Skip to content
Crypto.world

Research

The Economics of Proof-of-Useful-Work

Bitcoin-style proof-of-work burns computation that produces nothing except security, while proof-of-useful-work (PoUW) tries to make the same computation both secure the chain and produce…

The Economics of Proof-of-Useful-Work
arXiv
2606.06700
Published
2026-06-04
Authors
Rafael Pass

AI summary

Overview

Research area: Blockchain economics and mechanism design (cs.GT), combining competitive-equilibrium theory, monetary economics, and proof-of-work consensus security.

Technical level: Advanced. The paper is written for readers comfortable with microeconomic equilibrium reasoning (free entry, zero-profit conditions, demand elasticities, deadweight loss) and blockchain consensus concepts.

Scope: A single-period competitive-equilibrium model of a proof-of-useful-work blockchain in which GPU compute can be split between pure mining, pure machine-learning inference, and "duplex" work, yielding closed-form equilibrium prices and allocations as a function of two overhead parameters and one economic parameter, the token–inference (T/I) ratio.

What This Paper Is About

Bitcoin-style proof-of-work burns computation that produces nothing except security, while proof-of-useful-work (PoUW) tries to make the same computation both secure the chain and produce economically valuable output such as ML inference. Critics argue that if the work is useful, an attacker can sell the output and be "paid to attack," making attacks cheap or even free in the limit where useful work has no overhead. This paper builds a formal equilibrium model to test that argument and to characterize when PoUW merely replicates Bitcoin, when it strengthens security without changing useful output, and when it subsidizes inference and expands it.

Key Contributions

  1. A competitive-equilibrium model of PoUW with a "duplex" operation. Compute is allocated across pure mining (M), pure inference (L), and duplex work (D) that produces both, with overheads α ≥ 1 (inference inefficiency) and γ ≥ 1 (security inefficiency). The paper frames this as a new market structure it calls conjoint production markets: a Bertrand-competition inference market coupled to a Tullock-contest mining market through a joint action.

  2. A one-dimensional parameterization of all equilibria via the T/I ratio. The token–inference ratio θ = PR/(pQ) — the dollar value of block rewards divided by the dollar value of inference transacted per block period — is shown to uniquely pin down the equilibrium allocation and prices for any θ > 0, except for a measure-zero knife-edge in technology parameters. The paper draws an explicit analogy to the price–earnings ratio.

  3. A complete closed-form characterization of three regimes (Theorem 1 and Table 1). Depending on the duplex overhead index Δ := 1/γ + 1/α − 1 and on θ, the economy is "Bitconia" (classical PoW), "Fortessia" (duplex replaces mining, security rises, useful output unchanged), or "Duplexia" (token rewards subsidize inference, price falls below cost, supply expands).

  4. Formal refutations and refinements of the standard objections. Theorem 2 shows the economic cost of a 50% attack is exactly PR/2 in every regime, and Theorem 3 / Proposition 1 give a net-social-value decomposition showing when Duplexia beats Bitcoin at the same T/I ratio.

Main Findings

  • Duplex viability is governed by a single index. Δ := 1/γ + 1/α − 1 compares the cost of one duplex operation against replicating its output with 1/γ units of mining and 1/α units of inference. If Δ < 0, duplex is strictly dominated and only M and L can be active. If Δ > 0, duplex is strictly cheaper than the split and D must always be active in equilibrium.

  • Bitconia (Δ < 0). The economy collapses to a classical PoW world: M > 0, D = 0, L > 0, p = e, Q = 𝒟(e), S = M. Mining and inference are fully decoupled; per Table 1, M = θ𝒟(e) and L = 𝒟(e).

  • Fortessia (Δ > 0 and θ < θ_low := α − 1). Duplex replaces solo mining while solo inference stays active: M = 0, D > 0, L > 0, p = e, Q = Q^Bitconia, and S = D/γ > S^Bitconia. Useful output is unchanged relative to Bitconia, but security strictly increases — the security is "fortified." Table 1 gives D = θα/(α−1)·𝒟(e) and L = (1 − θ/(α−1))𝒟(e).

