When Goldman Sachs analyst Guini Lee published a note earlier this month projecting the high-bandwidth memory market will hit $30 billion this year, the figure landed with a thud that most coverage missed. The number itself is staggering — a near-100% compound annual growth rate from just three years ago. But buried inside the arithmetic is a fact that should keep every AI investor awake at night: memory has quietly grown to nearly two-thirds of the component cost of a state-of-the-art AI chip.

Not logic. Not silicon architecture. Not cooling or interconnects or any of the esoteric wizardry that chip designers love to keynote about. Memory. The stuff we used to think of as a commodity afterthought.

This is not a supply-chain hiccup. It’s a structural shift in who holds power in the semiconductor industry, and the implications are more interesting than the standard “shortage means higher prices” story that has dominated coverage since January, when CNBC reported that AI-grade memory was effectively sold out through year-end.

The BOM Tells a Story No One Wants to Hear

Look at the bill of materials for Nvidia’s latest Blackwell-class accelerator, and you see the inversion clearly. Logic — the actual compute silicon that Nvidia designs and TSMC fabricates — represents a shrinking share of total cost. HBM stacks, each one a three-dimensional tower of DRAM dies bonded with microscopic precision, eat up more than half the budget. In some configurations, approaching two-thirds.

That means the most expensive components in the world’s most coveted AI hardware are not made by Nvidia. They are made by three companies: Samsung, SK Hynix, and Micron.

An executive at one major cloud provider, asked about this dynamic in a Slack exchange last quarter, put it bluntly: “We used to negotiate with Nvidia. Now we find ourselves negotiating with Nvidia and praying to the memory gods.”

This is not how the AI narrative was supposed to play out. The story we have been telling for five years is about compute scaling — more FLOPS, more parameters, more everything. Jensen Huang owns the keynote stage. The memory companies? They operate in the background, shipping components, occasionally issuing press releases about bit density improvements that only electrical engineers pretend to read.

But compute scaling has hit a wall that memory hasn’t. Logic chips get more transistors per square millimeter with each process node shrink, yes, but the gains are slowing and the costs per wafer are climbing. HBM, meanwhile, improves with every generation not because of lithographic breakthroughs but because stacking more layers of DRAM directly increases bandwidth. The physics favors memory right now, and the economics follow the physics.

The Cartel We Are Not Supposed to Notice

There are only three companies in the world that can manufacture HBM at the volumes and yields required for AI accelerators. Three. That is not a market — it is a triopoly with a velvet rope.

When Samsung reported in January that its December-quarter operating profit would nearly triple, driven overwhelmingly by memory sales to AI customers, it was not a fluke. When SK Hynix began exploring a U.S. listing in December to capitalize on demand it described as pre-sold into 2027, that was not hype. These companies are not merely benefiting from the AI boom. They are structurally positioned to extract an ever-larger share of it.

The irony is that antitrust regulators, in their current frenzy to scrutinize everything Nvidia does, are looking at yesterday’s bottleneck. Nvidia’s dominance is real, but it is a dominance of design and ecosystem — the CUDA moat, the software stack, the developer relationships. Those are powerful but not unassailable. The HBM bottleneck is harder to circumvent because it is physical. You cannot code your way around a memory shortage.

What Happens When Memory Makers Realize What They Have

A contractor who installs server racks for a major data-center operator told me last month, standing next to a pallet of equipment that had been delayed six weeks waiting for memory-qualified nodes: “Nobody in this building talks about Nvidia anymore. They talk about whether SK Hynix shipped.”

The question now is whether the memory makers will act on their leverage. Historically, the DRAM industry has been its own worst enemy — boom-and-bust cycles, overinvestment during good times, brutal price wars during bad ones. But HBM is different from commodity DRAM. It requires custom design wins, multi-year qualification cycles, and close co-engineering with customers long before a chip tapes out. Once your accelerator design is locked to a specific supplier’s HBM stack, switching is not a matter of price-shopping. It is a matter of redesigning your entire chip.

That creates switching costs that make enterprise software look like impulse purchases. And it means the memory makers, if they are disciplined, can price accordingly.

The real story of the Goldman note is not the $30 billion figure. It is the quiet redistribution of pricing power from the chip designers to the component suppliers that nobody used to pay attention to. Nvidia will continue to dominate headlines. But the companies writing the invoices that matter may increasingly be the ones in Hwaseong, Icheon, and Boise.

Sources