A good strategy starts with a clear understanding of your current position. As a head of IT strategy at a large corporation, I’ve developed a habit of spreading a radar everywhere and gathering and organizing various news and information daily; today I want to discuss Chinese AI among them.
The same applies to other sectors, but news about Chinese AI never fails to surprise me with each update. Recently, it became a hot topic that Cursor is reported to be using the kimi2.5 model. This indicates that Chinese open‑source projects are performing at a level that even Silicon Valley acknowledges. However, China is not merely pursuing a strategy of boosting market dominance through open source. Its AI initiatives operate on a far larger and more formidable scale.
1. China is turning ‘containment’ into ‘self‑reliance’
U.S. export controls have been a clear shackle for China. The number of high‑performance GPUs a nation possesses has been a critical key to AI development, and those controls blocked that access. Yet, that very shackle has actually strengthened China’s AI ecosystem.
As of 2025, domestic chips account for 41% of China's AI server market. In other words, there is no longer a need to rely on companies like NVIDIA. Full independence will still require more time, but the pace is truly remarkable.
This was once a market dominated 95% by NVIDIA. Yet within just a few years, Huawei Ascend shipped over 810,000 units as a single vendor, capturing 20% of the market, while Cambricon, Kunlunxin, and T-Head have rapidly filled the remainder.
What’s more interesting is Huawei’s approach. When individual chip performance falls short of NVIDIA’s, they bundle 8,000 to 15,000 chips into a SuperPod. It’s a strategy that compensates for the “performance gap of a single chip” with “system integration capability.” The limitations of the 7 nm DUV process are mitigated by looser circuit designs and government subsidies.
From an engineer’s perspective, this is not merely a volume push; it represents a paradigm shift at the system‑architecture level.
2. The real game is played in 'inference' rather than 'training'.
A common question AI companies ask themselves is, “Should we develop our own LLM?” My recent answer is, “Inference, not training, is the core business.”
Projections indicate that by 2030, 80% of AI semiconductor market revenue will be for inference. While training requires extreme performance, inference competes on cost‑effectiveness and ecosystem. This is a segment where even a 7 nm process can remain competitive.
This is precisely the target China is aiming for. If it saturates its 1.4 billion‑person domestic market with shopping, education, healthcare, and administrative AI‑X, U.S. restrictions become irrelevant. The goal is not the “smartest AI” but the “most widely adopted AI ecosystem.”
3. So, where does South Korea stand?
By the numbers, South Korea is undeniably a powerhouse.
· Highest AI patents per capita worldwide (14.3 points, surpassing the US, China, and Japan)
· Ranked 3rd globally for notable AI models (5 models, following the US with 50 and China with 30)
· 4th worldwide in industrial robot installations
· Near‑monopoly in the HBM market
However, on the ground, the weak points are evident.
Four of the five notable models come from LG AI Research’s Exa‑One series, indicating an over‑reliance on a handful of large firms. Moreover, the net outflow of AI talent per 10,000 residents is –0.4, meaning our trained professionals are migrating to Germany and the United States. In addition, more than 80% of the talent pool is male, placing the sector among the least diverse.
4. The genuine strengths of South Korea as seen by field consultants
I believe the path Korea should take is neither the Chinese model of closed self‑sufficiency nor the American model of full‑stack dominance.
What we pursue is 'open technology sovereignty'.
We are already deeply embedded in the global value chain with the overwhelming bargaining chip of HBM, and at the same time we have a sovereign model that offers a 'cultural alternative' to non‑English‑speaking countries, like Naver HyperCLOVA X. Kakao is partnering with big tech to drive mass adoption. This dual‑track strategy aligns best with Korea’s reality.
It is also noteworthy that Reverion’s revenue jumped to 32 billion won last year, a three‑fold increase over the prior year. If the HBM leader combines with an NPU fabless company, Korea can create a position in on‑device AI and smart‑factory domains that is difficult for anyone to replicate.
5. So, what should we do?
If there’s one thing I’ve learned over nearly three decades in the IT field, it’s that strategy is ultimately decided by execution speed. The government’s announcements—expanding GPU capacity by 37,000 units, a 100‑trillion‑won National Growth Fund, 10‑trillion‑won AI R&D, and a shift to negative regulation—are all on the right track. The issue is timing.
The next two to three years, during which China completes its self‑sufficient ecosystem, constitute a golden window. From a field‑level perspective, the actions we need to take during this period are summarized as follows.
First, the timeliness of computing resource supply — the power grid and data‑center sites are the real bottlenecks.
Second, transition HBM leadership to NPU — the bridge from a memory powerhouse to a system semiconductor powerhouse must be built now.
Third, create an environment that attracts talent rather than merely blocking brain drain — a reputation that “working in Korea is the most challenging” must be established. Salary alone isn’t enough.
Fourth, broaden the base of AI adoption — a true G3 nation is one where not just five large corporations excel, but ten thousand midsize and small firms use AI.
This is also the reality I confront daily while running NoWbus.
If China is building a massive wall, we must become the highway that stretches over it.
Connection, not isolation,
Bridge, not self‑reliance.
Korea’s place is right there.
The next 24 months will decide the next 30 years.
Since I’m not a hardware specialist, my approach may be superficial or overlook certain aspects. There are experts from various fields on LinkedIn, so I would appreciate any feedback you’re willing to share.