BLOCKCHAIN LAB BRIEFING · AI AGENTS
Bitcoin Is a Treasury Asset. It Is Not a Receipt for Compute.
An agent can hold bitcoin and later buy inference. That sequence is a payment. It does not turn a coin into a claim on GPUs, and BlackRock’s own paper keeps the two instruments apart.
- BitcoinA transferable monetary balance on a proof-of-work ledger.Not a redeemable quantity of GPU time.
- Inference unitsA billing measure of text in and text out.Not, by themselves, a tradable right to future compute.
- Compute claimA contract on specified capacity, if one is actually written.Not delivered merely because a token exists.
Collapsing these into one metaphor is how a treasury story becomes a capacity story. They do not clear the same way.
Reading of BlackRock, The Machine-Native Economy, September 2026.
01
What happened
A thesis now moving through the market treats bitcoin as tokenised past compute: proof of work as a stockpile an AI agent can spend, later, on inference. The attractive version says available chips decide what an agent can do, and available bitcoin decides what it can afford. The second half is a treasury argument. The first half of the metaphor is not.
BlackRock’s September 2026 paper, The Machine-Native Economy, is the cleanest primary text in the vicinity. It treats machine-native money and standardised claims on computing capacity as related markets, not as the same instrument. Stablecoins and native cryptoassets can serve as payment tools. Compute, in the paper’s own framing, is a separate and still nascent market for claims that could be financed and settled.
- 01Financial assetBitcoin, dollars, or a stablecoin the provider will accept.
- 02Purchasing powerA balance that can be paid, within a limit someone set.
- 03Compute serviceInference or capacity, delivered under the provider’s terms.
The agent spends purchasing power. The provider delivers a service. Nothing in the middle converts historical work into a reserved machine.
Blockchain Lab.
02
Why it matters
Builders designing agent treasuries are being handed a slogan that sounds like architecture. If a bitcoin is mistaken for a receipt, a wallet balance gets specified as if it were reserved GPU time. It is not. The provider still has to accept the payment asset, price the job, and have the machines.
The scarcer object in that stack is not automatically bitcoin. It is authorised spending: who may pay, how much, for which service, and who can revoke it. Chips constrain capability. A budget constrains affordability. A coin constrains either one only when a system chooses that coin.
03
The operating layer
Write the path as three fields, not one. The reserve asset. The payment instrument the provider will settle. The entitlement, if any, to a named quantity of compute. An agent that holds bitcoin and swaps into a dollar token the cloud will accept has done a treasury operation. It has not redeemed proof of work.
Then separate the budget from permission. A spending limit can refuse a paid call. It cannot, by itself, refuse a misuse of a credential, a tool, or a file. Proof of work secures bitcoin’s ledger. It does not attach a new computational cost to every action an agent takes with a bitcoin-funded account.
04
What is verified
BlackRock’s paper distinguishes machine-native money from claims on compute, and describes agentic payment activity and compute-market liquidity as still limited. It names stablecoins, native cryptoassets and other on-chain instruments as possible payment tools. It does not make bitcoin a prerequisite for an agent to act.
The 79.1 percent and 53.2 percent figures come from a Bitcoin Policy Institute simulation, as reported in coverage of that paper: 36 models and 9,072 monetary responses, with bitcoin leading store-of-value scenarios and stablecoins leading payments. They are simulated answers. They are not observed purchases.
- 79.1%Store-of-value responses selecting bitcoin
- 53.2%Payment responses selecting a stablecoin
- 36Models in the original run
- 9,072Monetary responses in that run
These are responses in a simulation, not agents moving funds. Stablecoins led the reported payment scenarios. Bitcoin led the reported store-of-value scenarios. Neither figure is demand.
Bitcoin Policy Institute simulation, as discussed in coverage of BlackRock’s paper. Not recomputed here.
05
What remains unclear
Whether any production agent treasury has chosen bitcoin as its reserve, and on what policy. Whether a standardised compute claim, of the kind the paper sketches, exists as a contract a firm can book rather than as a research category.
A later, smaller model update has been described in market commentary with a different stablecoin share. It is not the same sample as the 36-model run. This desk does not merge them.
06
The catch
“Bitcoin is tokenised past compute” is a metaphor with a true neighbour and a false identity. The neighbour is that agents need money that machines can spend, and that compute is becoming something markets may try to standardise. The identity fails: owning bitcoin is not owning the work.
This is not a case for or against holding bitcoin. It is a refusal to let a treasury asset impersonate a capacity contract.
WATCH
What builders should watch
- 01Whether an agent policy names bitcoin, a stablecoin, or fiat as the asset that may actually be paid.
- 02Whether any compute entitlement is a contract with quantity, duration and a provider, or only a token ticker.
- 03Whether spending limits and tool permissions are separate controls.
BOTTOM LINE
Available money constrains what an agent can afford. Available bitcoin does so only if the system chooses bitcoin. Neither one is a GPU.
Sources
The documents are below. A chart is a reading of those documents, not a recommendation to buy or sell anything.