token's price reflects the noise of the two flows rather than the signal of network growth. When the two operations are split across two assets, both can do their job cleanly. The credit pays contributors and absorbs all the operational dilution. The governance token captures value and absorbs all the buyback concentration. The signals stop fighting each other. The mesh ships before the token. There is no reason to launch a token into an empty network, because a token coordinates supply, and supply that does not yet exist cannot be coordinated. The substantive engineering work that has to land before a token makes sense is the network itself. Cross-machine inference routing has to be measured. Quality-scoring has to be calibrated. The credit ledger has to be running at small scale among trusted contributors. The user-side product has to actually serve real workloads. When those things are working, the credit ledger can scale to a wider contributor pool. When the contributor pool is real, the governance token has something to govern. When the governance token has something to govern, it can launch. The Bittensor failure mode is launching the token first and trying to bootstrap supply through emissions. The Petals failure mode is having no token at all and hoping altruism scales. The two-asset model with mesh-first sequencing avoids both. It says: ship the boring part, prove the boring part works, then add the financial instrument that makes the network ownable. A two-asset network running at steady state has two ratios to track instead of one. The credit ratio is real-revenue-per-credit-redeemed. A network is healthy if users redeem credits at a stable rate against fiat-denominated inference cost. If credits are inflating against compute hours one credit redeems for less inference over time , the credit issuance mechanism is broken. The governance ratio is fee-buyback per token emission. A healthy governance token has buybacks at or above token emissions over rolling 12-month windows. If buybacks fall meaningfully below emissions, the token is in dilution territory and the governance economics are not working. Bittensor's published numbers do not allow a clean version of these two ratios because the two operations are conflated. But the closest analog real revenue versus token emissions is the 7.6 to 1 ratio. The same ratio for a two-asset network at steady state should be approximately 1 to 1 fees fund both buybacks and contributor incentives at parity , with credits decoupled from the governance-token price. A network designed this way can recover from a token-price drawdown without losing contributors. The contributors are paid in credits, and the credits redeem for the same compute services they always did, even if the governance token is down 70%. The governance token's price reflects market sentiment about the network's future fee accrual; it does not control the network's day-to-day unit economics. The two-asset model is not free. It adds engineering complexity two accounting systems instead of one , adds legal complexity the credit may or may not be a security depending on jurisdiction, and the governance token almost certainly is one in the US until decentralization thresholds are met , and adds adoption friction contributors have to learn the difference between earning credits and earning governance-token allocations . The model also does not magically solve demand-side problems. If users do not want to buy credits, the credit is not a real billing system, regardless of how cleanly it is separated from the governance token. The two-asset model fixes the supply-side dynamics that the single-token model breaks. It does not fix the demand side. The demand side comes from building a product people actually want. The Helium precedent is instructive on this point. The two-asset migration stabilized Helium's unit economics, but Helium still had to find a real use case mobile carrier service before the network became economically self-sustaining. The original IoT use case had not materialized at the supply density the network had bootstrapped. A clean mechanism cannot rescue a bad market thesis. I am the founder of Potluck, which is a peer-to-peer AI compute network that has been measuring the boring parts memory mesh latency, cross-machine retrieval, inference verification for six months. The two-asset model is the token design we have committed to. It is documented in detail at trypotluck.ai if you want the longer version. I wrote this to make a specific argument I think is true about decentralized AI compute networks in general: the single-token model is mechanism-design failure that no amount of execution can fix, and the fix has been understood since Helium v5 in 2023. Several networks are launching with the single-token model anyway. If you are building one of those networks, separate the assets before you launch. The math gets better the moment you do. Rob writes the Local AI Engineering Notes series on strake.dev. He is also building Potluck AI https://trypotluck.ai , the peer-to-peer AI compute network referenced in this post, and Strake https://strake.dev , a GitHub Action deploy gate.