The editorial frames Korea's bet as a deliberate, well-targeted move: it leans into the two layers of the AI stack where Korea already leads (HBM memory bandwidth and precision mechatronics) instead of trying to catch TSMC on leading-edge logic. The author explicitly compares Seoul's treatment of HBM to Washington's treatment of EUV — a chokepoint worth subsidizing into permanence.
The editorial argues the AI capex discourse obsesses over GPUs and parameters while ignoring the upstream constraint: quarterly HBM stack output. Citing the HBM content per accelerator (80GB on H100, 192GB on B200, 288GB+ rumored for HBM4-class chips), the author claims the compute die is the easy part — yielding 12-high stacks with sub-micron TSVs is what actually gates the industry.
The submitted Ars Technica piece reports the headline commitment of ~1,400 trillion won across Samsung, SK Hynix, Hyundai and chaebol-aligned suppliers, with the bulk going to HBM3E/HBM4 and GAA-node DRAM and a smaller tranche to humanoid robotics, training data centers, and the actuator supply chain. It positions the announcement as a coordinated state-backed bid to lock in Korea's lead in memory and build a domestic alternative to Japan's grip on robotics-grade actuators and reducers.
South Korea's government, working with Samsung, SK Hynix, Hyundai, and a consortium of chaebol-aligned suppliers, has committed roughly 1,400 trillion won (~$1 trillion USD) of combined public and private investment through 2030. The split, per the announcement covered on Hacker News and Ars Technica: the bulk goes to memory — specifically HBM3E, HBM4, and the GAA-node DRAM lines that feed them — with a second, smaller tranche earmarked for humanoid robotics manufacturing, training data centers, and the actuator supply chain.
The headline number is the kind of figure that's easy to dismiss as a press release. The composition is not. Korea is doubling down on the two layers of the AI stack where it already leads — memory bandwidth and precision mechatronics — and explicitly declining to chase TSMC on leading-edge logic or chase the US on frontier models. Samsung's Pyeongtaek P4 and P5 lines and SK Hynix's M15X expansion in Cheongju are the named anchor projects. On the robotics side, Hyundai (via Boston Dynamics) and a new Samsung-led humanoid joint venture take the lion's share, with the government underwriting a robotics-grade actuator and reducer supply chain that today is overwhelmingly Japanese (Harmonic Drive, Nabtesco).
The political framing is industrial-policy-as-defense: Seoul is treating HBM the way Washington treats EUV — a chokepoint worth subsidizing into permanence.
The AI capex conversation has been dominated by two numbers: GPU count and parameter count. Both are downstream of a third number almost nobody outside the supply chain tracks: HBM stack output per quarter. An H100 is 80GB of HBM3 wrapped in compute; a B200 is 192GB of HBM3E; the rumored next-gen training accelerators want 288GB+ of HBM4. The compute die is the easy part. The 12-high stack with sub-micron TSVs and a working yield is not.
SK Hynix shipped the first qualified HBM3E to NVIDIA in early 2024 and has been the de facto sole supplier for Blackwell-class parts. Samsung qualified later and has been fighting yield issues on the 12-high stack through 2025. Micron is a credible third, but its share is in the single digits. The Korean duopoly currently controls roughly 70% of global HBM revenue, and Seoul's $1T plan is engineered to make sure that share doesn't erode the way DRAM commodity share did in the 2010s. This is not speculative capex. It's defensive moat-building dressed up as industrial policy.
The humanoid leg of the announcement reads stranger until you look at where the actuator and reducer supply chain currently sits. A Boston Dynamics Atlas or a Figure 02 is, mechanically, a bag of harmonic drives and brushless motors. Today, the high-precision reducers are Japanese, the high-torque-density motors are split between Japan and Switzerland (Maxon), and the integration is wherever the cash is. Korea has decided that the *next* compute platform — embodied agents in factories and warehouses — has a supply chain worth owning end-to-end, the same way they own memory. The Hyundai/Samsung play is to be to humanoids what TSMC is to GPUs: the place you go when you need to actually ship units at scale.
Community reaction on the HN thread (177 points, ~280 comments at time of writing) split predictably. The skeptics noted that Korea's last mega-investment plan — the 622 trillion won Yongin cluster announced in 2023 — is already behind schedule on land acquisition and grid capacity. The optimists pointed out that Korea is the only country with a functioning chaebol-state coordination model that can actually deploy capital this fast; the US CHIPS Act has obligated ~$33B over three years and is still arguing about water permits.
If you train or fine-tune models, the practical read is: HBM supply, not GPU allocation, is the constraint that determines whether your 2027 training run happens on schedule. When you read that a hyperscaler "secured" 100,000 B200s, what they actually secured was a slice of SK Hynix's HBM3E output. If that output expands on the Korean timeline, frontier-model training costs continue their current trajectory. If it slips — and Yongin's slippage is a real signal — expect the spot price of inference-class GPUs to stay elevated through 2027 and for the gap between hyperscaler-only and self-hosted training to widen further.
If you're building anything that touches robotics, embodied AI, or factory automation, the medium-term implication is harder to ignore. The current humanoid stack is bottlenecked on hardware unit economics, not on policy or model capability. A Figure 02 reportedly costs north of $150k in bill-of-materials, dominated by actuators and reducers. A Korean-funded, vertically-integrated humanoid supply chain is the single most plausible path to a sub-$50k unit by 2029, which is the price point at which warehouse and last-mile economics flip. If you're building software for that platform — teleoperation, fleet management, manipulation policy — the hardware partner you pick in the next 18 months will likely determine your reachable market.
For backend and infra teams, the second-order effect is power and thermal. HBM4 stacks dissipate ~30W each at full bandwidth; a B200 board with 8 stacks is ~240W just for memory before you touch the compute die. The Korean capacity expansion assumes the data-center grid expands to match. It probably won't, on the same curve. Expect the location of training capacity to keep migrating toward cheap, abundant power — which is increasingly *not* the US Northeast or Western Europe.
The interesting question isn't whether Korea hits the $1T number. They probably won't — these plans never do in absolute terms. The interesting question is whether the bet's *composition* is correct. Memory is a defensible, ten-year moat. Humanoids are a coin-flip that won't resolve until 2028 at earliest. If both work, Korea owns the two physical layers of the next compute platform the way Taiwan owns the current one. If only the memory bet works, Seoul has still bought itself another decade of leverage over every AI capex line in the world. That's the play. It is not a subtle one.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.