Nobody touched the earnings, and 15% evaporated anyway

Here's the deal: on August 6, SK Hynix fell 10.37% to ₩1,495,000. The next day it dropped another 4.88% to ₩1,422,000. Nearly 15% of market cap gone in two sessions. The KOSPI closed August 7 down 0.60% at 6,258.77, and a large share of that index decline came from this single stock. Daishin Securities analyst Lee Kyung-min noted that SK Hynix, with its heavy HBM revenue exposure, took the direct hit from supply-terms concerns.

And in those two days SK Hynix reported no bad earnings. No production cut, no lost customer, no plant incident. Everything that moved the stock came from outside the company — three separate stories. A research note from TrendForce, a half-price supply rumor circulating in investor communities, and speculation about a US listing for its NAND subsidiary Solidigm. Stacked in sequence over two sessions, they took out 15%.

What makes this worth writing about isn't the price move. It's that two days exposed the structure the HBM business actually sits on. Customer concentration is extreme, and a single design decision by that customer moves both volume and price for the supplier at once. What the market priced this week wasn't SK Hynix's competitiveness — it was the sensitivity of that structure.

Three stories, three different weights

First, the TrendForce report. On August 4, TrendForce said DRAM supply will remain tight through 2027, and that as a result Nvidia is evaluating lower HBM configurations for its next-generation Rubin Ultra. Specifically: Nvidia held 12-Hi HBM4e as the baseline design from 2025 through the first half of 2026, and beginning in early Q3 2026 started evaluating alternatives including 8-Hi HBM4e, 12-Hi HBM4, and 8-Hi HBM4.

The heaviest sentence in that report is this one. The mainstream SKU being previewed to key customers retains peak theoretical FLOPS and HBM4, but cuts memory to 8-Hi and 192GB. Given that Rubin is 12-Hi and 288GB, that means the follow-on Ultra part would carry less memory than its predecessor. TrendForce gives two reasons: the 2027 DRAM shortage limits the wafer capacity memory suppliers can allocate to HBM, and uncertainty remains around the validation schedule and yield ramp for 12-Hi HBM4e. The report also carries an important qualifier — the final specification has not been determined.

Second, the half-price rumor. Claims circulated through Telegram and investor communities that SK Hynix was supplying HBM to Nvidia at roughly 50% below market price in exchange for SK Group affiliates receiving priority allocation of Nvidia data center GPUs. SK Hynix publicly called this "entirely baseless." The denial was unambiguous. The stock had already moved.

Third, Solidigm. As speculation grew about a possible US listing for the NAND subsidiary, SK Hynix filed a regulatory statement saying it is "reviewing various options to strengthen the competitiveness of its overseas subsidiary Solidigm, with nothing confirmed at present." Disclosures in that form neither deny nor confirm, and in practice they often read as confirmation that a discussion is real rather than as a reduction in uncertainty.

The three stories are categorically different: one is about industry structure, one is an unfounded rumor, one is an undetermined capital-structure matter. The market did not distinguish among them. That tells you how thin confidence in the HBM supercycle narrative had already become.

What happens when 12-Hi becomes 8-Hi

Item Detail
SK Hynix, Aug 6 −10.37%, ₩1,495,000
SK Hynix, Aug 7 −4.88%, ₩1,422,000
KOSPI, Aug 7 −0.60%, 6,258.77
Samsung Electronics, Aug 7 +0.22% (closed higher)
Rubin (current) HBM4, 12-Hi · 288GB
Rubin Ultra (original design) 12-Hi HBM4e baseline
Rubin Ultra (alternatives under review) 8-Hi HBM4e / 12-Hi HBM4 / 8-Hi HBM4
Mainstream SKU preview Peak FLOPS and HBM4 retained; memory cut to 8-Hi · 192GB
Stated reasons 2027 DRAM shortage · 12-Hi HBM4e validation schedule and yield uncertainty
Status Final spec undetermined
Half-price rumor SK Hynix: "entirely baseless"
Solidigm listing "Reviewing options, nothing confirmed" (filing)

The most consequential row is the gap between 288GB and 192GB. HBM pricing scales roughly with the number and capacity of stacked DRAM dies. Going from 12-Hi to 8-Hi cuts the die count per GPU to two-thirds, and revenue falls with it. Step the generation down from HBM4e to HBM4 and the price premium shrinks too. That's volume and ASP contracting simultaneously — a multiplication, not an addition, and that's what startled the market.

