Can compute become money?
The AI industry's computing demand could anchor a new idea of money, but this vision requires a banking system to run it – and big changes to AI itself.
It’s a Monday morning in the spring of 2030 and Marco, the treasurer of a large, AI-heavy multinational enterprise, starts his week by opening a liquidity dashboard that looks quite different to the one he used a decade ago.
Cash now comes in four categories. Marco surveys tokenized bank deposits in multiple currencies for payroll and bills. He checks his tokenized money-market fund shares, and his regulated dollar stablecoins for cross-border payments.
Then there’s the fourth category: compute tokens.
These are not cryptocurrencies in the usual sense. They are claims on a defined amount of computation at specific AI labs, such as Anthropic.
Marco holds them in two main forms. First, on‑demand compute balances: instantly usable claims that AI agents across the firm draw down every time they call a model. Second, term compute notes: futures contracts locking in a stream of compute over the next three years at an agreed price and service level.
Corporate treasurer’s day in the life
Every Monday, payroll drains tokenized deposits. Marco tops them up by redeeming some money-market tokens. He knows that next day, the corporation’s new AI-driven product line will require more compute, so Marco buys a tranche of term compute notes, paying with redeemed MMF shares.
Wednesday comes and a market wobble forces the multinational to raise intraday liquidity. Marco pledges some of its compute notes as collateral in a short-term repo with Alicia, a global dealer at a global bank (it could be a JP Morgan or Goldman Sachs, or 2030’s neo-bank darling). She sends him tokenized cash in return.
The next day, another AI provider raises its lab-access prices again, so Marco enters into a swap with Alicia to hedge his future compute costs. Marco ends his week by rebalancing, keeping enough on-demand compute to run the business, enough term notes to lock in strategic capacity, and the rest in conventional cash and equivalents (mostly RMB, HKD, SGD, JPY, AED, plus some dollars because the multinational has a huge US footprint, and EUR for croissants).
Compute tokens have become a working form of money for one crucial slice of the firm’s activities: everything the company does through AI. They are liquid, priced in familiar ways, accepted as collateral, and woven into the same digital rails that carry bank money and securities.
Compute as money
This idea recently surfaced by some tech insiders who think bitcoin will be replaced by “Anthropic Coin”. (Anthropic is the lab of choice because it is a pure play, independent of broader internet businesses such as Microsoft or Google.)
The argument goes like this: Large AI labs control scarce, capital-intensive infrastructure (chips, GPUs, model weights, data centers); they meter access to this infrastructure in tokens; they sell long-term capacity commitments to enterprises as well as to other machines or agents.
These tokens today have money-ish qualities: they serve as a unit of account within AI-heavy firms, with product teams budgeting in terms of tokens instead of hours or dollars; they can serve as a medium of exchange between agents; and this makes them a store of value for bots that need to plan for the enterprises’ future computing needs. Compute tokens don’t buy goods and services but they do buy cognition.
Are compute-token enthusiasts unveiling the future of finance? Not so fast. These tokens today operate more like prepaid utilities or air miles. They lack standardized legal form, robust backing structures, and integration with the broader financial system.
Some pretty big “ifs”
There are reasons why “AI money” is, for now, a fantasy. First, the argument for compute money is being framed as an alternative to bitcoin. This rather puts the cart before the horse, as bitcoin isn’t money. Scarcity alone doesn’t make it so. The flaws in thinking bitcoin is something we might all use to pay our debts apply to the argument about compute.
However, the idea of compute as money is more compelling than the bitcoin story. It’s rooted in actual supply and demand linked to the real world. Bitcoin may not be money but its invention has given rise to a new, alternative infrastructure for capital markets, one that is beginning to integrate with traditional networks, and these rails would indeed be a good fit for compute tokens.
Second, we are assuming current large-language models prove to be sufficiently reliable to anchor anything money-like. However, LLMs do bullshit, a feature that is baked into their models. Hyper-scaling compute hasn’t corresponded to similar gains in LLM robustness (perhaps it could, but the commercial models prefer BS to LLMs failing to provide a response). Trustworthy AI will need new architectures that will require time and a lot of investment to build. It’s therefore hard to put faith in compute tokens as claims on a service, let alone as money.
Third, we are also assuming the issuers of these tokens – the big AI labs – will survive. While there is real demand for AI, many serious analysts question whether these revenues will ever come close to the complex debt structures that Silicon Valley and Wall Street have concocted to keep the wheels spinning.
A huge financial crash would not end the AI story, but it could end OpenAI or Anthropic. In this scenario, compute contracts or deposits start to resemble tulips, at least until the industry recovers on a more stable footing.
