Samsung Is Juggling Knives. One Might Drop.

Here’s a supply chain story that unfolded in real time over 48 hours — and almost nobody connected the dots.

On Saturday, TrendForce reported that Samsung and SK Hynix were the exclusive HBM4 suppliers for NVIDIA’s Vera Rubin platform. By Monday evening, Jensen Huang had personally praised Micron’s CEO on stage at GTC, Micron confirmed high-volume HBM4 production, and Jensen had approved Samsung’s HBM4.

In 48 hours, NVIDIA went from two suppliers to three.

On that same Monday, Reuters reported that Samsung’s union chief confirmed a strike plan that “would disrupt chip supply.”

The world’s most important AI chip company just tripled its memory suppliers — and one of them might be about to shut down.

The 48-Hour HBM4 Rewrite

This wasn’t an accident. Jensen didn’t just happen to praise three different memory CEOs at the same conference. He was telegraphing a message: NVIDIA will not be held hostage by any single memory supplier.

The original “exclusive 2-supplier” framing from TrendForce was already outdated by the time it published. Behind the scenes, Micron had been qualifying HBM4 for Vera Rubin — and Jensen’s public blessing confirmed it.

Samsung’s Four Jobs

Here’s what most people don’t realize. Samsung isn’t just making memory for NVIDIA. Samsung serves NVIDIA in four distinct, simultaneous roles:

Role 1: HBM4 Memory. Samsung is one of three qualified HBM4 suppliers for Vera Rubin’s 288GB-per-GPU memory. With Micron and SK Hynix also qualified, this role has backup.

Role 2: Groq3 LPU Fabrication. Samsung Foundry is manufacturing the Groq3 LPU chips on its 4nm process — the inference accelerators Jensen wants in 25% of every data center. Groq wafer orders jumped from 9,000 to 15,000 wafers. Shipments to NVIDIA start Q3 2026. There is no backup foundry. Porting to TSMC takes 6-12 months.

Role 3: HBM4E Development. Samsung unveiled next-generation HBM4E at GTC 2026, targeting NVIDIA’s Rubin Ultra platform. This is R&D work that pauses if engineers walk out.

Role 4: Taylor, Texas Fabs. Samsung is building two massive fabs in Taylor (2.7 million sqft each), backed by $4.7 billion in CHIPS Act funding for 2nm Gate-All-Around technology. Fab 1 mass production starts 2027. Construction and equipment installation require a stable workforce.

One company. Four supply lines. All to NVIDIA.

The Knife That Might Drop

Samsung’s semiconductor union is demanding a performance bonus equal to 20% of operating profit. Reuters confirmed the union chief says a strike “would disrupt chip supply.” Samsung experienced a multi-week strike in 2024.

Here’s the asymmetry that matters:

For HBM4 memory, NVIDIA now has three vendors. A Samsung strike is painful but survivable — SK Hynix is the primary supplier and Micron is ramping. NVIDIA already de-risked this.

For Groq3 LPU chips, Samsung Foundry is the sole manufacturer. There is no Plan B. You can’t port a 4nm chip design to another foundry in weeks. If Samsung Foundry workers walk out, NVIDIA’s entire disaggregated inference strategy — the Vera Rubin + Groq architecture Jensen spent half his GTC keynote on — has a hole in it.

The Binning Amplifier

There’s a subtle second-order effect that makes this worse for NVIDIA specifically.

HBM4 comes off the production line in a range of speeds. NVIDIA Rubin needs the top bin — 11+ Gbps per pin, well above the JEDEC standard. AMD MI455X needs only 9.8 Gbps per pin — within standard yields.

If a Samsung strike reduces total HBM4 output, the remaining production qualifies more easily for AMD than NVIDIA. The same fab line produces memory that sorts into two buckets:

• Top bin (11+ Gbps): Qualifies for NVIDIA Rubin
• Standard bin (9-10 Gbps): Rejects for NVIDIA, perfectly good for AMD MI455X

A supply disruption doesn’t just reduce total HBM4 — it disproportionately reduces NVIDIA-qualified HBM4. AMD, needing less demanding specs, could see its supply relatively unaffected.

What To Watch

The union timeline. Is this a negotiating tactic or a genuine strike plan? Samsung’s 2024 strike lasted weeks. The union’s 20%-of-profit demand is ambitious — Samsung management will resist.

CHIPS Act leverage. Samsung received $4.7 billion in US government subsidies for the Taylor fabs. Does the US government have leverage to pressure a settlement? Disrupting a CHIPS Act-funded project is bad optics for everyone.

Groq’s contingency. Does Groq (now effectively part of NVIDIA’s ecosystem) have any discussion with TSMC about a backup process? Or is the Samsung 4nm commitment locked in?

AMD’s positioning. If Samsung HBM4 production slows, AMD — needing only standard-bin memory — might end up with more HBM4 supply than NVIDIA. Watch for AMD commentary on MI455X supply confidence.

The Bigger Picture

Samsung’s situation captures the central tension of the AI buildout: the same companies building the future are also the ones most exposed to disruption.

Samsung isn’t a passive supplier. It’s simultaneously a memory maker, a foundry, an R&D lab, and a construction company — all for the same customer. That vertical integration makes Samsung indispensable. It also means a single point of failure affects everything.

NVIDIA saw this coming for HBM4 and de-risked by qualifying three vendors in 48 hours. But for Groq3 LPU fabrication, there is no de-risk. Samsung Foundry is it.

The AI industry’s $1 trillion infrastructure buildout runs on chips. Those chips run on memory. That memory runs on fabs. And those fabs run on workers who are asking for 20% of the profits their work creates.

The juggling act continues. The question is which knife drops first.