‘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

by
Alisa Davidson

Printed: July 17, 2026 at 9:20 am Up to date: July 17, 2026 at 12:37 pm

by Anastasiia O

Edited and fact-checked:
July 17, 2026 at 9:20 am

To enhance your local-language expertise, generally we make use of an auto-translation plugin. Please be aware auto-translation will not be correct, so learn authentic article for exact data.

The AI race is often framed as a contest between fashions, chips, and capital. However a rising variety of voices argue the actual chokepoint is less complicated and extra bodily: electrical energy, grid entry, and the handful of companies that management the infrastructure all the things else runs on. When you don’t personal compute, the argument goes, your AI technique is only a request.

The implications of that logic are already taking part in out. Grid operators are sounding alarms, utility payments are climbing, and the query of who will get to entry AI — not simply use it, however run it, personal it, construct on it — is quietly turning into one of many extra consequential infrastructure questions of the last decade. For many of the world, the reply proper now could be: on another person’s phrases.

MPost sat down with Daniil and David Liberman, co-creators of Gonka, a decentralized AI compute community, to seek out out whether or not there may be one other manner.

At what level did the AI power drawback turn into an emergency? What breaks subsequent?

The emergency began when the limiting issue stopped being chips and have become electrical energy. For years, the AI trade handled knowledge facilities like monetary objects. You increase cash, purchase GPUs, and construct a large campus. However knowledge facilities are bodily objects tied to transformers, energy crops, and native grids. These issues can not scale on the velocity of software program. The current July orders present that grids like PJM are actively managing round knowledge heart demand, utilizing backup era as an absolute final resort. What breaks subsequent is your complete financial mannequin. You’ll be able to have billions of {dollars} in GPUs and actually nowhere to plug them in. Then the political layer breaks as households begin asking why their utility payments are funding an AI facility. The deeper mistake is assuming each AI computation requires a gigawatt scale campus. We have to transfer compute towards power: a decentralized strategy routes workloads to current machines and areas the place energy is already out there.

Who’re the lords and serfs of GPU feudalism? What does the hire appear to be?

The lords are the few actors who management the bodily and contractual layers, like superior chips, hyperscale clouds, and the power required to run them. The serfs are nearly everybody else. They’re the startups, builders, and whole nations that should import intelligence as a service. That is precisely what we name GPU feudalism, a future the place individuals turn into tenants on another person’s compute property. The hire they pay is partly monetary, however the extra harmful hire is permission. A supplier can change your quota, droop your account, or silently swap the mannequin you’re utilizing. If 5 corporations and two states management the infrastructure, you get eight billion tenants. AI is pure data and will be copied nearly with out restrict. The mannequin matches on a flash drive, however the monopoly lives within the knowledge heart. That bodily bottleneck is strictly the place synthetic shortage is imposed on one thing meant to create international abundance.

Why does infrastructure possession decide what AI will get constructed?

An concept with out compute just isn’t an organization, only a doc. When compute entry is pricey or politically restricted, builders bodily change what they construct. A startup would possibly abandon an open supply mannequin just because a large cloud supplier subsidizes entry to a proprietary API. The infrastructure homeowners dictate which {hardware} is supported, which areas get capability, and who will get precedence throughout shortages. Because of this open weights are by no means sufficient. Publishing mannequin weights with out inexpensive execution is like sharing blueprints for an plane when one single company owns each runway. You additionally want verifiability. A decentralized system should show that the anticipated mannequin was truly executed with out silent reductions in high quality. An agent that strikes cash wants greater than belief, it wants proof. If you don’t management compute, your AI coverage is a request relatively than a technique.

How shut are we to picking between permissions relatively than fashions?

We’re already residing within the first model of that world. Customers assume they’re selecting amongst dozens of AI merchandise, however many are simply interfaces constructed on the very same small group of mannequin suppliers. At present, permissions appear to be geographic restrictions, utilization tiers, and API insurance policies. Tomorrow, when AI methods turn into important for financial productiveness, these restrictions turn into structural. Think about one engineer accessing the most effective coding agent whereas one other doesn’t. The corporate or nation denied entry instantly falls behind. The final word consequence is a hierarchy of intelligence the place entry relies upon solely in your jurisdiction, employer, or id. Individuals will technically be free to reject these methods, however they may now not be economically aggressive with out them. We are able to nonetheless reverse this, however the window is measured in years. Each API shutdown or entry denial is simply one other commercial for decentralized AI.

Who wins as inference turns into the dominant battleground?

