Notes from Forbes Russia's interview with Daniil and David Liberman: "Почему ИИ сделает одних сверхбогатыми, а других — нищими?".
The useful part is not the crypto pitch by itself. The useful part is the frame: AI is becoming infrastructure. If intelligence becomes a basic input like electricity, internet access, or roads, then the important question is not only which model is smartest. It is who owns the compute, who sets access rules, and whether the price of useful intelligence trends toward marginal cost or toward maximum extractable value.
The interview is also more detailed than a simple "decentralized AI" slogan. It connects internet monopolies, advertising rents, pension-fund ownership, labor-market inequality, open-source agents, compute scarcity, Bitcoin-like incentives, and specialized AI chips into one thesis: the future of AI will be shaped less by chat UI and more by who controls the physical and economic substrate of intelligence.
The interview starts with "copyable products": things that are expensive to create once but cheap to reproduce. Search engines, social networks, marketplaces, operating systems, streaming catalogs, and AI models all have this shape.
This breaks the comforting version of market competition. If a product gets more useful when everyone uses the same one, the winner can become the default environment rather than one vendor among many. That is why internet markets keep producing one search engine, one dominant social graph, a few clouds, and a few model providers.
The problem is not that monopoly is always unnatural. The problem is that once one actor controls the common layer, it can slowly turn the layer toward its own economics. In physical markets, serving the whole planet is hard: logistics, customs, inventory, local operations. In software, the same service can be delivered to another continent at almost zero marginal cost. That makes global concentration much easier.
The Google/Facebook point is simple: a service can be free at the point of use and still expensive for society.
Advertising is a search market between producers and consumers. If one platform controls that search, it can capture the margin. Businesses then pay more to reach customers, and those costs show up inside the products people buy. The user does not see an invoice from Google for a jacket, a taxi ride, or a SaaS subscription, but the ad market's take rate is still embedded somewhere.
The user can also pay in less visible ways: attention, data, dependency, lower bargaining power, or the risk of being cut off from the platform. A business that depends on a marketplace, app store, social graph, search ranking, or model API is not using a neutral tool. It is building on somebody else's toll road.
This is the same reason "we give users AI for $20/month" should not be mistaken for the final economics of AI. Early platform pricing is often subsidized while the market is being captured.
One of the more useful parts of the interview is not technical at all. The Libermans point out that corporations do not maximize profit only because executives are cartoon villains. They are owned by shareholders, and large shareholders are often pension funds, university endowments, index funds, and other institutional investors.
That creates an uncomfortable loop. Society wants corporations to grow because millions of people depend on asset growth for pensions and savings. But the same profit pressure also pushes platforms to extract more rent from users, workers, suppliers, and younger people who do not yet own capital.
This reframes inequality. It is not only "rich versus poor." It is also "those who already own accumulated capital versus those who only sell labor." Younger people may create much of the technological progress, but the returns tend to flow toward the owners of the infrastructure and equity.
In AI this gets sharper. If AI raises productivity while ownership remains concentrated, the upside can accrue to capital while the downside — weaker labor bargaining power — is distributed broadly.
The strongest analogy in the interview is electricity. We do not want a world where electricity for a lamp costs one price but electricity for a life-support machine costs 1,000x more because the provider prices by the value of the use case.
AI providers may be tempted to do exactly that. If a token helps with a toy task, charge one price. If it helps cure disease, automate a factory, or replace a department, charge much more. That is rational for a profit-maximizing company, but it is a grim architecture for a society.
This is why the "AI is just SaaS" frame is too small. If AI becomes a universal productivity layer, access to it becomes labor power. Losing access does not just mean losing a tool; it means competing against people and firms who still have amplified intelligence.
The interview's darker scenario is not that AI becomes expensive entertainment. It is that AI becomes a privately metered layer of intelligence, available on the best terms only to selected partners, governments, or large enterprise customers.
The practical claim: model weights and techniques become more copyable over time. Papers are public, data is often public or scraped, and strong models can be approximated, distilled, or used to train weaker models. DeepSeek is used as an example of the way open or semi-open ecosystems can catch up faster than incumbents expect when enough compute is available.
But compute is physical. The control points are:
Open-source models matter, but they are not enough if only a few actors can run them cheaply at scale. "Open weights" without affordable inference is a library whose books only billionaires can afford to read.
So the durable power may move from "who has the best model file" to "who has cheap, scalable, reliable compute and the right to use it."
