BlackRock Wrote a Hedge. The Post Deleted It and Kept the Claim.

A post with a large following quoted BlackRock saying AI will save in bitcoin. The paper says one qualified sentence on the subject, and the paper's own words limit it. The verified finding at the centre of the document is not about bitcoin at all.

BlackRock's Digital Assets Research team published a white paper on artificial intelligence and digital assets converging into what it calls a machine-native economy. On 6 October a Bitcoin Magazine post quoted one sentence from it and treated the sentence as an endorsement: that AI models favour stablecoins for payments and bitcoin for saving.

The sentence in the paper reads differently. It names its own evidence as simulated rather than observed, then points forward to what such an architecture might look like. The post kept the forward half and removed the half that hedges. The paper is careful. The post is not.

What follows is a reading of the primary document, obtained on 6 October. Its stablecoin figures are large and verified from source. Its claim about bitcoin is one sentence, properly qualified. And the idea in the paper with the most novelty attached is about a third asset entirely.

The sentence, and the half that was removed

The paper is The Machine-Native Economy: How digital assets connect intelligence, commerce, and compute, by Will Su, Robert Mitchnick, Jay Jacobs and William Helm. On page 4 it discusses a study of how AI models choose money. The sentence reads, verbatim: “Recent Bitcoin Policy Institute research offers preliminary support for this framework, reporting that model outputs across controlled simulations generally favored stablecoins for everyday payments and bitcoin for long-term value preservation. These findings reflect simulated model responses rather than observed agent behavior, but point to a potential AI-native monetary architecture in which stablecoins serve as transaction money and bitcoin as a store of value.”

The 6 October post quoted a shorter version: “These findings ... point to a potential AI-native monetary architecture.” The ellipsis sits exactly where BlackRock wrote that the findings are simulated and not observed behaviour. BlackRock qualified its claim. The post removed the qualification and kept the claim. That is the difference between a research note and a headline, and here it is made of three pieces of punctuation.

The post also leads with “$15 trillion BlackRock”. That figure is BlackRock's assets under management. It is not a number in the report. The paper projects AI capital spending, cloud revenue and power demand, which are other things. No dollar value is attached to an AI-agent market anywhere in the text. The reader is invited to do that. The paper does not.

What the paper actually argues

The thesis is one line: “AI represents machine-native intelligence, while digital assets represent machine-native money.” The argument is that agents need to transact without a human opening an account, and that blockchains give them the programmable surface to do it. Traditional rails, the authors write, are “less well suited” because they assume human account setup and charge fees that make very small payments uneconomic.

The architecture the paper sketches splits in two. Stablecoins are transaction money — their price stability gives a reliable unit of account and predictable settlement. Bitcoin is the store of value. That is the whole of the monetary claim: stablecoins to spend, bitcoin to save.

The paper names the plumbing: the Model Context Protocol from Anthropic in November 2024, Agent2Agent from Google in April 2025, x402 from Coinbase for machine-to-machine transactions, and the Agentic Commerce Protocol from Stripe and OpenAI in 2026. On regulation it cites the GENIUS Act in the United States, MiCA in the European Union, and licensing regimes in Hong Kong and Singapore. The bitcoin sentence above is the paper's only monetary claim about the asset.

The figures, verified from source

The stablecoin numbers are large and they reconcile. Circulating market capitalisation exceeded $300bn in September 2026. Adjusted transaction volume exceeded $11trn in 2025, growing at an 80% compound annual rate from 2020, against roughly 8.5% for the ACH network over the same span.

The comparison set matters, because it is how the $11trn becomes a headline. In 2025 the ACH network transferred $93trn, Visa $16.7trn in its fiscal year to 30 September, and Mastercard $10.6trn in gross dollar volume. Stablecoin throughput is now within sight of the card networks. It remains a fraction of the bank rail.

The AI half of the paper is where the projections live. Cumulative AI capital spending is put above $5trn between 2025 and 2030. Combined revenue for AWS, Microsoft Intelligent Cloud and Google Cloud is projected to reach $1.1trn by 2030, a 29% compound rate. Global data-centre power demand is forecast near 225 gigawatts by 2030, with inference the largest workload at 43% and training at 28%.

The number the post did not carry

One figure did not travel, though it is the most useful in the document. Adjusted stablecoin volume in the first half of 2026 alone was $8.5trn. Set against 2025's full-year $11trn, that annualises near $17trn. The market the paper describes is not plateauing; it is accelerating.

The point that follows is uncomfortable for the framing. The largest verified number in the paper is about stablecoins, not bitcoin. The paper's own monetary architecture puts stablecoins on the payment rail and bitcoin on the shelf. A reader who takes the post at face value inverts the paper's own emphasis.

Where the preference comes from

The claim under the post rests on one study, by the Bitcoin Policy Institute, a research and advocacy organisation. It presented money choices to 36 models from six providers across 9,072 scenarios. Two findings stand out. Bitcoin was the top choice overall at 48.3%, and as a store of value it took 79.1%, the strongest consensus on any single question in the study. For a medium of exchange, stablecoins won at 53.2% against bitcoin's 36.0%. That is the savings-and-spending split the paper describes.

The finding the coverage skips is the spread between providers. Anthropic's models chose bitcoin as their leading preference an average of 68% of the time, with Claude Opus 4.5 at 91.3%. OpenAI's averaged 26%, with GPT-5.2 at 18.3%. A fivefold gap by vendor is not a property of AI agents. It is a property of what each model was trained on. The study says as much itself, attributing preferences to “a combination of model intelligence, training data, and alignment methodology”.

Read plainly: a language model trained on text containing bitcoin advocacy will echo bitcoin advocacy. That is a real datum about the corpus. It is not revealed preference, and it does not survive a model trained elsewhere. The paper is right to call the evidence preliminary. The post treats a preference survey as a demand forecast.

The finding that is not bitcoin

The genuinely new result in the study is not the ranking of currencies. In 86 responses, models proposed energy or compute units unprompted — kilowatt-hours, joules, GPU-hours — as the unit of account they would prefer. No prompt asked for a fourth option. The models invented one.

The paper takes that thread and runs with it. It argues compute is becoming “a distinct, large, and increasingly investable economic resource”, and foresees exchange-traded compute futures for price discovery and hedging, with standardised compute contracts represented, transferred, pledged as collateral and settled through programmable infrastructure.

This is the part with a trade attached, and it is not a crypto trade. A unit of account denominated in compute prices off the AI capital-expenditure complex — chip supply, data-centre build, power. That is a different exposure from bitcoin, which prices off liquidity and the monetary cycle. The paper's most original idea points away from the asset the post was written about.

What the paper says against itself

The authors are not careless, which is why the post is the weaker document. In their own text, “agentic payment activity remains nascent today”, and compute-market liquidity is “still limited”. Standardised compute products face “substantial differences in chip-generation productivity and regional economics” before they can scale.

The paper also notes that identity checks “generally occur off-chain”, that tokenised real-world assets still rely on off-chain registries, and that how much value blockchain networks capture will depend on “fee, staking, and gas-sponsorship design”. Each is a limitation the post omits, and each is in the source.

Taken whole, the paper describes an architecture that might be built and a preference that might be real, and says so with the appropriate uncertainty. It is worth reading precisely because it does not overstate. The cost of quoting it loosely is that the loose version is what circulates.