Samsung projects a 100 trillion won quarter on AI memory demand
The AI buildout's money keeps landing in the same place — memory — and Samsung just put the biggest number yet on it. Meanwhile, China's leading open-model lab walked through what its next models still can't do.
Samsung projected third-quarter operating profit of 107.4 trillion won — roughly $80 billion — which would be the first time any company has cleared 100 trillion won in a single quarter. The preliminary guidance, released Thursday in Seoul, compares with 12.17 trillion won a year ago, a jump of 782%, and comes with revenue of about 195 trillion won, up 127% — a fourth consecutive record quarter that also nudged past the consensus estimate of 106.1 trillion won. The driver is the one readers of this publication will recognize: AI servers are hoovering up DRAM and NAND, prices are climbing into a shortage analysts expect to persist into 2028, and HBM bit shipments are estimated to have risen about 50% from the previous quarter. The market's reaction was a 0.7% dip in the shares, because guidance that merely beats estimates is no longer enough when the whisper numbers are this high. This is the cleanest hard number yet on where AI capital expenditure actually converts into profit — the model labs argue over benchmarks while the memory makers bank the checks — and we tracked the squeeze's mechanics earlier this month in AI demand sold out all 2027 DRAM and HBM capacity. One caution: this is guidance, not an audited result, and the full report on October 29 will show how much of the windfall is actually memory.
Zhipu founder Tang Jie used a talk at the National University of Singapore to lay out where GLM goes next — and, more usefully, to name the two things his lab still cannot do. He described three scaling fronts — pre-training, post-training (with GLM-5.3, he said, proceeding on that path), and inference-and-agent scaling — and identified the open gaps as recursive self-improvement and robotics. Chinese media ran with a firmer claim: trillion-plus parameters for GLM-5.4/5.5 and "two new paths." That specification is extrapolation from the talk rather than an announcement, and the "fully self-training" framing traces back to Zhipu's August 31 earnings call, where it pointed at GLM-6.0 — a claim we covered when a filing gave it away early, Zhipu spoils GLM-6.0: a fully self-training model, leaked by a filing. The genuine signal in the talk is the admission: after years of chasing parameter counts, one of China's top labs is publicly saying its next leap is a capability problem, not a scale problem.
What to watch: Samsung's October 29 full results — specifically the memory segment's share of that record quarter — and whether Zhipu turns any of Tang Jie's roadmap into an actual release.
Whose earnings do you actually want to see next to judge the AI buildout — the chipmakers or the model labs? Tell us in the comments.
