The chokepoint nobody saw coming
A few years ago, memory chips were the boring corner of the semiconductor market, a commodity business dominated by price wars. Then artificial intelligence arrived, and the memory sitting beside each AI accelerator became one of the most valuable, most constrained products in the entire technology stack.
That product is high-bandwidth memory, or HBM. It is the reason the memory business went from a cyclical afterthought to the center of the AI hardware story. And it is the product at the heart of the “Elon Musk’s 10X Project” pitch, which we break down in our full review.
How HBM works
Ordinary DRAM is a flat chip that sits on a circuit board next to the processor, connected by a relatively narrow data path. HBM takes a different approach. It stacks DRAM dies vertically, one on top of another, and connects them with thousands of tiny vertical wires, then places the whole stack directly beside the AI accelerator.
That design solves a specific problem: bandwidth. A modern accelerator can crunch numbers far faster than a flat memory chip can deliver them, which leaves the processor waiting and wastes its speed. HBM’s stacked design widens the data path dramatically, so the chip gets fed as fast as it can think.
The trade-off is cost and complexity. Stacking memory dies is difficult to do without defects, and the process consumes far more of a supplier’s manufacturing capacity than ordinary DRAM does. That difficulty is a feature for the companies that have mastered it, because it keeps the product scarce and the prices high.
HBM versus GDDR: two ways to be fast
HBM is not the only fast memory in the industry. Consumer graphics cards use a different memory called GDDR, which is also quick but built differently. GDDR is a flat, wide memory chip that sits on the same board as the graphics processor, optimized for cost and simplicity.
The difference is the trade each makes. GDDR is cheaper and easier to make, which is why it shows up in consumer products where price matters. HBM is more expensive and harder to make, but it delivers far more bandwidth in a smaller footprint and uses less power, which is why it wins in data centers where performance and energy efficiency dominate. As AI workloads have grown, the data-center market has swung decisively toward HBM, and that is where the money has flowed.
Who makes HBM
HBM is made by the same three companies that control DRAM: Micron Technology (MU), SK Hynix, and Samsung. There is no fourth option, and that concentration is why the product is so valuable to its makers.
SK Hynix has led the market for the past few years, particularly in supplying the highest-end accelerators. Micron and Samsung are chasing hard, and each has announced its own roadmap for the next generation. Because all three are selling into a market that cannot get enough HBM, the segment has been more profitable than the commodity DRAM it descends from. Our semiconductor explainer places HBM in the wider chip supply chain.
The HBM3e to HBM4 transition
The current margin story in memory is the shift from HBM3e, the generation shipping today, to HBM4, the next one. Each new generation packs more layers, more bandwidth, and a higher price, which is exactly what the memory makers want: a product that is harder to copy and sells for more.
The transition matters because it is where the three suppliers fight for the next round of design wins. Whoever gets qualified into the leading accelerators first tends to lock in premium pricing for a generation. The pitch’s framing of an exclusive “silent partner” relationship, though, does not match reality. It has been widely reported that Tesla asked both SK Hynix and Samsung for HBM4 prototypes, which is the opposite of a secret one-supplier deal.
The $33 billion question
The number the pitch leans on is the industry estimate that the HBM market reaches roughly $33 billion by 2027. It is a real figure from the analyst community, and it has held up as AI demand has grown. That growth is what justifies the premium valuations the memory makers now carry.
The honest reading is that the growth is real but not guaranteed at the same pace. HBM demand tracks AI capital spending, which can pause. And because memory is a three-way oligopoly, all three suppliers will expand capacity into that growth, which is the classic setup for the next oversupply. The $33 billion is a target, not a floor. For how HBM differs from the memory in your laptop, see our DRAM explainer.
The bottom line
HBM is the tightest chokepoint in the AI hardware stack, a genuinely scarce product made by three companies into a market that cannot get enough of it. The 10X Project pitch points at real technology and a real trend, then adds a layer of mystery the market has already moved past.
Ready to see the research? Click here to access the Near Future Report.
NewsletterVetter is an independent publication. We receive compensation from some of the services we review through affiliate links. Nothing on this site is investment advice. Always do your own research.