Nvidia's Memory Panic Turns Samsung, LG, and Hyundai Into Chokepoint Players

The Number That Changes Everything

Nvidia doubled its supply agreements in a single quarter — from $119 billion to $279 billion. The $160 billion increase landed almost entirely on memory. That is not a procurement decision. It is a panic purchase from a company that has just realized its chips are useless without someone else’s silicon.

At the August 26 earnings call, Jensen Huang did not mince words. “People used to ask why a chip company was working with memory suppliers,” he said. “Now they understand we were quite clever to start earlier.”

The admission matters because it reframes the entire semiconductor cycle thesis. For years, the narrative was straightforward: Nvidia designs the GPU, TSMC fabs it, and everybody waits for demand to slow down. Memory suppliers were peripheral. Now they are the bottleneck.

Korea Is the New Bottleneck

The three major memory suppliers Collette Kress named — Samsung, SK Hynix, and Micron — are not equal players. Two of them are South Korean. That concentration alone makes Korea structurally powerful in any AI infrastructure conversation.

Samsung is not just supplying memory. It is co-developing the manufacturing process itself. Nvidia’s cuLitho software, which accelerates computational lithography by up to 20 times, is running inside Samsung’s fabs to sharpen the masks that define the circuits on every advanced chip. This is not a vendor relationship. It is a co-dependency. Samsung’s process nodes depend on Nvidia’s GPU-accelerated tools; Nvidia’s chip yields depend on Samsung’s ability to print defects out of existence.

SK Hynix occupies the most strategically sensitive node. It is co-developing next-generation AI memory for the Vera Rubin platform and the accompanying CPU and robotics computing stack. High-bandwidth memory is the oxygen supply for every Nvidia GPU cluster. Whoever controls HBM production controls the pace of AI deployment. SK Hynix currently leads the market, and Nvidia’s multi-year cooperation language in its earnings materials is a signal: it cannot afford to lose that lead to Samsung or Micron.

LG and Hyundai: The Physical AI Play

The Nvidia call also elevated two non-chip Korean companies into the AI infrastructure fold. LG is building what Nvidia calls an “AI factory” covering robotics, autonomous driving, data centers, and GPU cloud. Hyundai is deploying 50,000 Blackwell GPUs across vehicle AI, autonomous driving, smart factories, and robotics development.

This is the “physical AI” thesis that Huang has been promoting for two years. It is the next leg of demand after data-center training clusters. The catch: physical AI requires massive on-device inference compute, which means more memory, more HBM, and more packaging capacity. The Korean companies are not buying Nvidia chips — they are buying the future of Nvidia’s demand curve.

The Demand Gap Is Real

Nvidia reported $96.2 billion in Q2 revenue, up 106 percent year over year. Data center revenue alone hit $89 billion, up 117 percent. Management guided for roughly 70 percent growth in fiscal year 2028 — and Huang was blunt about why the number is conservative. “Demand is far greater than 70 percent,” he said. “We are confident we can deliver 70 percent based on current supply.”

Translation: the constraint is not demand. It is memory. If HBM supply does not expand, Nvidia’s revenue growth collapses regardless of how many orders pour in.

This directly challenges Gartner’s recent claim that the semiconductor industry is approaching a “peak.” Gartner is measuring installed demand, not constrained supply. The semiconductor cycle has not peaked — it has hit a memory wall. Until Samsung and SK Hynix bring new HBM capacity online, which is a multi-year construction problem involving TSMC CoWoS packaging, advanced packaging, and fab investment, Nvidia’s growth runway is shorter than the market assumes.

Who Wins, Who Loses

Winners: Samsung and SK Hynix. Their bargaining power has improved dramatically. They are no longer commodity memory vendors; they are strategic enablers of the most valuable company in semiconductors. Hyundai and LG gain access to cutting-edge AI infrastructure at scale, which positions them ahead of global peers in robotics and autonomous vehicles. SK Telecom and Naver, both mentioned in the release for building GW-scale AI clouds on Nvidia’s DSX platform, are securing sovereign compute capacity that will matter as regulation tightens around AI data residency.

Losers: Everyone else waiting for memory prices to normalize. The shortage is structural, not cyclical. New HBM fabs take three to four years to come online. TSMC’s CoWoS packaging capacity is already fully allocated through 2027. Startups and mid-tier cloud providers that cannot secure memory allocations will be priced out of the next generation of AI training clusters.

Uncertain: Micron. Named alongside two Korean competitors as a primary memory supplier, Micron faces a difficult position. It is the only non-Korean player in the trio, and Samsung and SK Hynix are simultaneously expanding HBM capacity and deepening their co-development ties with Nvidia. Micron’s margin depends on maintaining technical parity while losing ground on strategic access.

What Happens Next

Nvidia’s supply agreements will likely continue to grow through the rest of FY2027 as the company locks in memory for the Vera Rubin launch. Samsung and SK Hynix will face pressure to accelerate HBM4 and HBM4e production timelines. Any delay becomes a delay for every AI cluster built on Nvidia silicon.

Korean industrial policy will take notice. The government’s semiconductor subsidy discussions are already underway, and Nvidia’s public endorsement gives Seoul a stronger case for direct support of HBM expansion. Hyundai and LG’s AI factory commitments may trigger additional domestic incentive programs.

The broader implication is that the AI infrastructure buildout is no longer a software and chip story. It is a materials and memory story. The companies that control HBM supply — and the Korean firms that dominate it — will determine who builds the next wave of AI systems and who simply watches from the sidelines.