Training large models requires building or upgrading data centers. But centralized infrastructure is now running into hard physical limits. To improve this infrastructure, AI is being implemented to produce abundance and intelligence. However, the control over computing is becoming a critical point of power in the AI industry.
Here Gonka comes to the rescue. Gonka protocol is a permissionless and global network where anyone can join, and requests are routed programmatically across distributed participants.
In this exclusive conversation with Analytics Insight, Anastasia Matveeva, Co-Founder and Senior Product Manager at Gonka, discusses how they are bringing innovation to compute access for a more controlled and safe AI ecosystem.
Much of the public debate around AI focuses on the centralization of models, but far less attention is paid to the centralization of compute. Why is control over compute becoming a critical point of power in the AI industry, and what risks does this concentration pose for innovation and the market as a whole?
Much of the public debate around AI focuses on models because they are visible. But the real locus of power sits one layer below: compute – the substrate that determines who can build, deploy, and scale AI systems.
Control over compute is becoming critical for economic and physical reasons. The main bottleneck in modern AI is no longer algorithms – it is access to GPUs, energy, and data-center capacity.
Training large models increasingly requires building or upgrading data centers. But centralized infrastructure is now running into hard physical limits: energy density, cooling constraints, and the maximum power that can be delivered to a single location. The industry is experimenting with extreme solutions – redesigning chips, cooling, and new energy sources.
This concentration has systemic consequences. First, it erects structural barriers to innovation. Access to compute becomes an infrastructure privilege rather than a merit. Smaller teams, independent researchers, and entire regions are priced out, narrowing experimentation and pushing incremental innovation.
Second, compute centralization entrenches a rent-extraction model. AI has the potential to produce abundance – intelligence is inherently replicable – but when the underlying infrastructure is scarce and controlled, that abundance is artificially throttled. The market shifts toward subscriptions, lock-in, and pricing power rather than falling costs and broad access.
Third, it introduces systemic fragility. When advanced compute is concentrated in a small number of operators and geographic locations, regulatory, political, or physical disruptions propagate across the entire AI ecosystem. Dependence becomes structural rather than optional.
Most importantly, compute is not neutral. Whoever controls compute implicitly determines what is feasible, permissible, and economically viable in AI. When that control is centralized, governance over AI emerges by default rather than by design.
The risk is not just a monopoly. It is a long-term distortion of AI’s development trajectory: fewer builders, less diversity of applications, slower hardware innovation, and infrastructure that cannot scale alongside the ambitions of next-generation models. That is why compute must be treated as foundational infrastructure – and why architectures that scale economically and physically matter for the future of AI.
Many AI compute platforms – both centralized and decentralized – claim to be highly efficient. In your view, what metrics actually matter when evaluating efficiency in AI compute systems, and where do these models most often run into practical limitations?
Efficiency in AI compute is often treated as marketing. In practice, only a few concrete metrics truly matter, spanning user-facing performance, provider-side operational efficiency, and the systemic incentives that govern both.
For users, efficiency means speed and transparency of costs.
Speed is latency under real demand. Centralized hubs traditionally hold an advantage due to physical co-location. However, decentralized architectures can achieve comparable performance when the blockchain functions only as a security layer rather than part of the real-time execution path. If user requests remain off-chain, the protocol adds no latency.
Cost transparency is equally critical. While cost per token is a standard KPI, what often lacks clarity is model integrity. In centralized environments, the product can function as a black box. During peak demand, providers may adjust model configurations to manage margins. These changes are rarely visible but can affect output quality. True efficiency requires pricing to reflect consistent computational precision.
For providers, efficiency balances GPU utilization and elasticity.
Centralized operators excel at utilization. In controlled, co-located environments, GPUs can operate near full saturation with minimal coordination overhead. Their weakness is a lack of elasticity. Because these hubs are static and capital-intensive, they bear the cost of idle capacity during demand troughs.
Decentralized networks trade some utilization for elasticity, but must minimize consensus and verification overhead. This allows compute to be redistributed across workloads in response to demand.
