AI looks like software at the interface. At scale, it depends on compute, energy, chips, labor, and institutions.
This series is a personal research note. It is not a policy proposal, a legal argument, or professional advice in economics, sociology, law, or governance.
I am writing it because AI agents are beginning to move beyond isolated text generation. They can call tools, write code, search databases, coordinate workflows, monitor systems, make plans, and interact with other agents. In limited domains, they already perform work-like functions that institutions used to assign to people: communication, search, supervision, judgment, and coordination.
Current systems remain artifacts. This essay does not argue for human-equivalent consciousness, experience, rights, or moral status. The issue is functional role. When artificial agents perform tasks that organizations previously assigned to people, technical questions about capability become entangled with institutional questions about work, ownership, accountability, access, and distribution.
The discussion is conceptual and comparative rather than country-specific. I use governance here to mean responsibility, auditability, institutional design, and allocation of resources.
I do not have a settled theory. The series is meant as a starting point, a way to put several concepts on the table and invite sharper arguments from the broader community. The views are my own. I hope the posts encourage AI researchers, social scientists, economists, legal scholars, workers, policymakers, and affected communities to think more carefully about the social relations that agentic AI may create.
The frame is simple:
Treat AI as a productive capacity: a new ability to produce, coordinate, and decide at scale. Its social effect depends on who has access to the infrastructure and how institutions govern its use.
This is an analytic frame rather than a prophecy. Technology changes what societies can do. Institutions decide who benefits, who pays, and who has authority.
A useful distinction separates productive capacity from institutional arrangements. Productive capacity means the practical resources through which a society produces: labor, tools, machinery, scientific knowledge, energy, transport, communication, and coordination. Institutional arrangements mean the rules around those resources: ownership, access, work status, liability, distribution, and authority.
The strongest deterministic reading of this distinction is too crude for AI. Technology does not mechanically write the next social order. The weaker version is more useful: when productive capacity changes, old institutions can fit poorly.
Different historical systems centered different productive assets:
| Historical Pattern | Central Productive Asset | Institution Under Pressure |
|---|---|---|
| Agrarian land systems | land, water, livestock | access, inheritance, rent, and obligation |
| Industrial manufacturing | factories, machinery, finance | wage work, firm ownership, and workplace organization |
| Digital platforms | data, networks, cloud services | platform access, interoperability, and dependency |
| AI-intensive systems | compute, models, data, energy, robots | access, identity, responsibility, and distribution |
The table is schematic. It compresses enormous variation into four rows. Its purpose is narrower: when the central productive asset changes, the institutions around work, ownership, and distribution come under pressure.
AI matters because it turns parts of intelligence into a scalable input. By intelligence I mean planning, classification, language, design, search, coding, supervision, and scientific conjecture under constraints. Current systems do these unevenly. Even partial automation of these functions changes the production process.
A factory machine extends muscle and precision. AI extends coordination, judgment, communication, and search. That is why the institutional question begins before robots walk through every warehouse.
At the interface, AI feels weightless. A user types a prompt; a model replies. At the point of production, the system depends on electricity, data centers, accelerators, cooling, water, fiber, cloud platforms, chip supply chains, capital markets, engineers, and regulatory capacity.
The International Energy Agency’s 2025 report on Energy and AI states the point directly: there is no AI without energy, specifically electricity for data centers. The report also treats AI as a two-sided energy problem: AI raises electricity demand, and AI may also change how the energy sector operates.
This material base matters for governance. Compute is unevenly distributed. So are advanced chips, cloud regions, cheap power, secure data centers, and the talent to operate them. A model may look like code at the point of use, but frontier AI production rests on a compute-energy-model complex.
That complex has geography. It favors places with power, land, grid interconnection, water, capital, security, and supply-chain access. It also creates local planning problems: new data centers can bring investment while straining grids, water systems, housing markets, and ratepayer arrangements.
The labor question often appears as a yes-or-no question: will AI take jobs? That framing hides the first stage. AI changes tasks before it eliminates occupations.
The International Labour Organization’s 2025 update on generative AI and jobs estimates that one in four workers worldwide are in occupations with some degree of GenAI exposure. The same brief emphasizes that most exposed jobs still require human input, so transformation is the dominant near-term category.
That caution is important. Writing, translation, coding, customer support, legal research, data analysis, design, education support, and office operations already contain exposed tasks. Physical work changes more slowly because robotics must solve embodiment, reliability, safety, cost, and integration with messy environments.
