Samsung HBM4 hits $1 billion in sales within 4 months

Key Takeaways

• Samsung became the first company globally to mass produce HBM4 chips in February 2025
• HBM4 revenue is projected to exceed $1.2 billion by the end of June
• The chips are designed for next-generation AI accelerators including NVIDIA's Vera Rubin platform

Samsung Electronics crossed $1 billion in sales for its HBM4 memory chips just four months after launching mass production. The South Korean company became the first in the world to ship sixth-generation high bandwidth memory when it began deliveries in February, according to The Korea Herald.

That speed matters. SK Hynix has dominated the HBM market through its close partnership with NVIDIA, leaving Samsung scrambling to catch up. By beating its rival to HBM4 production, Samsung has a narrow window to reclaim market share before SK Hynix ships its own next-generation chips.

Why HBM4 sales are accelerating this fast

Industry sources told The Korea Herald that Samsung expanded HBM4 shipments rapidly through the spring. The company expects revenue to hit $1.2 billion by the end of June. That trajectory suggests either strong NVIDIA qualification progress or successful diversification to customers like AMD, Google, or Amazon's chip teams.

Most data centers still run on fifth-generation HBM3E products. But industry observers expect HBM4 to become the primary growth driver going forward. The newer chips specifically support the hardware that runs complex generative AI applications, where bandwidth bottlenecks can cripple training performance.

What makes HBM4 different from previous generations

High bandwidth memory stacks multiple memory dies vertically, connecting them with thousands of tiny wires called through-silicon vias (TSVs). This architecture delivers massive bandwidth that traditional memory cannot match. HBM4 pushes this further with 36GB capacity in 12-high configurations, up from 24GB in HBM3E, and bandwidth exceeding 2TB per second.

Samsung designed its HBM4 chips to work with next-generation AI accelerators. That includes NVIDIA's Vera Rubin platform, which relies on high-performance graphics processing units. NVIDIA GPUs power most generative AI training workloads, making HBM qualification from NVIDIA essential for any memory vendor's AI ambitions.

Samsung's competitive position against SK Hynix

The HBM market has become a two-horse race between Samsung and SK Hynix, with Micron trailing as a distant third. SK Hynix held the lead through 2023 and 2024, winning early qualification for NVIDIA's H100 and H200 accelerators while Samsung struggled with yield issues on its HBM3E production.

Being first to market with HBM4 gives Samsung a chance to reset that dynamic. Data centers require higher bandwidth to process massive AI workloads, and the transition from older generation memory to HBM4 is expected to accelerate through the remaining months of the year.

The $1 billion milestone in four months signals strong customer demand. For comparison, it took longer for HBM3E products to reach similar revenue figures after launch.

What this means for AI infrastructure spending

HBM has become a chokepoint for AI infrastructure. Every high-end GPU requires multiple HBM stacks, and supply constraints have delayed data center buildouts throughout 2024. Samsung's HBM4 ramp could ease those constraints, assuming yields remain stable and NVIDIA qualification holds.

The broader HBM market is projected to exceed $25 billion by the end of 2025, driven almost entirely by AI and machine learning demand. Companies building AI training clusters care deeply about which memory technology their accelerators use, since memory bandwidth often determines how fast models can train.

Frequently Asked Questions

What is HBM4 and how is it different from HBM3E?

HBM4 is Samsung's sixth-generation high bandwidth memory. It offers higher capacity (up to 36GB vs 24GB for HBM3E) and greater bandwidth (over 2TB/s), specifically designed for next-generation AI accelerators.

Which companies use Samsung HBM4 chips?

Samsung designed HBM4 for next-generation AI accelerators including NVIDIA's Vera Rubin platform. The rapid sales growth suggests successful qualification with multiple AI chip customers.

Is Samsung ahead of SK Hynix in HBM4 production?

Samsung became the first company globally to mass produce and ship HBM4 in February 2025, giving it a lead over SK Hynix in sixth-generation memory, though SK Hynix dominated the HBM3E generation.

Why is HBM important for AI applications?

AI training requires massive data throughput. HBM's vertical stacking architecture delivers bandwidth that traditional memory cannot match, making it essential for high-end GPUs and AI accelerators.

Source: Tech-Economic Times / ET