South Korea committed $919B to an AI programme targeting 8.4GW by 2029

That is enormous revenue with no strategic position attached. Samsung and SK Hynix supply the high-bandwidth memory that every accelerator needs, which is a genuine chokepoint — but a component supplier captures a fraction of the value of the systems it enables, and has no say in what gets built.

The $919 billion is an attempt to convert a supply chain position into a compute position, on the reasoning that a country which makes the memory should be able to run the models.

Power is the part that is hard to buy

Targeting 8.4GW by 2029 is the ambitious element, not the compute. Accelerators can be purchased. Generation, transmission and grid interconnection take years and cannot be accelerated with capital alone.

The global evidence on that is consistent. Five gigawatt-scale AI data centres are expected online this year. Loudoun County in Virginia, with roughly 250 data centres, now draws more revenue from them than its operating budget requires. Australia has softened renewables rules for new AI data centres against a forecast sevenfold rise in electricity demand. SpaceX runs a foundry producing gas turbine blades to address a backlog constraining data centre power. Crusoe raised about $3 billion at a $30 billion valuation on the proposition that power, not chips, is the binding constraint.

Korea is densely populated, has limited land, and imports most of its energy. Adding 8.4GW of dedicated load in three years is a harder problem there than in Texas or Inner Mongolia.

The third way, and who is funding it

The strategy has an external limb. Samsung led Mistral's €3 billion Series D this month at a valuation above €21 billion, with President Macron describing the round as reflecting France and South Korea's aim to build "a third way in AI" — neither American nor Chinese.

That is a coherent alignment. Korea supplies the memory, France supplies models that European buyers can host under their own control, and neither depends on a US hyperscaler for the deployment layer. Mistral has committed to 1GW of European compute by 2030.

The comparison that constrains it is scale. China's Ministry of Industry and Information Technology has set a target of 9,800 EFLOPS by 2030 backed by 3.8 trillion yuan. Alphabet alone carries $811 billion in infrastructure purchase commitments, and OpenAI has a partnership to deploy at least 10GW of Nvidia systems.

A third way at 8.4GW is a real position and not a competitive one at the frontier. Whether that matters depends on whether frontier scale continues to determine capability — and this month's releases, where post-training refinement produced more than added scale, are the first evidence in a while that it might not.