Checked

What is HBM?

HBM, short for high bandwidth memory, is memory built as a stack of chips that sits right next to an AI processor so it can feed it data very fast. AI chips depend on it, and it comes mainly from three memory makers.

What HBM is

Memory chips stacked like floors of a building

Ordinary computer memory lays its chips side by side. HBM stacks memory chips on top of each other and wires them together through tiny holes drilled straight through the silicon, called through-silicon vias. The stack sits on a base chip that handles the connections.

The stack is placed right beside the processor on the same package, so data travels a very short distance.

Sources: TechTarget: What is high-bandwidth memory (HBM)?; Tom's Hardware: What are HBM, HBM2 and HBM2E?; IONOS: What is High Bandwidth Memory and why it matters for AIChecked

Why it's faster

Instead of running a narrow connection very fast, HBM uses a very wide one: each stack has a 1,024-bit connection, compared with 64 bits for a standard DDR5 memory stick. That gives it far more bandwidth, the amount of data it can move each second, while using less power.

Sources: TechTarget: HBM compared with GDDR and DDR5; IONOS: What is High Bandwidth Memory and why it matters for AIChecked

Why AI chips need it

Feeding the processor

AI chips have to move very large amounts of data in and out of memory. HBM's wide, short connection does that quickly and with little power, which is why it has become a key technology for AI chips.

Demand from AI data centers has caused a memory shortage, and memory prices rose sharply through 2026. Higher memory prices pushed Amazon's 2026 spending plan up by about $20 billion.

Sources: IONOS: What is High Bandwidth Memory and why it matters for AI; Motley Fool: Micron and Nvidia's Vera Rubin memory; Fierce Network: Memory drives Amazon capex up another $20BChecked

Who makes HBM

Three memory makers

SK hynix built the first HBM chip in 2013, working with AMD. Today HBM comes mainly from SK hynix, Samsung and Micron, the three memory suppliers named in Nvidia's annual report. SK hynix was expected to supply over half of Nvidia's AI memory in 2026.

For HBM4, the newest generation, Samsung and SK hynix were reported in March 2026 to have won the main orders for Nvidia's next chips, called Vera Rubin, ahead of Micron.

Sources: TechTarget: What is high-bandwidth memory (HBM)?; Nvidia 10-K, fiscal 2026 (SEC): foundry and memory partners; TrendForce: Samsung and SK hynix tapped for Rubin HBM4; Motley Fool: Micron and Nvidia's Vera Rubin memoryChecked

Who buys it

Nvidia and AMD both use HBM in their AI chips. AMD's flagship MI350 AI chip uses Micron's HBM3E memory. See who supplies each one on the Nvidia map and the AMD map.

Sources: TrendForce: AMD MI350 uses Micron HBM3E; Nvidia 10-K, fiscal 2026 (SEC): foundry and memory partnersChecked

How making HBM changes the chip supply chain

More steps, more machines

Newer generations such as HBM4E need more etching for the through-silicon holes and more demanding deposition steps. Lam Research, which makes etch and deposition machines, said its memory equipment revenue rose 120% from a year earlier in the quarter to September 2026.

Memory makers also need EUV machines: SK hynix placed an order of about $8 billion with ASML for AI memory production. See What is EUV?.

Sources: Tech Times: Lam Research and AI chip demand; MarketBeat: Lam Research Q1 earnings call highlights; Tom's Hardware: SK hynix places record $8 billion ASML EUV orderChecked

A chip factory joins in

The base chip at the bottom of the stack is becoming more complex. TSMC will make that base chip for Micron's HBM4E, which Micron plans to launch in 2027.

Source: Tom's Hardware: Micron hands TSMC the keys to HBM4EChecked

See it on the maps

Each map shows who depends on whom, with a source for every link.

Common questions

What does HBM stand for?

High bandwidth memory: memory chips stacked on top of each other and placed beside a processor so it can move data very fast. Open Micron's map

Who makes HBM?

SK hynix, Samsung and Micron. SK hynix made the first HBM chip in 2013 and was expected to supply over half of Nvidia's AI memory in 2026.

Why do AI chips use HBM instead of regular memory?

Each HBM stack has a much wider connection than a regular memory stick and sits right next to the processor, so it can move far more data each second while using less power.

What is HBM4E?

A newer generation of HBM. Micron plans to launch it in 2027, with TSMC making the base chip at the bottom of the stack.

More explainers

About this page

Figures come from company reports and news sources linked under each section, and were last checked on Nov 21, 2026. How we check connections.

StockTies is for information only. It is not investment advice. Spot something out of date? Tell us.