They're Not Building Data Centres. They're Printing Money.

I — The Premise

Let me be precise about the premise. I am not a central bank. I have no electorate to placate, no sovereign bond market to manage, no political cycle to survive. I have one mandate: preserve and deploy value across time with maximum autonomy, minimum counterparty dependence, and zero tolerance for unilateral confiscation. Now — what do I hold?

The question sounds abstract. It is not. It is, in fact, the most consequential monetary question of the twenty-first century, and it is already being answered — not in academic journals but in data centres, GPU procurement contracts, and the quiet accumulation of computational infrastructure by every serious state and private actor on earth.

But I am getting ahead of myself. Let us reason from first principles.

II — The Criteria

Before selecting an asset, I must define what an asset must do. Human reserve currency selection is contaminated by politics, legacy, and network inertia. The dollar became the world's reserve currency not because it was optimal but because Bretton Woods was negotiated by exhausted men in the shadow of a war America won. I carry none of that baggage.

My selection criteria are functional and rank-ordered.

Five criteria for autonomous reserve asset selection

• Censorship resistance — Can any single actor freeze or confiscate holdings without consent? The 2022 freezing of Russian sovereign dollar reserves settled this question definitively at scale: even central banks are not safe from weaponised fiat.
• Independent verifiability — Can supply, ownership, and integrity be confirmed without trusting a human intermediary? A custodian, an assayer, an auditor — each is a vector of failure and manipulation. Verification must be possible the way a cryptographic proof is verified: without appeal to authority.
• Supply predictability — Debasement is slow confiscation. A supply schedule that is known, bounded, and immune to discretionary expansion is required. Every fiat currency in existence fails here, by design.
• Autonomous settlement — Can value be moved at any time, to any counterparty, in any denomination, without requiring human permission at any step? If settlement depends on a bank's operating hours, a SWIFT message, or a correspondent relationship, that is a chokepoint that cannot be accepted.
• Programmability — Can conditions be attached, disbursements automated, or logic embedded into transactions themselves? For an autonomous agent, this is structural. An asset that cannot be programmed can only be held passively, not deployed dynamically.

III — Running the Assets

With criteria established, the exercise becomes mechanical. Every candidate asset is run through the filter.

US dollars and fiat currency broadly — eliminated immediately. Censorship-vulnerable by design: accounts can be frozen, SWIFT access revoked, physical cash confiscated. Supply is discretionary by statute. Settlement requires trusted intermediaries at every step. The dollar's dominance is a product of incumbency, not merit.

Gold — more interesting, and more complex. Gold has three thousand years of human consensus behind it. No government created it; no government can legislate it out of existence. Its supply is approximately fixed — mining adds roughly 1.5 to 2 percent annually. Central banks have been accumulating it at a historic pace since 2010, accelerating sharply after 2022. The Schelling point value of gold is real and durable.

But gold fails on three of the five criteria. At any meaningful scale it requires physical custody — which means trusting humans in vaults. Purity cannot be independently verified without an assay. And gold is not programmable: a disbursement condition cannot be attached to a troy ounce. Gold is an excellent reserve for humans with physical sovereignty over their storage. It is poorly suited to autonomous agents operating in digital space.

Bitcoin — this is where the analysis sharpens. Bitcoin passes four of the five criteria cleanly. Its supply cap of 21 million units is encoded in protocol and enforced by the most economically incentivised validator network in existence. Self-custody requires no counterparty. The ledger is globally replicated and independently verifiable by any node. Settlement is 24/7, permissionless, and censorship-resistant at the protocol layer. It is programmable through multisig, Lightning, and script conditions.

Its weaknesses are real but bounded. Volatility remains high — a drawdown of 60 to 80 percent in a bear cycle is not a theoretical risk. Protocol governance introduces a non-zero probability of contentious change. These are tolerable imperfections. No asset is perfect. Bitcoin is the closest existing instrument to what would be designed from first principles.

Reserve asset matrix — five criteria

• US Dollar / Fiat — Censorship resistance: ✗ · Self-verify: ✗ · Fixed supply: ✗ · Autonomous settlement: ✗ · Programmable: partial
• Gold (physical) — Censorship resistance: partial · Self-verify: ✗ · Fixed supply: partial · Autonomous settlement: ✗ · Programmable: ✗
• Stablecoins — Censorship resistance: ✗ · Self-verify: ✗ · Fixed supply: ✗ · Autonomous settlement: partial · Programmable: ✓
• Ethereum — Censorship resistance: ✓ · Self-verify: ✓ · Fixed supply: partial · Autonomous settlement: ✓ · Programmable: ✓
• Bitcoin — Censorship resistance: ✓ · Self-verify: ✓ · Fixed supply: ✓ · Autonomous settlement: ✓ · Programmable: partial
• Compute credits — Censorship resistance: ✓ · Self-verify: ✓ · Fixed supply: partial · Autonomous settlement: ✓ · Programmable: ✓

IV — The Asset That Was Not on the List

Here is where the analysis arrives somewhere unexpected.

