Samsung will build Nvidia's custom memory base dies at TSMC, report says

AI chips need huge amounts of fast memory sitting right next to them. The standard answer is high-bandwidth memory, or HBM: several memory chips stacked on top of each other like a tiny skyscraper.

That skyscraper has two kinds of floors:

• Core dies: the memory chips in the stack, where the data lives.
• Base die: the logic chip at the very bottom, which connects the whole stack to the processor.

The core dies are classic memory, which Samsung, SK hynix and Micron know inside out. The base die is closer to a regular processor chip, so it benefits from the most advanced logic manufacturing. That is where TSMC, the world's largest contract chipmaker, comes in.

Nvidia unveiled NVHBM in a blog post on August 26. In normal HBM, the memory controller, the logic that directs traffic between memory and processor, sits on the processor itself. NVHBM moves Nvidia's own controller down into the base die.

Nvidia's numbers, compared with standard HBM4E:

• Bandwidth: up to 30% more.
• Power: 15% less for the memory.
• Space: up to 25% more room on the processor for actual computing.

Counterpoint Research adds some context. It notes that standard HBM4 talks to the processor over more than 2,000 signal pins across 32 channels. With the controller inside the stack, Nvidia can replace that wide bus with a narrower, faster link of its own design.

In real life Think of a busy restaurant. Normally the head waiter stands in the dining room and runs back and forth to the kitchen. NVHBM moves that person into the kitchen doorway, so orders move faster and the dining room gets an extra table.

According to the Seoul Economic Daily, Nvidia asked all three memory makers picked for NVHBM to build it the same way:

• Base die: made on TSMC's 3nm process.
• Core dies: standardized on the sixth-generation DRAM process known as 1c.
• Suppliers: Samsung, SK hynix and Micron.

Why it matters: the newspaper reports two reasons. Different manufacturing methods across suppliers make it hard to get the same performance from a single Nvidia design. And one standard process means Nvidia doesn't have to validate each supplier's product separately.

That lines up with what Nvidia said in August. It is "establishing a standard NVHBM implementation, available from multiple memory providers," which it says cuts the work needed to qualify memory from several suppliers.

Until now, Samsung made both the core dies and the base die for all the HBM it supplies. The industry calls that a turnkey approach: one company, start to finish.

Samsung's Device Solutions division, which also runs its foundry business, reviewed Nvidia's request and accepted it, the Seoul Economic Daily reports. This is the first time Samsung and TSMC have partnered from the HBM manufacturing stage.

They aren't strangers, though. About four years ago, the two worked together to make Samsung's HBM compatible with TSMC's packaging process for connecting memory to GPUs. That was about how the chips fit together, not how they're made.

No. The change covers NVHBM only, according to the report. Here's what stays in-house:

• Standard HBM4: Samsung began mass-producing it earlier this year, as the first company in the world to do so. Its base die uses Samsung's own 4nm process.
• HBM4E: the next generation, now in mass-production validation, also on Samsung's 4nm base die.
• Other customers: custom HBM for buyers such as AMD keeps the turnkey approach.

The core dies for HBM4 and HBM4E are made on 1c lines at Samsung's Hwaseong and Pyeongtaek campuses in South Korea.

Samsung is also keeping a backup plan. "Samsung is validating whether it can also make the NVHBM base die at its own foundry," an industry official told the newspaper. "This is a hedge in case NVHBM demand exceeds TSMC's production capacity."

They were already heading the same way, Crypto Briefing reports:

• Micron: has confirmed plans to outsource both its HBM4E and NVHBM base dies to TSMC, and expects volume production by 2027.
• SK hynix: used TSMC's 12nm process for earlier HBM4 base dies and is now prioritizing 3nm for future HBM, including custom versions.

The big picture: memory makers are concentrating on the memory itself, while foundries take on the advanced logic layer underneath. For TSMC, that means making Nvidia's processors and, more and more, the base of the memory that feeds them. We saw how much AI demand already means for TSMC in its record third quarter.

Not completely, and the analysts are upfront about it. Counterpoint Research notes that chip designer Marvell described the same kind of design, with the controller in the base die, back in December 2024.

What Nvidia adds, according to Counterpoint, is a named customer and a rack standard to attach it to. Five days after the NVHBM news, MediaTek adopted NVLink Fusion and will offer it to its own chip customers, the research firm says, with $3.5 billion of Nvidia money attached.

By the numbers: Counterpoint says the one claim it can check directly is a 67% smaller area for the connection circuitry, which follows from the lower pin count. The bandwidth and power figures, it notes, are measured against HBM4E, which isn't shipping yet.

Nvidia, for its part, says NVHBM is built on the same technology it will use for its own future GPUs.

NVHBM is part of NVLink Fusion, Nvidia's program for companies that design their own AI chips but want to plug into Nvidia's rack systems. Nvidia says Amazon's Annapurna Labs is the first partner, starting with its Trainium4 chip.

"NVHBM represents a new architectural approach to advancing high-bandwidth memory performance and efficiency," said Nafea Bshara, vice president of Annapurna Labs at Amazon, in Nvidia's announcement.

What's next: the report cites unnamed sources, and we found no public comment from Nvidia, Samsung or TSMC on the request. NVHBM is still months away from volume. More on the company on our Nvidia topic page.

• Nothing changes on your desk: NVHBM is for data center AI chips, not phones or gaming PCs.
• AI could get cheaper to run: memory that is faster and uses less power helps the data centers behind the AI apps you use.
• One factory matters even more: TSMC now sits under both the processors and part of the memory, so its capacity shapes how fast new AI hardware arrives.