The AI memory market rebalancing

Samsung lost half its market share in AI memory chips in 12 months.

In semiconductors, where nothing moves fast.

• Q4 2024: SK Hynix held 51% of the HBM market. Samsung held 40%. Micron was a distant 9%.
• By Q4 2025, Micron is at 21%. Samsung at 22%.

Here's what actually happened.

Samsung's HBM3E chips kept failing Nvidia's qualification tests. Not once. Repeatedly, across 2024 and into 2025. The reported reason was heat and power consumption. Samsung's stacks ran too hot for Nvidia's requirements, while SK Hynix had already set the bar using a superior bonding process called MR-MUF. Samsung couldn't match it. NVIDIA's orders went elsewhere.

Micron, meanwhile, got its yields up. HBM3E 8-layer hitting around 75%. 12-layer around 70%. Quiet, unglamorous execution while the attention was on the Korea vs Korea rivalry.

That's the past story. A product problem that cost Samsung two years of the most important memory cycle in a generation.

The forward story is harder.

South Korea imports virtually all of its helium. That matters because helium isn't optional in fab operations. It goes into lithography cooling, wafer handling, and process chambers. When supply tightens, Korean fabs absorb that pressure before anyone else.

Micron's fabs are located in the US and Taiwan. Both had more stable helium access over the last 18 months. That's not the reason Samsung fell behind. But it is the reason catching up is harder than the recovery narrative suggests.

Samsung has to fix a process problem, rebuild Nvidia's confidence, and do it inside a supply chain that has structural input exposure its main competitors don't share.

I keep thinking about how tidy the "Samsung comeback" headlines will sound when they come. The underlying math is messier than that.

There's a longer piece on South Korea's helium exposure and its implications for the HBM roadmap. Read the complete breakdown here: https://www.vizmaya.fyi/story/south-korea-gpu-hour