privacy · proofs · compute
v2.0 · checksummed

§9. Compute Through the "AI Money" Lens

v2.0
Download PDF Table of Contents Cite / Checksums
Cite this section

Copy/paste (plain text): 

Jason St George. "§9. Compute Through the "AI Money" Lens" in Next Generation Stores of Value: Privacy, Proofs, Compute. Version v2.0. /v/2.0/read/part-ii/9-compute-as-ai-money/

Compute Through the “AI Money” Lens

Compute has always been an economic input: first as muscle, then as steam and electricity, now as FLOPs. The AI boom made this explicit: GPUs, TPUs, and datacenters are priced like oil fields.

Raw compute is not even a standardized service claim:

  • Most compute markets are opaque capacity rentals (cloud contracts, colo deals).

  • There is no standardized unit of verified work; only hours and instance types.

  • There is no cheap, public way to verify that a claimed computation actually ran correctly.

Research Maturity Caveat

Verified inference remains technically immature. Only Tier A cryptographic verification (full ZKML) should support high-assurance capacity issuance without an additional trust haircut. Probabilistic or TEE-backed inference verification (Tiers B and C) receives haircuts or lower-assurance service status. The service-value claim weakens if verified compute trades at no durable premium to ordinary cloud compute; that result would say nothing by itself about monetary premium. See §21: The Modular Stack for the tier taxonomy.

The “AI Money” analytical lens reframes verified-compute demand by insisting on:

  • Canonical workloads: Matrix multiplications, FFTs, and core model primitives.

  • Proofs of useful work (PoUW): Miners/provers earn rewards for producing proofs that these workloads ran correctly, with verification asymmetry.

  • Verification economics: VerifyPrice turns “one unit of verified MatMul at dimension nn” into something that can be priced and traded.

In that context, verified compute capacity becomes a standardized service commodity:

“Rights to future, standardized, cheaply verifiable units of useful compute (verified FLOPs).”

Compute Deflation and What Remains Scarce

A skeptical reader will object: “Compute gets cheaper every year. Moore’s Law drives raw FLOP cost down. How can rights to compute retain service value when the underlying commodity deflates?”

This is a serious objection. Here is the response:

What deflates:

  • Raw FLOPs per dollar (secular decline).

  • Cost of unverified, permissioned compute from hyperscalers.

  • Commodity inference for non-sensitive workloads.

What remains scarce:

  • Verified FLOPs with receipts: Compute where correctness is cryptographically proven, not vendor-asserted.

  • Policy-constrained capacity: Compute that runs under specific jurisdictional, privacy, or compliance constraints.

  • Censorship-resistant access: Compute that cannot be denied by TOS changes, sanctions, or platform bans.

  • Priority access under congestion: During demand spikes, priority slots are scarce even if raw capacity is abundant.

  • Specific hardware profiles: Compute on L0-C or L0-D grade hardware may remain scarce even as closed alternatives proliferate.

The instrument should reference capacity share, not “one timeless FLOP”:

Under the “AI Money” lens, the relevant service claim is not “rights to 1 FLOP forever.” It is:

“Rights to XX% of verified compute throughput under SLA YY on hardware profile ZZ.”

This is analogous to how oil futures reference “barrels of WTI crude delivered at Cushing”—not “energy” in the abstract. The specificity (verification, SLA, profile) is what creates persistent scarcity.

Telemetry that detects deflation risk:

If raw compute deflation outpaces the scarcity premium of verification/censorship-resistance, the following metrics will signal it:

  • VerifyPrice for verified FLOPs converges toward unverified market rates.

  • Capacity utilization on verified networks falls below thresholds.

  • Fee revenue from compute workloads declines as a share of total fees.

These are observable. The thesis predicts that a verification service premium may persist because demand for trustworthy AI is structural and the compliance/sovereignty premium may grow as AI becomes more consequential. If telemetry shows otherwise, the compute service case weakens. Neither outcome determines whether the separate base asset earns monetary premium.

Verified-compute service instruments can include:

  • Work Credits: typed claims minted against verified-compute contributions and redeemable or transferable for specified future workloads.

  • Compute futures: for specific workloads (e.g., “XX verified inferences at model MM with SLO YY”).

  • Stake in proof factories: whose entire business is converting energy + silicon into verified FLOPs with transparent VerifyPrice.

Why can these be valuable service claims rather than arbitrary utility tokens? Because:

  • Demand is global and secular: As long as AI is an input to value creation, there is demand for FLOPs.

  • Verification is public: Anyone can check that a Work Credit corresponds to a valid proof and a stated DVC envelope.

  • Terms are explicit: Workload, tier, delivery window, redemption, and expiry determine value rather than an implied monetary promise.

Tip: hover a heading to reveal its permalink symbol for copying.