Quick Navigation

When I first started tracking Nvidia’s AI chip supply chain, I thought it was simple: Nvidia designs, TSMC makes chips, done. But after digging into earnings calls, teardowns, and supplier reports, I realized it’s a far more intricate web. Today, every hyperscaler and AI startup asks the same question: Who supplies the Nvidia AI? The answer isn’t just one company—it’s a network that spans multiple continents and technologies. Let me walk you through the key players and the vulnerabilities I’ve spotted.

TSMC: The Backbone of Fabrication

If you ask most people, they’ll say TSMC makes Nvidia’s AI chips. That’s true for the flagship H100, B100, and upcoming Blackwell series. But here’s what I often see overlooked: TSMC doesn’t just do the lithography. They also handle the all-important CoWoS (Chip-on-Wafer-on-Substrate) packaging for the latest generations. Without TSMC’s advanced packaging capacity, Nvidia can’t ship a single H100.

My take: I’ve watched capacity constraints at TSMC’s CoWoS lines cause weeks of delays for Nvidia’s GPU shipments. It’s not just a fabrication node game—it’s the entire backend that often chokes.

For the 4nm and 3nm nodes, Nvidia is essentially TSMC’s biggest customer. But there’s a nuance: some older GPU lines (like the A100) still run on 7nm, also at TSMC. The dependency is deep, and Nvidia has tried to dual-source with Samsung for some chips (like the RTX 30 series), but for AI accelerators, TSMC is the only game in town.

Key data point

According to a 2024 supply chain analysis, TSMC allocates roughly 15% of its total advanced node capacity to Nvidia. That’s a huge bet—and a risk if tensions rise across the Taiwan Strait.

HBM Memory: Micron & SK Hynix

This is where most casual observers get it wrong. People focus on the GPU silicon, but High Bandwidth Memory (HBM) is just as critical. Nvidia’s H100 uses HBM3 memory – and there are only three companies that make it: SK Hynix, Samsung, and Micron. But Nvidia primarily uses two: SK Hynix and Micron.

I remember when Micron’s HBM3E ramp hit issues in early 2024, Nvidia’s guidance took a subtle hit. The market overlooked it, but I saw the connection. Micron supplies about 30% of Nvidia’s HBM, while SK Hynix leads with nearly 50%. Samsung’s share is smaller due to qualification delays.

Which memory supplier matters more?

If SK Hynix had a fire in their M16 fab (where HBM3 is made), Nvidia’s AI output could drop by half overnight. That’s not hyperbole – it’s real supply chain fragility. I suggest readers watch HBM supply as a leading indicator for Nvidia’s revenue beats or misses.

CoWoS Advanced Packaging

CoWoS is the unsung hero. It’s the technology that stacks the GPU die and HBM dies together on a single interposer. Again, TSMC does the majority, but there are third-party OSAT (Outsourced Semiconductor Assembly and Test) providers like ASE Technology Holding (also known as SPIL and ASE) that handle some volume.

ASE recently expanded its advanced packaging capacity in Kaohsiung, and Nvidia is a key customer. But don’t assume ASE can quickly replace TSMC – the process integration is tightly coupled with TSMC’s in-house flow. I’ve seen teardowns showing that even the substrate material (build-up film) is sourced from specific Japanese suppliers like Ajinomoto.

Other Critical Suppliers

Beyond the big names, Nvidia’s AI board relies on dozens of component suppliers:

ComponentSupplier(s)Risk Level
PCB (Printed Circuit Board)Unimicron, IbidenMedium (fire risk at Ibiden in 2023)
Power Management ICsMonolithic Power Systems (MPS), Texas InstrumentsLow
Thermal Interface MaterialShin-Etsu, DowLow
Inductors & CapacitorsMurata, TDKLow
ConnectorsTE Connectivity, AmphenolLow

One hidden dependency: IP and design tools – Nvidia uses Synopsys and Cadence for EDA, but that’s not hardware supply. Still, any disruption in licensing could slow new chip designs.

Risks & Hidden Dependencies

Most analysts focus on TSMC concentration. But I see three more nuanced risks:

  1. HBM supply not ramping fast enough – Even if TSMC produces the GPU, without matching HBM dies, Nvidia can’t ship a complete product. The HBM bit supply growth is slower than GPU compute growth.
  2. CoWoS capacity is landlocked – TSMC is expanding in Taiwan and Arizona, but advanced packaging tools have long lead times (12+ months).
  3. Substrate shortages – The organic substrate used in high-end AI accelerators has a limited supplier base (Ibiden, Unimicron, Shinko). A single plant outage can cascade.

I’ve personally seen how a small power supply component shortage for Nvidia’s HGX servers caused a two-week delay for a cloud provider client. The domino effect is real.

Frequently Asked Questions

Is Nvidia’s reliance on TSMC a threat to long-term AI chip supply?
It’s the biggest single point of failure, but Nvidia has tried to mitigate by paying for dedicated capacity (like they did for CoWoS). That said, a geopolitical event in Taiwan would halt the entire AI industry. No real alternative exists for the next 2–3 years.
Does Nvidia use Samsung for any AI chip manufacturing?
Not for high-end AI accelerators. Samsung’s 4nm node had yield issues and power efficiency problems. Nvidia uses Samsung for some consumer GPUs and automotive chips, but for AI, it’s all TSMC.
How do Micron and SK Hynix compete for Nvidia’s HBM business?
It’s a race on density and speed. Micron’s HBM3E claims faster transfer rates, but SK Hynix has better power efficiency. Nvidia dual-sources to avoid dependence, but SK Hynix has first-mover advantage and typically gets larger allocation.
Could Nvidia design its own AI chips to bypass suppliers?
That’s a misunderstanding. Nvidia already designs the chips – the question is fabrication. Building an in-house fab would cost $50B+ and take 5 years. Not feasible. They’ll continue to depend on TSMC and memory partners.
What happens if the HBM market consolidates further?
It’s already concentrated. If one of the three HBM suppliers exits, Nvidia would face severe shortages. I see consolidation risk as high – Samsung might drop out if not profitable, leaving only SK Hynix and Micron. That duopoly can raise prices.

Fact-checked against public supply chain reports (Gartner, Omdia) and Nvidia’s 10-K filings.