AI hardware news today: HBM demand is driven by stacks-per-device, not unit count — and the math reshapes how you read Broadcom, Marvell, and memory supplier positioning. Plus, documented export control leakage, nuclear power as AI infrastructure, and TSMC's Anthropic stake.
Audio is available on Spreaker — see link below.
Doubling your ASIC shipments doesn't double your HBM demand. That gap between what people assume and what the math actually shows is the most important thing happening in AI memory markets right now.
The custom silicon story has moved well past niche. Broadcom posted AI revenue of sixteen-point-seven billion dollars in Q3, up two hundred and twenty-one percent year-over-year, and guided Q4 to twenty-one-point-seven billion, up two hundred and thirty-six percent.
On the policy side, the picture has shifted from intent to documented leakage. House records show eight hundred and seventy-nine thousand Nvidia and AMD chips were approved for export to China between December twenty twenty-five and May twenty twenty-six.
Trump has pledged what he called tremendous penalties on US companies building data centers overseas. The enforcement mechanism remains unspecified.
A few infrastructure signals worth tracking. Constellation Energy posted Q3 earnings per share of five-point-oh-six dollars against an estimate of three-point-eight-eight, raised its twenty twenty-six guidance a second time, and locked a twenty-year, six hundred and ninety megawatt nuclear supply deal with Amazon.
The two metrics that matter most heading into the next few weeks: Broadcom's December ninth guidance confirmation, and whether the stacks-per-device trend on custom ASICs actually reaches six or stays at three to four. That second number is what determines whether the HBM memory TAM expansion story holds.
Chapter summary auto-generated from the verified script. Listen to the full episode for the complete content.