#BernsteinSaysMemoryBullMarketToLastUntil2027

Bernstein Calls for Memory Bull Run Into 2027 as HBM Scarcity Rewrites the Cycle

Bernstein just threw out the old playbook for memory. Analysts Gautam Chhugani and Mahika Sapra say the DRAM and HBM upcycle won’t peak until 2027, breaking the four-year boom-bust pattern that defined chips for decades. The reason is simple: AI demand is growing into a supply wall, and long-term contracts are locking in prices years out.

The numbers behind the call
HBM is sold out through 2026. Bernstein expects 2 to 2.5 times year-over-year gains in HBM average selling prices into 2027, covering HBM3, HBM3E, and the jump to HBM4. Jefferies models 40 to 50 percent quarter-over-quarter increases in Q3 2026 alone, with peak ASP hitting $2.23 per GB by 2027. That’s not a spike. It’s a structural repricing.

Long-term agreements change the game. About 35 percent of total bits for FY27-29 are already under LTAs at $0.26 per GB, only 10 percent below projected spot. SK Hynix removed price caps on 2027 contracts, shifting leverage to suppliers. Hyperscalers are paying to guarantee supply because training clusters are doubling every 6 to 9 months, and missing HBM means idle GPUs.

Who benefits
SK Hynix owns 62 percent HBM share and is the sole qualified HBM3E supplier for Nvidia’s Vera Rubin platform. Revenue went from an annualized $9 billion in late 2025 to $30 billion by March 2026. Micron is sold out of HBM3E through 2026 and just guided margins higher on a $200 billion fab build. Samsung still leads commodity DRAM and is qualifying HBM3E, so it catches the upswing when spot flips.

Equipment and power feel it too. ASML’s EUV orders stretch to 2027 because each HBM stack needs more layers. Infineon and other power names see demand as AI racks pull higher watts per GPU.

Why this cycle is different
Fabs take years. Micron’s build, SK Hynix’s Yongin cluster, and Samsung’s Pyeongtaek expansion won’t add meaningful bits until late 2027. In the past, oversupply killed pricing within 18 months. Today, contracts set a floor. Bernstein’s point: even at the next peak, the floor is higher than any prior top. The downcycle, if it comes, starts from $0.26 per GB, not $0.08.

Risks on the board
China’s domestic HBM push, export shifts, or a sudden AI capex pause could cool orders. A macro slowdown would hit consumer DRAM first. But AI training and inference are now line items in every hyperscaler budget. Cutting memory means cutting capability, and that’s a trade few will make.

Market read
Memory stocks already moved. SK Hynix is up eightfold in 2026. Micron hit new highs. ETF flows are broadening from Bitcoin into semis and infrastructure. Bernstein sees trading volumes peaking in 2027, with earnings growth through 2026 and a re-rate as investors treat memory as secular AI infrastructure, not a cyclical trade.

For Gate traders watching tech rotations, the signal is clear. Chips were the first AI bet. Power was second. Memory is the third leg, and Bernstein says it runs until 2027. The old cycle is dead. The new one is contracted, priced, and booked.