Coaching fashions generates the largest headlines, however inference is the place AI turns into an actual financial system. A mannequin is educated periodically, however inference occurs repeatedly each single time a consumer asks a query or a robotic comes to a decision. This shifts the aggressive panorama. Whereas coaching requires large and tightly related clusters, inference can regularly be processed independently. You don’t want each request to run inside a single gigawatt campus. The winners will likely be methods that scale back the associated fee per helpful token, keep excessive {hardware} utilization, and assure mannequin integrity. The losers will likely be companies completely counting on sponsored API pricing and nations importing each unit of intelligence. That is the precise opening for distributed methods. Decentralized compute can make the most of international provide and native power with out requiring one company to personal each facility.

What do 100 million requests throughout 26 nations reveal about untapped compute?

They show that geographically distributed AI compute is now not only a theoretical idea. Earlier than working an actual community, individuals assumed the world contained a homogeneous pool of idle GPUs ready to be activated. The truth is that {hardware}, networking, and availability differ drastically. The actually scarce useful resource is coordination. A community should uncover capability, measure efficiency, confirm execution, and create incentives that hold operators on-line. Because of this Gonka makes use of a transformer-based Proof of Work and Dash mechanisms to confirm actual AI {hardware}. The untapped provide is very large, together with enterprise reserves, regional clouds, and college {hardware} that merely lacked a connection to international AI demand. Processing these workloads exhibits that builders can devour decentralized inference by way of a typical interface with no need all compute to reside inside a company perimeter.

Can an open different nonetheless emerge, and at what scale?

Sure, but it surely won’t emerge mechanically from the present market leaders. An open different wants to succeed in large bodily scale to turn into a real structural choice. We’re speaking about gigawatts of capability and tens of millions of superior GPUs. Bitcoin supplies the right lesson right here. It didn’t increase billions of {dollars} to construct one monumental knowledge heart. As a substitute, it created an financial protocol that mobilized international bodily infrastructure. AI wants a comparable normal for helpful computation. Open fashions require open execution infrastructure. Financial rewards should circulation on to the individuals who present and enhance {hardware}. If AI will get a compute-aligned protocol, {hardware} innovation will transfer from sluggish company roadmaps to international competitors.

Is shared compute a path to AI sovereignty for rising areas?

It completely will be, supplied sovereignty is outlined appropriately. AI sovereignty doesn’t imply each nation should fabricate its personal chips and construct a home hyperscale cloud. Let’s be reasonable — it’s not attainable. A extra sensible definition is {that a} nation can’t be unilaterally disconnected from intelligence. Renting an API from a overseas firm is pure dependence. A rustic with out compute will negotiate with AI precisely like a tenant negotiates with a landlord. Connecting home power, engineers, and knowledge facilities to an open community is true participation. No medium sized nation can outspend america or China individually. But when 200 nations every contribute ten thousand GPUs on common, they create a shared community of two million GPUs. This coordination ensures sovereign entry to the intelligence layer, giving rising areas in Asia, Africa, and the Center East an actual exit choice from centralized management.

Is decentralized AI structurally aggressive, or solely filling Massive Tech’s gaps?

At present, it’s doing each. Most decentralized networks begin in areas the place centralized suppliers are costly, restrictive, or just uninterested. That’s utterly regular. Nonetheless, it turns into structurally aggressive when it creates an open normal that coordinates international power, rewards higher {hardware}, and provides builders a reputable exit technique. The related metric isn’t the market capitalization of a token. The decisive metric is beneficial AI computation processed per day. Decentralized methods don’t want centralized suppliers to vanish. The web didn’t eradicate personal networks, and Bitcoin didn’t eradicate banks. As soon as builders have a viable different, incumbents are pressured to compete way more severely on worth, entry, and transparency.

Why would decentralized AI keep away from reconsolidation?

Reconsolidation is an actual threat. Decentralization just isn’t a everlasting state simply because somebody wrote the phrase in a white paper. AI compute has large economies of scale, which means bigger operators can negotiate higher electrical energy contracts and purchase {hardware} extra cheaply. A permissionless community may simply consolidate round a number of industrial operators. That is precisely why Gonka is constructed on compute-weighted governance, making certain that affect is related to verified computational contribution relatively than easy token possession. The way forward for AI infrastructure should be ruled by the individuals who truly deliver intelligence into the community, not by idle capital. The objective is to not forestall massive operators from current. Giant operators are extremely environment friendly. However we now have to make sure they continue to be operators inside an open protocol relatively than turning into absolute homeowners of the protocol itself.

Alisa, a devoted journalist on the MPost, focuses on crypto, AI, investments, and the expansive realm of Web3. With a eager eye for rising traits and applied sciences, she delivers complete protection to tell and have interaction readers within the ever-evolving panorama of digital finance.

Extra articles