Korea's $550 Billion Bet on the Power Grid
The Bottleneck Moved
For two years the AI story was export controls. In October 2023 the US Bureau of Industry and Security, the agency that polices what technology can leave the country, issued a rule keeping Nvidia's most advanced training chips out of mainland China. Everyone watched that door. This week Seoul put $550 billion behind a completely different problem, and the difference matters because it changes who wins. Korea's chip crunch has nothing to do with who can buy an H100 or a B200 NVL72 rack, Nvidia's current top training server. It comes down to something more basic: who has enough electricity and factory capacity to make the chips at all. A training chip is useless if there's no power to run it, or no factory that can build it in volume. Here's the part that's easy to miss. SK Hynix and Samsung supply the high-bandwidth memory that sits inside every one of Nvidia's training chips. Without enough of that memory, the chip itself is a shell: a fast processor with nowhere to store the data it needs. Seoul's separate $1 trillion pledge for chips and robotics, announced the same week, confirms what's really going on. This is industrial policy for the physical layer, meaning the power grid and the factories underneath the AI industry, not the software sitting on top of it. Seoul's $550 billion is a grid and factory commitment, not a chip order. What that means in practice: SK Hynix and Samsung's memory output through 2028 will decide who trains the next model. Not who signs with Nvidia first.
Compute Bought, Not Grown
OpenAI's new inference chip with Broadcom, and HP's fresh compute deal, point at the same fact. The labs are not building factories. They're signing contracts. That's a much faster and cheaper way to get capacity than Korea's approach, but it leaves them dependent on whoever they signed with. It helps to know the two chips aren't the same bet. Inference chips are the hardware that runs a trained model when you actually use it. Training chips are the hardware that builds the model in the first place. A lab that owns its inference silicon controls its own cost per query instead of paying Nvidia's margin on every API call. So every question a user asks gets a little cheaper to answer. That's why OpenAI went to Broadcom: a company that has never sold a chip directly to consumers, but has spent a decade building custom silicon for Google's TPUs, and so already knows how to design chips for exactly this kind of buyer. It's a bet on distribution economics, meaning the cost of serving the model to millions of users, not a bet on building a bigger model. One more thing happened the same week: Anthropic filed its complaint against Alibaba the same week Sonnet 5 shipped. Different companies, same target. Both moves are aimed at the pipe the model runs through, not the model itself.
Korea's $550 billion buys memory capacity that ships in 2028 at the earliest. OpenAI's Broadcom chip and HP's compute deal are bets placed years ahead of that supply landing. If the memory arrives on schedule, the labs that signed early end up with the cheapest cost per query. If it slips, the next training run will expose whoever bet wrong first.