Korea IT Times interviews Cheng Wang, Akash’s COO and CFO, on NVIDIA-led AI infrastructure, sovereign AI investment, export controls, inference demand, and the evolving case for decentralized compute.
Korea IT Times interviews Cheng Wang, Akash’s COO and CFO, on NVIDIA-led AI infrastructure, sovereign AI investment, export controls, inference demand, and the evolving role of decentralized compute
As the global AI race intensifies, control over compute infrastructure is emerging as a strategic issue for both governments and industry. In an interview with Korea IT Times, Cheng Wang, COO and CFO of Akash, said the growing dominance of NVIDIA-led stacks has accelerated adoption but also heightened concentration risk, while sovereign AI investment, export controls, and energy constraints are reshaping how AI infrastructure is built, financed, and utilized.
NVIDIA-led standardization and the growing risk of concentration
Wang said NVIDIA has built a unified hardware and software layer that made AI accessible at scale, lowering barriers to entry and accelerating adoption across the ecosystem. That level of standardization, he noted, played a critical role in compressing what would otherwise have been a much longer development cycle.
But he argued the industry is increasingly conflating standardization with concentration. In his view, the market is moving toward a structure in which a small number of hyperscalers, operating largely on a single vendor’s stack, exert disproportionate control over the compute layer behind AI.
He added that the risks are compounded by structural constraints, including the concentration of advanced semiconductor manufacturing and persistent GPU supply shortages. In such an environment, pricing power, access, and even the direction of development can become centralized, while geopolitical risks further expose the fragility of supply chains.
Wang said Akash’s approach is to keep the infrastructure layer open and distributed by aggregating underutilized GPU capacity, including NVIDIA hardware, from a global network of independent providers. That model, he said, is designed to prevent any single entity from dictating access to AI compute.
Sovereign AI investment and the strategic case for decentralized compute
Wang said the global race for AI leadership is reinforcing the idea that compute infrastructure is becoming a geopolitical asset. Export controls and sovereign GPU investment programs are accelerating the emergence of a multipolar AI landscape defined by regional compute zones rather than a single global market.
He cautioned, however, that hardware availability is only part of the constraint. Even where chips are secured, energy infrastructure often becomes the limiting factor, as existing power systems may not be able to absorb large-scale compute demand reliably.
That dynamic shifts the core challenge from acquisition to utilization. Wang said that once governments commit significant public capital to domestic GPU infrastructure, the question becomes how to maximize the productivity of resources that are geographically distributed and operationally heterogeneous.
In that context, he argued that decentralized compute networks can complement sovereign infrastructure. They enable access to a broader pool of distributed compute while remaining compatible with jurisdictional requirements and existing energy constraints.
Inference demand and near-term openings in power-constrained Asian markets
Wang identified inference as the most immediate and underappreciated opportunity for decentralized compute. He emphasized that inference is not a single market, but spans multiple environments and workload types, each with distinct performance and cost metrics.
For user-facing and agent-driven applications, he said time to first token and time to last token are critical. For operators running inference at scale, cost per token and performance per watt become the defining economic factors.
That fragmentation creates an opening for decentralized models. In power-constrained markets, distributing inference workloads across smaller nodes can meet demand without the concentrated energy requirements of hyperscale data centers.
Wang pointed to Akash’s Homenode initiative, which allows users with high-performance consumer GPUs, such as RTX 4090 and 5090 hardware, to contribute idle capacity from homes or offices. He said this approach is particularly relevant in markets like South Korea, where underutilized GPU resources can be mobilized close to demand.
He added that without an effective coordination layer, much of the GPU capacity being accumulated across Asia risks remaining fragmented and underutilized. Decentralized compute, in his view, provides both cost efficiency and the orchestration needed to make distributed infrastructure productive.
How export controls are reshaping AI infrastructure and capital formation in Asia
Wang said export controls are changing how AI infrastructure is financed and deployed. When access to advanced chips becomes uncertain, the investment case for large centralized data centers becomes significantly more fragile.
He noted that it is difficult to justify large-scale capital commitments when hardware procurement and upgrade paths cannot be guaranteed. This uncertainty is driving a shift toward more modular and distributed infrastructure models.
Under such models, smaller deployments are aggregated through marketplace frameworks, with investment spread across multiple operators rather than concentrated in single facilities. At the same time, hardware constraints are pushing software teams to improve efficiency, accelerating demand for hardware-agnostic infrastructure.
Wang said the fragmentation created by export controls is precisely the environment in which decentralized compute marketplaces gain relevance.
Where centralized cloud retains its edge and decentralized models are gaining ground
Wang acknowledged that centralized cloud providers continue to dominate large-scale foundation model training, enterprise managed services, and workloads integrated with proprietary toolchains. For these use cases, he said centralized infrastructure remains the most effective solution.
However, he argued that inference presents a different competitive landscape. Because inference workloads vary widely in latency, throughput, memory, and power requirements, the segment is less conducive to winner-take-most dynamics.
In that environment, decentralized providers do not need to outperform hyperscalers across every dimension. Wang said that achieving comparable performance, while offering flexibility, resiliency, and lower costs, can be sufficient for a growing number of use cases.
He added that GPU compute on Akash can run two to three times cheaper than centralized alternatives, as providers monetize idle capacity rather than relying on hyperscaler margins. As inference demand expands and energy costs rise, he said this cost advantage is likely to become more significant.
Burn Mint Equilibrium and the push to align stable pricing with token utility
Wang said Burn Mint Equilibrium, or BME, was designed to address a common tension in decentralized infrastructure networks. Stable pricing is necessary for enterprise adoption, but can reduce the role of the native token, while token-based pricing introduces volatility that complicates long-term workloads.
Under the model, tenants burn AKT to mint ACT, a non-transferable, dollar-pegged compute credit, allowing for stable pricing without direct exposure to token fluctuations. At settlement, consumed ACT is burned and AKT is minted back to providers at the prevailing oracle price.
If AKT appreciates during the lease period, fewer tokens are minted back than were originally burned, effectively removing supply from circulation. Wang said this mechanism ties network usage directly to token dynamics.
He said key performance indicators include the net AKT burn rate relative to minting and the level of active compute usage. While revenue can provide directional insight, deployment and lease activity offer a clearer measure of real utilization.
Akash’s 2026 roadmap and the next phase of platform expansion
Wang identified support for virtual machines and bare metal as a central priority for Akash’s next phase. While the platform has historically focused on container-native workloads, he said this has limited its ability to serve industries with legacy systems that are not easily containerized.
Expanding to full virtual machine support would allow those workloads to migrate without extensive restructuring, significantly broadening the platform’s addressable market. He described this as a key step in advancing enterprise adoption.
He also pointed to the scaling of Homenode as a major priority. Following the open beta launch, Akash plans to expand GPU support and increase participation, particularly in markets with dense hardware availability.
A third priority is migrating to a shared security model, which Wang said would allow the team to focus more directly on the compute marketplace, improve capital efficiency, and remove structural constraints on growth.
Taken together, Wang’s view is that AI infrastructure is entering a more fragmented phase shaped by geopolitics, energy constraints, and utilization dynamics, as much as by hardware supply and model demand. In that environment, decentralized compute is evolving from an alternative concept into a practical complement to centralized cloud infrastructure.
