How Difficult Is It to Build Decentralized AI? Gonka Founder David Liberman's First-Hand Account

null

From GPU computing power from zero to ten thousand, to dozens of DDoS attacks, to secretly discussing AI sovereignty with government officials - this is a true story of faith, decentralization and human gifts.

Source: DeAI Nation’s State of DeAI 2026 report | Finishing: Gonka.ai

David Liberman is the co-founder of Gonka, a decentralized AI inference and training network. The network went live in August 2025 and has accumulated more than 10,000 GPUs (in NVIDIA H100 equivalents) in just a few months. This article is compiled from an exclusive interview with David in DeAI Nation’s “State of DeAI 2026” report, covering his in-depth analysis of the “AI version of Bitcoin” thesis, decentralized boundary disputes, numerous attacks on Gonka, and negotiations with governments on AI sovereignty.

1. Everyone wants to be the Bitcoin of the AI world

When observing the decentralized AI ecosystem, there is a phenomenon that stands out: almost every project and blockchain is trying to shape itself as the new Bitcoin in the AI world. Is this due to the inertia of the crypto industry, or is there some deeper structural reason behind it?

David gave his judgment: this is a superposition of two motivations, but the focus of different projects is different.

From a comparative logical point of view, this analogy is not an exclusive habit of the crypto circle, it is spread throughout the technology community and even the entire entrepreneurial ecosystem. When a disruptive new thing emerges in a field, the pioneers must prove the feasibility of the underlying logic from scratch, while the latecomers can stand on the shoulders of giants and endorse themselves with existing successful cases. Just as Silicon Valley investors ask entrepreneurs to introduce themselves at the beginning of the financing PPT: ‘We are Airbnb for pet dog owners’ - this short sentence saves the tedious time of repeatedly demonstrating the feasibility of the platform economic model.

Bitcoin is not the first decentralized project in history, nor is it the first open source project, and even before it, BitTorrent was already a typical model of a decentralized network. What Bitcoin really proves is that the incentive model based on tokenomics can operate on its own in the real world. The value of this proof allows all subsequent projects built on tokenomics to confidently skip this demonstration link.

“We used the Bitcoin analogy in part to save ourselves the trouble of reproving the viability of tokenomics. There are still some skeptics who believe that Bitcoin will eventually return to zero, although there are fewer and fewer of them.” ——David Liberman

However, for Gonka, this analogy has a deeper meaning. While most crypto projects are turning to proof-of-stake (PoS), Gonka sticks to proof-of-work (PoW) and builds its computing infrastructure with it at its core. David made it clear that Gonka is following the path of Bitcoin, not modern Ethereum. Ethereum also initially adopted PoW, which also spawned the development of infrastructure such as mining machines, but later shifted to PoS and gradually moved away from this infrastructure incentive system.

His judgment is that PoW can create stronger infrastructure incentives. Of course, it is understandable that other projects choose to use the Bitcoin analogy to express themselves - the point is that no one is claiming to reach the size of Bitcoin’s market capitalization, but saying that the underlying propositions that have been verified by Bitcoin also apply to us, and the only new variable is AI.

2. How does Silicon Valley view decentralized AI?

When the concept of “decentralized AI” came to Silicon Valley, it provoked a reaction that was far more complex than the outside world imagined - not only from the enthusiastic endorsement of crypto investors, but also from the deep reflection of the AI security research circle, and the silent wait-and-see from the practitioners of large model companies.

David mentioned two representative voices: a16z partner Chris Dixon has long spoken out in support of decentralized AI and has a layout in this field; Shaun Maguire, a partner at Sequoia Capital, wrote that crypto and AI are inherently a pair. Although some believe that Dixon’s stance stems from his crypto background, these voices still constitute a positive footnote to decentralized AI within Silicon Valley.

What is more noteworthy is the quiet turn of the AI security research circle. David pointed out that almost all the founding scientists of modern AI were born out of the AI safety research community. The birth of OpenAI, which itself stemmed from concerns about Google’s monopoly on AI, was an alternative aimed at checks and balances - but when OpenAI itself gradually approached its monopoly position, this original intention quietly collapsed.

“The AI security community used to oppose decentralization and were unwilling to release AI capabilities to ordinary people. But when computing power was highly concentrated by a few giants, the community began to realize that without enough computing power, no AI security research could be promoted. As a result, their attitude towards decentralization is undergoing a fundamental shift.”

At the same time, the appeal of decentralized AI is becoming increasingly clearly linked to cost among the broader developer community. David’s observation is that when a project is just starting out and has VC funds, there is no cost pressure to use centralized inference services; But as the scale expands, the bill will suddenly wake up. He gave a vivid example: a large number of developers connected their AI Agent to Claude Opus, and the next morning found that the agent was running all night, and the token consumption was shocking, so they began to eagerly look for alternatives.