There's a force pushing the other way, though, and it's easy to miss: the stated cause of the spec change is a 2027 DRAM shortage. Shortage means conventional DRAM prices rise, which is positive for all three memory makers' profitability. MoneyToday published an analysis the same week arguing that Nvidia's HBM cost burden is rising while Samsung and SK Hynix remain on solid footing, for exactly this reason. If somewhat less HBM volume frees wafers for higher-priced commodity DRAM, the earnings impact may be far smaller than the share price implies.

One more piece of context: this isn't SK Hynix's first adjustment. In April 2026 the company was reported to have revised down its HBM4 volume plan for the year and shifted production weighting toward HBM3E and other products. Supply plans wobbling with customer roadmap changes is a constant in this business, not an anomaly. What differed this time is only that the news reached the market first.

Who gains, who loses

The clearest beneficiary is, ironically, Samsung Electronics, which closed up 0.22% the same day. Two reasons. First, HBM is a smaller share of Samsung's total revenue than of SK Hynix's, so the same headline hits it less. Second, and more important: lower specs favor the challenger. 12-Hi HBM4e is the configuration where stacking difficulty and thermal management are hardest, and where the leader's technology gap is widest. If the requirement drops to 8-Hi HBM4, the entry barrier falls and Nvidia — which wants supplier diversification anyway — gets more options.

Nvidia's gain is straightforward. HBM is among the largest cost components in an AI accelerator, and that cost burden has been climbing. Lower stacks mean lower bill-of-materials cost and less exposure to supply shortfall. A SKU that keeps peak theoretical FLOPS while trimming memory is also a rational segmentation aimed at workloads where compute matters more than memory bandwidth. That said, for large-model inference, memory capacity directly sets batch size and context length — so how customers receive the tradeoff is a separate question.

What SK Hynix loses is near-term volume expectations, not market position. That's why sell-side analysts pointed to HBM4 and long-term agreements as evidence that fundamentals remain solid. HBM moves on multi-year contracts and advance orders; a spec change doesn't erase a contract. What can get repriced is the slope of the growth curve embedded in the valuation. The stock reacted to the slope, not to earnings.

For retail investors, these two days left a different lesson. A denied rumor, an unconfirmed corporate action, and real industry information all got priced in together. Those three have completely different verifiability — and in a selloff, nothing distinguishes them. It's a clean demonstration that price follows narrative direction, not the credibility of the input.

Custom HBM is the next round

The structural problem this episode exposed — a single customer design decision moving supplier revenue — is not news to the memory industry. Which is why the direction of the HBM roadmap for the past few years has been clear: move from standard part to custom part.

The core change under discussion from the HBM4 generation onward is customization of the base die. At the bottom of an HBM stack sits a logic base die, which memory makers have built to a standard spec. Once customers start putting their own functions into it, that HBM becomes a part specific to one customer's one chip. From the supplier's side, that means it becomes hard to swap — and that is negotiating leverage. Standard parts compete on price; custom parts are locked in at the design stage.

If that transition completes, an episode like this one changes character. Today, Nvidia saying "we might go 8-Hi" costs the supplier volume and ASP simultaneously. In a custom structure, the design is already committed jointly, so changes impose costs on both sides. And this isn't only good for suppliers — customers benefit too, since memory tuned to a specific workload beats a standard part on power efficiency and bandwidth.

The problem is time. Custom HBM has long design and validation cycles, so what's being discussed now takes years to show up in revenue. In the meantime SK Hynix remains inside the standard-part negotiation structure and stays exposed to news exactly like this. The structural fix is in progress and the near-term vulnerability is unchanged — that's the accurate description of where things stand.

There's a compounding variable. HBM consumes far more wafer than conventional DRAM: the stacked structure and lower yields mean several times the wafer input per unit of capacity. So every quarter memory makers run the calculation "does this wafer go to HBM or commodity DRAM?" If a 2027 DRAM shortage lifts commodity prices, the answer to that calculation can change — which paradoxically makes HBM supply tighter. That's the other face of the causality TrendForce identified: the reason Nvidia is trimming specs and the reason memory makers struggle to add HBM volume grow from the same root.

Memory has watched this movie twice: 2018 and 2022

The memory industry has a recurring cycle called customer inventory correction. The sharpest example is the second half of 2018. During the 2016–2018 DRAM supercycle, when SK Hynix and Samsung were printing record profits, server customers decided they had enough inventory and orders simply stopped. Prices halved over eighteen months — and that episode also started not with earnings but with industry chatter that North American data center operators were cutting orders. The lesson: in memory, news of a customer's plan change reaches the stock several quarters before it reaches the income statement.