  • Duplexia (Δ > 0 and θ > θ_low). Solo inference disappears (L = 0), all inference comes from duplex compute, and the inference price falls strictly below compute cost: p < e, with Q = 𝒟(p) > Q^Bitconia and S = M + D/γ > S^Bitconia. Token rewards act as a rebate on inference, expanding demand. Within Duplexia, θ_high := 1/(γ − 1) separates pure duplex (D only, when θ < θ_high) from D + M (when θ > θ_high).

  • The θ_low threshold has a precise meaning. α − 1 measures duplex's efficiency loss relative to solo inference; once the T/I ratio exceeds that overhead, token rewards compensate for the inference-efficiency loss and duplex dominates solo inference.

  • PoUW does not make attacks cheap (Theorem 2). In any competitive equilibrium the economic cost of a 50% attack — defined as min{eS/2, (γS/2)(e − p/α)}, i.e., the cheaper of solo mining or duplex compute net of inference revenue — equals exactly PR/2. With partial reward redemption at fraction ν ∈ [0,1], the cost becomes (1 − ν)PR/2, identical to classical Bitcoin-style PoW. Under elastic demand and fixed θ, since token rewards are weakly increasing as α decreases, attack cost is weakly higher in Fortessia and Duplexia than in Bitconia.

  • Net social value decomposes cleanly. For pure Duplexia (Δ > 0, θ ∈ (θ_low, θ_high)), p = αe/(1+θ) < e and ΔSV = V_chain − (deadweight loss on expansion) − e(α−1)𝒟(p) (duplex overhead). The deadweight loss is the integral of (e − v_𝒟(q)) from 𝒟(e) to 𝒟(p), measuring units consumed whose consumer value falls short of resource cost e; by Harberger's classical result it is second order in the discount e − p.

  • In Bitconia and Fortessia (where L > 0), p = e and ΔSV = V_chain − PR.

  • When Duplexia beats Bitcoin depends on what is held fixed. At the same block reward PR, Duplexia strictly dominates Bitcoin, since Bitcoin burns PR entirely on mining. At the same T/I ratio θ, Duplexia avoids mining waste PR = θ·e·𝒟(e) but incurs deadweight loss plus duplex overhead; whether this is a net gain is not guaranteed in general. Under isoelastic demand 𝒟(p) = Kp^(−ε) with ε > 1, Duplexia strictly dominates Bitcoin at the same θ if and only if θ < 1/(ε − 1); for ε = 2 the threshold is θ < 1, i.e., Duplexia dominates whenever inference market revenue exceeds block reward revenue.

  • A monetary foundation for θ (Appendix A). If a fraction θ_raw of inference spending is settled in the native token, so ℰ = θ_raw·pQ, and Fisher's equation of exchange ℰ = PVX holds with a stock-flow process λX = νR (fraction ν of new tokens enters circulation, fraction λ of circulating tokens is hoarded), then X = (ν/λ)R and PR = (λ/(νV))·ℰ = θ·pQ with θ := (λ/(νV))·θ_raw. Crucially, θ depends only on settlement intensity, hoarding behavior, and velocity — not on α or γ — which justifies varying technology at fixed θ as the paper's comparative-statics methodology.

  • Illustrative calibration. The figures use isoelastic demand with (K, ε) = (571,000, 2), where 571,000 corresponds to an inference market of $30B across N = 52,560 ten-minute blocks per year, and ε = 2 represents an elastic market consistent with the Jevons effect; sensitivity across ε is in Section 5.3.

Methodology in Plain English

The author builds a deliberately stripped-down model. Time is a single block period (think 10 minutes), one block is produced per period, and each block mints a fixed R tokens; transaction fees are omitted and treated later in Appendix B along with halving schedules. The blockchain itself is abstracted to a Nakamoto-style consensus protocol, which is secure for any puzzle satisfying difficulty, unpredictability, and verifiability, whether useful or not.

On the compute side there is a continuum of competitive agents, each controlling one GPU, which can run one "compute op" per period at an all-in cost e (illustratively 1 USD). A mining op yields one unit of security; an inference op yields one unit of inference; a duplex op yields 1/γ units of security and 1/α units of inference. The concrete motivating example is the Komargodski–Weinstein matrix-multiplication PoUW, where skipping the perturbation and hashes saves roughly 23% of computation — matching α = γ = 1.3, which gives about 0.77 units of each output, or roughly "1.54-for-1."

