Lisa Su in Seoul: AMD's HBM4 Crisis, Samsung Foundry Pitch & Nvidia RTX Spark Ships

Jamie Cole (0:00)

A Hardware & Chips, Daily News.

I'm Jamie Cole, thanks for joining me. Today, HBM4 shortage reshaping AMD supply strategy.

AMD CEO, Lisa Su flew to Seoul on October 7th to secure HBM4 memory supply from Samsung. That visit tells you something important. The shortage is no longer theoretical.

The Mi455X accelerator, AMD's flagship AI chip, requires 432GB of HBM4 per unit. That's 50% more than the prior generation needed.

When you scale that across production volumes, the math becomes acute very quickly. Supply constraints are now expected to run through 2027, and AMD can't afford to wait for the market to sort itself out. This wasn't a courtesy call. The October meeting marks a clear escalation from the exploratory Memorandum of Understanding signed earlier in 2026, which covered HBM4 for Mi455X, DDR5 for Venice EPYC, and the Helios platform. What changed is that AMD now has a specific capacity requirement, a tight timeline, and a chip that won't ship in volume without the memory to back it up. The important distinction here is that Samsung also used the sole meeting to pitch its foundry services to AMD. Memory partnership is one thing. Manufacturing is another. Samsung wants to move into the higher margin foundry business. And AMD is a target worth pursuing. No deed is confirmed. And the risks are real. Process validation that Samsung for advanced accelerator nodes carries yield uncertainty, scheduling risk, and engineering effort that nobody is yet quantifying publicly. That's the underlying pressure this story surfaces. AMD, like almost everyone in high-performance compute, runs critical production through TSMC.

Samsung pitching for manufacturing work doesn't change that overnight, but it raises a legitimate question about single-supplier dependency. The signal here is that the conversation is now formal, not exploratory. Foundry switching isn't a fast option. It's a hedge that takes years to validate. But the fact that AMD is entertaining it, even at this stage, says something about how seriously they are managing supply chain concentration risk going forward. Here's the thing, meanwhile. Samsung is restructuring its own geography. The company is investing the equivalent of around $4 billion US dollars in chip testing facilities in Vietnam, with those facilities expected to open in the fourth quarter of 2027

Standard package DRM and NAND testing will shift there, freeing capacity at the Korean churn implant for higher value work, including HBM production. The logic is clean, push volume commodity testing south, protect advanced memory production at home. What remains unresolved is whether the Vietnamese facilities supplement Korean output, or displace some of it, and whether logistics costs of moving package chips back and forth add meaningful friction. Those details haven't been disclosed. On the Nvidia side, something more concrete landed this week. Pre-orders opened for the Surface Laptop Ultra and RTX Spark Windows PCs with an October 16th confirmed ship date. That's a real product milestone, and it matters as a signal of Nvidia's push into consumer-grade AI compute, not just data center infrastructure.

Contrast that with Nvidia's early stage discussions with Reflection AI, which remain unsigned and unresolved. Multiple deal structures are reportedly on the table, including a stake increase, compute capacity provision, or full acquisition.

None of it is confirmed. Until terms are settled, investors should treat it as optionality, not a revenue event.

Consider this. One more thread worth tracking. Upscale AI's token fabric platform is set to begin rollout in the fourth quarter of 2026 It links multi-vendor AI chips, including AMD, Intel, and ARM-based designs, through Nvidia's Spectrum-X networking infrastructure. If adoption follows, it starts to shift AI cluster design away from pure Nvidia-only stacks, toward heterogeneous architectures. The key question is whether the software stack and pricing make it competitive. That answer comes in 2026 The through line across all of this is constraint. Memory constraints, foundry constraints, supply chain concentration risk. The companies managing that most effectively in the next two years will have the clearest path to scaling AI hardware. Watch the HBM4 allocation numbers when they surface, and watch whether Samsung and AMD move from pitch to agreement. Those are the signals that matter most right now. Thanks for listening. This podcast was built using AI technology. A YesWe production.