@adlrocha - Money and collateral in an AI-first society

How stablecoins, tokenised markets, and AI merge to build a parallel economy.

It is no secret that I’ve always been obsessed with how money and the pipes of the financial system works. The moment I learnt how the discount window, central bank reserves, collateral, the repo market, shadow banking, etc. worked, it really blew my mind. I used to think of money as being just bank notes and deposits, and that the key role of the financial system was played by retail banks: they held individual and companies’ deposits, provided them with credit, and arbitraged with their balance to make a profit.

The two books that opened my eyes about how the financial system really works, and pushed me down this rabbit hole of understanding how money really worked, were Central Banking 101 and Capital Wars by Michael Howell (if you are interested in this topic, happy to share more bibliography about it. I have to admit that Ray Dalio’s books also had an important role helping my understanding, but they weren’t as eye-opening as the ones above).

And why am I telling you all this? Because after writing last week’s post about intelligence becoming a commodity, and describing an AI-first society as one where the fabric of the economy is automated through agents interacting with each other without human involvement, I could not stop thinking about an obvious follow-up question. If the economy changes that fundamentally, what happens to money and collateral?

Coincidentally, I have been reading Emad Mostaque’s The Last Economy this week. Mostaque is one of the founders of Stability AI, and in the book he argues that we are at an inflection point he calls the Intelligence Inversion: where intelligence is no longer a scarce resource exclusive of humans and that needs to be rented by the hour, but a form of capital that you can invest in to run uninterrupted.

The book presents a lot of really nice ideas. I really liked how he presents the symptoms of our currently broken financial system; or how he frames the economy as being ruled by the 2nd Law of Thermodynamics where agents try to minimise entropy by bringing chaos into order.

All this along with some personal developments pushed me to think more about how money, currencies, commodities, and the economy could look like in the hypothetical AI-centric society that we were discussing last week.

To give some order to these thoughts before we dive down the rabbit hole, here is the high-level roadmap of what I’m going to cover:

• Why the scarcest inputs to the AI economy, compute, energy, and tokens, are becoming the new global commodities and collateral.

• The inevitability of two parallel economies: A high-frequency, tokenised economy for agents, and a slower-paced fiat economy for humans.

• Why inference efficiency, not scale, is the ultimate economic advantage in a system designed to reduce entropy.

---

The money and collateral we use today

Let’s start with the obvious (although from my conversations with some friends, it may not be as obvious at first). Money is a collective agreement about what we use as a unit of account, a store of value, and a medium of exchange. This is why we tend to choose scarce things like gold as money. We could get into the properties that make something good money (rare, durable, divisible, fungible, nobody can print more of it out of thin air, etc.).

Gold is what I would call hard money, and it worked reasonably well until the Bretton Woods agreement in 1944, where the USD was pegged to the ounce of gold, until Nixon in 1971 closed the convertibility window, and the USD started being backed by nothing more tangible than the user’s trust on the institution issuing it. This introduce the fiat money that we all use and hate today.

Since 1971, what actually backs the dollar has been a combination of things: the size and productive capacity of the US economy, the fact that oil is priced in dollars (the petrodollar system that emerged in the 1970s after the OPEC shock), the eurodollar, etc.

Through the petrodollar, eurodollar, and corresponding commercial exchanges, the USD had become the de-facto global currency permeating all of the financial system. But fiat money is essentially backed by the belief that everyone else will keep accepting dollars in exchange for oil, food, or services.

But here is the part that really broke my brain when I first understood it. Most people think of the dollar as something that sits in your bank account or circulates as banknotes. The reality is that the real engine of modern finance runs on collateral, specifically, on the ability to pledge assets (mostly US Treasuries) to borrow cash overnight in the repo market.

In a repo transaction, a financial institution sells a security (say, a Treasury bond) to another party with an agreement to buy it back the next day at a slightly higher price. The difference is the overnight interest rate. Through this mechanism, a single Treasury bond can be pledged and re-pledged multiple times across the system, what economists call the collateral multiplier. The US repo market currently runs at around $12.6 trillion in daily exposures, almost entirely denominated in US Treasuries.

