Akash Founder and CEO Greg Osuri discusses why semiconductors alone will not determine AI leadership, how sovereign AI should be defined, why compute accessibility matters, and what South Korea must do to translate its hardware strengths into long-term AI competitiveness.
South Korea has firmly established itself as one of the world’s most important players in the artificial intelligence supply chain. The country is home to Samsung Electronics and SK hynix, which together dominate the global market for high-bandwidth memory (HBM), a critical component powering AI accelerators from companies such as NVIDIA. At the same time, the Lee Jae Myung administration has placed AI at the center of its industrial strategy, pledging billions of dollars toward AI infrastructure, supporting the National AI Computing Center, and accelerating sovereign AI initiatives to strengthen the country’s technological competitiveness.
Yet around the world, the conversation around AI is rapidly evolving. While governments continue investing heavily in semiconductor manufacturing and next-generation AI models, another question is emerging: who controls the infrastructure that allows artificial intelligence to function at scale?
Access to GPUs, reliable energy supplies, cloud infrastructure, and affordable compute are increasingly becoming strategic concerns alongside semiconductor manufacturing. For startups, researchers, and enterprises alike, the ability to access AI infrastructure may prove just as important as the ability to design advanced chips.
Greg Osuri believes this is where the next phase of global AI competition will be decided. As Founder and CEO of Akash Network, a decentralized compute marketplace, Osuri has become one of the industry’s strongest advocates for rethinking how AI infrastructure is built and distributed. Earlier this year, he testified before the U.S. Congress on AI’s growing compute and energy constraints, arguing that concentration within today’s cloud infrastructure poses long-term risks to innovation and economic competitiveness.
In an exclusive interview with KoreaTechToday, Osuri shared his views on South Korea’s AI ambitions, sovereign AI, infrastructure resilience, startup accessibility, and why democratizing compute may ultimately become one of the defining factors of AI leadership.
South Korea has the ingredients for AI leadership, but leadership requires more than chips
South Korea enters the AI era from a position of considerable strength. The country’s semiconductor industry remains indispensable to the global AI ecosystem, while its advanced manufacturing capabilities and world-class engineering talent provide a solid foundation for future innovation.
However, Osuri argues that semiconductor leadership alone should not be confused with AI leadership.
“Korea’s advanced semiconductor and manufacturing base gives it a real structural advantage, but the gap with TSMC remains considerable,” he told KoreaTechToday. “Ultimately, raw fabrication capabilities don’t translate directly to AI leadership without a considerable transition period and ramp-up. AI bottlenecks extend beyond chips to energy, datacenter infrastructure, land, and water, all of which pose challenging choke points for Korea.
“Korea can clearly build the hardware necessary for AI expansion. The key questions are: how long will it take, where will they do it, where will the power come from, and what is the interim stopgap? I think the countries that lead the AI race will be the ones that make compute easily accessible and reduce concentration risk that defines the current market structure. Korea has the inputs to lead, but that leadership depends on diversifying strategies and ensuring access to the infrastructure it builds.”
For much of the past decade, semiconductor manufacturing was widely viewed as the primary determinant of AI competitiveness. Today, governments are increasingly recognizing that advanced chips represent only one layer of a much larger ecosystem that includes cloud platforms, GPU clusters, networking infrastructure, data centers, software frameworks, and energy systems capable of supporting increasingly complex AI workloads.
For South Korea, this distinction is becoming increasingly important as policymakers seek to convert the country’s manufacturing strengths into a sustainable AI ecosystem that supports startups, universities, enterprises, and public-sector innovation.
Sovereign AI is about control, not just ownership
As countries race to establish national AI capabilities, the concept of “sovereign AI” has become a central theme in technology policy. Governments across Asia, Europe, and the Middle East are investing heavily in domestic AI models, national cloud infrastructure, and sovereign computing resources.
Osuri believes the concept extends far beyond building local data centers.
“Sovereign AI looks beyond domestic data centers or national AI models,” he explained during the interview. “It is a country’s ability to develop, deploy, and govern AI without depending on infrastructure it does not control. That means thinking about the full stack, from data and models to compute and deployment.
“Dependency creates leverage. If a country’s researchers, startups, and enterprises rely on services run by a small number of foreign hyperscalers, its AI capability exists at someone else’s discretion. In this case, access runs the risk of being repriced, restricted, or prioritized according to external commercial interests. Domestic infrastructure alone cannot solve the problem, but it is the first step in creating an open and diverse compute layer that augments concentrated networks, where capacity comes from many providers rather than a single centralized operator.”
That perspective has particular relevance for South Korea. While the country is investing aggressively in sovereign AI infrastructure, much of today’s global cloud ecosystem remains concentrated among a handful of multinational providers. At the same time, demand for advanced GPUs continues to outpace supply, prompting governments worldwide to reconsider how national AI capacity should be developed and managed.
Rather than viewing sovereign AI purely as a question of domestic ownership, Osuri argues that resilience, openness, and infrastructure diversity should become equally important components of national AI strategies.
