Memory now makes up 63% of an AI accelerator's cost - does the margin dollar flow to Samsung and SK Hynix or stay with Nvidia?

The worry, stated plainly, is that memory is dismantling Nvidia's margins from inside the bill of materials. Memory has gone from about 52% of an AI accelerator's component cost in early 2024 to 63% by the end of 2025. Public per-chip estimates put HBM and advanced packaging at 60% to 70% of a leading accelerator's total cost today. If suppliers can take that much of the cost of the box, the argument runs, Nvidia's industry-leading profitability must be the next thing to give.

That framing gets the direction of the value transfer right and the identity of the loser wrong. This is not a zero-sum squeeze between Nvidia and the memory trio. The constraint in the AI value chain has migrated from GPU compute to HBM, and the maker that holds the constraint sets the price. In 2026 that means memory is not eating Nvidia's margin; it is being passed through to hyperscalers, and both ends of the pass-through are printing record numbers at the same time. The question worth asking is not who wins this quarter. It is whether the redistribution is durable — and that is falsifiable as a divergence between two margin series on the same reporting calendar.

The pass-through is the mechanism

The load-bearing detail is a reported pricing decision, not a margin forecast. Nvidia is said to have told the contract manufacturers that assemble its AI servers to raise prices by more than 15%, and as much as 17%, on systems shipping in early 2027, covering the Vera Rubin and Grace Blackwell platforms. Nvidia has not publicly confirmed the move — it comes from Bloomberg and Fortune via supply-chain sources — but the reported cause is specific: not a tariff, not a redesign, but DRAM, NAND and HBM costs. The natural reading is that Nvidia intends to hand the memory bill to Microsoft, Google, Amazon, Meta and Oracle rather than absorb it into its own gross margin, which management has said it expects to hold in the mid-70s for 2026.

The pricing backdrop supports the move. DRAM, NAND and HBM contract prices rose 80% to 90% quarter-on-quarter in Q1 2026, the sharpest leg of the shortage. SK Hynix sold out its entire 2026 memory output before the year began, and reported long-term commitments to U.S. customers reached $750 billion, led by Nvidia. A table of public per-chip estimates shows why HBM dominates the box: HBM is $1,350 of the $3,320 H100 bill of materials, $2,900 of the $6,400 B200, and $5,800 of the $13,500 GB200, each well over 40% of the accelerator's component cost before packaging is added.

Both margins rising is the evidence

What matters for the allocation question is that the two ends of the pass-through are simultaneously near record levels, on different basis. Nvidia reported both GAAP and non-GAAP gross margins of exactly 75.0% for the quarter ended July 26, 2026, essentially flat against prior quarters even as memory prices surged. That is the leg proving Nvidia is not absorbing the cost.

The other leg belongs to the memory makers. SK Hynix posted a 76% operating margin in Q2 2026 on revenue up 257% year over year and operating profit up 557%, describing it as a record quarter — even before a one-off gain from its Kioxia stake sale, and a print that still missed the buy-side bar badly enough to knock the stock down 10%. Samsung's chip division delivered its most profitable quarter on record on another all-time-high consolidated quarter, a warning shot after being lapped by Hynix through the HBM3 era. Whether the incremental dollar sticks with the suppliers, or reverts to Nvidia, is exactly the question this pair of trends is designed to answer.

A falsifiable divergence

Treat the allocation as a test with a clean observable trigger. Confirmation of the "dollar flows to memory makers" thesis requires both conditions to hold at the same disclosure window: Nvidia keeps printing ~75% gross margin — proof it is not absorbing the bill — while SK Hynix and Samsung hold operating margins at or near their records. That pairing means the value in this supply chain has genuinely moved to the constraint holder, and Nvidia has kept its own take by pricing through.

The thesis fails on a specific divergence: Nvidia reports roughly 75% gross margin while the memory makers' margins roll over — SK Hynix off its 76% peak, Samsung's chip profitability off its record. That combination is the tell that Nvidia kept the pricing power and that the shortage premium memory makers pocketed at the peak was competed or co-designed away. The candidate mechanism is already on the tape in decelerating price momentum: contract-price gains are projected to cool from the 80% to 90% QoQ spike to a 13% to 18% rise in Q3, even as Nvidia's pass-through holds.

The HBM4 shift to custom co-design is why this is live now rather than settled. Custom HBM bespoke to a GPU generation — SK Hynix's cHBM showcased at CES 2026, Samsung's eight-layer NVHBM build specifically for Nvidia — locks each supplier into the Nvidia roadmap with multi-year allocation, which is what sustains their pricing power at the constraint. But co-design also hands Nvidia co-control of the specification, and the three-way qualification race among SK Hynix, Samsung and Micron is the pressure valve. The custom stack that guarantees a supplier's margin today is the same mechanism Nvidia could eventually use to engineer that margin down. Watch the disclosure window where Nvidia's next gross-margin print lands against the memory makers' next operating margins: if the one holds near 75% while the other steps down, the margin dollar never really left Nvidia.