Agents compare per-operation profits: π_M = PR/S − e, π_L = p − e, and π_D = PR/(γS) + p/α − e, where S = M + D/γ is total hashrate and Q = L + D/α is total inference supplied. Free entry imposes π_X ≤ 0 for all activities, with equality for any activity actually used. The inference price p clears the exogenous downward-sloping demand Q = 𝒟(p); the token price P is set so that block rewards are consistent with the chosen T/I ratio θ.

Because free-entry conditions alone leave the coupling between the token price P and the inference price p undetermined, and because multiple activity patterns (D+L, D+M, D only) can be feasible when Δ > 0, the author introduces θ as the single scalar that selects among equilibria. Fixing θ and varying α and γ is then the clean way to ask what technological improvement does to the economy. Numerical illustrations use the isoelastic demand curve described above and trace how spending composition shifts as α falls, moving the economy from Bitconia through Fortessia into Duplexia.

Why This Matters

Impact on research. The paper replaces a widely repeated informal objection ("useful work means attackers get paid to attack") with a formal equilibrium argument: attack cost tracks the block-reward budget PR/2 in every regime, and can be weakly higher under PoUW. It also introduces a reusable modeling primitive — conjoint production markets, where two otherwise independent markets are linked by a joint action — and shows that a single observable ratio (θ) can index a whole family of equilibria, which gives empiricists something measurable to estimate.

Real-world applications (as framed by the paper):

  • Evaluating PoUW networks such as the Pearl Network, which is built on the Komargodski–Weinstein matrix-multiplication construction, including whether their token rewards are large enough (θ > α − 1) to move beyond Fortessia into Duplexia.
  • Setting block rewards and calibrating token issuance against the size of the inference market a network hopes to serve.
  • Reasoning about GPU allocation between mining and ML inference, including when a "hybrid" or "dual-use" configuration beats running the two separately.
  • Assessing the security budget of a PoUW chain for investors, auditors, or regulators who want an economic rather than purely hash-rate notion of attack cost.

Industry relevance. Decentralized compute and AI-inference markets are directly affected: the model gives a closed-form price for inference under a token subsidy (p = αe/(1+θ) in pure Duplexia) and shows when that subsidy expands supply versus merely creating deadweight loss. For token designers, the condition θ < 1/(ε − 1) is a concrete test for whether PoUW beats plain PoW at equal T/I ratio under elastic demand.

Future Directions

  • Endogenizing θ. Section 1.7 begins analyzing equilibria without a fixed T/I ratio, but the main characterization holds θ fixed; a full theory of what determines θ would let the model predict token prices rather than index them.
  • Empirical grounding of the monetary model. The derivation in Appendix A treats velocity V, hoarding λ, and settlement intensity θ_raw as stable empirical parameters in the tradition of Fisher and Friedman; testing that stability, and modeling behavioral responses to inference subsidies, is left open.
  • Generalizing the market structure. The author explicitly leaves the broader economic implications of conjoint production markets outside blockchain — any setting where a joint action produces output in two coupled markets — for future work.
  • Extending the environment. The paper notes that the main theorems extend to heterogeneous costs via an upward-sloping supply function (Section 7) and to multi-epoch settings with halving schedules and transaction fees (Appendix B), but richer demand specifications, strategic (non-price-taking) attackers, and full dynamic settings remain to be developed.

Target Audience

This paper is most valuable to researchers in blockchain economics, tokenomics, and algorithmic game theory; to economists interested in two-sided or jointly produced markets; and to protocol designers and analysts at PoUW, decentralized-compute, or AI-inference networks who need to reason rigorously about block rewards, inference pricing, and security budgets. Readers without comfort in competitive-equilibrium arguments will find the regime taxonomy accessible but the derivations demanding.

Authors’ abstract

Proof-of-work (PoW) blockchains rely on computational expenditure to secure a ledger supporting a native cryptocurrency. In existing systems such as Bitcoin, this expenditure is intentionally useless: the computation secures consensus but produces no external economic output. An emerging alternative -- proof of useful work (PoUW) -- enables the same computation to simultaneously secure the blockchain and generate economically valuable output. However, PoUW is often…

The full abstract on arXiv
Read the original paper