AMD Needs 31TB of Memory Per Rack. Supply Is Already Gone.

Why is everyone calling this a chip shortage when it's actually a memory problem?

AMD's newest AI hardware, a rack called Helios, is not short on chips. It's short on the memory that sits next to the chips.

Helios packs 72 of AMD's MI455X accelerators into one rack. Together they carry 31 terabytes of a memory type called HBM4, moving data at 1.4 petabytes per second. That memory, not the processor itself, is the part of the AI supply chain that's hard to get right now.

What HBM4 actually does inside an AI chip

HBM4 stands for high bandwidth memory, fourth generation. Think of it as the chip's short-term memory, the place it holds the data it's actively working on. A large AI model needs to move enormous amounts of information in and out of memory to answer a question or generate text. The faster that memory moves data, the faster the whole system runs.

AMD's own numbers put it plainly: each MI455X accelerator carries up to 432 gigabytes of HBM4. Multiply that across a 72-accelerator rack and you get the 31 terabytes Helios needs just to turn on.

How much memory one AMD rack actually needs

That 31 terabytes isn't a nice-to-have. AMD has said Helios can't ship without it. The rack's entire performance case, roughly 2.9 exaflops of inference compute per unit, depends on having that full memory load in place. A rack short on memory isn't a slower rack. It's an unfinished one.

What happened with AMD and Samsung this month?

AMD's CEO flew to Samsung's headquarters this month to lock down more HBM4 memory supply for next year's hardware.

On October 7, 2026, Lisa Su met with Samsung's Device Solutions chief, Jeon Yong-hyun, at Samsung's Seocho office in Seoul. The agenda: HBM4 and DDR5 supply, and what both companies are calling a broader partnership.

The March 2026 deal that set this up

This wasn't a cold call. Samsung and AMD signed a memorandum of understanding in March 2026 naming Samsung the primary HBM4 supplier for the MI455X. Su's October trip was the follow-up, and the subject wasn't just more memory. According to Hardware Busters and Tech Times, Samsung is now asking AMD for chip manufacturing business in return, using its HBM4 supply position as the lever.

Why is 2027 shaping up to be the worst supply year on record?

SK Hynix already sold out its 2026 memory capacity, and its own CEO says next year will be worse still.

The company's chief, Kwak Noh-jung, told Reuters on the day SK Hynix listed on the Nasdaq that 2027 will be the worst year in the memory industry's history from a supply standpoint, with demand expected to keep outrunning production capacity past 2030.

Scale that up and the picture gets starker. Data centers are projected to consume about 70% of every memory chip made worldwide in 2026. That leaves 30% for phones, laptops, cars and every other device that needs memory. HBM alone now eats up 23% of total DRAM wafer capacity, up from single digits two years ago.

What Samsung is asking AMD for in return

A memory supplier doesn't usually get to tell a chip designer how to run its manufacturing. Samsung can right now, because AMD needs Samsung's HBM4 more than Samsung needs AMD's order. That's the part of this story that a stock headline doesn't capture: the leverage runs backward from where most buyers assume it does.

What does this mean if you're buying AI infrastructure right now?

Every account of this story explains what it means for AMD's stock. None explain what it means for your hardware order.

Tech Times, TokenPost and Hardware Busters all cover what this does to AMD's growth or Samsung's share price, not to a company actually trying to build or buy AMD-based AI hardware on a real-world 2026 or 2027 timeline.

Here's the gap they're leaving open. A delivery date on an AI hardware order is only as solid as the memory commitment behind it. Spot capacity, not a locked contract, means a bigger customer can bump your order whenever supply tightens. And 2027 is shaping up to be the tightest year yet.

Three questions to ask before you sign an AI hardware order

Three questions get past the marketing slide and into the part of the contract that actually determines whether your hardware shows up on time.

• Which memory supplier is behind this specific order? Not which supplier the vendor generally works with. Ask for the supplier tied to your allocation.
• Is that memory capacity contracted or spot? A contracted allocation survives a tightening market. Spot capacity doesn't.
• What happens to your delivery date if that supplier reallocates capacity to a bigger customer? If the vendor can't answer this clearly, that's the answer.

What should you do next?

Check whether your AI hardware order names a specific memory allocation, before the delivery date slips, not after.

If you're mid-negotiation on an AMD or Nvidia-based AI cluster for next year, the chip was never the risk everyone told you to watch. The memory inside it is. Pull up your current order or quote. If it doesn't name a memory allocation at all, that's the first question to put back to your vendor this week.

Nexairi Dispatch tracks the AI buildout where it meets physical and financial limits: power, memory, land and money. Subscribe here for the operator read, three times a week.