The memory tutorial track at Hot Chips this year was mostly one argument made from several directions, and the numbers are worth having in one place.

\- Accelerator FLOPS have grown about 3× every two years. HBM bandwidth has grown less than 2× over the same period. (Micron tutorial)
\- Memory's share of an AI chip's component cost went from 52% in Q1 2024 to 63% in Q4 2025. (Samsung, citing Epoch AI)
\- HBM takes roughly 3× the die area of DDR per bit, DRAM wafer capacity has been flat for over a decade, and a new fab has a 2+ year lead time. Spot DRAM is up \~7× per GB, and most of the memory vendors' revenue growth is price, not bits. (Jim Handy, Objective Analysis)
\- HBM4 doubles the interface to 2,048 I/O for \~2.8 TB/s nominal per stack, against 128 GB/s for HBM1. SK hynix quoted >2 TB/s in practice and \~40% better power efficiency than HBM3E. (SK hynix and Micron tutorials)

At 63% of BOM, halving the bytes moved is worth more than a node shrink, which is why the responses on stage were so different from each other:

\- \\Samsung, custom HBM base die.\\ Move the base die to 4nm logic, replace the 8 × 4 mm HBM PHY with a compact die-to-die link, offload the memory controller off the GPU, and eventually stack DRAM directly on the processor ("zHBM") for a claimed 70% lower I/O power and \~100 W saved per package.
\- \\Oxmiq, High Bandwidth Flash.\\ NAND in an HBM-style stack: 8 to 16× the capacity per dollar at \~3 TB/s per cube. Their rack sim at equal cost and power came out at \~14× aggregate capacity for \~0.6× per-GPU bandwidth. Wins on MoE expert weights and cold KV, loses on dense models. No production inference engine has a backend for it yet.
\- \\Samsung, LPDDR5X-PIM.\\ MAC trees in all 16 banks, 614 GB/s internal against 76.8 GB/s external. The clever part is "Address Align Mode," which maps MAC instructions onto ordinary DRAM addresses so an unmodified controller still works. Measured 3.01× tokens/sec on Llama 3.1 8B on real silicon.

My read is that PIM is the only one of the three whose gating problem is already solved, since it needs no host redesign, while HBF's gate is entirely software and custom HBM's is a business-model change for DRAM vendors. Interested in whether people closer to this see it differently.

Disclosure: I wrote these up in more detail in a free primer on this year's conference. The memory chapter is at https://allaboutchips.com/#memory, and every figure there is sourced to the specific talk and slide.

And it's only that expensive because they are willing to pay that price.

63% of the BoM being just the memory is wild. A couple years ago everyone was obsessed with compute die area and now the real bottleneck is shipping enough gigabytes close enough to the chip.

The PIM stuff skipping the host redesign is huge, that's usually the part that kills these ideas before they leave a lab. If you can drop it in and get a 3x bump on real silicon with no software rework, that's basically cheating.

Everyone gets greedy and keep raising prices on each other. Going to be an epic collapse

\> and the numbers are worth having in one place

:(

Honestly a little heart warming to see east asian suppliers to get to fleece investors just as hard. Only thing stopping US gov from stepping to limit their rends is many of them have 50% foreign shares, but wouldn't be surprised they'll be donating back 100s of billions in antitrust in af ew years.

RAM has increased in price by at least 600%, but its BOM share has grown from 52% to 63%? How strange is that?

So the RAM makers are grossing as much as TSMC from these cards? That's wild.

Have you considered that the computer part might also have increased in price?

- Samsung, custom HBM base die. Move the base die to 4nm logic, replace the 8 × 4 mm HBM PHY with a compact die-to-die link, offload the memory controller off the GPU, and eventually stack DRAM directly on the processor ("zHBM") for a claimed 70% lower I/O power and ~100 W saved per package.

So where would the memory controller go? Onto the base die? The last time someone did that (Micron, Hybrid Memory Cube [HMC]), people lost their minds.

The kind of memory matters. Expensive HBM increases BoM much higher than DDR, yet the hyperscalers still buy it. If it was just about shitting enough gigabytes it would be different landscape.

Nice to see global suppliers getting share of generous US AI bubble wealth transfer. Until in a few years US anitrust fines memory producers a few hundred billion for non market behaviors.

I’m pretty sure that has always been the case. Even before memory prices really took off.