Memory just became the expensive part of a data center

Nvidia's B200 costs an estimated $5,700 to $7,300 to manufacture, midpoint around $6,400, and sells for $30,000 to $40,000. Epoch AI's teardown puts roughly half of that production cost in one place: high-bandwidth memory. HBM and the advanced packaging that surrounds it account for close to two-thirds of the chip's total unit cost. The compute logic, the part everyone assumes is the expensive part, is a minority line item.

That single fact should reorder how development teams think about a data-center pro forma. It has not yet.

How did memory eat the chip?

Epoch AI tracked the component mix across Nvidia, AMD, Google, and Amazon accelerators from Q1 2024 through Q4 2025. Memory's share of total chip component cost went from 52% to 63% in eight quarters. Logic held almost flat, 14% down to 13%. Packaging fell from 19% to 15%. HBM spending across the industry went from about $12 billion in 2024 to about $32 billion in 2025, the fastest-growing line of any component category, by a wide margin.

Micron's own numbers confirm the shift from the supply side. In fiscal Q2 2026, reported in March, Micron's cloud memory and core data-center units combined for $13.4 billion of $23.9 billion in total revenue, 56.1%. CEO Sanjay Mehrotra told analysts the quarter set records "across revenue, gross margin, EPS, and free cash flow," driven by "a strong demand environment, tight industry supply, and our strong execution." A memory company is now, functionally, a data-center company that also sells to phones.

The price side is where the story gets ugly for anyone buying, rather than selling.

Just how tight is supply?

Tight enough that the word "shortage" undersells it. Counterpoint Research clocked DRAM contract prices up 90 to 95 percent quarter over quarter in January 2026. DDR4, the workhorse memory standard for ordinary servers, rose 1,360% since April 2025. TrendForce's late-March forecast for the second quarter of 2026 projected DRAM contract prices up another 58% to 63% and NAND flash up 70% to 75%, with suppliers explicitly prioritizing server DRAM under long-term agreements because that segment carries the best margin.

Here is the part that should worry anyone underwriting a five-year hold. Samsung raised its 32GB DDR5 module price from $149 to $239, a 60% jump, and Counterpoint's MS Hwang and Neil Shah told Network World that a 64GB DDR5 RDIMM module could cost twice as much by the end of 2026 as it did in early 2025. That is not a frontier-chip story. That is the memory that goes into every commodity server, doubling in price inside 24 months.

HBM tells the same story at a different altitude. Samsung and SK Hynix reportedly raised HBM3E prices roughly 20% for 2026 orders, a move TrendForce itself called unusual, since HBM3E prices were expected to ease as the next generation ramped. They did not ease. China's approval to buy Nvidia's H200 (six HBM3E stacks per chip), plus Google's next TPU generation (eight stacks) and Amazon's Trainium3 (four stacks), pulled demand forward faster than anyone modeled. SK Hynix's own 2026 market outlook cites Bank of America pegging the HBM market at $54.6 billion in 2026, up 58% year over year, with DRAM average selling prices projected up 33%.

Most of the capacity that would fix this is already spoken for.

IDC frames the underlying supply picture at 16% year-on-year DRAM growth and 17% NAND growth for 2026, both below the historical norm the industry needs to keep pace with data-center demand alone, before a single phone or laptop gets built. IDC calls this a permanent reallocation of the world's silicon wafer capacity toward accelerator memory, not a cyclical mismatch that self-corrects, and it expects PC and smartphone prices to rise 4% to 8% by the end of 2026 as a direct consequence. Read that carefully. The hyperscale industry's memory appetite is now large enough to move retail electronics prices for everyone else on the planet.

Who actually has 2026 supply?

Nobody who isn't a hyperscaler with a signed agreement. SK Hynix holds roughly 57 to 62% of HBM shipments and revenue depending on the quarter measured, Samsung runs 35 to 40%, and Micron trails at single digits, but Micron's entire calendar-2026 HBM output is already committed under binding price and volume agreements. Most of SK Hynix's and Samsung's 2026 capacity is reportedly sold out too. If you are a mid-tier cloud operator or a corporate buyer without a multi-year supply contract signed in 2025, you are shopping in a market where the shelves were cleared before the year started.

Half the cost of building a Blackwell-class chip is now memory, not compute. That single line has grown faster than any other component category for two straight years.

What does this do to a rack, and to a campus?

An Nvidia GB200 NVL72 rack runs $2.8 million to $3.4 million at the ODM level, before the $200,000 to $500,000 in liquid-cooling hardware it needs to run. The next generation moves the number again: Tom's Hardware reported Vera Rubin NVL72 rack pricing reaching as much as $8.8 million apiece, more than double the outgoing generation, and memory is a large part of why.

Zoom out to the campus and the proportions stay the same, or get worse. Epoch AI's cost model for a one-gigawatt data center puts annualized servers and networking at $5 billion of an $8.5 billion total, 60%. Energy, land, water, taxes, and labor combined come to $0.9 billion, barely a tenth. Our case file on Crusoe's Abilene campus shows what that looks like at real scale: eight buildings, a $15 billion joint venture, roughly 400,000 GPUs targeted across the full build. Somewhere in that number is a shell and a slab of concrete. It is not where the money went.

Applied Digital's Ellendale, North Dakota campus makes the same point from the tenant's side. The site carries a roughly $7 billion, 15-year CoreWeave lease across 400 MW of contracted capacity, and CoreWeave's own business is reselling access to exactly the GPU and memory stack this piece is describing. The rural North Dakota land under that campus was never the asset anyone was pricing.

So does site selection still matter?

Yes, but for a narrower reason than it used to. Power availability, interconnection timing, and the physical ability to run high-density liquid cooling now matter because they gate how fast you can turn very expensive silicon into revenue, not because the land itself carries much cost. A site that adds six months to energization is not a modest inconvenience anymore. On a campus where the racks alone run into the billions and depreciate on a clock measured in a few years, six months of an idle, fully-paid-for GPU fleet is real money walking out the door, a subject worth its own read once you have the memory numbers in front of you.

Here is my falsifiable read. The consensus, visible in SK Hynix's own outlook, is that HBM pricing eases after 2026 as new capacity comes online. I don't buy the timing. Fab capacity for leading-edge memory takes years to bring up, most 2026 output is already contracted, and every new accelerator generation (Rubin, the next TPU, Trainium4) asks for more stacks per chip, not fewer. My bet is that memory stays the largest single component of frontier chip cost through the end of 2027, and that at least one hyperscaler publicly cites memory cost, not GPU allocation, as its binding constraint on a 2027 build before this year is out. If DRAM and HBM prices are flat or falling by mid-2027, write me back and I will take the loss in public.

This analysis is a source-cited research summary drawn from public records, not legal advice. It can contain errors and should be verified independently before any investment decision.