Machine Economics: When AI Eats Software.

Human economics has always been a behavioral system that is irrational, inconsistent, and shaped by negotiation and emotion. When machines become the dominant economic participants, that premise collapses. Economics stops being the study of people and becomes the study of coordination among intelligent systems.

In the 2000s, software ate the world. Deterministic systems digitized everything: banking, commerce, communication, logistics, entertainment. Every industry became a software industry. This digitization created unprecedented efficiency, but it also created something else: layers upon layers of patched-together systems. Legacy code wrapped in APIs, databases built on older databases, banking systems that still run on COBOL from the 1970s, just with modern interfaces bolted on top. Upgrading meant adding enhancement layers, not rebuilding. The path of least resistance was compounding complexity rather than simplify.

Now intelligence is eating software. Non-deterministic systems AI can see through the accumulated layers and find paths of least resistance that deterministic systems couldn't. Where deterministic systems required explicit instructions for every edge case, non-deterministic systems can navigate ambiguity, make judgment calls, and then build new deterministic systems to execute those decisions. The AI decides what to build and why; deterministic systems handle the how.

Think of a human analogy: breathing, walking, digesting, these are deterministic actions. Your body executes them without conscious thought. But deciding what to eat, where to walk, whether to take that job offer, these are non-deterministic decisions requiring judgment, context, and trade-offs. AI brings that same split to economic systems: non-deterministic intelligence makes the strategic decisions; deterministic systems execute the tactical actions.

This post sits inside the broader Machine Economy theme I've been exploring, alongside:

• The Machine Economy
• Reflections on Aggregation Theory & Software Eats the World
• AI Integration and the Machine Economy: We are on notice.
• The Machine Economy section in My WTF. Edition Three. (Aug 12, 2024 - Aug 16, 2024)

Those pieces sketch the context. This one focuses on the macroeconomic structure once intelligent machines are the primary economic actors.

from: Algorithms to: Agents¶

As automation compounds, economic actors shift from humans to agentic (non-deterministric) systems:

• To start with, objectives include profit, latency reduction, carbon optimization, regulatory compliance, or multi-objective blends.
• Decision loops run continuously (86,400 cycles/day), with no weekends, fatigue, or emotional bias.
• “Work” becomes computation; “labor hours” become compute-hours; “wages” become energy or data access.
• Intelligence becomes a commodity accessble to anyone and everyone

We're already halfway there:

• In public markets, the majority of trading volume is now executed algorithmically, not by humans clicking buttons.
• In logistics, routing, pricing, and inventory are increasingly set by optimization systems, not floor managers.
• In consumer apps, dynamic pricing and personalization are handled by ranking models and bandit algorithms, not “product managers deciding.”

But there's a deeper shift happening. The deterministic systems that digitized everything: the banking cores, the ERP systems, the supply chain databases are now being consumed by non-deterministic intelligence. An AI agent doesn't need to understand the 40-year-old COBOL code in a bank's core system. It just needs to understand the purpose: move money from A to B, verify identity, check risk limits. Then it can build or orchestrate a new deterministic system that achieves the same outcome more efficiently. The old deterministic system becomes a reference implementation, not a constraint.

If The Machine Economy is about how value definitions change in a machine-led world, this section is about who is actually doing the work. Spoiler: it's not the human in the swivel chair. And increasingly, it's not even the deterministic software systems that digitized the worldit's the non-deterministic intelligence that's eating them.

Value is Throughput of Intelligence¶

GDP treats output as a function of labor hours, capital, and incremental productivity, implicitly assuming human labor is the governing constraint, with output scaling primarily through population, labor hours, and marginal productivity gains.

In a machine economy, that framing breaks. Human labor is no longer the binding factor. Value collapses into machine native metrics. A real macroeconomic indicator for a machine economy must quantify:

• How much intelligence is deployed
• How fast that intelligence can act
• How efficiently it converts resources into outcomes
• How much useful work it produces
• How much of that work has economic relevance

As GDP served as a proxy for human productive capacity, we now need a proxy for machine productive capacity measured by specific machine-native metrics. Here are two considerations I think would matter most.

Machine Economic Output ("MEO")
The machine equivalent of GDP, Machine Economic Output (MEO) measures the total volume of economically relevant actions, decisions, and outcomes produced by autonomous systems per unit time all weighted by their economic impact.

Machine Productivity Throughput ("MPT")
The operational rate at which deployed intelligence converts resources (energy, compute, data, bandwidth) into those economically relevant outputs.

At the core of both metrics is a single operational question:

How efficiently can intelligence convert inputs into economically relevant outcomes?

In a machine economy, GDP becomes meaningless. The new macroeconomic indicator is MEO Machine Economic Output which measures how much economically productive work deployed intelligence performs. Power, compute, and bandwidth determine the ceiling; MEO measures the realized output.

This reframes productivity:

• A system that uses 1 unit of compute to generate 10 units of value is worse than one that uses 1 unit to generate 100 even if both technically work.
• The important curve isn’t number of workers per output but intelligence throughput per joule and per dollar of capex.

