‘Your AI Policy Is A Request, Not A Strategy’: Gonka’s Liberman Brothers On GPU Feudalism, The Energy Bottleneck, And Who Controls The Intelligence Layer
In Brief
AI’s real chokepoint is electricity and grid access, not chips. Gonka’s founders on decentralized compute, GPU feudalism, and open AI infrastructure.
The AI race is usually framed as a competition between models, chips, and capital. But a growing number of voices argue the real chokepoint is simpler and more physical: electricity, grid access, and the handful of corporations that control the infrastructure everything else runs on. If you don’t own compute, the argument goes, your AI strategy is just a request.
The consequences of that logic are already playing out. Grid operators are sounding alarms, utility bills are climbing, and the question of who gets to access AI — not just use it, but run it, own it, build on it — is quietly becoming one of the more consequential infrastructure questions of the decade. For most of the world, the answer right now is: on someone else’s terms.
MPost sat down with Daniil and David Liberman, co-founders of Gonka, a decentralized AI compute network, to find out whether there is another way.
At what point did the AI energy problem become an emergency? What breaks next?
The emergency started when the limiting factor stopped being chips and became electricity. For years, the AI industry treated data centers like financial objects. You raise money, buy GPUs, and build a massive campus. But data centers are physical objects tied to transformers, power plants, and local grids. Those things cannot scale at the speed of software. The recent July orders show that grids like PJM are actively managing around data center demand, using backup generation as an absolute last resort. What breaks next is the entire economic model. You can have billions of dollars in GPUs and literally nowhere to plug them in. Then the political layer breaks as households start asking why their utility bills are funding an AI facility. The deeper mistake is assuming every AI computation requires a gigawatt scale campus. We need to move compute toward energy: a decentralized approach routes workloads to existing machines and regions where power is already available.
Who are the lords and serfs of GPU feudalism? What does the rent look like?
The lords are the few actors who control the physical and contractual layers, like advanced chips, hyperscale clouds, and the energy required to run them. The serfs are almost everyone else. They are the startups, developers, and entire countries that must import intelligence as a service. This is exactly what we call GPU feudalism, a future where people become tenants on someone else’s compute estate. The rent they pay is partly financial, but the more dangerous rent is permission. A provider can change your quota, suspend your account, or silently swap the model you are using. If five companies and two states control the infrastructure, you get eight billion tenants. AI is pure information and can be copied almost without limit. The model fits on a flash drive, but the monopoly lives in the data center. That physical bottleneck is exactly where artificial scarcity is imposed on something meant to create global abundance.
Why does infrastructure ownership determine what AI gets built?
An idea without compute is not a company, just a document. When compute access is expensive or politically restricted, developers physically change what they build. A startup might abandon an open source model simply because a massive cloud provider subsidizes access to a proprietary API. The infrastructure owners dictate which hardware is supported, which regions get capacity, and who gets priority during shortages. This is why open weights are never enough. Publishing model weights without affordable execution is like sharing blueprints for an aircraft when one single corporation owns every runway. You also need verifiability. A decentralized system must prove that the expected model was actually executed without silent reductions in quality. An agent that moves money needs more than trust, it needs proof. If you do not control compute, your AI policy is a request rather than a strategy.
How close are we to choosing between permissions rather than models?
We are already living in the first version of that world. Users think they are choosing among dozens of AI products, but many are just interfaces built on the exact same small group of model providers. Today, permissions look like geographic restrictions, usage tiers, and API policies. Tomorrow, when AI systems become essential for economic productivity, these restrictions become structural. Imagine one engineer having access to the best coding agent while another does not. The company or country denied access immediately falls behind. The ultimate outcome is a hierarchy of intelligence where access depends entirely on your jurisdiction, employer, or identity. People will technically be free to reject these systems, but they will no longer be economically competitive without them. We can still reverse this, but the window is measured in years. Every API shutdown or access denial is just another advertisement for decentralized AI.
Who wins as inference becomes the dominant battleground?
Training models generates the biggest headlines, but inference is where AI becomes a real economy. A model is trained periodically, but inference happens continuously every single time a user asks a question or a robot makes a decision. This shifts the competitive landscape. While training requires massive and tightly connected clusters, inference can frequently be processed independently. You do not need every request to run inside a single gigawatt campus. The winners will be systems that reduce the cost per useful token, maintain high hardware utilization, and guarantee model integrity. The losers will be businesses permanently relying on subsidized API pricing and countries importing every unit of intelligence. This is the exact opening for distributed systems. Decentralized compute can utilize global supply and local energy without requiring one corporation to own every facility.