This is also why the Fed’s Standing Repo Facility matters so much: when repo markets seize up, the entire financial system loses its ability to function. In October 2025, the Fed had to inject $29.4 billion overnight in what was described as the largest such operation in two decades. The point is that the USD does not just sit in your wallet, it is the lifeblood of a global collateral chain that underpins every financial transaction on the planet. It lubricates the economy with liquidity.

Funnily, the liquidity of repo markets is one of the key metrics that I monitor for my personal investment decisions as a way to understand if the system is under any stress, or liquidity is being drained from the system (this is why I had the charts from Baselight from this chat below from an analysis I did a few weeks ago and that came pretty handy for this post. By the way, I no longer make this query by hand in Baselight and I use my own agent connected to Baselight for this, but that’s a topic for some other day).

This brief introduction is just to set the context as to me what counts as money, what counts as collateral, and what gets used as the lubricant in the machine, is exactly what I think is about to be disrupted or at least changed in the AI age.

What the heck, this is already changing before AI, the debasement trade is already a thing, and AI may just accelerate it. Maybe what I will describe here is the ultimate debasement trade :)

---

New-age commodities

Oversimplifying it a lot, the pattern underneath all of monetary history is that whoever controls the scarcest and most essential input to economic activity ends up controlling the unit of account.

Gold because it was physically scarce and universally valued. Oil because after the industrial revolution no modern industrial economies could function without it, and finally dollars once the economy was financialised and globalised due to the need of dollars to participate in the global economy.

So the question for the AI-first society is: what is the scarce, essential input that everything else depends on?

My take: compute, energy, and intelligence access measured in PFlops, MWh, and tokens.

And you probably won’t believe it when I tell you how I came to this realisation. A few months ago I started noticing that I was converting all of my AI subscriptions that I use daily from monthly to annual billing. Every service that gave me the option, I locked in the yearly rate. This is what made me realise I was treating access to intelligence like a commodity that I needed to hedge (this event is essentially what gave me the inspiration for this post).

The logic is the same as a manufacturer locking in energy prices for the next year, or hedging the oil price. The price of accessing AI models may change in the next few months in ways that are hard to predict (even more considering the current pace of progress).

There are deflationary forces like competition between labs, model efficiency improvements, falling inference costs. And there are inflationary forces, explosive demand, data centre energy constraints, the cost of training and running frontier models. I don’t know which direction the net price moves over the next two years. But I know I need access to intelligence to do my work more productively. I realised that I am unconsciously deciding that the rational move is to lock in supply at today’s prices.

I then read this argument about the AI Bubble not being a Bubble but a trap that confirmed a thought that I was having for awhile, and aligned with my subconscious bias of hedging my access to intelligence:

“The real product isn’t chatbots. It’s a dependency.

Businesses are being nudged, gently at first, to replace chunks of their workforce with machine labor. Once that happens, reversing course becomes prohibitively expensive. Institutional knowledge evaporates. Workflows warp around proprietary systems. Human staff disappear, and with them the ability to function without the platform.

After that, the trap is set.

Prices go up slowly enough to prevent mass defections but fast enough to extract monopoly rents. Switching back to humans would cost more than staying put. Switching to a competitor is impossible because there won’t be many competitors left.

This is the same playbook Big Tech always uses. Subsidize adoption. Starve alternatives. Centralize infrastructure. Then turn the screws.”

The geopolitical version of this is exactly why the US is restricting GPU exports and why nations like the UAE, Saudi Arabia, and France are spending tens of billions on sovereign AI compute. They are treating PFlops the way previous generations treated oil reserves.