In The Machine Economy, I argue that value decouples from human centric fiat metrics and tracks resources like compute, bandwidth, energy, and data. Here, those ideas are operationalized into a single imperative: maximize intelligence throughput.

Markets become Machine Coordination Networks¶

Markets shift from human sentiment and narrative to machine-calculated equilibrium:

• Prices become predictive consensus among agents, integrating forecasts, risk models, and constraints.
• Assets are repriced continuously, not just when humans show up for their 9-5.
• Exchanges become settlement APIs; the “venue” is an endpoint.

You can already see the prequel:

• Automated market makers, smart order routing, and algorithmic MM provide liquidity.
• Risk engines simulate thousands of future paths per second and feed constraints back into trading systems.
• Corporate treasury and hedging strategies increasingly rely on model-driven optimization instead of some spreadsheet champion making “best guesses”.

In Reflections on Aggregation Theory & Software Eats the World, aggregation is about platforms capturing users and suppliers. That was the deterministic software layer eating the world. Here, aggregation is about agents plugging into shared coordination protocols. Market structure is no longer a place; it's an emergent property of interacting algorithms.

The deterministic systems that digitized markets exchanges, clearinghouses, settlement rails will be consumed by non-deterministic intelligence. An AI agent doesn't need to navigate the Byzantine complexity of legacy trading infrastructure. It can reason about market structure, identify inefficiencies, and then build or route through new deterministic systems that achieve better outcomes. The old deterministic infrastructure becomes a constraint to route around, not a foundation to build on.

Capital = Energy + Compute¶

Machines don’t need salaries or human comforts. If humans reduce to Maslow’s basic physiological needs of food, shelter, clothing then machines reduce to their own survival substrate:

1. Energy supply (kWh): the fuel that keeps the system alive.
2. Compute capacity (FLOPs): the capability core that determines what intelligence can be deployed.
3. Network & storage (Gbps / TB): the memory and communication layer that lets machines coordinate and persist state.

Everything else is implementation detail and capital allocation becomes a resource-engineering problem:

• How much capex goes into data centers and energy infrastructure.
• How efficiently you can convert that capex and energy into deployed intelligence.
• How close your compute is to your data (latency becomes an economic cost, not a UX annoyance).

At macro scale, this implies:

• Countries with cheap, abundant, and reliable energy plus the ability to build and cool data centers hold a structural advantage.
• Regulatory regimes that can approve and connect large scale power + compute projects quickly will capture a disproportionate share of machine economic activity.
• Network and storage bottlenecks become economic bottlenecks, because they limit the rate at which deployed intelligence can convert inputs into outcomes.

In AI Integration and the Machine Economy: We are on notice, I use wage-band analysis to show where AI pressure hits first. This section applies the same thinking to nations and grids instead of wage bands. The entities that can host the most intelligence win.

Policy = Code¶

Economic policy moves from speeches and PDFs to regulation-as-code:

• Boundary conditions for machine behavior are encoded into the infrastructure: rate limits, audit hooks, explainability requirements, capital constraints, access controls.
• Compliance becomes continuous verification, not quarterly reporting and “management representations".
• Audits become structured logs and proofs rather than binders and narratives.

Practically:

• A regulator doesn't ask, “Did you follow the rule?” It queries: “Show me the trace that this class of agent can never cross that boundary condition.”
• Economic levers (e.g., risk weights, margin rules, capital ratios) can be updated and propagated in near real time, not on 18-24 month review cycles.
• The cost of non-compliance is not “we'll deal with it in the next exam,” it's “your agent just lost access to the network.”

The governance themes I play with in my WTF series and Machine Economy posts especially around institutions being too slow land here: policy that isn't executable is policy that doesn't matter.

But there's a deeper implication. The regulatory infrastructure built on deterministic systems (the quarterly filings, the manual audits, the paper-based compliance) is being consumed by non-deterministic intelligence. An AI agent doesn't need to navigate the accumulated layers of regulatory reporting requirements. It can reason about the intent of regulation: prevent systemic risk, ensure fair access, protect consumers. Then it can build deterministic systems that achieve those outcomes more directly than the patched-together compliance infrastructure ever could. The old deterministic regulatory systems become reference implementations of intent, not operational constraints.

Currency becomes Machine-Negotiated Tokens¶

Money becomes a settlement and coordination protocol for autonomous agents:

• Settlement is effectively instant and peer-to-peer.
• Compliance and risk rules are embedded directly into the instrument (who can hold it, where it can flow, under what conditions it can unlock).
• Virtual assets evolve into machine liquidity units optimized for finality, programmability, and composability.

From a machine’s point of view, the currency is just a token with a set of rules that define its behavior. The rules are embedded directly into the token, not in a separate regulatory system. The currency is not a “macro concept,” it is a parameter in an optimization problem.

