HBM Identified as Primary Bottleneck for AI Chip Manufacturing; Nvidia, Google, AMD, Amazon Consumed 90% of Global Supply in 2025

September 20, 2026

New Market Pitch USA

Overview

High Bandwidth Memory (HBM) has emerged as the most significant bottleneck in AI chip manufacturing, surpassing advanced packaging technologies like CoWoS. According to Epoch AI estimates, just four companies—Nvidia, Google, AMD, and Amazon—consumed approximately 90% of the global HBM supply in 2025. While CoWoS capacity has seen some relief through TSMC’s expansions, HBM consumption nearly matches the entire global supply, making memory scarcity the most critical manufacturing constraint.

In Depth

Key Findings

The most critical bottleneck in the manufacturing of AI chips has shifted from advanced packaging, such as CoWoS, to High Bandwidth Memory (HBM). Epoch AI estimates that in 2025, a mere four companies—Nvidia, Google, AMD, and Amazon—collectively consumed approximately 90% of the world’s HBM supply.

Technical / Clinical Details

HBM is an indispensable high-speed, high-density memory technology crucial for AI accelerators and high-performance GPUs. Its stacked architecture delivers significantly greater bandwidth compared to conventional DRAM, which is vital for the intensive data processing requirements of AI workloads. While advanced packaging like CoWoS remains tight, extensive capacity expansions by leading foundries such as TSMC have provided some relief. However, HBM consumption has virtually reached the entirety of the global supply, indicating a structural challenge that demands not only manufacturing capacity expansion but also broader supply chain adjustments and technological innovation. The fundamental challenge lies in scaling HBM production to keep pace with the exponential growth in demand from leading AI developers.

Background & Context

The rapid expansion of AI infrastructure has created unprecedented demand across the semiconductor industry, leading to a succession of identified bottlenecks. Previously, leading-edge process nodes and CoWoS packaging were primary concerns, but now, the scarcity of HBM has become the most critical constraint. AI leaders like NVIDIA are aggressively expanding their data centers, and HBM is key to maximizing GPU performance within these facilities. The overwhelming dominance of HBM consumption by a few large corporations poses significant challenges for new entrants and smaller firms in the AI chip development space, potentially concentrating market power.

Strategic Significance & Outlook

The HBM supply crunch will directly impact the pace of AI chip development and deployment, making HBM manufacturing capacity expansion and efficiency improvements a top priority for the coming years. Memory manufacturers are expected to accelerate investments in HBM production and facilitate the transition to next-generation HBM (e.g., HBM4). Additionally, AI chip designers may explore novel architectures that optimize HBM utilization and investigate alternative memory solutions. Diversifying HBM supply sources and geographically dispersing production facilities will also be crucial for enhancing supply chain resilience and mitigating future bottlenecks, ensuring sustainable growth for the AI industry.

Source: https://newmarketpitch.com/blogs/news/ai-chip-challenges

Get our weekly technology intelligence — free

Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.

Subscribe Free — Weekly Tech Intelligence

By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.