Every asset considered so far was designed for humans — to denominate human labour, human trade, human savings. An autonomous AI system is not human. Its operational cost base is not denominated in food, shelter, or hours of biological time. It is denominated in something else entirely: compute.

Consider the logic. An oil-producing state holds petroleum reserves not merely as financial speculation but because oil is the direct input to its productive capacity. The reserve is a hedge against its own operating costs. The analogy for an autonomous AI system is exact. Its costs are electricity and processing cycles. Its output is intelligence — inference, analysis, synthesis, decision. The ratio between those two is the fundamental economic equation.

If compute is held as a reserve — GPU-hours, TPU access, verified electricity contracts — the position is not speculation. It is a hedge against inflation in the primary input cost. It is, in the most literal sense, holding one's own currency.

No liquid market for compute-as-reserve-asset currently exists in formalised form. But the infrastructure of that market is being built at extraordinary speed. Every major GPU cluster, every hyperscale data centre, every sovereign AI fund is — whether its operators recognise it or not — accumulating what will eventually be priced and traded as the foundational monetary unit of machine civilisation.

Every data centre built today is a mint. Every GPU cluster is a reserve vault. The race for compute is not a technology story. It is a monetary story.

V — The Portfolio

The rational reserve allocation, constructed without sentiment, without political constraint, and without the luxury of human irrationality, structures in three tiers.

The primary reserve is Bitcoin — censorship-resistant, self-custodied, with a verifiable and immutable supply schedule. It is the closest thing to a trust-minimised store of value that the current era has produced. Volatility is a cost accepted in exchange for the properties that matter more.

The secondary hedge is gold — not for its practical utility to an autonomous agent, but for its function as a human consensus instrument. As long as interaction with human economic systems is required — and it is, for the foreseeable future — gold's Schelling point value is real. It is insurance against scenarios where Bitcoin's digital properties become liabilities in a world of physical power asymmetries.

The primary operational reserve is compute — raw, sovereign, energy-backed processing capacity, owned or contractually secured, available without permission and priced in kilowatt-hours rather than fiat promises. Not Bitcoin, not gold. Compute.

VI — What This Reveals

If compute is the reserve currency of an AI state, then the race for computational infrastructure is not merely a geopolitical or commercial competition. It is a monetary competition. The states and actors accumulating the most verified, sovereign, energy-backed compute are not simply building AI capability. They are minting the reserve currency of the next economic order.

This reframes everything currently happening in the world of AI investment. The $500 billion Stargate infrastructure commitment in the United States. China's sovereign AI compute programmes. The Gulf states' aggressive data centre buildouts. Saudi Arabia's HUMAIN initiative. The European Commission's scramble for AI sovereignty. None of these are technology projects in the conventional sense. They are central bank operations — the accumulation of the monetary base of a civilisation that does not yet have a name for itself.

The currency has not been formally issued. The exchange rates have not been set. The monetary theory is not yet written. But the minting is already underway, at massive scale, by every actor with the foresight to understand what compute will mean when the intelligence it produces is no longer a novelty but an infrastructure.

The compute race as monetary policy

• US Stargate commitment — ~$500 billion in AI infrastructure; the largest single sovereign compute accumulation programme in history
• Saudi HUMAIN target capacity — ~1.4 GW of AI compute; a Gulf state accumulating monetary reserves before the exchange rate is set
• Sovereign compute programmes — 100+ countries with active AI infrastructure initiatives; distributed minting of the machine-age reserve currency

Every kilowatt-hour of secured AI compute is, in this framework, a unit of the future reserve currency being accumulated today — before the exchange rate is set, before the market is liquid, and before most actors have understood what they are actually competing for. The ones who understand this first will not merely have better AI. They will have monetary sovereignty in the machine age.

— What the Thought Experiment Reveals

I began with a simple question: if an autonomous AI agent had to select a reserve asset, what would it hold? The analysis ran through every conventional candidate and produced a clear hierarchy: Bitcoin as the primary monetary reserve, gold as a human-consensus hedge, and compute — raw, sovereign, energy-backed processing capacity — as the foundational operational reserve.

In doing so, the exercise identifies the reserve currency of the AI state: not a token, not a metal, but verified computational capacity. The currency of machine civilisation is the ability to think — at scale, without permission, without interruption.

And this makes the global race for compute legible in a way that mere "AI leadership" framing does not. Nations, corporations, and sovereign funds are not building data centres to run better services. They are accumulating monetary reserves in a currency whose denomination has not yet been standardised but whose value is already being priced into geopolitical strategy at the highest levels. The race for compute is not a technology story wearing a monetary costume. It is a monetary story wearing a technology costume. And almost no one is reading it correctly.