Bernstein: HBM Prices to Rise 2 to 2.5x Next Year, Cloud Providers' AI Capex May Increase 30%
Miles Bennett
Bernstein expects HBM prices to rise 2–2.5x next year; combined with surging conventional memory prices and GPU vendors' markup pass-through, hyperscaler AI capex may need to increase roughly 30% — a supply-chain-wide cost rebalancing is now unavoidable.
Why is HBM suddenly due for a massive price hike?
Conventional DRAM prices have surged roughly 4.5x since Q3 2025, while HBM prices — locked into annual contracts — barely moved.
This means → deploying a wafer to conventional DRAM now generates about 2x the revenue and nearly 3x the gross profit versus HBM. Suppliers have no incentive to keep allocating capacity to HBM.
Samsung confirmed the shift on its Q1 2026 earnings call: non-HBM DRAM margins already exceed HBM, and the company signaled the gap will "narrow significantly" in 2027.
In plain terms = memory makers are voting with their fabs — capacity flows to whichever product pays more, and HBM can't compete without a reprice.
Why 2–2.5x and not higher?
To match conventional DRAM on a per-wafer-revenue basis, HBM would theoretically need a 3x increase.
Bernstein's 2–2.5x forecast sits below that level. This reflects suppliers' concern that aggressive pricing could damage overall AI ecosystem demand.
This means → even after the hike, HBM profitability will still trail conventional DRAM next year — the gap just narrows meaningfully. Suppliers are balancing margin recovery against the risk of scaring off customers.
Why does the cloud bill jump far more than HBM alone?
HBM accounts for only about 5% of an Nvidia VR200 rack's selling price. A 2–2.5x HBM hike directly lifts rack cost by roughly 6%.
But Nvidia, protecting its 75% gross margin, amplifies the HBM cost increase by roughly 4x when passing it through — potentially pushing rack prices up about 24%.
In plain terms = for every dollar HBM goes up, Nvidia adds four dollars downstream. The GPU vendor's "markup multiplier" is the real cost amplifier.
Adding HBM pass-through (~4%), Nvidia's full markup (~15%), and conventional DRAM/NAND increases (~14%), hyperscaler AI capex needs to rise roughly 30% to absorb the memory cost surge.
Who stands to gain from this repricing?
Bernstein sharply raised FY2027 EPS forecasts for the big three memory makers, running 25–40% above consensus: Samsung +~26%, SK Hynix +~32%, Micron +~38%.
Target prices were lifted across the board: Samsung to KRW 440,000 (~26% upside), SK Hynix to KRW 3,300,000 (~20%), Micron to $1,300 (~15%) — all rated outperform.
This means → Bernstein believes the market has not yet priced in HBM repricing upside. As price negotiations conclude over the coming months, consensus estimates will be forced higher.
Any surprise winners — or losers?
If hyperscalers bypass GPU vendors' markups by purchasing HBM directly, Asian ASIC — custom chip — service providers are well positioned to facilitate that model.
Bernstein names MediaTek as a potential beneficiary; the stock has already rallied roughly 130% over the past two months. Rating: outperform, price target NT$4,380.
On the losing side, pure NAND supplier Kioxia has no HBM exposure and cannot share in the repricing windfall. Bernstein maintains an underperform rating with a target of ¥40,000.
What is the single biggest test for this thesis?
Bernstein switched all three memory makers' valuation framework from price-to-book to price-to-earnings — arguing that margins and ROE are heading to historically unprecedented levels, making historical P/B benchmarks meaningless.
This means → if HBM repricing fails to materialize on schedule, both the earnings forecasts and the new valuation framework face simultaneous downside risk.
In plain terms = every bet here rests on one premise: that HBM prices actually rise. The outcome of price negotiations is the single biggest validation point for the entire thesis.
Content is for reference only, not financial advice.