Bernstein has revised upward its valuation targets for major memory manufacturers, projecting a pronounced jump in high-bandwidth memory prices that should materially alter profit expectations across the sector.
The broker raised its price objective on Samsung Electronics to KRW440,000 from KRW225,000; lifted SK Hynix to KRW3,300,000 from KRW1,150,000; and increased Micron Technology to $1,300 from $510. Bernstein kept Outperform ratings on all three names, while maintaining an Underperform rating on KIOXIA, which the firm notes has no exposure to HBM.
Bernsteins central thesis is that HBM prices must rise significantly to reduce a growing profitability discrepancy relative to conventional DRAM. Since the third quarter of 2025, conventional DRAM prices have risen by roughly 4.5 times, whereas HBM prices - which are largely fixed by annual contracts - have shown minimal movement over the same period. That divergence, Bernstein says, is creating a strong economic incentive to reprice HBM.
Using wafer-level economics, Bernstein finds that allocating capacity to conventional DRAM currently produces more than twice the revenue and nearly three times the gross profit per wafer compared with deploying that capacity to HBM. The brokers analyst, Mark Li, summarizes the dynamic simply: "This is why memory suppliers and GPU/XPU companies are negotiating 2027 HBM price now to narrow the gap."
In its modeling, Bernstein assumes HBM prices will increase by 2 to 2.5 times next year. The firm notes this uplift would be meaningful but still somewhat short of the level required to completely close the profitability spread between HBM and conventional DRAM. The forecast rests on the assumption that suppliers will acknowledge the strategic importance of keeping HBM priced at levels that sustain the broader AI ecosystem.
Bernstein highlights a specific transmission mechanism for these cost shifts. Unlike conventional DRAM and NAND, which cloud providers can purchase directly, HBM typically arrives packaged inside GPUs and becomes part of the chip suppliers cost base. If Nvidia operates at a 75% gross margin and seeks to preserve that margin, Bernstein calculates Nvidia would need to mark up any HBM cost increase by roughly a factor of four when transferring the cost to customers.
Combining that potential GPU markup with rising conventional DRAM and NAND prices, the analyst estimates hyperscalers could face data center capital expenditures that are "higher by 30% just to cover higher memory costs." Bernstein nevertheless expects AI investment to continue; however, the firm warns that "a re-calibration is inevitable," with cost burdens potentially redistributed along the supply chain and weaker suppliers likely to be squeezed.
On the earnings front, Bernsteins fiscal 2027 EPS projections now sit 25% to 38% above consensus for Samsung, SK Hynix and Micron, respectively. The broker attributes that gap mainly to the fact that broader sell-side models have not yet incorporated the anticipated HBM price increase. As the ongoing negotiations around 2027 HBM pricing reach resolution in the coming months, Bernstein expects consensus forecasts to move upward, which could provide near-term support to these stocks.
The firm also notes a potential secondary market effect. If hyperscalers seek to avoid the GPU-level markup, they could move to source HBM directly from memory suppliers. Bernstein highlights MediaTek as a potential beneficiary should that sourcing shift occur, because direct procurement could circumvent the added GPU pass-through costs.
Implications for markets and corporate buyers
- Memory producers with HBM exposure stand to see meaningful upward revisions to earnings forecasts if Bernsteins price assumptions materialize.
- Hyperscalers and data center operators may face materially higher capital spending requirements to support AI deployments if memory costs rise as modeled.
- GPU vendors that embed HBM in their products may need to reprice their offerings to preserve margins, which could shift cost burdens across the hardware and cloud ecosystem.
This article reports Bernsteins projections and the firms stated assumptions. It reflects the brokers published targets, modeling assumptions, and quoted commentary without addition or extrapolation beyond those items.