AI chatbots like ChatGPT and Claude triggered a scramble for high-bandwidth memory (HBM). The result: a profit supercycle that will take memory chip makers from $42 billion to $94.5 billion in operating income by 2029. The substrate layer of the AI economy is printing money.
HBM Profit Supercycle — Operating Income Forecast
What the Data Shows
New data from Fiscala (as of June 18, 2026) charts the operating income trajectory of the global memory chip industry — and the curve is unmistakable. From near-zero profitability in 2023, when the AI race ignited, the memory sector is forecast to reach $94.5 billion in annual operating income by 2029.
The driver is HBM — high-bandwidth memory — the specialized chips that sit next to GPUs in AI data centers. Every AI model that runs inference needs HBM. Every training cluster needs more of it. And there are only three companies on Earth that can make it at scale.
The Three-Player Oligopoly
SK HYNIX — $35.5B BY 2029
The HBM market leader. First to ship HBM3E to Nvidia, first to qualify for next-gen HBM4. SK Hynix doesn’t just supply the AI boom — it enables it. Nvidia’s H100 and B200 GPUs are built around SK Hynix memory.
SAMSUNG — LARGEST TOTAL STACK
The broadest memory portfolio — DRAM, NAND, HBM. Samsung was late to HBM qualification but is catching up. Its sheer manufacturing scale makes it the volume play in a market that needs more capacity than any single player can provide.
MICRON — $16B BY 2029
The American entrant. Micron’s Idaho fabs give the US a domestic HBM supply chain — critical as AI chips become a national security asset. Smaller than SK Hynix and Samsung, but strategically irreplaceable.
The key insight: Everyone talks about Nvidia’s GPU monopoly. But behind every GPU is memory — and the memory oligopoly is even tighter. Three companies. No substitutes. $94.5 billion in operating income by 2029. This is the layer beneath the layer everyone watches.
The Structural Read
In the AI Supercycle framework, the substrate layer — rare earths, chips, memory, power — is the foundation everything else is built on. This data confirms what the framework predicts: the substrate layer captures value first and most reliably, because demand for AI compute is inelastic. Every model needs memory. Every inference cycle consumes HBM bandwidth. And unlike software, you can’t copy a memory chip — you have to fabricate it.
The 2.25x growth from 2025 to 2029 is not speculative. It is the direct mechanical consequence of every AI company — OpenAI, Anthropic, Google, Meta, Amazon — building out data center capacity simultaneously. The memory makers don’t need to win the AI race. They just need the race to continue.
The Substrate Thesis
You don’t need to pick the winning AI model.
You just need the race to keep running.
The memory makers always get paid.
This is the same logic that made Nvidia a $3 trillion company — but applied one layer deeper. Nvidia sells the shovels. SK Hynix, Samsung, and Micron sell the steel the shovels are made from.
Business Engineer Framework
The AI Supercycle — Nine Layers, From Rare Earths to Governance
Memory sits in the substrate layer — Layer 1 of the AI Supercycle. The complete framework maps all nine layers, shows where value concentrates, and explains why the bottom of the stack captures profit first.
Source: Fiscala / Leverage Shares, as of June 18, 2026