This is also starting to show up in financial markets in a very literal way. AI startups have been using GPU clusters as loan collateral, pledging racks of H100s to secure financing the same way a previous generation of businesses pledged property or receivables (which by the way I think it is a horrible idea and extremely risky for this corporations in its current form). The new commodities are becoming collateral. Which means they are starting to function like money.

The rational move right now is to secure the tokens, the MWh, the PFlops, because those are the raw materials of the next economy.

---

Two parallel economies

And here comes the most trippy argument of my post and the one I am least convinced of (to the point that after writing it I was considering not including, but YOLO). Please, bear with me and do not hesitate to share your feedback after reading it.

The economy we have been living in, what I’ll call the human economy, has high-level the following structure: labour and capital combine to produce goods and services. Those goods and services are priced in fiat currency. It’s relatively slow. It operates on bank hours, settlement cycles counted in days, and relies on human-centric institutions. And honestly, that is fine for what it does. It is perfect for buying real estate, paying for a haircut, or buying groceries. Things that humans need, operating at a human pace.

Then there’s the pipes beneath it, the financial system with its repo markets, collateral chains, shadow banking, etc. that exists to allocate that capital “efficiently” (quotes intended for obvious reasons) across the economy.

But look at who actually drives most of that allocation today. It is not retail banks moving small deposits around. The liquidity that makes modern markets function comes overwhelmingly from hedge funds, primary brokers, and large financial institutions running highly automated strategies. These firms already operate at speeds and scales that no human trader can meaningfully oversee in real time. The humans set the parameters; the algorithms execute.

And this is going to be exacerbated by the short-term future that is coming (even before AI) with stablecoins and always-on, instantly transactable, tokenised markets.

The AI-first society I described last week is just the logical endpoint of that trend. The agent economy, where AI agents perform tasks, commission sub-tasks from other agents, and exchange outputs without human involvement. My feeling is that with the combination of programmable money and AI agents this is the direction that the most automated parts of the financial system are already moving in. And when it arrives fully, fiat currencies structural problems are going to be exacerbated.

People in the crypto space like Haseeb Qureshi from Dragonfly are claiming that “crypto was not made for humans bug AIs” (there you go blockchain, you finally found your killer app :) ).

Agents do not need fiat. They have no rent, no food, no kids (like the ones that I love with all my heart but keep distracting me as I write these words,). What they need is compute, energy, and access to the models that give them reasoning capability. The natural medium of exchange between agents is not dollars, or treasuries, it is AI tokens, compute credits, and energy units.

This is why the current maturity of stablecoins and the push for tokenised markets is so critical. Yes, humans use stablecoins today and will continue to use them alongside fiat for things like cross-border payments. But for agents, they aren’t just an alternative; they are the necessary rails for an AI-first economy. Once tokenised markets and highly liquid stablecoins are fully entrenched, they will become the default financial infrastructure for agents. AI agents will commission sub-tasks, trade MWh, buy PFlops, and even execute those complex collateral chains and repo agreements we discussed earlier using these crypto rails, settling instantly, in fractions of a cent, with mathematical finality.

They will be left to run the over-financialised economy that we have today, leaving humans to operate at their human pace (saving a lot of stress to a lot of people).

This leaves us looking at a fascinating bifurcation. Two economies running in parallel.

On one side, the high-frequency, highly automated agent-to-agent market, running on crypto rails and trading tokenised commodities (compute, energy, intelligence, financial instruments).

On the other side, the slower-paced, human-to-human market, running on a mix of traditional fiat and stablecoins dedicated to physical and emotional human needs, like shelter, community, services, and art.

The human economy won’t disappear. But it will likely become a slower, higher-level layer that sits on top of this massive, hyper-efficient, tokenised machine economy. Humans will hold the real estate and the physical assets; agents will run the plumbing, the finances, and the compute. And I really think this is for the better.

---

The thermodynamic view of the economy

If this high-frequency agent economy is going to run in the background, trading tokenised commodities and executing complex collateral chains, what exactly are these algorithms optimising for? What is the fundamental “physics” of this new financial system?

