Gonka Protocol is positioning itself around a straightforward but ambitious idea: global AI compute should not be controlled solely by centralized cloud providers when vast amounts of GPU capacity exist around the world. In a recent podcast conversation, co-creator David Liberman outlined why he believes decentralized infrastructure can become a meaningful alternative for AI workloads, especially if incentives are tied directly to real hardware contribution.
According to the discussion, Gonka is building a decentralized network for high-efficiency AI compute that aims to give builders and researchers permissionless access to computing resources. Participants that contribute hardware are rewarded through the network’s native token, GNK. Liberman framed the project as an attempt to apply lessons from Bitcoin’s decentralized infrastructure model to the rapidly expanding market for AI computation.
Why Gonka Looks to Bitcoin for Inspiration
One of Liberman’s central arguments is that Bitcoin demonstrated something many industries once considered unlikely: a decentralized network can coordinate massive, global infrastructure without relying on a central operator. Gonka, in his view, is trying to bring a similar logic to AI compute by organizing GPU owners into a transformer-based Proof-of-Work network.
He contrasted this model with the centralized cloud approach that currently dominates GPU provisioning. In centralized environments, access, pricing, and allocation are often controlled by a relatively small group of providers. Liberman argues that this structure can create inefficiencies in how GPUs are utilized and may limit open participation. A decentralized model, by comparison, is designed to broaden access while making better use of globally distributed hardware.
Inference Over Training
A notable part of Liberman’s thesis is Gonka’s emphasis on AI inference rather than model training. While training typically captures headlines because of its scale and cost, inference is where ongoing real-world usage happens. By focusing on inference, Gonka appears to be targeting a class of demand that could be more continuous, more distributed, and potentially easier to align with a broad network of independent hardware suppliers.
In practical terms, this means the protocol is less concerned with competing directly against the biggest centralized players on frontier model training and more concerned with creating an open compute layer that can serve meaningful AI workloads at scale. Liberman suggested that this focus improves the fit between decentralized infrastructure and market demand.
Why PoW Matters in the Gonka Model
Liberman also made a strong case for a PoW-based architecture over Proof-of-Stake systems in the context of compute networks. His argument is rooted in incentives: a PoW system rewards actual hardware deployment and operational contribution, while PoS systems tend to favor token ownership and capital concentration. For a network whose core purpose is to expand available compute, he believes hardware-aligned incentives are more effective than stake-aligned ones.
That distinction matters because infrastructure growth depends on attracting providers willing to commit real machines, electricity, maintenance, and technical effort. In Liberman’s framing, a network grows more robustly when rewards flow to those increasing practical compute capacity, not merely to those locking up tokens. This, he argued, creates a healthier loop between token economics and infrastructure expansion.
He further suggested that a PoW-based design can reduce the risk of artificial value inflation because token value remains more closely connected to the real cost of compute. In other words, if the network’s economic layer reflects actual production costs, the ecosystem may be less vulnerable to purely narrative-driven distortions.
From GPUs to ASICs
Another key theme in the conversation was hardware specialization. Liberman drew parallels to the evolution of Bitcoin mining, which moved from general-purpose hardware to increasingly specialized systems. He believes AI compute could follow a similar trajectory, with ASICs eventually playing a larger role in optimizing performance and efficiency for specific workloads.
This point is significant because it reframes decentralized AI not just as a marketplace for idle GPUs, but as a long-term infrastructure thesis. If the network can align incentives early, specialized hardware may emerge to serve it more efficiently over time. That, in turn, could improve cost structures and strengthen the network’s competitive position relative to conventional cloud offerings.
Liberman’s view implies that token design should prioritize hardware providers and innovators who push efficiency forward. Rather than treating tokenomics as a separate financial layer, he presented it as a mechanism that should accelerate better compute infrastructure.
AI Valuations and Bubble Risk
The conversation also touched on the current state of the AI market. Liberman acknowledged that AI’s long-term impact is likely to be enormous, but he cautioned that some company valuations may be in bubble territory. He compared today’s environment, at least in part, to the dot-com era: a period when the underlying technological transformation was real, even though many individual valuations later proved unsustainable.
That distinction is important. Liberman is not arguing against AI’s significance. Instead, he is questioning whether current market pricing always reflects durable infrastructure value. In his framework, sustainable upside is more likely to come from systems that improve hardware efficiency, align incentives correctly, and expand access to compute in economically grounded ways.
Decentralized AI as a Competitive Necessity
Beyond economics and hardware, Liberman described decentralized AI as strategically important for global competition. If compute remains concentrated among a handful of national champions or large centralized platforms, smaller countries, startups, and independent builders may face growing structural disadvantages. Open, distributed compute networks could serve as a counterweight by lowering barriers to access.
He argued that this matters not only for innovation, but also for autonomy. A world in which a small number of centralized entities control advanced AI infrastructure may also become a world in which economic power and decision-making become increasingly concentrated. Decentralized compute, in this view, is not simply an engineering preference; it is part of a broader attempt to preserve pluralism in the AI era.
Liberman extended that logic to a more expansive vision of the future. Rather than a landscape dominated by a few giant AI systems, he suggested society may benefit more from billions of independent, decentralized AIs. Such an outcome, he implied, could help prevent extreme concentration of wealth and power while supporting a more abundant and competitive ecosystem.
Who Is Behind Gonka
The podcast also provided background on David and Daniil Liberman. Based in Los Angeles, they were described as futurists, serial entrepreneurs, investors, and former directors of product at Snap. They are also associated with ventures including ProductScience.ai and Humanism Co. by Libermans Co., alongside their work on the Gonka protocol.
The broader Liberman family entrepreneurial track record spans computer graphics, finance, and AI. Among their prior projects, they co-founded Frank Money, focused on radical financial transparency, and later launched Kernel AR, which was acquired by Snap in the same year. The report also noted that the family signed the Founders Pledge and channels its economic output through Libermans Co., which was cited at a valuation of $400 million. Investors mentioned in the article include Marc Andreessen, Josh Kushner, and Arielle Zuckerberg.
The Bigger Picture
Gonka’s pitch arrives at a moment when demand for compute is becoming one of the defining bottlenecks of the AI economy. Liberman’s contribution to the debate is clear: if AI is going to reshape industries and societies, the infrastructure behind it should not be left entirely in centralized hands. A decentralized network, especially one built around PoW and hardware-first incentives, could offer a different path—one aimed at improving utilization, rewarding real contributors, and broadening access to critical compute resources.
Whether that model can scale remains an open question, but the strategic argument is gaining clarity. In Liberman’s telling, decentralized AI compute is not just a niche crypto experiment. It is a bid to build an alternative infrastructure layer for an AI-driven future, one in which efficiency, openness, and competition matter as much as raw model capability.