Nvidia's H200 Reached China and HBM Memory Prices Jumped 20 Percent

Every Nvidia H200 carries six stacks of HBM3E. Hold onto that number, because it explains everything that happened to memory prices this month.

H200 chips are now reaching China following export approval. Samsung and SK hynix responded by raising HBM3E supply prices by close to 20 percent for 2026. No new factory opened. No shortage was declared. A policy decision in one country opened a new block of demand, and the global price of a component moved within weeks.

Why the Nvidia H200 China export decision moved HBM prices

The mechanism is not complicated once you count the parts. China is a large market that had been closed to this specific accelerator. Reopening it does not create six stacks of memory per chip out of thin air; it just adds every unit sold there to a queue that was already full.

Supply cannot flex to meet it. Three companies on the planet make high bandwidth memory at volume: SK hynix, Samsung and Micron. Stacking and bonding DRAM dies vertically is slow, yield-sensitive work, and you cannot spin up a new line the way you can shift wafer allocation. Two of those three suppliers just repriced.

Memory is the bottleneck now, not compute

The instinct is to treat GPU supply as a logic problem, as though the constraint sits in the fab that prints the processor. It does not. You can fab more GPU silicon faster than you can stack and bond the memory that goes next to it, which is why HBM availability, rather than die output, is what sets delivery dates for an entire data centre build.

That distinction matters for anyone reading capacity announcements. When a company says it has secured chips, the question worth asking is whether it has secured memory. It is the same question that sits underneath the enormous data centre commitments being signed this year, and it is the reason the split between training silicon and inference silicon keeps shifting.

What to watch next

Two things. First, whether Micron follows the other two upward, which would confirm this as a market-wide reset rather than two suppliers testing the ceiling. Second, whether the increase shows up in list prices for accelerators or gets absorbed in margin.

Either way it lands on the same pile as the foundry price increases Samsung pushed through earlier this year. Wafers up, memory up, and roughly nine to eighteen months of lag before any of it reaches the price of a finished device.