OpenRouter’s data changes confirm this trend: two months ago, the top models on the platform were almost entirely closed-source; Today, the proportion of open source models has increased significantly. David’s judgment is: “Every financial crisis pushes more people to Bitcoin, and the mass adoption of decentralized AI will unfold in the same way - wave after wave, each wave will leave more users behind. The first wave will be driven by price.”

3. Where is the boundary of decentralization?

While the entire industry is shouting the slogan of “decentralization,” the word itself is quietly losing its precision. David admits that this concept has been diluted to varying degrees - partly because true decentralization is extremely difficult to achieve at the engineering level, and partly because some projects have maintained control over core power for a long time in the name of “progressive decentralization”.

He understands the trade-offs that make trade-offs: "Whenever you claim to be completely decentralized, you get in the way at every step. Some projects say, ‘We are not decentralized here, only there’, especially in terms of AI infrastructure, many early projects have made too big compromises. For me personally, too much compromise sometimes undermines the credibility of the decentralized idea itself. "

Gonka’s choice on this point is particularly clear: from the beginning, the team chose not to retain control for itself and to hand over governance to the community. This drew a lot of criticism from the outside world, but David insisted: “Why should anyone trust a centralized authority? Decentralization is what really attracts trust.” The price is real – every change must be negotiated with everyone.

In David’s view, there is an imprecise but generally established law in this industry: projects with higher decentralization tend to carry greater value. The market capitalization of Bitcoin and Ethereum has long been higher than that of XRP and even Solana; Projects that are found to have founders and foundations actually controlling the entire ecosystem often lose a significant amount of market capitalization.

“Decentralization is not a marketing label, but a mechanism for long-term trust accumulation. In this industry, the filters around power structures are real, although they don’t always work in a timely manner.”

He also clearly expressed his respect for Prime Intellect, believing that it is a team with outstanding capabilities and dares to face the hardcore challenge of decentralized training head-on. However, he also pointed out that there is still no clear answer to the business model of decentralized training - because free and open source models with stronger capabilities continue to emerge, making the competition in the training market more difficult. Gonka’s final choice to focus on reasoning is based on a sober judgment of business reality: reasoning generates continuous demand, gives birth to real infrastructure, and is also the direction in which capital is really willing to flow in.

4. Attack, Collapse, and Resilience: Gonka’s life-and-death test

Since its launch in August 2025, Gonka has undergone a much more intense stress test than expected.

David admitted that Gonka suffered not one DDoS attack, but dozens. Attacks began in the first month of launch and were initially small and simple, but by late December 2025 to January 2026, the scale and sophistication of the attacks had significantly escalated. Attackers continue to look for every possible vulnerability and constantly test the limits of the system.

This has caused Gonka to suffer from a strong decentralized design: in a centralized system, attacks can be uniformly scheduled and directly responded to by the core team; But in a decentralized network, every miner must secure their own infrastructure. The network is home to both veteran crypto miners and new players attracted by the idea of decentralized AI, who do not have the experience and tools to combat cyberattacks, making security education at the community level a priority.

During the peak attack period, several nodes are still offline due to attacks every day. But the more serious problem comes from Gonka’s original incentive design: when a miner is attacked and cannot prove availability, its reward for the day is confiscated and redistributed to the rest of the miners - which means that defeating 30% of miners can increase their earnings by 30%. attack, become profitable.

“We have experienced a paradox firsthand: decentralization makes us more vulnerable to attacks, but it also makes our defenses stronger because of the participation of the community.”

The community then voted to modify this mechanism so that attacking others no longer directly yields financial gain. The attack did not disappear, but the motivation behind it was greatly compressed. David admits that he now understands why some projects choose to centralize API access - a distributed, publicly available API node that is far more difficult to protect than a centralized architecture. But Gonka’s position remains the same: APIs should remain open and decentralized, as this is at the heart of the entire project philosophy.

At the same time, the downturn in the macro crypto market has also put pressure on them. Bittensor’s GPU count has fallen, and Gonka’s peak GPU count has also declined. But David characterizes this period as a “breathing period”: “If Bitcoin were at $120,000 today, the number and scale of attacks could be several times higher than it is now. Now is the best time to take advantage of the market quiet and build a stronger line of defense before the next bull market comes.”

Even after all this, the Gonka network still runs about $200 million worth of hardware assets online, and the number of GPUs is still significantly ahead of other similar projects. David sees this as a concrete manifestation of the community’s beliefs.