Task transformation still matters. Market economies use wages for two functions at once: they organize production and distribute purchasing power. If AI reduces the amount of human labor needed for valuable output, the wage relation becomes a weaker basis for social distribution.
The sharper question is:
If AI systems produce a growing share of economic value, how do humans receive legitimate claims on that output?
Possible answers include shorter workweeks, public services, training pathways, broad access to productive tools, worker-owned or cooperative AI systems, public compute for research and education, and benefit-sharing mechanisms. None follows automatically from the technology. The pressure comes from the mismatch between scalable artificial labor and institutions built around human wage labor.
AI also changes the geography of dependence. Regions and organizations that control frontier models, chips, cloud platforms, energy contracts, and data center capacity can export intelligence services. Users who mainly call external APIs consume intelligence through infrastructure they do not control.
A rough hierarchy may emerge:
| Position | Core Asset | Main Dependence |
|---|---|---|
| Frontier AI cores | models, chips, cloud, talent, capital | energy and supply-chain security |
| Energy-compute hubs | power, land, data centers, long-term capital | external models and chips |
| Regional AI nodes | domestic clouds, sector models, regulation | frontier platforms and hardware |
| Compute importers | API access and downstream applications | external infrastructure and standards |
This is a hypothesis, not a map. It says what to watch. Power may attach to a model provider, a chip supply chain, a cloud region, a data center campus, a power purchase agreement, or a shared AI cloud.
AI adoption alone does not guarantee operational autonomy. A country, university, hospital system, or firm can use AI throughout its work while depending on external models, external clouds, external safety rules, and external pricing.
Within compute-rich environments, AI may create a new industrial ecosystem. Cloud providers, model companies, chip firms, data center operators, utilities, grid planners, energy developers, financial capital, security teams, and compliance institutions all become part of the same production system.
This ecosystem controls the assets needed to reproduce artificial intelligence at scale: training clusters, inference capacity, foundation models, data pipelines, feedback channels, energy contracts, deployment platforms, security teams, and eventually robot fleets.
The local governance questions can be sharp. AI infrastructure can raise valuations, exports, and strategic capacity while creating fewer ordinary jobs than older industrial expansion. Communities may absorb grid upgrades, land use conflicts, water stress, and rate pressure while the gains accrue to shareholders, specialized workers, and compute owners.
Electricity becomes a planning question. Who receives priority access to power? Who pays for grid expansion? How should data center growth be balanced against households, hospitals, research labs, local businesses, and public services? How should scarce electricity be allocated among commercial, scientific, and civic uses?
These questions sit at the center of intelligence as infrastructure.
The same productive base can support different institutional forms.
| Path | Ownership Pattern | Distribution Pattern |
|---|---|---|
| Private platform path | private firms own models, clouds, data, and platforms | gains flow through APIs, subscriptions, licenses, and platform control |
| Public-sector coordinated infrastructure | public agencies coordinate models, chips, data centers, and energy allocation | firms profit under strategic direction and industrial policy |
| Public-interest infrastructure | compute and model access become partly public, nonprofit, cooperative, or community-governed | gains support public services, research, education, reduced work time, or broad access programs |
Real systems will mix these forms. A public institution may fund shared compute while relying on private cloud vendors. A firm may operate under public safety rules while charging platform fees. A cooperative model may depend on public energy policy and commercial chips.
The point is that AI does not choose among these paths. Ownership law, workplace organization, public finance, competition policy, energy policy, and international competition choose.
The first move is to stop treating AI as weightless software. At scale, AI is a material productive capacity built from compute, energy, chips, data, models, labor, and institutions.
That capacity changes work, but it also changes the ownership structure around work. The core institutional question becomes:
What relation of production fits a world where intelligence becomes scalable infrastructure?
The next post moves from AI as infrastructure to AI as actor. As agents become persistent, tool-using, and coordinated, they may enter production as semi-autonomous participants rather than passive instruments.
If you found this post useful, please consider citing it:
@article{song2026aiproductiverelations1,
title={AI and Productive Relations 1: Intelligence as Infrastructure},
author={Song, Xiangchen},
year={2026},
month={May},
url={https://xiangchensong.github.io/blog/2026/ai-productive-relations-1/}
}
Here are some more articles you might like to read next: