SK Hynix: Customer HBM Demand Doubles Available Supply, Revenue and Profits Expected to Surge by 2027
nashnova research
SK Hynix disclosed that customer demand for HBM exceeds its supply capacity by more than two times, and has already locked in 2027 supply volumes and pricing — signaling that the HBM seller's market will hold, with margin expansion gaining contractual backing.
Demand at two times supply — what does that number actually mean?
On its Q3 earnings call, SK Hynix said customer HBM demand exceeds supply by more than 2×.
Management clarified: this is customer demand intent, not confirmed orders, and is not broken down by customer or HBM generation.
This means → even with a discount for soft demand, the supply-demand gap remains massive. HBM — high-bandwidth memory, ultra-fast memory built specifically for AI chips — is still in severe undersupply.
Why lock in 2027 pricing this early?
The company has completed 2027 HBM supply-volume and pricing negotiations with key customers. Pricing reflects the higher value of HBM4.
Management expects HBM revenue and profit to grow significantly in 2027, driven by a rising HBM4 mix, price increases, and manufacturing efficiency gains.
In plain terms = price is agreed, volume is locked. The 2027 profit story is no longer just a "forecast" — it is a high-certainty call backed by contracts.
This reflects a profitability trajectory that will far exceed the peak of any prior memory cycle.
Fewer DRAM layers per stack — is demand weakening?
The opposite. Some customers are voluntarily cutting HBM stacks from 12 DRAM layers to 8, so they can build more AI accelerators from the same DRAM output.
In plain terms = DRAM capacity is fixed. Fewer layers per stack → fewer chips per HBM unit → more HBM packages from the same output → more AI accelerators. The bottleneck is supply, not demand.
Using HBM4E (8-layer) as an example, management noted that higher per-die DRAM capacity offsets most of the capacity loss from fewer layers. The medium-to-long-term HBM demand outlook is unchanged.
Fewer layers — how does the cost structure shift?
Every HBM stack, regardless of layer count, requires one base die — the chip at the bottom that communicates with the processor. Fewer layers → more HBM packages → base-die demand and back-end processing volume actually rise.
This means → 8-layer products carry a higher base-die cost per bit than 12-layer products. Cost per GB may increase.
Additionally, 8-layer and 12-layer products require separate qualification. Nvidia's Rubin Ultra accelerator has not publicly confirmed its final memory configuration.
Where does HBM4 mass production stand?
HBM4 is expected to surpass HBM3E in shipment volume and revenue by Q4 2026. The M15X fab is on track to reach full capacity by year-end.
Manufacturing yield has exceeded internal targets; the company has raised its yield target. Both front-end and back-end yields remain stable.
Next-generation HBM4E samples were sent to customers in Q2 2026. Development and qualification are proceeding on customer timelines. SK Hynix is also co-developing custom HBM products with key customers.
What should investors watch next?
A 2×-plus supply-demand gap + locked-in 2027 pricing → the HBM seller's market is unlikely to reverse near-term.
This means → two milestones will determine whether 2027 margin expansion materializes: sustained HBM4 yield performance and M15X capacity ramp pace.
In plain terms = demand is not the question, and pricing is settled. Everything now hinges on whether SK Hynix can execute the production plan on schedule.
市场有风险，内容仅供研究参考，不构成投资建议。