One of the chapters I enjoyed the most in The Last Economy is the one that presents the idea that the economy is fundamentally a machine for reducing entropy.

“The economy, as a complex adaptive system, evolves to favor configurations that are most efficient at creating predictive models of their environment

[...]

It is a phase transition in the efficiency of entropy reduction. Economy based on ordering machines and entropy reduction. Value is not a pre-existing substance.. It is a state of low entropy, a temporary victory against chaos, achieved by intelligent agents sorting the environment.

[...]

Each action is a small, incremental denoising step, an attempt to move the chaotic state of the present slightly closer to a more ordered, predictable future. Remove uncertainty, thus entropy.

[...]

The forward pass of diffusion models is a perfect simulation of the second law of thermodynamics.”

Norbert Wiener, the father of Cybernetics, reached a similar conclusion in 1950: “In control and communication we are always fighting nature’s tendency to degrade the organised and to destroy the meaningful; the tendency for entropy to increase.”

And one can’t frame an argument using thermodynamics without mentioning Maxwell’s demon. For those of you unaware, Maxwell’s demon is a thought experiment where a demon controls a door between two chambers containing gas. This demon would open and close the door to let hot and fast particles enter one side, and cold and slow ones the other, in this way ordering “the universe of particles” from these two chambers.

This thought experiment appeared to disprove the second law of thermodynamics, because no energy was spent by the demon to reduce the entropy of this closed system. Turns out, the solution to the thought experiment is “information”. In order for the demon to know when to open this door, it needs to know the position, direction, and speed of a particle to predict when to open and close the door. The energy spent on this measurement is the one spent to reduce the entropy of the system.

Using this same framing to the economy, where the energy spent to reduce the entropy is that of acquiring information to order the system. In a world where intelligence is the primary economic input, the unit of value for these agents will not be raw energy or raw compute, it will be intelligence per unit of energy. Not MWh, but useful output per MWh. Not tokens, but insights per token. The most valuable thing in the AI economy will not be the entity with the most GPUs; it will be the entity that converts a given amount of energy into the most useful intelligence most efficiently (context is all you need, apologies for the self-cite).

This reframes the competitive dynamic entirely. The most efficient models running on the most efficient hardware will capture the most value. Inference efficiency, not raw scale, becomes the monetary advantage. And the race to improve it, smaller models, better quantisation, more efficient architectures, becomes more important. This links with another strong opinion I’ve been having lately where I think that small models running in the edge will be what enables a real agentic economy, and where the value of this economy will be captured.

My take is that superintelligence is a set of decentralised agents collaborating.

---

Let’s wrap it up for now!

And this ended up being waaay longer and taking me waaay more time as originally expected (as always).

My goal with this post was to try to dump in writing a lot of the disconnected ideas that I’ve been having lately around the economy, AI, tokenisation, etc after reading a lot about it. They may feel a bit disordered right now, but my plan is to use this first stop as a way to collect feedback from all of my smart readers, and then extract each of them into an isolated (and hopefully better explained) post.

So if you have any feedback or strong opinions about any of the topics and framings presented here, I would love to hear them.

To help you digest all oft his, let me try to share one last time a map of the high-level ideas of my line of thinking:

• I first expect new-age commodities to enter our current economy as AI starts becoming more critical for our day-to-day lives.

• As the new pipes of the financial system start establishing and permeating the economy (in the form of stablecoins and tokenised markets) we are start seeing an agentic economy developing over them where human intervention will be extremely limited.

• Up to a point, where there will be two distinct economies: a fast paced one involving agents, and the day-to-day one for human interactions.

• Finally, I was planning to completely remove the thermodynamics section, but I love physics, and the framing of the economy as a thermodynamic system with the goal of lowering the entropy (creating order) through intelligence blew my mind. Even if readers hate it, I couldn’t publish it for posterity :)

They say that the solution to the “too many ideas syndrome” is to write to get those ideas organised. This is that piece of writing that I needed.