Most importantly, incentive design becomes decisive. When earnings are tied to delivering faster, cheaper, and verifiably correct AI workloads, optimization becomes structural rather than optional. Participants are motivated to improve hardware efficiency, reduce latency, and experiment with specialized chips. Performance follows economic self-interest.
By contrast, when rewards or governance influence are tied primarily to capital ownership or passive positioning, optimization shifts away from infrastructure performance. In such systems, inefficiency is not accidental – it becomes embedded in the incentive structure itself.
This is where architectural differences become decisive. In Gonka, efficiency is embedded at the protocol level: nearly 100% of computational resources are allocated to real AI workloads, primarily inference. Earnings and governance weight are tied to measured computational contribution rather than capital ownership.
In practice, most systems hit limits in centralized capacity or decentralized overhead. True efficiency emerges only when the majority of compute is directed toward real AI tasks, incentives reward verified contribution, and internal overhead does not scale faster than the network itself.
Do you think it is fundamentally possible for decentralized AI compute networks to approach a model where most computational resources are directed toward real AI workloads rather than maintaining the network itself? What architectural choices are critical to making this viable?
Yes, it is fundamentally possible – but only if overhead is treated as a core architectural constraint rather than an inevitable byproduct of decentralization.
Most decentralized compute networks allocate a significant share of resources not to AI workloads, but to maintaining consensus and network security. This happens because productive work and security are separated, leading to duplication. As a result, a portion of GPU capacity is diverted from real AI demand toward sustaining the system itself.
To approach a model where the majority of computational resources are dedicated to real AI tasks, several architectural principles are essential.
First, security and measurement mechanisms must be time-bound rather than continuous. Instead of consuming computational resources at all times to maintain consensus, proof mechanisms can be concentrated into short, clearly defined intervals. In Gonka, this principle is implemented through Sprint – structured, time-limited cycles that measure computational contribution. Outside these intervals, hardware resources remain available for real AI workloads, primarily inference.
Second, duplication must be minimized through selective and reputation-adjusted verification rather than full task replication. If every task is validated through repeated execution, overhead grows proportionally with activity. A more efficient model combines an honest-majority assumption with randomized verification of a subset of results, reinforced by a dynamic reputation system.
When a new participant joins the network, nearly 100% of their work may be validated. As they demonstrate consistent and honest behavior over time, the share of tasks subject to verification can decrease – in some cases to approximately 1%. This adaptive verification rate allows the overall portion of network compute dedicated to validation to remain below roughly 10%, while preserving strong security guarantees.
Participants who attempt intentional fraud – such as submitting manipulated outputs or misrepresenting completed computational work – simply do not receive rewards for those tasks. As a result, deliberate cheating becomes economically irrational without requiring constant full duplication of computation.
Third, rewards and governance weight must be tied to verified computational contribution rather than capital ownership. If influence is determined by staking or financial positioning, participants optimize for capital efficiency instead of infrastructure performance. When reward distribution and governance weight are directly linked to measurable compute, the network naturally evolves toward higher productive utilization.
Decentralized AI compute can serve real workloads if consensus is lightweight, verification adaptive, and incentives aligned with productive computation.
Decentralized AI compute networks often promote open participation, yet infrastructure requirements can create high barriers to entry. How can such systems scale while remaining accessible to participants with very different levels of compute capacity?
While decentralized networks aim to lower entry barriers for AI infrastructure, long-term survival also requires competing with centralized providers and meeting real-world demand. Hardware constraints ultimately come down to one requirement: the ability to host models that are actually in demand.
To scale while remaining accessible, several principles matter. First, permissionless access to infrastructure. Any GPU owner – from a single-device operator to a full-scale data center – should be able to join without approval processes or centralized gatekeeping. This removes structural barriers to entry.
Second, proportional rewards and influence based on verified compute. In a compute-weighted model, greater computational contribution naturally results in a larger share of tasks, rewards, and governance weight. This does not equalize small and large participants – nor should it. What matters is uniform rules: influence is determined by actual computational contribution rather than capital, delegation mechanisms, or financial leverage.
