Another New Bottleneck

Markets were up slightly this week, with the S&P 500/Nasdaq 100 +1%/+1%. Strong performance was driven primarily by continued strong corporate earnings, notably NVDA’s “beat and raise” release on Wed, and continued hope for progress in US-Iran talks. WTI crude fell ~5% on the week to $99/bbl. Our NPM Private Market Tracker*, which shows the average price performance of the 50 largest names in our NPM data, is +22% YTD vs. SP500/Nasdaq 100 +9%/+16%. (Bloomberg; NPM)

High Bandwidth Memory & the AI Supply Chain

As the AI industry grows, it continues to face new constraints. Two weeks ago we discussed one new bottleneck: power. This week we will discuss another: memory.

Specifically, High Bandwidth Memory (“HBM”) has emerged as one of the most strategically important components in the AI stack. As LLMs continue to scale, the ability to move data rapidly between processors and memory has become increasingly critical to performance. Modern AI systems are now fundamentally memory-bound, meaning that processors often sit idle waiting for data delivery, rather than lacking computational capability. This dynamic has elevated HBM from a niche, high performance computing technology into a critical piece of the global AI economy.

What is HBM?

High Bandwidth Memory is an advanced memory architecture designed to provide dramatically greater speed and efficiency vs. conventional DRAM memory. Unlike traditional memory systems, which position DRAM chips separately on a circuit board, HBM vertically stacks multiple memory dies using through-silicon vias (“TSVs”) and places them adjacent to AI accelerators. For example, if a customer orders a new Nvidia Blackwell B200 chip, the GPU processing unit is delivered permanently bonded to most commonly eight vertical stacks of HBM memory as a single piece of hardware. This structure enables substantially greater physical density and significantly faster data transfer rates. HBM therefore allows GPUs and AI accelerators to process massive AI workloads with less risk of becoming bottlenecked by data movement constraints. (NPM; Company Reports)

While memory has long been the sleepy, boring part of the chip world, the importance of memory, specifically HBM, has risen with the emergence of modern AI workloads. LLMs continuously move enormous parameter sets between memory and compute units. As models become larger, memory bandwidth requirements increase exponentially. This creates a situation in which raw compute performance alone is insufficient. Most of the industry’s most advanced AI processors – from leaders like NVIDIA, Google and AMD – depend heavily on HBM to achieve higher speeds. In many cases, AI performance improvements are now driven more by memory advances than by improvements in transistor density. (NPM)

This dynamic has exposed a new bottleneck in the AI supply chain, mostly because HBM manufacturing is more difficult than commodity DRAM production. The process requires advanced stacking techniques, precision bonding and sophisticated thermal management. In addition, HBM must be tightly integrated with advanced semiconductor packaging (i.e. the bridge between HBM and adjacent chips) systems such as TSMC’s CoWoS (chip on wafer on substrate) platform. As a result, bottlenecks exist not only in the memory fabrication itself, but also in system design and packaging.

Supply Overview

The HBM industry is highly concentrated, with SK Hynix, Samsung Electronics, and Micron Technology collectively controlling nearly the entire market. Counterpoint Research estimates SK Hynix at 57% market share, with Micron at 21% and Samsung at 22%. Samsung has been the largest global memory manufacturer for decades but has experienced execution challenges in HBM. Micron, meanwhile, has emerged as a credible third supplier, with CEO Sanjay Mehrotra stating that the company’s HBM capacity for CY26 is “fully booked.” (Bloomberg; Trendforce)

While supply remains constrained, building new HBM fabrication capacity requires billions of dollars in capex and multiple years of lead time. HBM also carries lower yields than commodity DRAM. As a result, the industry has experienced persistent shortages of HBM despite aggressive capacity expansion plans by all major suppliers.

Demand Overview

Demand for High Bandwidth Memory has increased at an extraordinary pace due to hyperscaler AI spending. Microsoft, Amazon, Google, Meta, OpenAI, Anthropic etc. are deploying increasingly large AI clusters that require enormous quantities of advanced memory. The HBM market itself is expected to grow from approximately $38bn in 2025 to ~$58bn in 2026, a >50% y/y expansion driven almost entirely by GPU deployment. TSMC’s CEO has described HBM demand as “growing exponentially” with a “tight supply-demand balance through 2026 and beyond.” (Bloomberg)

Pricing Overview

The supply/demand imbalance has materially altered pricing for HBM. Historically, memory markets have behaved like commodity markets: highly cyclical and characterized by severe pricing swings. HBM, which typically costs 5-10x DRAM per gigabyte, behaves more like a specialty product. Prices of HBM are privately negotiated (i.e. there is no “Bloomberg” price for HBM as there is for DRAM), but Samsung and SK Hynix are reported to have raised HBM3 supply prices by nearly 20% for 2026 contracts, an unusual hike given that HBM3 is already being phased out in favor of next-gen HBM4. In our view, pricing strength could persist for multiple years given the scale of anticipated AI infrastructure deployment and the difficulty of rapidly expanding supply. SK Hynix stock is +186% YTD. (Bloomberg; Counterpoint Research)

Why is HBM relevant to private markets?

The memory bottleneck has spurred a crop of new memory and memory performance enhancing startups. For example, Lightmatter ($4.4bn EV) is developing photonic interconnect technologies intended to reduce bandwidth bottlenecks within AI systems. Additional companies such as d-Matrix ($2bn EV) and Tenstorrent ($3.2bn) are also exploring alternative AI compute architectures with differentiated approaches to data movement and memory. SambaNova Systems focuses on dataflow architectures optimized for AI memory movement ($5bn EV). (Pitchbook; Bloomberg)

The memory bottleneck is also relevant to investors in that it is a source of data center cost inflation. Microsoft’s CFO has attributed approximately $25bn of its $190bn 2026 capex plan specifically to component price inflation, with memory cited as a primary driver. (CNBC; Bloomberg)

Performance by Sector

The chart below shows performance by industry/sector within the top 100 names we track in the private secondary market. (NPM)