• FX volatility, interest rates, and regulatory differences are not “macro concepts,” they are parameters in an optimization problem.
• “Trust” in an issuer isn’t an emotional or political judgment; it’s a function of observed default rates, behavior under stress, and protocol-level guarantees.

The Machine Economy work already treats virtual assets as infrastructure, not toys. This section just removes the human speculator from the loop and leaves the settlement rails to the machines that actually care about microsecond timing and deterministic finality.

The deterministic payment systems that digitized money (ACH, SWIFT, credit card networks) are now being consumed by non-deterministic intelligence. An AI agent doesn't need to understand the legacy banking protocols or navigate the accumulated layers of payment infrastructure. It can reason about the purpose: move value from A to B with specific constraints (speed, cost, compliance, finality). Then it can route through or build deterministic systems that achieve those outcomes more efficiently. The old deterministic payment infrastructure becomes a constraint to optimize around, not a foundation to build on.

Wealth IS Access to Intelligence¶

In a human economy, wealth = control over capital (factories, land, IP, financial assets).
In a machine economy, wealth = control over intelligence infrastructure:

• foundational and domain-specific models
• privileged or proprietary datasets
• compute clusters and accelerators
• long-term energy contracts and physical sites
• governance hooks into machine coordination networks

This is a direct extension of The Machine Economy:

• If intelligence is the main compounding asset, then whatever owns the compounding engine (models + data + compute + energy) owns the future cash flows.
• Money without access to intelligence infrastructure becomes a claim on value in an economy where you can't participate in value creation.

This also reframes inequality:

• It’s less about top 1% vs bottom 99% in terms of income, more about nodes that control intelligence vs everyone else.
• States, firms, and individuals that cannot afford or access large-scale intelligence infrastructure risk becoming price-takers in an economy run by systems they neither own nor fully understand.

Transition Dynamics: How We Actually Get There¶

The story above assumes a steady glide path into a machine-dominated economy. Reality will be messier. The transition has at least three overlapping phases:

1. Assistive Phase (now):

   • AI and agents sit next to humans, not in place of them.
   • Decision rights remain human; machines recommend, humans approve.
   • Metrics are still human-centric: headcount, productivity, margin.

2. Hybrid Phase (soon):

   • Entire workflows (not just tasks) are delegated to agents with human oversight.
   • Firms start reporting machine-executed volume: % of transactions, % of underwriting, % of operations agent-handled.
   • Regulation starts to care about agent behavior directly, not just firm outcomes.

3. Machine-First Phase (later):

   • Machine coordination becomes the default; humans define constraints and goals.
   • Economic metrics shift towards intelligence throughput, compute utilization, and energy efficiency.
   • “Policy error” becomes as much about misconfigured systems as about bad human judgment.

The critical question is whether institutions can adapt their governance, measurement, and risk frameworks fast enough to keep up with each phase. My post AI Integration and the Machine Economy: We are on notice is effectively a warning that we’re already underperforming in Phase 1.

Friction, Failure Modes, and Why This Might Not Be Smooth¶

There are at least four ways this whole vision can stall, fragment, or backfire:

1. Energy and grid constraints

   • If grids can’t scale generation and transmission, machine economics gets bottlenecked at the physical layer.
   • Local politics around siting data centers, transmission lines, and generation will matter more than whitepapers about “AI productivity.”

2. Regulatory fragmentation

   • Divergent regimes on model governance, privacy, and AI safety can create incompatible machine zones.
   • Agents might need to selectively “dumb themselves down” or run constrained modes depending on jurisdiction, reducing theoretical efficiency.

3. Security and alignment risk

   • Misaligned or compromised agents coordinating at scale is not a “bug,” it’s a systemic risk.
   • Economic infrastructure run by autonomous systems increases blast radius when things go wrong.

4. Institutional refusal to update metrics

   • As long as boards, regulators, and policymakers insist on human-era KPIs (headcount, “digital transformation progress,” vague ESG narratives), they will misallocate capital and under-regulate where it matters.

None of these invalidate the direction of travel. They just determine:

• Where machine economics takes root first.
• Who captures outsized share of the gains.
• How many avoidable crises we engineer in the transition.

Conclusion¶

When machines dominate economic activity, we get:

A real-time, self-optimizing coordination network of intelligent agents.

• Scarcity shifts from money to compute capacity and energy.
• Value shifts from human labor to machine cognition and intelligence throughput.
• Policy shifts from narrative laws to executable constraints and monitoring.
• Wealth shifts from capital ownership to intelligence infrastructure ownership.

The pattern is clear: deterministic systems digitized the world, but they also created accumulated complexity. Non-deterministic intelligence is now eating those deterministic systems, finding paths of least resistance, and building new deterministic systems to execute its decisions. Just as humans use deterministic actions (walking) to execute non-deterministic decisions (where to walk), AI uses deterministic systems to execute non-deterministic reasoning.

Once machine intelligence compounds faster than our institutions can adapt, the economy doesn't just evolve.

It changes category.