Narrow banking redux
These “ifs” all rest on our final, and biggest, assumption: that compute tokens rapidly evolve into a new form of narrow banking that operates alongside TradFi.
Narrow banks are basically like today’s stablecoin issuers: deposit-taking institutions that only invest in safe, liquid assets, and do not engage in risky lending and maturity transformation.
A “Compute Bank” would accept fiat or tokenized cash from users, issue compute deposits and compute notes, and hold its assets in contracted capacity with AI labs or data centers plus fiat cash and short-term government debt, with some risk management rules resembling capital reserves, but in compute terms. The bank would earn money on transaction fees and float.
Just as we are now getting comfortable with licensed stablecoin issuers with fully backed reserves (although perhaps we are too comfortable), compute banks would have to prove their tokens are stable, predictable, and instantly redeemable for AI work.
AI labs would use compute banks as wholesale customers, selling them capacity up front, locking in that revenue, and relying on the bank to distribute capacity to the market. For traditional banks and dealers, compute deposits and notes would appear as a new class of high-grade, infrastructure-linked paper, potentially eligible for collateral in repo and derivatives.
This would not replace traditional money markets: everyone will still pay taxes, wages, and non-AI contracts in their fiat currency. But over time, if AI lives up to its promise across industries, corporate treasurers like Marco would come to hold compute deposits alongside cash and treasury bills, and dealers like Alicia would make markets in compute notes and compute-backed securities. MSCI would be rolling out compute indexes, and enterprises might begin trading in compute terms, creating a currency akin to petrodollars. All backed by productive capacity rather than faith in a government or hand-wavy belief bitcoin.
Not ready for prime time…yet
For this to happen, though, these assets have to be safe, cheap, and predictable – as T-bills or repo markets are today. Compute is none of those things. It is expensive, with mind-bogglingly up-front capex on hardware and data centers; it’s a voracious consumer of energy; and we have no model for cost per unit of useful work – how do we measure someone prompting their way to protein folds to cure cancer, versus someone burning carbon to ask their LLM about Hollywood gossip or sports trivia?
Moreover, compute carries its own operational risks: outages, cyberattacks, misuse, sudden regulatory changes, Iranian drones blowing up Microsoft data centers in the Middle East, chip makers in South Korea losing access to cheap electricity because of geopolitical hazards.
Finally, when we bring Wall Street practices into the equation, we end up importing Wall Street culture. That implies new layers of opacity, complexity, and leverage around speculative expectations of future AI demand. Prediction markets would be all over this asset class, bringing chaos as much as they might foster liquidity. The more the world comes to depend on AI, the more vulnerable it would be to a financialized compute market always vulnerable to a meltdown.
Therefore, the notion that compute deposits are comparable to Circle’s licensed stablecoin USDC has a long way to go. And even if a compute bank could claim full and reliable backing, users are still exposed to “money” that could go poof! because an LLM screws up, is lethally prompt-injected, or manipulates the human for its own arcane agenda.
What has to happen
This doesn’t mean compute money is a non-starter. It simply means a lot of things need to happen before it can take off. The incentives are real.
On the technology side, this implies the development of LLMs that don’t hallucinate or lie on important tasks – a goal that new AI labs in Silicon Valley are beginning to address. It implies strong evidence that AI systems can augment or replace human labor across many industries and activities. And it implies that humans have better control over AI, which allows regulators to regard it as imperfect but manageable.
On the economic side, we’d need to see AI leading to sustainable productivity gains across many industries, so that AI labs enjoy healthy margins. We need these companies to be bedrocks of stability, not debt-fueled tinderboxes.
On the institutional side, we’d need transparent, simple, and auditable structures for compute claims. This goes hand in hand with clear legal terms and sensible backing. The stablecoin example is probably insufficient, as it remains riddled with weak attestations: this isn’t good enough if compute tokens are to become the new money. We’d also need regulatory regimes that know how to supervise compute banks and enforce capital, liquidity, and risk-management standards. The upside for the compute banks would be the opportunity to integrate with traditional deposits and commercial-bank money.
These developments are all possible, although they require multiple changes of direction in Silicon Valley, on Wall Street, and in the halls of regulators and policymakers. But it’s an optimistic vision, for the emergence of compute money implies a successful adoption of AI across industries and society. The horror story is also possible, of compute money gone haywire, but it’s more likely that AI will self-destruct (either in economic and valuation terms, or something worse) before computer money has a chance to take hold.
Indeed, the idea that AI companies can morph into bedrock financial institutions, but only under the right conditions, might be the incentive we need to get them on a more sustainable, less dangerous track.