A chat interaction is small compared with an agentic workflow. A human writes a prompt and receives an answer. An agent can search, plan, call tools, generate code, retry failures, inspect logs, summarize files, run tests, and ask another model to verify the result. A short user request can become hundreds of thousands or millions of internal tokens.
That changes product architecture. A serious agent stack should not be welded to one model provider. It should be able to route by task, price, latency, privacy, and failure mode.
The Libermans make a sharp observation here: agents do not have human brand loyalty. If you tell an agent to find a cheaper usable model, it can compare providers and move. Humans may say "I like Claude" or "I trust OpenAI." Agents will increasingly behave like cost-sensitive buyers of compute.
This is where decentralized compute becomes interesting even if one is skeptical of the token story. A competing compute market gives agents somewhere else to go.
The project they are pointing to is Gonka: an open-source decentralized network for AI compute. The name is easy to miss in the transcript because it is auto-transcribed as "гонка," but the mechanism matters more than the branding.
The idea is Bitcoin-like, but not identical:
The key difference from Proof of Stake is philosophical and economic: buying tokens should not be enough to control the network. To gain influence, you need to provide compute. The key difference from Bitcoin Proof of Work is that the work is intended to be useful AI work rather than pure hash grinding.
This is the strongest version of their claim: do not invest in a company that owns the infrastructure; participate in the infrastructure. Rent GPU, connect it, mine, or buy coins from miners. If the network grows, the compute layer becomes a public market rather than a corporate asset.
They claim the network grew from a small set of initial participants to thousands of GPUs and a much larger community. The point is not the exact number at the time of publication — it is the bootstrap pattern. If many people see the same opportunity at once, attempts by one actor to dominate can be countered by others joining with more hardware.
The analogy works because Bitcoin showed that a protocol can coordinate strangers into building global infrastructure without a company owner. Mining created a direct economic reason to add hardware. No board of directors had to approve every participant.
But AI compute is harder than Bitcoin mining. A Bitcoin hash is trivial to verify. Useful AI work is not. For inference and training, the network has to care about:
So the right stance is: promising thesis, not magic. A decentralized AI network has to earn its place by being cheaper, reliable enough, and easy enough for agents and developers to use. Ideology does not replace SLA.
The interview also gestures at a hardware transition. Bitcoin did not remain a CPU/GPU mining network; it created a market for specialized ASIC miners. The Libermans expect AI compute to go through a similar phase.
The important nuance: Gonka itself is not necessarily "making chips." The stronger claim is that an open compute protocol can create demand for independent chip makers. If a startup can build a better inference/training accelerator, it does not need to sell only to Meta, Google, Amazon, or Microsoft. It can connect hardware to an open network and monetize compute directly.
They mention or gesture toward companies such as:
The thesis is that if Nvidia GPUs are too expensive and too supply-constrained to become the entire substrate of public AI, specialized AI chips may be the pressure valve. The Bitcoin world produced engineering knowledge around high-density, efficient, specialized compute. Some of that talent and mindset may migrate into AI accelerators.
The interview sketches two futures.
In the first, AI is owned by a few corporations and geopolitical blocs. Everyone else rents access, loses bargaining power, and perhaps receives some kind of subsidy after their labor becomes less valuable. In this version, much of the world outside the US and China becomes a customer base or rent source rather than a co-owner of the infrastructure.
In the second, AI becomes a public productivity layer. Individuals and small organizations get their own agents and eventually robots. Productivity gains spread outward instead of only flowing to the owners of the infrastructure.
Stopping AI is not a serious option. The live question is who owns the substrate.
For AI products, the architectural lesson is clear:
For market analysis, the lesson is similar: do not look only at model labs. Watch inference infrastructure, GPU clouds, energy, data centers, AI ASIC startups, decentralized compute protocols, open-source agent frameworks, and the tooling that lets agents route between providers.
For users, the lesson is more personal: do not let your work become trapped inside one provider's interface. Keep prompts, workflows, notes, automations, and knowledge portable. A future where intelligence is metered by a few companies is much less scary if your own operating system can switch models.
The blog-post version of the whole interview is this:
> Future power in AI will belong not only to those with the best models, but to those who control cheap compute, access to it, and the rules for using it. If AI becomes as basic as electricity, the public question is whether it should be a private metered monopoly or a competitive infrastructure layer.
For agent systems, the practical consequence is even simpler: do not build as if there will be one permanent model provider. Build model/provider-agnostic systems that can choose the best compute route for the task, price, privacy, latency, and availability.