Samsung Electronics reported quarterly operating profit of KRW 107.4 trillion on 29 October 2026, on revenue of KRW 195.3 trillion, according to Samsung's third-quarter results release. On the earnings call, Samsung said its 2027 supply of high-bandwidth memory is already sold out. If your roadmap includes AI hardware, on premises or reserved in a cloud, plan the delivery date around memory supply, not around which model you pick. Order early and keep a fallback.
What is HBM4?
HBM4 is the sixth generation of high-bandwidth memory, the stacked memory that sits next to an AI accelerator and feeds it data fast enough to be useful. Modern AI chips are often limited less by how fast they compute than by how fast they can read their weights. HBM is the part that decides that. No HBM, no accelerator.
What did Samsung report?
The headline figures come from the results release.
- Revenue. KRW 195.3 trillion for the quarter, up 14% on the previous quarter.
- Operating profit. KRW 107.4 trillion.
- Chips. The Device Solutions division, which includes memory, made KRW 106.5 trillion of that operating profit on KRW 147.1 trillion of revenue.
- Investment. Capital spending was KRW 29 trillion in the quarter. Samsung expects about KRW 84.4 trillion for the full year.
In other words, almost all of Samsung's profit this quarter came from chips. The release says the memory business set all-time records for quarterly revenue and operating profit, and that Samsung "significantly scaled up sales of HBM4."
What did Samsung say about HBM on the call?
More than the release did. According to the Investing.com transcript of the earnings call, Jaejune Kim, executive vice president of the memory business, said three things.
- HBM4 sales "increased by more than three-fold QOQ in the third quarter."
- "HBM4 mix increased to 60% of total HBM sales."
- "Our 2027 HBM supply volume is already sold out at an agreed target price."
The release itself is broader but points the same way. It says that in 2027, "increasing AI-driven demand will far exceed the industry supply in comparison with 2026."
Read that last sentence again. It is a major memory maker telling its investors that next year, demand for memory will be well ahead of what the industry can make. That is a supply warning dressed as good news.
The capital spending tells the same story from the other side. Samsung is spending heavily to make more. New memory capacity takes time to build and qualify, so money spent this year helps supply later, not now. For anyone buying in 2027, the warning matters more than the investment.
Why does memory decide delivery dates?
Because it is the part you cannot swap.
When a team plans an AI system, most of the debate is about models. Which one, which size, open or hosted. Those choices matter for quality and cost. They rarely decide when the system goes live. Models can be changed in a week. A rack of accelerators cannot be built without the memory that goes on them, and that memory is booked a year ahead by the largest buyers in the world.
I will say it plainly. Memory supply, not model choice, is what slips a hardware delivery date. If your project needs dedicated AI hardware in 2027, you are competing for the same HBM as the biggest cloud providers and chip designers, and they tend to order first.
This reaches cloud users too. If you rely on reserved GPU capacity from a cloud provider, that provider is buying from the same constrained supply. Reserved capacity can be delayed, rationed or priced up when supply is tight.
What should a team that ships AI products do?
Treat hardware like any other long-lead dependency, not like a purchase you make when the software is ready.
If you plan to buy hardware
- Order when you decide, not when you are ready. Lead times are set by memory, and memory is sold a year out.
- Get delivery dates in writing. A quote without a date is not a plan.
- Design for a smaller first batch. If half the hardware arrives, can you go live with half the load?
If you rely on the cloud
- Reserve capacity early for anything with a fixed launch date.
- Keep a second region or provider tested and ready, not just named in a slide.
- Keep a hosted API fallback for the parts of your product that can run on someone else's model while your own capacity catches up.
Either way
- Make models swappable. If you cannot get the hardware for the large model, a smaller one on what you have is better than nothing on time.
- Put hardware on the risk register. Give it an owner and a review date, like any supplier.
Is this only a problem for big companies?
No. Smaller teams feel it later, but they feel it. The big buyers take the first allocation. Everyone else gets what is left, at whatever price the market sets that month. A client who wants an on-premises AI system for privacy reasons may be told by their hardware supplier that delivery is months away. That is a project risk on day one, and it should be in the plan on day one.
What would I do this week?
If any project on your roadmap needs dedicated AI hardware in 2027, ask your supplier for a delivery date today, in writing. If it runs on reserved cloud capacity, check when the reservation renews and what happens if it is not available. Then make sure your software can run, in a reduced form, on less.
Where Alongside fits
Planning a delivery date around hardware lead times, not model choice, is the kind of work covered by Alongside's technical advisory.