Democratizing compute could determine who participates in the AI economy
One theme surfaced repeatedly throughout the conversation: accessibility. According to Osuri, the greatest risk facing today’s AI ecosystem is not simply a shortage of GPUs but the concentration of compute within a relatively small number of infrastructure providers. This challenge is particularly relevant for startups.
While large technology companies can often secure long-term GPU contracts and absorb rising infrastructure costs, smaller companies frequently face limited access to the computing resources needed to build and scale AI products.
“Korea has a strong startup and developer ecosystem,” Osuri said. “Alternative compute models could be significant for Korean startups, developers, and researchers because compute access is increasingly becoming a barrier to AI development. Access to high-end GPUs is often gated behind enterprise contracts, long procurement cycles, and pricing structures that favor large companies. That creates an uneven playing field, where smaller teams may have the talent and ideas to build AI products, but may not have access to the necessary infrastructure.
“For Korea, alternative compute models could help turn existing technical strength into broader participation. While Korea has world-class developers, a strong hardware culture, and a thriving startup ecosystem, access to compute remains expensive, scarce, and concentrated. A more open model would give smaller teams and developers a better path to build and scale, while also helping Korea turn its existing technical strength into broader participation in the AI economy.”
Osuri believes democratizing access to compute will require more than market competition alone.
“It won’t happen by default,” he said. “The current trajectory points straight at concentration, so avoiding it requires deliberate choices. The Big Three hyperscalers, AWS, Microsoft Azure, and Google Cloud, now control roughly two-thirds of global cloud infrastructure spend, and mega-deals like Nvidia’s $100 billion commitment to OpenAI keep deepening that pattern. Each deal reinforces a structure where a handful of companies secure priority access to compute, leaving everyone else to compete for the remaining capacity.
“Open, distributed marketplaces can be an alternate path forward. We need models that bring more underused capacity online, make pricing and availability more transparent, and give developers practical ways to access compute. However, market innovation alone is not enough. Governments also have a responsibility to ensure public investment in AI infrastructure expands access for startups, researchers, and local developers, rather than reinforcing a market where compute is only available to the largest players.”
For South Korea, where AI has become both an economic priority and a national strategic objective, the challenge extends beyond expanding GPU capacity. It also involves determining who can access those resources and whether future AI innovation will be driven by a broad ecosystem of startups, researchers, and enterprises or remain concentrated among a relatively small number of large technology companies.
Infrastructure resilience is becoming a strategic advantage
Beyond accessibility, Osuri believes another challenge deserves far greater attention: the concentration of AI infrastructure within a small number of cloud providers.
Recent outages affecting major cloud platforms, including AWS, Azure, and Cloudflare, have exposed how dependent modern digital services have become on centralized infrastructure. While these disruptions were temporary, they highlighted the potential risks of concentrating critical AI workloads within only a handful of providers.
For countries investing heavily in AI infrastructure, Osuri argues that resilience should be considered alongside scale.
“It all comes back to concentration,” he told KoreaTechToday. “A large share of critical digital infrastructure now depends on a small number of centralized providers. When one of those providers goes down, the impact is widespread because too many services are built on the same chokepoints. The aftermath of recent AWS, Cloudflare, and Azure outages are prime examples of the risks we face as a result of over-reliance on centralized infrastructure.
“The more resilient path is to distribute compute resources across a network of independent providers, so no single operator becomes the master switch for an entire ecosystem. Countries and enterprises should be designing infrastructure that reduces single points of failure rather than increasing dependence on them.”
He believes the risks extend beyond operational outages.
“Absolutely. The real risk is that AI is being built on top of a very narrow infrastructure layer,” Osuri said. “Every new model, application, agent, and enterprise AI product may look like innovation at the surface, but underneath, much of it still depends on the same small group of cloud providers and GPU supply chains.
“That creates two problems. The first is resilience. If critical AI systems are routed through the same few providers, then any outages, pricing changes, capacity issues, or policy decisions made by them can directly impact a significant portion of the market at once. The second is market structure. When a handful of companies control the compute layer, they also influence who can build, who can scale, and who gets priced out before they ever reach the market.
“A more distributed compute layer gives developers and enterprises a foundation where capacity can come from many providers, closer to where the energy and demand already exists.”
For South Korea, these observations align with ongoing efforts to diversify AI infrastructure while strengthening domestic computing capacity. As the country expands public AI infrastructure and invests in sovereign computing initiatives, policymakers are increasingly balancing scale with resilience, security, and long-term accessibility.
Energy may become AI’s biggest constraint
Although GPUs often dominate discussions around artificial intelligence, Osuri believes the industry’s next challenge lies elsewhere. According to him, compute cannot be separated from the physical infrastructure that powers it.
“Energy is already becoming one of the defining constraints in AI, but I would not separate it from compute,” he explained. “Every GPU, LLM model, and AI agent ultimately depends on power, cooling, and grid capacity. So when people talk about a compute shortage, they are often also talking about an energy and infrastructure shortage underneath it.
“Take the US as an example, data centers already consume around 4.4% of US electricity, more than double the share from 2018, and the Department of Energy projects that figure could reach 12% by 2028. We can buy more chips, but if we cannot power them economically, cool them efficiently, or connect them to reliable infrastructure, they do not translate into usable AI capacity.