Third, the role of pools. In systems with meaningful infrastructure requirements, resource aggregation emerges naturally. Pools allow smaller participants to combine resources, reduce volatility, and participate in larger workloads.
However, the architecture must avoid granting structural advantages to large pools or incentivizing excessive concentration of influence. Pools should function as coordination tools, not as mechanisms of re-centralization.
Ultimately, scaling a decentralized AI compute network should not mean raising entry barriers. It should mean increasing aggregate computational capacity while preserving neutral, transparent, and consistent participation rules – and while maintaining the real economic value the network delivers to users. Open access, proportional economics, and controlled concentration determine whether a system remains decentralized as it grows.
Why has the question of decentralizing AI compute become especially urgent at this moment? And what do you see as the long-term consequences for the industry if this issue is not addressed in the coming years?
The urgency reflects AI’s shift from experimentation to infrastructure.
As discussed earlier, compute has become a physical bottleneck. Scaling is increasingly constrained not only by capital but also by energy, power density, and data-center limitations. At the same time, access to advanced GPUs and hyperscale infrastructure is shaped by long-term contracts, corporate concentration, and national priorities.
This combination deepens structural asymmetry. Those who control large-scale infrastructure consolidate their advantage, while entry barriers rise for smaller teams and emerging regions. The risk is not just market concentration, but a widening global compute divide.
If this continues, innovation will depend more on infrastructure access than ideas. The AI market could harden into a rent-based model, where intelligence is accessed under the terms of a limited number of dominant providers.
For that reason, decentralizing compute 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.
AI agents are increasingly booking GPU resources autonomously. How does Gonka's architecture support seamless integration for self-regulating AI compute economies?
The rise of agentic AI means systems increasingly make autonomous decisions – including procuring computational resources. In this model, compute becomes a core asset in economic interactions between agents.
Such an ecosystem requires programmatic access, transparent economics, and reliability.
First, integration must be seamless. Gonka offers an OpenAI-compatible API, allowing most AI agents to connect without changing their architecture or workflows.
Second, compute economics must be transparent and system-driven. Pricing adjusts dynamically to network load rather than being contractually fixed. In the network’s early stages, inference costs are designed to remain significantly lower than centralized providers, as participants are compensated through both user fees and Bitcoin-style issuance rewards proportional to available computational capacity.
This structure enables agents operating within defined budgets to execute workloads efficiently. As the network evolves, pricing parameters remain subject to community governance.
Third, reliability is reinforced at the protocol level. In centralized environments, reliability comes through certifications and service-level agreements. In a decentralized infrastructure, it is supported through open-source code, third-party audits, and measurable on-chain proofs of completed computational work and network performance.
Together, these elements allow AI agents to request compute and allocate budgets within a transparent framework. In this way, Gonka provides the infrastructural foundation for self-regulating AI compute economies, where agents can not only execute tasks, but dynamically optimize the resources they depend on.
Regulatory uncertainty around decentralized tech is intensifying. How is Gonka proactively addressing data sovereignty and AI governance compliance across fragmented global markets?
In the case of decentralized compute, the main challenge is balancing network openness with diverse and evolving jurisdictional requirements.
Gonka is a permissionless and global network – anyone can join, and requests are routed programmatically across distributed participants. At the current stage, users do not have deterministic control over the geographic location where their request is processed. For use cases with strict data residency or regional processing requirements, this can currently be a limitation.
From a privacy standpoint, however, the architecture reduces data concentration. Each request is handled by randomly selected participants and routed independently, preventing the accumulation of full user histories. So far, this model has covered most practical use cases while allowing the network to scale.
As the network grows and market demand becomes clearer, governance mechanisms enable participants to propose and vote on architectural changes to support specific regulatory requirements. These could include specialized subnets with additional participation criteria, jurisdiction-specific operational constraints, or hardware-level guarantees such as Trusted Execution Environments (TEE) for enterprise workloads.
Decentralization does not remove compliance obligations. It provides architectural flexibility. Gonka’s design allows the network to evolve in response to regulatory and market demands, rather than being locked into a single compliance model from the outset.