SK hynix Is Talking to Intel About Making Memory in Ohio

On September 16, Reuters reported that SK hynix is in talks with Intel about manufacturing memory in the United States. Two structures are reportedly under discussion: leasing part of the fab Intel is building in New Albany, Ohio, or forming a joint venture with Intel and AI hyperscalers that want high-bandwidth memory.

Intel rose about 5% on the report. SK hynix rose about 2% premarket.

The talks are exploratory. Neither a final agreement nor the type of memory to be produced has been decided. SK hynix said it is exploring options to strengthen global competitiveness and that nothing has been finalized.

Why this is being discussed at all

High-bandwidth memory is the tightest constraint in the AI hardware stack, and it is made almost entirely in Asia.

Every frontier accelerator — Nvidia's, AMD's, the custom silicon Google and Amazon build — needs HBM stacks sitting next to the compute die. Those stacks come from three companies: SK hynix, Samsung, and Micron. SK hynix has led on HBM share through the current cycle and has been the primary supplier into Nvidia's highest-volume parts.

The result is a supply chain where the United States designs the accelerators, TSMC fabricates the logic, and Korean memory makers supply the component that determines how fast the finished part can actually feed itself. Two of those three steps happen outside US borders, and the memory step is the one with the least domestic capacity.

Intel, meanwhile, has a partially built fab in Ohio and a balance sheet that has spent three years being asked how it will fill it. The New Albany site has slipped repeatedly. A tenant that brings its own demand is worth more to Intel right now than a marginal logic customer.

The joint venture option is the one to watch

Leasing fab space is a real estate transaction with extra steps. The joint venture structure — Intel, SK hynix, and cloud companies — is a different animal, because it would put the buyers of HBM inside the entity that makes it.

That is the same equity-linked supply logic Amazon just applied to Generac, moved up the stack to the scarcest component in the chain. Hyperscalers have spent this cycle discovering that when a supplier is capacity-constrained, a purchase order does not guarantee delivery. Ownership does.

It would also solve a financing problem. Memory fabs are brutally expensive and brutally cyclical. DRAM and NAND prices swing on supply-demand imbalances in a way foundry wafer contracts don't, which is exactly why memory names sold off hardest when Dario Amodei's pacing essay hit the tape last weekend — SK hynix closed down 6.4%, Samsung 4.1%, Micron 3.9% in US overnight trading. A JV with committed offtake from hyperscalers converts some of that cyclicality into contracted revenue, which is what makes the capex financeable.

Seoul has a veto

The part of this story that gets least attention is the one most likely to determine the outcome.

South Korea's trade ministry said production involving "national core technology" would be subject to review under the Industrial Technology Protection Act. HBM process technology is exactly the kind of capability that designation exists to protect.

Korea's position on semiconductor onshoring has been consistently uncomfortable. Its two largest companies are also two of its largest strategic assets, and every fab that opens in Ohio or Arizona is capacity that does not open in Icheon or Pyeongtaek. The government has generally supported overseas expansion where it serves customer relationships and resisted it where it looks like technology transfer.

An HBM line in Ohio, operated inside an Intel facility, potentially with American cloud companies as equity partners, sits on the wrong side of that line by default. Whether it clears depends on how the deal is structured — what process nodes move, who owns the IP, and whether the advanced packaging steps that make HBM difficult stay in Korea.

What Intel gets, and what it doesn't

Intel gets a tenant, revenue against a stranded asset, and a strategic role in the memory supply chain it exited in 2021 when it sold its NAND business to — as it happens — SK hynix.

What it does not get is a path back into memory as a product business. In a lease, Intel is a landlord. In a JV, it is a minority participant in a business whose technology belongs to its partner. Neither restores the position Intel gave up.

That may be the correct trade anyway. Intel's problem is utilization, not ambition. A fab running someone else's product at a margin is better than a fab running nothing, and the Ohio site has been a symbol of the latter for long enough that filling it has independent value for the stock.

What to watch

Whether the memory type gets named. HBM and commodity DRAM are different businesses with different politics. A DRAM line is an onshoring story. An HBM line is a strategic-technology story, and it will move slower.

Whether a hyperscaler is named. The JV structure only works if buyers commit capital. The first named cloud partner turns this from a Reuters report into a transaction.

The Korean review. If the trade ministry opens a formal Industrial Technology Protection Act review, the timeline extends by quarters, and the scope of what can move gets negotiated in public. That review, more than anything Intel or SK hynix says, will set the ceiling on this deal.

For now it is talks. But the fact that an American logic fab, a Korean memory maker, and unnamed hyperscalers are in a room together at all is the clearest signal yet that HBM scarcity has become a policy problem rather than a procurement one.