Washington wants Samsung and SK hynix to put memory fabs in the US, but a 100% tariff threat won't make chips appear faster. If you're budgeting servers in 2026, you should treat memory as the bottleneck now.
Commerce Secretary Howard Lutnick made the pressure public at Micron's first concrete pour in Clay, New York, on July 9. According to Aju Press, he said he wanted Samsung Electronics and SK hynix to build memory chip production facilities in the United States. Other reports on the same remarks had him putting the choice more bluntly: pay a 100% tariff or build in America.
That is a hard message. It also lands at a bad time.
SK hynix listed American depositary receipts on Nasdaq the next day, July 10, raising about $26.6 billion, according to Yonhap. Washington was sending its signal just as one of the world's most important memory suppliers was tapping US capital markets. You can understand the politics. You still have to ask whether it helps supply.
A tariff can't build a fab #
The memory market is already stretched thin. More than 70% of high-end memory chips produced globally in 2026 are expected to go to data centers, according to industry analysis cited by Fox News and other outlets. The Washington Post reported in February that Amazon, Google, Microsoft, Meta and Oracle were collectively on track to spend about $700 billion this year on AI infrastructure, almost double their 2025 spending.
That demand has pulled Samsung, SK hynix and Micron toward high-bandwidth memory and high-capacity DDR5, the chips AI servers actually need. IDC put the problem plainly in its December memory shortage analysis: DRAM and NAND supply growth in 2026 is expected to be below historical norms, at 16% and 17% year-over-year. The firm also said the shortage could persist well into 2027.
Prices show you what polite forecasts hide. DRAMeXchange data cited by Edaily showed PC-grade DDR4 8Gb prices jumping from $2.10 at the end of May 2025 to $20.00 a year later, almost a tenfold move. Server DDR5 64GB RDIMM prices more than doubled over the same period. That isn't a normal cycle. It is a repricing of the memory stack underneath AI.
So Lutnick's threat is not a supply-side solution. It is a cost shock with a flag on it. If Samsung and SK hynix face a 100% duty on Korean-made memory, they don't produce a US fab overnight. Fabs take years. The companies reprice product, customers scramble for allocations, and buyers with weak purchasing power get pushed to the back of the line.
That last group includes startups.
Founders feel it first #
If you're building an AI product this year, you're buying infrastructure in a market where hyperscalers set the pace and everyone else accepts the invoice. Hetzner said its April 1 price increases were driven by sharply higher infrastructure and hardware costs, then published another June 15 price adjustment tied to standardization and procurement pressure. A recent StartupFortune analysis found some Hetzner cloud plans had nearly tripled for new customers after the June repricing. That is not abstract. A seed-stage company running inference workloads can't negotiate like Microsoft. It can't buy memory quarters in advance like Amazon. It rents servers from someone whose own bill is moving under their feet, then calls the result cloud inflation.
Investors have noticed the choke point. TechCrunch reported in May that XCENA raised $135 million in a Series B at a $570 million valuation. The startup's pitch is direct: AI's problem is memory movement - compute you can rent, memory you cannot if it isn't being made in enough volume. Its MX1 chip is designed to handle data work closer to DRAM, including KV-cache management, so systems waste less time and power shuttling information between CPUs, GPUs and memory.
The investors aren't wrong. You can rent GPU time. You can't rent memory that isn't being made in enough volume.
The shortage is now leaking into consumer politics too. Fox News reported in April that former House Financial Services Committee chairman Patrick McHenry said the AI-driven chip crunch is "hurting Republicans" because it cuts against the party's affordability pitch. Microsoft also attributed Surface price increases to "recent increases in memory and component costs," according to MacRumors. IDC later forecast global PC shipments would decline 11.3% in 2026, with no meaningful relief from the memory shortage before the end of 2027.
Frankly, a 100% tariff on companies that dominate global memory supply is not a chip strategy. It is leverage. Leverage can move an announcement, but it can't qualify a cleanroom, hire process engineers, or turn a Korean DRAM line into a New York one before buyers run out of supply.
Samsung and SK hynix may eventually put more memory capacity in the United States. Micron already has the political advantage there, with its New York project and its existing US footprint. Until more capacity actually ships, every founder signing a server contract, every PC maker setting fall prices, and every customer buying hardware is dealing with the same fact: the AI memory shortage is here, and tariffs make it more expensive before they make it better.
Also read: John Ternus takes the Apple CEO role in September with a hardware empire and a broken AI story to fix • Managing AI agents is now the most valuable engineering skill you can have • The AI industry is repricing itself around intelligence per dollar and Amazon is showing how