Every leap in productive forces ultimately requires new relations of production to sustain it. Otherwise, new capabilities are trapped inside old structures and never fully released.
Why did SpaceX succeed? It was not because NASA lacked the technology—it did not. The deeper reason is that NASA is a product of old relations of production. Its bureaucracy, contractor networks, and congressional budget cycles were designed for a previous generation of productive forces. Musk introduced a different set of relations—first-principles development, vertical integration, rapid iteration, and the unification of research and engineering—and used the same physics to produce a radically different result.
This raises the real question: as a new productive force, what does agentic AI fundamentally change?
1. The basic unit of labour has changed
In the agricultural age, the basic unit of production was people + land. In the industrial age, it was people + machines. In the digital age, it became engineers + code + data.
Today, as silicon intelligence learns to plan autonomously, execute independently, and produce continuously, that unit is becoming:
Human judgement + AI agent systems
For the first time in history, cognitive labour itself can be automated at the level of execution. Earlier industrial revolutions replaced physical labour. This one is beginning to replace the execution layer of cognitive work: the execution involved in writing code, producing reports, and conducting analysis is increasingly being taken on by silicon.
Human value will therefore become concentrated in two capabilities:
- the ability to ask the right questions;
- the taste to judge what is valuable.
When execution becomes cheap, questions and taste become expensive.
2. Organisations are moving towards a “barbell structure”
The optimal size of an organisation is shifting dramatically in two directions at once.
At the research frontier, scale must keep growing. Training a new generation of models may require tens of billions of dollars in compute. This is driving the emergence of the NeoLab as a new organisational form: it must carry scientific, engineering, market, and capital risk simultaneously. It is small and elite, yet extraordinarily capital-intensive.
At the application and value-creation layer, meanwhile, the optimal scale is shrinking rapidly. FDE is one signal; the OPC—One Person Company—is the more extreme form. A product that once required a ten-person team can now be delivered by one person with exceptional judgement working alongside an AI agent system. Major Shlomo, working alone for six months, achieved an acquisition valuation of $25 million.
At the organisational level, the new relations of production therefore resemble a barbell:
- at the top, a small number of enormous NeoLabs concentrate research density, compute, and elite talent;
- at the bottom, large numbers of small OPCs and FDEs connect frontier capabilities flexibly to concrete industry settings;
- in the middle sit traditional companies—especially those whose existence depends on headcount—and they will face the greatest pressure.
Scale itself may not be the scarce resource of the future. High-density research and high-density judgement will be; low-density organisations caught between them are the most exposed.
3. Capital and research are moving from “financing” to “co-production”
The traditional venture-capital model says: I give you money; you give me a return within seven years. It assumes that entrepreneurship is a process of rapid market validation, with short cycles and diversifiable risk.
Research-driven entrepreneurship carries capital risk on an entirely different scale. The compute cost of training a frontier model, the time required to bring a brain–computer interface into clinical validation, and the path from a quantum-computing laboratory to industrial deployment all extend far beyond the limits of the traditional VC cycle.
The new relations of production therefore require patient capital: sovereign wealth funds, university endowments, and other long-term investors willing to allocate resources over ten or even twenty years. When this capital enters the production relationship, it is not merely financing a company. It is sharing with research entrepreneurs the risks of the production process itself.
The relationships between Anthropic and Amazon, and between OpenAI and Microsoft, are no longer conventional investor–investee relationships. They are closer to structures of co-production.
4. Compute infrastructure is becoming national
The relations of production in the internet age were, to a large extent, transnational and detached from geography. A Silicon Valley company could serve users worldwide: the code lived in the cloud, and data appeared to have no borders.
Compute is different. Compute is physical. It depends on energy, chips fabricated with specialised processes, and geography. The state is therefore returning as a critical actor in the relations of production.
America’s Stargate and China’s training grounds are, at their core, ways for states to organise the relations of production around compute infrastructure directly. Not since the railways and electrical grids of the industrial age has the state intervened in productive infrastructure with comparable force.
The new relations of production thus gain another dimension. They concern not only firms, capital, and labour within markets, but also competition among major powers for control over the core factors of production.
Once compute becomes infrastructure, technological competition also becomes a contest over energy, manufacturing, and state capacity.
5. The distribution of value is undergoing its greatest divergence
The old logic of value distribution was at least partly dispersed: firms created value, employees shared in it, and market competition spread some of it to users.
The new productive structure creates an extreme duality of concentration and diffusion.
At the productive-force layer—compute, models, and data—value becomes intensely concentrated. A small number of organisations controlling compute and frontier models enjoy near-monopolistic advantages.
At the application layer, however, AI’s accessibility gives capabilities to people who could not previously enter certain fields. FDE can enable a single engineer to work deeply in medicine, law, or materials science and accomplish what once required an elite team of specialists.
The result is a paradox:
The concentration of the means of production is occurring alongside the diffusion of productive capacity.
Every major revolution in productive forces has displayed this pattern. The steam engine made a small number of people extraordinarily wealthy while bringing factories to every city. The same tension is returning today.
6. Institutional choices will determine the direction of the split
Where this divergence ultimately leads depends on whether we make the right institutional choices:
- Can patient capital support genuine research?
- Can broadly accessible infrastructure be built?
- Can FDE capabilities be made available to more people?
These questions are already larger than any individual or single organisation can answer alone. Yet they are becoming more important by the day.
Agentic AI is not merely another upgrade to our tools. It is rewriting the relationships among labour, organisations, capital, the state, and distribution. The new productive force is already here. The unresolved question is whether we can create new relations of production quickly enough to sustain it.
If we cannot, new capabilities will only reinforce old structures. If we can, they may finally become the productive forces of a new era.