“AI demand is scaling faster than the physical infrastructure needed to support it. That is why the next phase of AI infrastructure has to be designed around energy, not just compute. The industry cannot keep assuming that the answer is to build larger data centers in the same energy-constrained locations and expect the grid to absorb it. A more sustainable path is to move compute closer to where energy already exists, including underused and renewable power sources, and coordinate that capacity through distributed networks.”
The discussion has growing relevance for South Korea, where demand for AI infrastructure is expected to increase rapidly alongside investments in sovereign AI, advanced manufacturing, and next-generation data centers. Ensuring sufficient electricity generation, cooling capacity, and resilient grid infrastructure is becoming an increasingly important part of national AI planning, alongside semiconductor production itself.
Akash’s alternative vision for AI infrastructure
While much of the interview focused on broader industry trends, Osuri also outlined how Akash seeks to address some of these structural challenges through a decentralized compute marketplace.
Rather than relying exclusively on hyperscale data centers, Akash aggregates underutilized GPU capacity from independent providers, allowing developers to access computing resources through an open marketplace.
“We’re here to break up AI’s infrastructure concentration, offering an alternative to the centralized model hyperscalers have normalized,” Osuri said. “Traditional cloud providers enforce arbitrary corporate premiums and rigid multi-year contracts, pricing independent developers out of the AI race. Hyperscalers also depend on centralized infrastructure that is vulnerable to supply chain shocks. By unlocking idle hardware and making that resource available in an open marketplace, Akash re-engineers the cloud economy.
“From enterprise data centres down to ordinary laptops and desktops, Akash pulls in underused hardware and routes it straight into production. In places like Seoul, where land is at a premium and public backlash against data centre development is growing, this model is particularly fitting.
“Beyond geography, distributed marketplaces like Akash shift the economics of AI for independent developers. Hyperscalers control prices and access to compute, with access provided to the highest bidder. Akash standardizes costs and makes pricing transparent, with far less exposure to single-point-of-failure shocks. The result is cheaper, more predictable compute that is much harder to knock offline than any cloud provider can guarantee.”
He pointed to Akash’s recent collaboration with gaming company Razer as an example of how distributed infrastructure can operate at commercial scale.
“Beyond significantly undercutting traditional cloud pricing, Akash’s joint pilot with Razer showed that distributed compute is a viable alternative to centralized models,” he said. “The dominant industry narrative implies that enterprise-grade AI applications can only run on specialized, multi-million dollar infrastructure. The pilot closes that gap by running a 4-billion parameter multimodal model on consumer gaming GPUs, RTX 4090s and 5090s, at commercial concurrency.
“Resilience was the other proof point. When demand spiked during the campaign, the network simply scaled, spinning up more capacity across independent providers worldwide rather than slamming into a fixed ceiling with a centralized provider. That’s the real lesson. As AI moves into always-on agentic workloads that demand far more compute, the networks that win will be the ones that can draw on capacity wherever it already exists, not the ones with the biggest data center. The compute power needed to keep AI scalable and accessible is already out there. We just need to be smart about how we tap into it.”
Looking beyond today’s AI race
Looking ahead, Osuri expects the AI industry to move toward a more distributed model, although he acknowledges the transition will not happen overnight.
“I am optimistic about where AI infrastructure can go over the next five years but at the same time, I am also realistic about the scale of the challenge in front of us,” he said. “The current developmental model is heavily concentrated around massive data centers, hyperscaler contracts, limited access to GPU capacity and limited power infrastructure. My hope is that five years from now, AI infrastructure will be far more distributed, energy-aware and accessible than it is today.
“The bigger opportunity will be to bring AI infrastructure closer to where energy and users already are. That could mean more distributed data centers, infrastructure at the community level, and eventually homes that contribute compute from their own renewable power. Picture a homeowner with rooftop solar helping to train the next model and offsetting their mortgage in the process. That’s the distributed future I’m optimistic about, one where participation is accessible to everyone.”
South Korea’s AI strategy has largely focused on strengthening semiconductor manufacturing, expanding sovereign AI capabilities, and investing in next-generation computing infrastructure. Those priorities remain essential as countries compete for leadership in an increasingly AI-driven economy.
Greg Osuri’s perspective adds another dimension to that conversation. Rather than measuring AI leadership solely by the number of chips produced or the size of data centers built, he argues that long-term competitiveness will depend on how widely computing resources can be accessed across an innovation ecosystem.
Whether governments ultimately embrace decentralized infrastructure or continue investing primarily in large-scale cloud platforms, the broader questions he raises are becoming increasingly relevant. As AI adoption accelerates across research, manufacturing, healthcare, finance, and public services, accessibility, resilience, and openness are likely to become as important as computing power itself.
For South Korea, which already possesses many of the technical foundations needed to compete globally, the next phase of AI leadership may ultimately depend not only on building more infrastructure, but on ensuring that startups, researchers, enterprises, and future innovators can all participate in the AI economy it is creating.
