Tech Bipolarization and the AI Chip

Six years ago, I covered the US-China contest for tech supremacy.

Seoul National University's Institute of International Affairs ran a four-week expert forum on the "post-COVID era," and JoongAng Ilbo carried it as the sponsoring media outlet. Having majored in international relations myself, I'd listened to my old professors grumble that this subject was genuinely important and yet the press never touched it — so I resolved to run a series come what may, and I got as much of it on the record as I could. The professors' conclusion back then was singular: "No matter who becomes US president, the US-China conflict will be protracted." At the time the front line was 5G, Huawei, and semiconductor equipment.

Now that front line has moved to the AI chip. And Korea is still standing in the same spot — leaning on the US for technology, looking to China for market, taking investment from both sides. A sentence one expert used back then to describe Korea's position still holds, word for word.

All HBM, but What Lies Beyond?

These days the domestic news is all HBM. Who's ahead in high-bandwidth memory, Samsung or SK hynix; who ships to Nvidia first and most. The pride of a memory powerhouse runs through it. It's not wrong.

But the scene feels familiar to me. We've always been good at making components. The question lies beyond that. The system the chip goes into, the model running on top of it.

Where Is the Center of Gravity?

The center of gravity in the AI-chip market is shifting. If 2023-2025 was the era of training giant models, 2026 is the era of inference — actually running them. The industry expects two-thirds of this year's AI compute spending to come from inference.

The scale is beyond imagination. The inference-chip market is projected to grow from $39 billion in 2024 to $475 billion in 2030 — more than twelvefold in six years.

Why does the distinction matter? Training is something a handful of giants do once, in a big burst. Inference is something everyone does, every day, over and over. Every time you throw a question at a chatbot, every time a company generates an answer from its own data, inference spins. So inference isn't a one-off investment — it's infrastructure, like electricity or the telecom grid. And with infrastructure, whoever holds it holds power. In a sense, this is closer to geopolitics.

A Hostage Drama Called "Standards"

A passage from that forum six years ago comes back to me. One presenter called the essence of US-China competition "the problem of standards." Whoever dominates the market holds the standard, and whoever holds the standard takes compatibility hostage. "In the world of relationships you can keep both sides; in the world of technology you must pick one and conform to it."

Nvidia's real moat isn't the chip — it's CUDA. Developers worldwide write their code on top of CUDA. Once you climb on, it's hard to climb off. This is the modern version of the standards problem, a lock-in that goes by the name of compatibility.

Around that same time, in another room, a scholar said something more fundamental. In an age where scientific knowledge itself becomes a source of power, he called it "intrinsic power" (內因的 힘). Computing capability in the AI era is exactly that. If the judgments of defense, intelligence, healthcare, and finance are made on top of someone's chip and someone's model, that "someone" is power.

Sovereign AI Is Not a Slogan — It's Survival

So sovereign AI is not a fancy slogan; it's a matter of survival. The familiar tightrope of "security from the US, economy from China" yields no answer. The harder the two poles harden, the smaller the spillover gains for a middle power like Korea — and the greater the pressure to pick a side. The experts warned of exactly this six years ago.

The essence of sovereign AI isn't grand. It's us deciding "which data and which inference must never leave our borders." Sensitive judgments — defense, information security, critical infrastructure, medical and legal — at minimum must run on domestic chips and domestic infrastructure. The point isn't to build everything. It's that there are things we need to keep in our own hands, the things that must not leave.

Beyond HBM, the People Who Carve Chips

Back to HBM. Memory is still our strength, and in the inference era HBM has become more than a component — it's a "strategic asset that determines inference efficiency." Because inference, in the end, is a fight over memory bandwidth and power efficiency. That's good news.

But components alone don't make sovereignty. What we lacked was the "inferring brain" that connects chip and model — the NPU. And startups have begun to fill that gap.

FuriosaAI and Rebellions. Companies that became unicorns side by side, just five to eight years after founding. They're said to deliver five to seven times the cost efficiency of Nvidia in real-world use. Rather than ramming the wall called CUDA head-on, they found a new path — a detour through open-source ecosystems like PyTorch and vLLM. It's a clever tactic for the underdog in a standards war.

The combination of "domestic chip — domestic model." The government, too, is weaving together capabilities that had been scattered — K-Cloud, the carriers' infrastructure, the startups' chips, and the data on industrial shop floors — in search of an answer.

And our real arena isn't the general-purpose LLM. Competing with the US and China on model size is a game of scale — lost from the start. I felt this again on a recent business trip to Shenzhen. Instead we should go for "industrial inference," specialized in domains where Korea owns both the data and the field: semiconductors, shipbuilding, batteries, manufacturing. This is the asymmetric strategy that sidesteps scale. It echoes another conclusion from that forum six years ago: "AI is not a field where you must pick a single ecosystem. We have to discover individual services that can compete globally."

A Small Window of Opportunity

As a journalist I covered the prologue to the US-China tech-supremacy war; now I'm watching that war enter its main act in AI chips. The only thing that changed is the location of the front line. The structure is the same. The giant grows faster, and whoever holds the standard writes the rules.

To make a chip in between. That is no longer merely an industrial question — it's a question of sovereignty. Can we run our own inference on top of our own hands?

Fortunately, we're finding the answer. We have memory as a foothold, startups have emerged to make the chips, and we've begun to ask what to protect. It's not too late. But as always, opportunity passes in an instant.

In the end, the question is this: can we really become one of the world's top three AI powers? The answer may be decided within the short few years we've been given right now. There's still a chance.