ConsultingAI

Samsung’s HBM4 sold at “target price” is the memory bill for Rubin

By Shreenivas Sangoram ·

Short answer

Jukan is right that this is the tell in Samsung’s quarter, and it is worth putting the unit math beside it: mid-to-high $4 per gigabit is roughly $36–40 per gigabyte, which is about $9,000–$10,500 of HBM in a single 288 GB Rubin GPU. Sold out at target price does not mean the shortage eased. It means the price of the shortage is now contractual, and NVIDIA is going to sell it as a product feature rather than absorb it.

Samsung’s earnings next year are going to be phenomenal... Take a look at this comment from Samsung’s earnings call today. Jaejune Kim, Executive Vice President of the Memory Business: “Next year’s HBM supply is already sold out, with pricing negotiations concluded at our target levels,” adding that “requests for additional volume continue to come in.” ... Korean media previously reported that Samsung had quoted HBM4 prices in the mid-to-high $4 range per gigabit (Gb). ... We believe Rubin variants equipped with Samsung’s HBM4 will be sold as a separate, higher-priced SKU compared with those using SK hynix or Micron memory. @jukan05 on X

The unit math nobody does out loud

Memory is quoted per gigabit and sold per gigabyte, and the two differ by eight. A “mid-to-high $4” quote is $36–40 per gigabyte, and the same arithmetic appears independently in the replies: “at 288GB per rubin gpu that’s ~$10k of hbm alone” (@anikatekaw). Scale that to a 72-GPU rack and the memory inside it is roughly $0.7M before a single die of logic. That is the number a lab should be budgeting against, not the per-gigabit quote, because it is the one that shows up twice: once in the GPU price and once in the rental rate on the same GPU.

It also explains Samsung’s tone. A company that had lost HBM share for two generations is now telling investors that the hardest product in its catalogue is sold out at its own price, with more requests queueing. That is a margin event, not a demand event, and it lands on the P&L the same quarter the DS division printed an operating margin near 72% (@NEWS2082680).

Why NVIDIA can live with the premium

Because it has been buying back on the spec sheet. TrendForce and ZDNet, relayed on X, have Rubin Ultra moving from 12-Hi HBM4E toward 8-Hi HBM4 for 2Q27, with Vera Rubin bandwidth trimmed by roughly 13% per GPU and up to 26% for HGX NVL8 systems (@EhrmantrautCap_). The pin-speed detail in that thread is the useful part: 11.7 to 9.3 Gbps. A lower spec at a higher price is not a contradiction if the alternative was fewer stacks. Fewer dies per stack is cheaper per stack, more stacks fit the same bit volume, and the package, not the DRAM cell, is what is scarce.

“A spec cut usually reads like a demand problem. This one isn’t,” as @JSCC2020Lee puts it, adding that TSMC has been raising CoWoS prices by up to 20% with lead times of six to eight months. When the package costs more and takes longer than the memory, paying Samsung’s number is the cheap half of the problem.

The premium lands on whoever buys that SKU

This is the part of Jukan’s read that deserves to be repeated in procurement meetings: “If Rubin ships as a separate higher-priced SKU, the premium lands on the buyer of that configuration, not on Samsung’s margin, and whoever blinked first just set the floor for SK hynix and Micron’s next pricing round” (@Rubys_MoneyFlow). Three suppliers and one price floor is a much better market structure for the suppliers than three suppliers competing on price.

The supplier-side caveat is honest and worth keeping: “for MU, the test is same price and yield, not just shortage. If Samsung slips on delivery, the premium reprices” (@alluxio1120). HBM4 moves the logic base die to TSMC, which flattens some of the historical advantage; qualification volume, not the announced quote, is what converts a price into revenue.

The spread is the real story

Independent tallies put HBM4 quotes across roughly $3 to $5 per gigabit, a wide spread for a part that is supposed to be a commodity (@MaximeRai, @WaferBull). A note from 27 October put names on it: “Samsung is quoting HBM at $4.50 — 12–25% higher than Micron ($4.00) and SK Hynix ($3.60–$3.80).” (@MojoTricks). If those levels hold, two identical Rubin boards can carry a materially different memory bill. Buyers respond by diversifying, and suppliers respond by treating allocation as a loyalty product. Neither of those is a price collapse.

What it means for labs and for the cycle

Two practical consequences. First, HBM is now large enough to change the shape of a purchase: 288 GB per GPU at these prices is a five-figure component, so multi-year contracts with deposits are rational for a buyer who expects to be short, not a favour to the seller (@firstadopter quoting Needham on the SK hynix print). Second, demand will start routing around HBM at the margin: @LiDaosen argues hyperscalers are shifting inference KV cache onto high-density QLC enterprise SSDs and testing LPDDR5x pools precisely to cap HBM exposure. That does not break the shortage, but it is where the first elasticity appears.

And the cost lands on operators whose schedules were priced before the increase: a $27B Meta agreement that depends on Vera Rubin arriving in early 2027 is exposed to the same allocation, as @bmcneely08 noted while staying long the name. Price is negotiated once; delivery is negotiated every quarter.

What we could not verify

NVIDIA does not publish HBM pricing, Samsung has not published a per-gigabit figure, and the “mid-to-high $4” number traces to Korean media reporting rather than a filing. The exact split between Samsung, SK hynix and Micron in Rubin’s allocation is not public. Treat the per-GPU dollar figures as arithmetic on reported inputs, not as disclosed invoices.

Sources

FAQ

Is “sold out at target price” the same as a price increase?

No. It means the negotiated level cleared without Samsung having to discount to fill capacity. The increase over HBM3E is already in the quoted level, not added on top of it.

Does an 8-Hi shift mean fewer bits?

It means more stacks for the same delivered bit volume, which raises package, base-die and test demand. That is why packaging capacity, not DRAM wafer starts alone, sets the ceiling.

Who absorbs the premium?

Whoever buys the higher-priced configuration. If it is sold as a separate SKU, the cloud or lab renting that rack pays it, and it becomes a line item in inference economics.

What should a lab do about it now?

Budget per gigabyte rather than per gigabit, assume the HBM line is a five-figure component per GPU, and treat contract coverage as a delivery guarantee first and a price guarantee second.

Building something like this?

I help teams ship agents, RAG and model evaluations that hold up in production.

Work with me →
More in AI newsAI news · 29 Oct 2026Between two Samsung lines, the eyes post: the shortage is now a payout storyJukan's wordless 👀👀 post sat between Samsung's 2027 shortage line and its buyback remarks: what repriced that morning was the KRW 55–75 trillion return plan.AI news · 29 Oct 2026Powertech’s $400M panel fab is where Broadcom’s ASIC volume goes when CoWoS is fullTaiwan cleared Powertech's US$400M Singapore panel-level packaging fab on 28 October: the second queue for AI packages when TSMC CoWoS is full, with Broadcom ASIC volume.AI news · 29 Oct 2026Samsung’s 60–70% LTA target is a mortgage on 2030 capacitySamsung wants 60–70% of memory on long-term agreements through 2030. The contracts lock volume, not price: a demand schedule with deposits that now justifies the capex.
nextPowertech’s $400M panel fab is where Broadcom’s ASIC volume goes when CoWoS is fullkeep scrolling