The AI Chip Shortage Everyone's Missing: It's Not the GPU Anymore

I'm David Bellairian, founder and CEO of Glencore Associates and founder of AIDiscover. A couple of weeks ago I wrote here about the roughly 700 billion dollars hyperscalers are spending on AI infrastructure this year. Since then I've spent more time in the weeds of where that money is actually getting stuck, and it isn't where most coverage points. Everyone talks about Nvidia and GPU supply. That's mostly not the constraint anymore. The real bottleneck sits one layer up, and almost nobody outside the industry is watching it.

Where people think the shortage is

The GPU shortage that dominated headlines in 2023 and 2024 was real, and it was mostly about raw chip fabrication, Nvidia and TSMC simply not making enough advanced logic dies. That part of the problem has largely been solved. TSMC scaled leading-edge wafer capacity aggressively, and the industry's 2nm capacity is already booked out through 2028. If you're picturing the 2026 shortage as Nvidia not making enough chips, that picture is out of date.

Where it actually is: memory

Every modern AI accelerator needs high-bandwidth memory, HBM, stacked directly onto it, specialized DRAM that moves data ten to thirty times faster than standard memory. Without it, a GPU is largely useless for AI workloads no matter how good the chip itself is. Only three companies make it: SK Hynix, Samsung, and Micron, controlling well over 95 percent of global DRAM production between them. Demand for HBM has been growing somewhere around 80 to 100 percent a year. Supply has been growing more like 50 to 60 percent. That gap doesn't close itself, and building a new memory fab takes eighteen to twenty-four months once you break ground.

SK Hynix, the dominant HBM supplier to Nvidia, has reportedly already sold out its entire 2026 output. Its revenue from AI memory has more than tripled since 2024. HBM pricing rose roughly 30 percent in a single quarter at the end of 2025. And SK Hynix itself has warned the shortage could persist past 2030, though I'd treat any forecast that far out as directional rather than precise.

And then there's the packaging

Even once you have the GPU die and the HBM stacks, someone has to physically bond them onto a single substrate. That process, TSMC's CoWoS packaging, is used in nearly every major AI chip on the market and is reportedly booked out 52 to 78 weeks. TSMC's own CEO has said packaging capacity is running roughly three times short of HBM-driven demand. The good news, if there is any, is that this specific gap has reportedly been narrowing over the course of 2026 as new capacity comes online, from roughly 20 percent earlier in the year toward something closer to 10 percent by year end. Packaging isn't solved. It's easing. Memory is the layer getting tighter, not looser.

Why the layer matters more than the headline number

A 700 billion dollar capex number is abstract. A three-supplier chokepoint controlling the one component every AI chip on earth physically requires is not. When Samsung struggled to meet Nvidia's qualification standards for its newer HBM chips earlier this year, that wasn't a minor supplier hiccup, it removed a meaningful chunk of theoretical supply from a market that was already sold out. Concentration like that cuts both ways. It's a genuine risk to the pace of AI buildout if any one of the three stumbles. It's also, for the companies sitting inside that chokepoint, an unusually durable pricing position, which is a different thing to underwrite than a hyperscaler's capex guidance.

How I actually think about it

This is the same instinct I wrote about with the capex piece, real estate and power as the physical constraint underneath the AI story, just one layer further into the hardware itself. The pattern that keeps showing up is that the most crowded trade is the model or the chatbot everyone's heard of, and the least crowded trade is the unglamorous chokepoint three companies happen to sit on top of. I don't own individual semiconductor names through Glencore, the portfolio is built around five broader asset classes rather than single-stock bets, but the framework travels: when an entire industry is fighting over a fixed input, the input is usually the more interesting story than whoever's currently winning with it.

The GPU shortage headlines will probably keep running for another year or two after the GPU shortage itself has actually eased, because it's the simpler story to tell. The real constraint, for now, is memory. That's worth knowing whether you're allocating capital, building on top of AI infrastructure the way we do at AIDiscover, or just trying to understand why compute still feels scarce even as everyone insists the chip shortage is over.

I'm David Bellairian, founder and CEO of Glencore Associates, a Los Angeles private investment and asset management firm, and founder of AIDiscover, an AI-powered marketing and brand discovery agency. This is my own read of public reporting, not investment advice.