Agentic AI era demands state-backed industrial strategy

Park Sung-hyun 

The author is the CEO of Rebellions. 

Last week, Wall Street was electrified by the public debut of a single company. AI inference chip startup Cerebras repeatedly raised its initial public offering price range as investor demand surged far beyond available shares. Its final valuation reached $56.4 billion, and the stock jumped 68 percent on its first trading day. Just eight months earlier, the company had been valued at $8.1 billion. Some may dismiss the frenzy as another chapter in the AI bubble. Still, from the front lines of the AI semiconductor industry, the market’s reaction signals something more significant: investors are beginning to recognize the next phase of AI.

That next phase is agentic AI. Until now, most AI services have functioned as systems that answer questions posed by humans. Agentic AI operates differently. Instead of simply responding, AI agents can independently create plans, call external tools, make decisions and complete complex tasks with minimal human intervention. These systems are already being deployed in industrial settings. Coding agents, for example, have become indispensable tools for developers, generating software continuously and autonomously. When Nvidia CEO Jensen Huang declared earlier this year that “agentic AI has reached an inflection point,” it was not merely a marketing slogan but a warning that a new era of computing had arrived.

This explains why Cerebras has attracted such intense attention. A chatbot answering a simple question may consume hundreds of tokens, the basic units AI systems use to process language. But an autonomous agent performing an extended task can consume tens of thousands of tokens without pause. Because these systems operate continuously without human interruption, processing speed becomes critical. If inference slows, the entire workflow bottlenecks.

Cerebras drew attention partly because of its partnership with OpenAI to build large-scale inference infrastructure designed to overcome these constraints. OpenAI itself acknowledged when announcing the collaboration that “optimized systems are needed for each workload.” The era when one general-purpose chip could handle every AI task is fading. In its place is a new competitive landscape built around specialized architectures optimized for specific workloads. Capital markets are beginning to recognize that transition.

This shift is also creating a new economic unit: the token. Tokens are the fundamental units AI models use to interpret and generate language. In multi-agent systems, where numerous AI agents collaborate simultaneously, token consumption grows exponentially. In the AI economy, tokens increasingly function like currency. The competitiveness of an AI service depends on how many tokens it can process, how quickly it can process them and at what cost.

Inference is no longer an abstract technical concept. It has become a bill. The size of that bill creates markets, and markets attract capital. Tokenomics, once associated primarily with cryptocurrency, is emerging as a core concept of AI infrastructure economics. Instead of subscription fees or advertising revenue, token consumption itself is becoming the basis of AI profitability.

Many assume exploding token demand simply means building more data centers. That is only partly correct. The greater constraint increasingly lies in electricity. Roughly half the operating cost of AI infrastructure comes not from GPUs themselves but from power, cooling and facility operations. In the United States, securing electricity for AI data centers has already become more difficult than finding suitable land. Korea faces similar pressures as the growing concentration of data centers around the Seoul metropolitan area strains the national power grid. OpenAI recently underscored the issue in a letter to the White House, describing electricity as “the new oil.”

Ultimately, the defining metric of the AI era may become “tokens per watt.” The chips that process more tokens using less electricity will determine the competitiveness of future AI services. Efficiency is no longer secondary to performance; it has become central to the industry’s future and to broader questions about sustainability and national energy infrastructure. The competition is shifting from building faster chips to building chips that deliver more AI output while consuming less power.

Cerebras’ success, therefore, represents more than the rise of a single semiconductor startup. The age of agentic AI is no longer a future possibility, but a present reality, and the race to design its infrastructure is already underway. Korea possesses startups spanning the full AI semiconductor spectrum, from server inference chips to on-device AI processors. Few countries have startups capable of mass production and commercial deployment at this level.

Long before Wall Street embraced Cerebras, the U.S. Department of Energy’s Argonne National Laboratory adopted its chips. At a time when private industry hesitated, the state became the company’s first customer. That early trust helped build today’s Cerebras. Korea’s public investment initiatives in AI semiconductor firms reflect a similar logic. Governments must become more than regulators or financial supporters. They must act as patient investors and early customers willing to validate emerging technologies through real-world deployment. Countries capable of playing that role will shape the standards of the AI era rather than follow them.

This article was originally written in Korean and translated by a bilingual reporter with the help of generative AI tools. It was then edited by a native English-speaking editor. All AI-assisted translations are reviewed and refined by our newsroom.