SK Hynix’s latest earnings came in below market expectations, and the miss is being framed as more than a simple reporting disappointment. It signals a structural turn in the memory industry. As each generation of AI accelerators carries more high-bandwidth memory, or HBM, commodity DRAM is losing share, and the old setup in which HBM expansion also lifted DRAM is starting to crack. Pure-play DRAM producers that cannot meet HBM quality requirements may be the first to feel the pressure.
SK Hynix miss tied to product mix, not just the headline number
The trigger, according to the source article, was a shift in SK Hynix’s product mix. Shipments are still moving away from commodity DRAM and toward HBM, and that slowed overall ASP growth.
Analyst Bubble Boi said HBM pricing has been largely flat in recent periods, while spot pricing for commodity DRAM has kept climbing. In practice, that means every step SK Hynix takes to allocate more capacity to HBM also means giving up DRAM bit share at a time when DRAM is appreciating quickly. From an ASP perspective, the company is swapping a product that has stalled for one that is still rising. Bubble Boi said that gap helps explain why the quarter missed expectations.
He added that this is not a SK Hynix-only issue. Samsung and Micron, in his view, are likely to show the same pressure in their own results.
Why memory makers were able to profit from both HBM and DRAM
Bubble Boi traced the past two years back to a market mechanism that worked unusually well for memory suppliers. Producing one equivalent HBM wafer consumes roughly the same capacity as three commodity DRAM wafers.
- HBM demand rises and eats into available commodity DRAM supply.
- DRAM supply tightens, pushing up spot and contract prices.
- Memory makers collect high HBM margins while also benefiting from DRAM price gains.
That is why SK Hynix had enough confidence to expand DRAM capacity aggressively. The logic was simple: even if conditions turned, the company could still redirect DRAM capacity into HBM production, and HBM demand appeared close to open-ended.
Bubble Boi described that framework as “memory and chill.” It supported bullish positions across the sector and fed the market consensus that memory stocks would remain AI beneficiaries.
More HBM per accelerator also means less server DRAM demand
That positive loop, however, has a blind spot. Bubble Boi argued that as each AI accelerator carries more HBM, the system around it needs less server DDR DRAM. Over time, that substitution effect is beginning to outweigh the supply squeeze created when HBM production consumes DRAM capacity.
NVIDIA’s roadmap is the clearest data point cited in the article. H100 carries 80 GB of HBM. B200 moves to 148 GB. The next-generation Rubin platform is expected to use 288 GB as a standard configuration, or 1.5 times the Blackwell generation. With each step, more bits move into the HBM layer and less room is left for commodity DRAM.
Bubble Boi put it bluntly: no engineer would actively choose DDR if there were a better option. Server DRAM is still used because HBM bandwidth is not yet enough. If technology allowed it, he said, engineers would rather place 1 TB of HBM on each accelerator than add more DRAM modules across the server.
He also noted that advanced packaging above 8-high stack becomes much harder, and that remains a bottleneck for further HBM expansion. In his view, though, that is a speed constraint rather than a change in direction.
Commodity DRAM is being squeezed from both ends
The pressure on commodity DRAM is not coming only from HBM. The article also points to flash offload technology, pushed mainly by Chinese AI labs, as another threat moving toward mainstream use. The approach relies on high-capacity NAND flash to replace part of DRAM working memory. Its latency is worse than DRAM, but it offers clear cost and capacity advantages, especially for KV cache management in inference workloads.
That leaves commodity DRAM under pressure from both directions. High-end compute demand is being pulled upward into HBM, while cost-sensitive use cases are seeing flash-based techniques move in from below. In that environment, suppliers that can survive over the long run are likely to be those with strong HBM manufacturing capabilities as well.
Bubble Boi said SK Hynix and Samsung are in a better position than Micron across the HBM supply chain. Micron, he wrote, still trails its competitors on technology maturity and yield in HBM, making it the vendor with the most urgent need to defend its position.
LTA timing and ADR issuance suggest SK Hynix moved early
Bubble Boi also argued that SK Hynix management may have seen the shift ahead of the market. The company’s LTAs, or long-term supply agreements, with key customers were signed at a time that could help cushion reported numbers when DRAM spot pricing starts to roll over.
He also pointed to SK Hynix’s recent ADR issuance in the US market as a capital raise that landed close to an industry turning point.
For investors in the memory sector, the message is less about one quarter and more about positioning. If the market consensus is still built around the idea that AI will keep driving both HBM and DRAM higher at the same time, some participants may already be adjusting to a different setup.