5. Governments talk about AI sovereignty: computing power is power

In Gonka’s journey, another parallel storyline is also striking: David and Daniil frequently meet with government officials and executives of large enterprises to discuss the possibility of decentralized AI at the national strategic level. These conversations reveal a bigger picture that goes beyond business logic.

David observes that governments’ interest in decentralized AI ultimately stems from three levels of motivation.

Motivation 1: Computing power sovereignty

At present, government services in many countries are deeply dependent on AI, but the computing power behind them is in the hands of external service providers. This dependency poses not only a cost issue, but also a strategic risk: once an external vendor controls access, pricing power, or infrastructure, it can be used as a bargaining chip to limit or even cut off critical services. This structural fragility is the issue that government officials are most wary of.

Motivation 2: Local industrial development

Governments want their data center industry to truly take root locally, not just as a “cloud access point” for foreign-funded enterprises. They expect local jobs, local capital accumulation, and long-term technology capacity building — not giving up data and profits to a handful of hyperscale cloud service providers.

Motivation 3: Participate in the chip industry chain

Some countries have extended their attention to the upstream link: not only to operate data centers, but also to participate in semiconductor manufacturing. This is not a dream, because the entry point is not the most advanced 3nm process, but more mature process nodes such as 16nm - which is realistic for more countries.

At the intersection of these triple motives, the narrative of decentralized AI networks begins to show its unique persuasive power.

“What we are showing them is not just sovereignty, but a viable economic model. If a country participates in a decentralized computing power network, it can build a 20,000-GPU data center and get endogenous demand from the global market — rather than counting on Microsoft or a hyperscaler willing to rent your computing power at a reasonable price.”

David used Bitcoin to draw an analogy: Bitcoin does not require any single country to occupy a structural advantage, and the natural growth of computing power has been completed on a global scale. Tokenomics creates distributed economic incentives, allowing countries to choose to participate independently without being attached to a centralized ecosystem. He believes that the same logic can be transplanted to the global distribution of AI computing power.

Of course, there are practical complexities to this: local infrastructure is often difficult to run at full capacity around the clock, and idle rates are a stubborn economic problem. David’s solution is a hybrid model of “local + distributed”: while the local cluster handles the basic load, the idle computing power is connected to the global decentralized network, turning idle resources into continuous income; During peak hours, additional computing power is called from the network in reverse to respond to sudden demand. He cited the logic of the birth of Amazon Web Services - it is precisely because of the huge demand for elastic computing power on e-commerce platforms during peak holidays that the business form of cloud computing was born, and today’s AI computing power scheduling is facing the same structural problems.

6. The other side of training: a gift to mankind

As a future vision, Gonka proposes to allocate 20% of inference revenue to decentralized model training. David not only maintained sincere expectations for this, but also did not shy away from the difficulties.

He bluntly said that decentralized training is still an unsolved engineering problem, and almost no one has found the answer to its commercial feasibility. The reason is simple: the open source community is constantly popping up with more powerful, completely free foundation models, which has almost killed the market space for independent training. Any project that tries to commercialize through decentralized training will have a hard time competing with free open-source alternatives for its output — unless your goal is to become a cutting-edge AI lab.

Gonka chose another path: first focus on reasoning, build infrastructure and token economics to form a real scale effect, and then use part of the network’s capabilities for training. The logic of this road is: there is a scale of computing power first, and then there is the possibility of training, not the other way around.

“Training may not be our growth engine, but it can be our gift to humanity. Why not? No one has lost anything for this, and we have the opportunity to give something really valuable to the world.”

David admits that there are many prerequisites for getting to this point: engineering physics, collective consensus from the community, and the continued growth of Gonka’s overall network. He knew very well that this would not happen anytime soon. But he also points out that any breakthrough made by teams that have poured tens of millions of dollars into this direction and worked around the clock will ultimately belong to all of humanity - because it is often much easier to replicate an achievement than to achieve it the first time. He admires these teams and positions Gonka’s primary mission as building a decentralized computing infrastructure that truly competes with top cutting-edge labs and hyperscalers.

Epilogue

David Liberman’s story is about entrepreneurs navigating dangerous shoals – dealing with attacks from the internet, proving the value of decentralized AI to government officials who are still on the sidelines, and maintaining the community’s belief in an uncertain crypto market cycle.

However, behind all this, there is a clear thread running throughout: decentralization is not a marketing slogan, but an infrastructure construction philosophy that accumulates trust over time. Gonka chose the most difficult path, and because of this, it has come to this day.

This experiment on decentralized AI is far from the moment when the coffin is closed. But as David said, the price of every first mover will be the starting point for those who come later. And those who persevere in the most difficult moments will eventually see the meaning of what they do in the next wave.