Core Summary: Training large models requires building or upgrading data centers. However, centralized infrastructure is now facing hard physical limits. To enhance infrastructure capabilities, AI is being used to create greater scale and intelligent output. Yet, control over computing power is becoming a critical power node in the AI industry. At this time, Gonka has emerged. The Gonka protocol is a permissionless global network that anyone can join, with requests routed programmatically among distributed participants. In an exclusive conversation with Analytics Insight, Anastasia Matveeva, Co-founder and Senior Product Manager of Gonka, discussed how they are innovating in the way computing power is accessed to build a more controllable and secure AI ecosystem.
Q: Public discussions about AI often focus on the centralization of models, but there is less attention on the centralization of computing power. Why is control over computing power becoming a critical power node in the AI industry? What risks does this concentration pose to innovation and the overall market?
A: Public discussions often focus on models because they are visible. But the real core of power lies deeper—at the computing power layer, which is the foundational layer that determines who can build, deploy, and scale AI systems.
Control over computing power is becoming critical due to economic and physical reasons. The main bottleneck for modern AI is no longer algorithms, but the ability to access GPUs, electricity, and data center capacity.
Training large models increasingly requires building or upgrading data centers. However, centralized infrastructure is encountering physical limits: energy density, cooling constraints, and the maximum power supply capacity a single location can handle. The industry is attempting extreme solutions—redesigning chips, cooling systems, and exploring new energy sources.
This concentration has systemic consequences.
First, it creates structural barriers to innovation. Access to computing power becomes a matter of infrastructure privilege, not competition based on capability. Small teams, independent researchers, and even entire regions are priced out, experimental space shrinks, and innovation becomes conservative.
Second, the centralization of computing power reinforces a "rent extraction" model. AI has the potential to create "abundance"—intelligence is essentially replicable—but when the underlying infrastructure is scarce and controlled, this abundance is artificially suppressed. The market shifts towards subscription models, lock-in effects, and pricing power, rather than cost reduction and broad accessibility.
Third, it introduces systemic fragility. When advanced computing power is concentrated in the hands of a few operators and locations, disruptions—regulatory, political, or physical—can ripple through the entire AI ecosystem. Dependence becomes structural, not optional.
More importantly, computing power is not neutral. Whoever controls computing power implicitly decides what is feasible, permissible, and economically sustainable. When this control is centralized, AI governance is formed by default, not by design.
The risk is not just monopoly, but a long-term distortion of AI's development trajectory: fewer builders, less application diversity, slower hardware innovation, and infrastructure unable to match the ambitions of next-generation models.
Therefore, computing power must be seen as foundational infrastructure—an architecture that can scale economically and physically, crucial for the future of AI.
Q: Many AI computing platforms—whether centralized or decentralized—claim to be efficient. What metrics truly matter when evaluating the efficiency of an AI computing system? Where do these models typically encounter practical limitations?
A: Computing efficiency is often used as a marketing concept. In reality, only a few specific metrics truly matter, covering user-side performance, provider operational efficiency, and the incentive structures governing both.
For users, efficiency means speed and cost transparency.
Speed refers to latency under real demand. Centralized hubs often have an advantage due to physical co-location. But if the blockchain acts only as a security layer and does not participate in the real-time execution path, decentralized architectures can achieve similar performance. As long as requests are processed off-chain, the protocol itself does not add latency.
Cost transparency is equally critical. While "cost per token" is a common KPI, model integrity often lacks transparency. In centralized environments, the product can be a black box. During peak periods, providers might adjust model configurations to maintain profits; these changes are often invisible but can affect output quality. True efficiency requires pricing to reflect consistent computational accuracy.
For providers, efficiency is a balance between GPU utilization and elasticity.
Centralized operators excel in utilization; GPUs in co-located environments can run near full capacity. But they lack elasticity, bearing idle costs during demand troughs.
Decentralized networks sacrifice some utilization for elasticity but must minimize consensus and verification overhead so that computing power can be reallocated across different workloads as demand changes.
Most critical is incentive design.
When rewards are tied to faster, cheaper, verifiable AI workloads, optimization becomes structural. Participants are incentivized to improve hardware efficiency, reduce latency, and experiment with specialized chips.
Conversely, if rewards or governance weight are primarily tied to capital holdings, the direction of optimization shifts away from infrastructure performance, and inefficiency becomes entrenched.
In Gonka, efficiency is embedded at the protocol layer: almost 100% of computing power is used for real AI workloads (primarily inference). Rewards and governance weight are based on measured computational contributions, not capital holdings.
True efficiency only emerges when the majority of computing power is used for real tasks, incentives reward verified contributions, and internal overhead does not grow uncontrollably with network scale.
Q: Is it possible for decentralized AI computing networks to dedicate most of their computing power to real AI workloads, rather than maintaining the network itself? What are the key architectural choices?
A: It is possible—but only if overhead is treated as a core architectural constraint, not an inevitable byproduct of decentralization.
Most decentralized computing networks use significant resources for maintaining consensus and security, not AI workloads. This is because productive work and security mechanisms are separated, leading to redundant computation.
To dedicate most computing power to real AI tasks, several key principles are needed:
First, security and measurement mechanisms must be "time-bound," not continuously running. Proof mechanisms should be concentrated in clear, short cycles, not constantly consuming resources. In Gonka, this is achieved through Sprints (structured, time-bound cycles). Outside these cycles, hardware resources are available for real AI workloads.
Second, reduce redundancy through selective and reputation-based dynamic adjustment of verification, rather than fully replicating verification for every task. New participants' work might be 100% verified; as reputation is established, the verification ratio can be reduced to about 1%. Overall verification computing power can be kept below about 10% while maintaining security.