What do 100 million requests across 26 countries reveal about untapped compute?
They prove that geographically distributed AI compute is no longer just a theoretical concept. Before operating a real network, people assumed the world contained a homogeneous pool of idle GPUs waiting to be activated. The reality is that hardware, networking, and availability differ drastically. The truly scarce resource is coordination. A network must discover capacity, measure performance, verify execution, and create incentives that keep operators online. This is why Gonka uses a transformer-based Proof of Work and Sprint mechanisms to verify real AI hardware. The untapped supply is massive, including enterprise reserves, regional clouds, and university hardware that simply lacked a connection to global AI demand. Processing these workloads shows that developers can consume decentralized inference through a common interface without needing all compute to reside inside a corporate perimeter.
Can an open alternative still emerge, and at what scale?
Yes, but it will not emerge automatically from the current market leaders. An open alternative needs to reach massive physical scale to become a genuine structural option. We are talking about gigawatts of capacity and millions of advanced GPUs. Bitcoin provides the perfect lesson here. It did not raise billions of dollars to build one enormous data center. Instead, it created an economic protocol that mobilized global physical infrastructure. AI needs a comparable standard for useful computation. Open models require open execution infrastructure. Economic rewards must flow directly to the people who provide and improve hardware. If AI gets a compute-aligned protocol, hardware innovation will move from slow corporate roadmaps to global competition.
Is shared compute a path to AI sovereignty for emerging regions?
It absolutely can be, provided sovereignty is defined correctly. AI sovereignty does not mean every country must fabricate its own chips and build a domestic hyperscale cloud. Let’s be realistic — it’s not possible. A more practical definition is that a country cannot be unilaterally disconnected from intelligence. Renting an API from a foreign company is pure dependence. A country without compute will negotiate with AI exactly like a tenant negotiates with a landlord. Connecting domestic energy, engineers, and data centers to an open network is true participation. No medium sized country can outspend the United States or China individually. But if two hundred countries each contribute ten thousand GPUs on average, they create a shared network of two million GPUs. This coordination guarantees sovereign access to the intelligence layer, giving emerging regions in Asia, Africa, and the Middle East a real exit option from centralized control.
Is decentralized AI structurally competitive, or only filling Big Tech’s gaps?
Today, it is doing both. Most decentralized networks start in areas where centralized providers are expensive, restrictive, or simply uninterested. That is completely normal. However, it becomes structurally competitive when it creates an open standard that coordinates global energy, rewards better hardware, and gives developers a credible exit strategy. The relevant metric is never the market capitalization of a token. The decisive metric is useful AI computation processed per day. Decentralized systems do not need centralized providers to disappear. The internet did not eliminate private networks, and Bitcoin did not eliminate banks. Once developers have a viable alternative, incumbents are forced to compete much more seriously on price, access, and transparency.
Why would decentralized AI avoid reconsolidation?
Reconsolidation is a real risk. Decentralization is not a permanent state just because someone wrote the word in a white paper. AI compute has massive economies of scale, meaning larger operators can negotiate better electricity contracts and buy hardware more cheaply. A permissionless network could easily consolidate around a few industrial operators. This is exactly why Gonka is built on compute-weighted governance, ensuring that influence is connected to verified computational contribution rather than simple token ownership. The future of AI infrastructure must be governed by the people who actually bring intelligence into the network, not by idle capital. The goal is not to prevent large operators from existing. Large operators are highly efficient. But we have to ensure they remain operators inside an open protocol rather than becoming absolute owners of the protocol itself.
Disclaimer
In line with the Trust Project guidelines, please note that the information provided on this page is not intended to be and should not be interpreted as legal, tax, investment, financial, or any other form of advice. It is important to only invest what you can afford to lose and to seek independent financial advice if you have any doubts. For further information, we suggest referring to the terms and conditions as well as the help and support pages provided by the issuer or advertiser. MetaversePost is committed to accurate, unbiased reporting, but market conditions are subject to change without notice.
About The Author
Alisa, a dedicated journalist at the MPost, specializes in crypto, AI, investments, and the expansive realm of Web3. With a keen eye for emerging trends and technologies, she delivers comprehensive coverage to inform and engage readers in the ever-evolving landscape of digital finance.
More articles
Alisa, a dedicated journalist at the MPost, specializes in crypto, AI, investments, and the expansive realm of Web3. With a keen eye for emerging trends and technologies, she delivers comprehensive coverage to inform and engage readers in the ever-evolving landscape of digital finance.
