A new supply chain conversation
SK Hynix, the South Korean memory chipmaker that has become one of the biggest beneficiaries of the AI buildout, is reportedly in early discussions with Intel about producing memory chips inside the United States for the first time. According to Reuters reporting from September 16, the companies are weighing two structures: SK Hynix leasing manufacturing space at Intel's planned Ohio facility, or forming a joint venture that could involve cloud service providers directly. SK Hynix has publicly said no specific plans or arrangements have been finalized.
The caution in that statement is standard for early-stage talks, but the substance of what is being discussed is not routine at all. Memory chips, specifically the high bandwidth memory that feeds AI accelerators, have quietly become as scarce and as strategically important as the GPUs themselves. A deal that puts US-based memory production on the table, with cloud providers potentially at the table as financiers, would mark a meaningful shift in how hyperscalers are choosing to secure supply.
Why memory became the bottleneck
For the past two years, most conversations about AI compute scarcity have centered on GPUs, and specifically on Nvidia allocation. That framing missed a parallel constraint building underneath it. High bandwidth memory, the stacked DRAM that sits directly next to AI accelerators and determines how fast a chip can actually move data, requires specialized manufacturing capacity that has not scaled anywhere near as fast as GPU production. SK Hynix has ridden that demand directly, and its 3.8 billion dollar investment in an Indiana advanced packaging facility reflects how central memory has become to its own growth story.
That Indiana facility will not begin production until 2029, a timeline that tells you how far out any real relief sits. If memory supply is genuinely constrained through the rest of this decade, GPU allocation stops being the whole story for anyone planning AI cloud capacity, because a fully allocated GPU is worthless without the memory bandwidth to feed it. Procurement conversations that focus exclusively on chip counts are missing half of the actual constraint.
Cloud providers moving upstream
The detail that stands out most in the reporting is the possibility that cloud service providers could be structured directly into a joint venture with SK Hynix and Intel, rather than simply signing supply agreements as customers. That would represent a further step in a pattern already visible across the industry: hyperscalers and neoclouds increasingly financing chip design and manufacturing capacity directly, rather than relying purely on arm's length purchasing relationships with Nvidia or memory suppliers.
We see this as a rational response to genuine scarcity rather than a speculative land grab. When a critical input is capacity-constrained for years rather than quarters, the companies that depend most on that input have strong incentives to take an equity or financing stake in expanding supply, locking in access ahead of competitors who are still purchasing on the open market. Expect more of this pattern across both logic and memory chips as the AI buildout continues.
The policy backdrop
These talks are unfolding against a policy environment actively pushing for domestic chip manufacturing. The Trump administration has signaled it may impose broader tariffs on semiconductor imports while offering relief to companies that commit to US-based production, adding a direct financial incentive for exactly the kind of arrangement SK Hynix and Intel are reportedly discussing. Intel's Ohio site, long delayed and underutilized relative to original plans, stands to benefit significantly if a major memory partner commits capacity there.
There is also a notable historical echo here. Intel sold its NAND flash memory business to SK Hynix for 9 billion dollars back in 2020, a deal that effectively marked Intel's exit from a segment of the memory market it once helped define. A new partnership focused on advanced memory manufacturing would bring the two companies back together in a market Intel previously chose to leave, this time with Intel as landlord and manufacturing partner rather than as a memory maker in its own right.
Where Samsung and Micron fit in
SK Hynix is not the only memory maker racing to keep pace with AI demand. Samsung and Micron are both expanding high bandwidth memory production, and both face the same underlying constraint: advanced packaging capacity, the specialized process that stacks DRAM dies into the high bandwidth modules AI accelerators require, is harder to scale quickly than standard chip fabrication. A US-based joint venture involving SK Hynix, Intel, and potentially cloud providers would give one memory maker a geographic and political advantage that its competitors would likely feel pressure to match.
That competitive dynamic matters for buyers because it suggests memory pricing and allocation will remain a live negotiation for years rather than settling into a predictable market. If SK Hynix secures preferential US manufacturing capacity and cloud provider financing, expect Samsung and Micron to pursue comparable arrangements, potentially with different hyperscalers or governments. Enterprises relying on any single cloud vendor's memory supply chain should watch which memory maker that vendor is aligned with, since the alignment increasingly determines delivery timelines as much as price.
What this means for capacity planning
For enterprises building AI roadmaps that extend into 2027 and beyond, the practical implication is straightforward: memory availability deserves the same procurement attention that GPU allocation has received for the past two years. Ask cloud vendors directly what memory configuration underlies the instance types you are being sold, and whether their supply commitments extend far enough to cover your planned scaling. A GPU instance with constrained memory bandwidth will underperform on the workloads that matter most for large model inference and training.
It also means treating any vendor's long-term capacity promises with more scrutiny when memory supply is this tight. A cloud provider can commit to GPU allocation on paper far more easily than it can guarantee the memory supply chain behind it, particularly when that supply chain runs through joint ventures and facilities that will not reach production until the end of the decade. Build contract flexibility around memory-dependent instance types now, while the constraint is still being negotiated rather than fully priced into every contract you sign.