The second case is 2022. When pandemic demand ended and PC and smartphone consumption collapsed, all three memory makers went into losses simultaneously. What that taught was the value of diversification. When revenue concentrates in one application — then PCs and mobile — that application's demand cycle becomes the company's cycle. HBM sits in exactly that position today: an overwhelming share of HBM revenue is tied to a small number of AI accelerator customers, effectively to one company's roadmap.

There's a case of navigating this well, too. TSMC has run for years with heavy exposure to a single enormous customer in Apple and still handled the cycles, via two moves: broadening the customer portfolio (Nvidia, AMD, Qualcomm alongside Apple), and maintaining process leadership deep enough that customers can't easily find a second source. SK Hynix is doing the same things — widening the technology gap through custom HBM and next-generation stacking, while growing the share of non-Nvidia customers such as the in-house silicon camp behind Google TPUs and Amazon Trainium. This week explained in two days why that diversification matters.

How the board moves from here

Nvidia will keep using its leverage. "Final spec undetermined" also means price and volume negotiations with memory suppliers are still live, and publicly evaluating an 8-Hi alternative is itself leverage. That isn't malice; it's what large buyers always do. But it imposes a real cost on suppliers in the form of reduced roadmap predictability.

Samsung has strong incentive to push into the opening. Winning volume share in a window where the spec requirement is lower, then using that record to qualify for the next generation, is the standard challenger playbook. How much of the HBM4 generation Samsung captures is the single biggest thing to watch in this industry over the next few quarters.

Micron sits quietly well-positioned, with a geopolitical advantage as a US supplier and sustained HBM capacity investment — an attractive third option for customers seeking diversification. Trailing the two Korean makers on total DRAM capacity, its realistic goal is share within specific generations rather than volume leadership.

AMD and the in-house silicon camp are indirect beneficiaries. If Nvidia is trimming specs because of HBM cost and supply constraints, competing products gain room to differentiate on memory configuration. Google, Amazon, and Meta accelerators will chase HBM allocation on the same logic — which, from the memory makers' side, is a customer diversification opportunity.

The decisive variable is 2027 DRAM supply and demand, since that's the root cause TrendForce cites. How tight DRAM actually gets in 2027 determines the ending here. If it's tight, commodity DRAM pricing rises and memory profitability can improve even as HBM volume trims; if supply loosens faster than expected, the HBM spec can revert toward the original plan. The most practical thing to hold onto right now is that qualifier: the final spec is not determined.

So what actually changes

For general readers, nothing changes directly. But this news tells you where the physical constraint on the AI boom sits. The bottleneck today isn't GPU design — it's the wafers to build the memory that goes next to the GPU. That a 2027 DRAM shortage forecast can alter Nvidia's product spec means software roadmaps are downstream of fab capacity.

For developers and infrastructure teams, there's a concrete implication. If next-generation accelerator memory capacity doesn't grow as expected, then efficiency per gigabyte stays valuable in large-model inference. Read quantization, KV cache compression, and memory offloading not as stopgaps that disappear when better hardware ships, but as optimizations with a long shelf life.

For Korean investors, these two days re-established the risk structure. Given the weight of two semiconductor names in the KOSPI, a single Nvidia design decision moves the entire index. That much of the August 7 decline came from SK Hynix alone illustrates it precisely — holding the index rather than the individual stock does not exempt you from this exposure.

For investors generally, the takeaway is to read the three inputs separately. A denied rumor is the part of the price that can come back. An unconfirmed corporate action carries a discount until it's resolved. Industry-structure information takes several quarters to show up in reported numbers. This drawdown blended all three, so which parts revert and which persist won't be clear until Q3 results and the final spec decision.

One sentence: SK Hynix's two days revealed a problem of customer concentration, not a problem with the company — and that structure will keep moving the stock until the final spec is locked.

🥄 Three Things You're Probably Wondering

— So what does this mean for me? If you hold Korean equities or an index tracking them, quite a lot. The weight of two semiconductor names in the KOSPI is large enough that news of an Nvidia spec review drags the whole index down. Much of the August 7 KOSPI decline traced to SK Hynix alone.

— Why now? TrendForce published its 2027 DRAM outlook along with the Rubin Ultra HBM configuration review on August 4, and the half-price supply rumor plus Solidigm listing speculation landed on top of it within two sessions. Any one of those alone would likely have been absorbed. Three at once amplified.

— Is the HBM supercycle over? Too early to say. The key detail is that the spec change is driven by DRAM supply shortage, not demand weakness. Shortage means prices rise, and wafers that don't go to HBM go to commodity DRAM, which can actually improve profitability. Plus the final spec isn't decided. What did change is that expectations about the slope of growth got reset once.

Sources

Numbers and criteria are as of announcement and may change. Investment calls are yours to make!