America's One Memory Maker

Happy Tuesday.

The build this week: the AI boom’s biggest checks went to memory, which looks like a genuine bottleneck for the AI (and future) economy.

Before the news, a quick one: we just launched the Ironclad Maritime Summit. If you work in maritime, you can apply to attend or sponsor on the event page at ironcladmaritime.com.

Below the fold, the money moved fast elsewhere. Raytheon booked $1.1 billion for another 1,653 Sidewinders, Hitachi Energy broke ground on a $457 million transformer plant in Virginia, and Enlight closed $2.6 billion to build a 1.2-gigawatt solar complex in the Arizona desert. South Carolina went the other way, idling its newest $1.2 billion port terminal as trade softens, while Safran agreed to spend about $2.5 billion on French sea drones and Williams lined up a $5.5 billion gas-pipeline deal.

Every AI accelerator on earth runs on memory that is mostly made in Korea. Two companies, SK Hynix and Samsung, build the bulk of the high-bandwidth memory (the stacked DRAM that feeds data into an AI chip) that the AI buildout depends on. The only American maker is Micron, a distant third. This week the gap got a price tag. SK Hynix filed to raise as much as $29 billion through American depositary receipts, the vehicle a foreign company uses to list in New York, in what would be the largest such offering ever, past Alibaba's 2014 debut. Days later, Seoul unveiled an 800-trillion-won plan, (~ $520 billion) to have SK Hynix and Samsung each build two new memory fabs in Korea.

Worth saying why memory is the thing everyone’s fighting over. For two decades, raw compute has roughly doubled every couple of years while the memory bandwidth that feeds it has crawled, and the gap keeps widening. A modern Nvidia chip can do the math far faster than it can be handed the data, so the bottleneck moved off the processor and onto the memory sitting beside it. HBM is the fix: a tower of DRAM stacked die-on-die and bonded right next to the logic, and every top-end accelerator now ships with several stacks of it. That’s why a memory shortage reads straight through to AI capacity. You can fab all the Nvidia silicon you want, but without the HBM to feed it, the chip spends its day waiting.

On the US side, the response is splitting into two bets. The first is to match Korea head-on. Micron is ramping HBM4 at about twice the pace of its last generation and, for the first time, moving HBM packaging onto US soil instead of just the raw DRAM, backed by a $6.1 billion CHIPS award for its Idaho and New York fabs. The catch is timing: the Idaho line opens in mid-2027 and New York’s output doesn’t land until around 2030. The second bet skips HBM altogether. US chip startups Groq and Cerebras hold an entire model’s weights in fast on-chip SRAM, so the data never has to commute to external memory, and Qualcomm is now bonding compute directly beneath the DRAM to chase the same effect.

Set that $520 billion next to the $52 billion the CHIPS Act committed to American chipmaking, and the gap is widening. Washington spent its money on logic: the processors from Nvidia, the leading-edge lines at TSMC and Intel, the part of the chip it could point to. Korea is spending roughly ten times as much to lock in the memory beside it, with the government clearing land, approvals, and now capital to pull four fabs forward by more than a decade. Micron is racing the same direction, lifting its own budget above $25 billion, but it is one outmatched company against a duopoly their government now treats as national infrastructure. For anyone underwriting a US-onshoring thesis, the story is impossible to ignore. The country secured the chips it designs and rented the memory it cannot live without, and the price of that memory is set in Seoul.

Back Thursday.