Participants attempting to cheat do not receive rewards, making cheating economically irrational.
Third, rewards and governance weight must be tied to verified computational contributions, not capital holdings.
When consensus is lightweight, verification is adaptive, and incentives are aligned with productive computation, decentralized computing can truly serve practical workloads.
Q: Decentralized AI computing networks often emphasize open participation, but infrastructure requirements can create high barriers to entry. How can such systems scale while remaining accessible to participants with vastly different levels of computing power?
A: While decentralized networks aim to lower the barrier to entry for AI infrastructure, long-term survival also requires competing with centralized providers and meeting real-world demands. Hardware constraints ultimately boil down to a core requirement: the ability to host models that have genuine market demand.
To scale while maintaining accessibility, several principles are crucial.
First, permissionless infrastructure access. Any GPU owner—whether a single-device operator or a large data center—should be able to join the network without an approval process or centralized gatekeeping mechanism. This eliminates structural entry barriers.
Second, proportional rewards and influence based on verified computing power. In a model weighted by computing power, higher computational contributions naturally lead to a larger share of tasks, rewards, and governance weight. This does not make small participants completely equal to large ones—nor should it. The key is uniform rules: influence is determined by actual computational contribution, not by capital, delegation mechanisms, or financial leverage.
Third, the role of computing Pools. In systems with real infrastructure requirements, resource aggregation naturally emerges. Computing pools allow smaller participants to consolidate resources, reduce volatility, and participate in larger-scale workloads.
However, the architecture must avoid giving large computing pools structural advantages or incentivizing excessive concentration of influence. Pools should exist as coordination tools, not re-centralization mechanisms.
Ultimately, scaling a decentralized AI computing network should not mean raising the barrier to entry. It should mean increasing overall computing capacity while maintaining neutral, transparent, and consistent participation rules, and preserving the real economic value the network creates for users. Open access, proportional economic mechanisms, and controlled concentration levels determine whether a system remains decentralized as it grows.
Q: Why has the issue of decentralized AI computing become particularly urgent at this moment? If this problem is not solved in the coming years, what do you think the long-term consequences for the industry will be?
A: This urgency reflects AI's transition from an experimental phase to an infrastructure phase.
As mentioned, computing power has become a physical bottleneck. Scalability is increasingly constrained not just by capital, but by energy, power density, and data center limitations. Simultaneously, access to advanced GPUs and hyperscale infrastructure is influenced by long-term contracts, corporate consolidation, and national strategic priorities.
This combination deepens structural asymmetries. Those controlling large-scale infrastructure continue to consolidate their advantages, while entry barriers for small teams and emerging regions keep rising. The risk is not just market concentration, but the widening of a global computing power divide.
If this trend continues, innovation will depend more on infrastructure access than on ideas themselves. The AI market could solidify into a rent-based model where intelligence is accessed under conditions set by a few dominant providers.
Therefore, decentralized computing power is not an ideological debate. It is a response to visible structural constraints—and a choice that will shape the long-term architecture of the AI industry.
Q: AI agents are increasingly autonomously booking GPU resources. How does Gonka's architecture support seamless integration for a self-regulating AI computing economy?
A: The rise of agentized AI means systems are increasingly making autonomous decisions—including acquiring computational resources. In this model, computing power becomes a core asset in economic interactions among agents.
Such an ecosystem requires programmatic access, transparent economic mechanisms, and reliability.
First, integration must be seamless. Gonka provides an OpenAI-compatible API, enabling most AI agents to connect without changing their architecture or workflow.
Second, the computing economy must be transparent and system-driven. Pricing adjusts dynamically based on network load, not fixed by contracts. In the network's early stages, inference costs are designed to be significantly lower than centralized providers because participants are compensated not only through user fees but also through rewards from a Bitcoin-like issuance mechanism proportional to available computing capacity.
This structure allows AI agents operating within budgets to execute workloads efficiently. As the network evolves, pricing parameters will remain subject to community governance.
Third, reliability is reinforced at the protocol level. In centralized environments, reliability comes from certification and service level agreements. In decentralized infrastructure, reliability is supported by open-source code, third-party audits, and on-chain verifiable proofs of computational completion and network performance.
Together, these elements enable AI agents to request computing power and allocate budgets within a transparent framework. In this way, Gonka provides the infrastructural foundation for a self-regulating AI computing economy, allowing agents not only to execute tasks but also to optimize the resources they depend on dynamically.
Q: Regulatory uncertainty around decentralized technology is increasing. How is Gonka proactively addressing data sovereignty and AI governance compliance in a global fragmented market?
A: In the context of decentralized computing, the main challenge is balancing network openness with diverse and evolving jurisdictional requirements.
Gonka is a permissionless global network—anyone can join, and requests are routed programmatically among distributed participants. At this stage, users cannot deterministically control the geographic location where their requests are processed. For use cases with strict data residency or regional processing requirements, this may currently be a limitation.
However, from a privacy perspective, this architecture reduces data centralization. Each request is processed by a randomly selected participant and routed independently, preventing the accumulation of complete user histories. So far, this model has covered most practical use cases while allowing the network to scale.
As the network grows and market demands become clearer, the mechanism allows participants to propose and vote on architectural changes to support specific regulatory requirements. These changes might include: dedicated subnets with additional participation criteria, operational constraints for specific jurisdictions, or hardware-level guarantees for enterprise workloads, such as Trusted Execution Environments (TEEs).
Decentralization does not eliminate compliance obligations. It provides architectural flexibility. Gonka is designed to allow the network to evolve according to regulatory and market demands, rather than being locked into a single compliance model from the outset.
