Now Playing
Google Rations Compute, Intel's 90% Yield & Korea's $518B Fab Risk
Audio is available on Spreaker — see link below.
Google couldn't deliver Gemini compute capacity to Meta despite a $460B backlog — proof that power grids, not chips, are now the binding constraint in AI infrastructure. Plus Intel's shock 18A-P yield, South Korea's fab cluster risks, and what TSMC's Q2 earnings will reveal.
Audio is available on Spreaker — see link below.
Google told Meta it couldn't deliver the Gemini compute capacity Meta had requested. That happened in March, despite a four hundred sixty billion dollar backlog sitting on Google's books.
Here's what that means in practice. You can have a four hundred sixty billion dollar order book, you can have leading-edge silicon in production, and you still can't serve your largest customers if the grid behind your data centres can't carry the load.
Shift to Intel. On June sixteenth, Intel's advanced eighteen A-P packaging node entered risk production with ninety percent first-month yields.
In South Korea, Samsung and SK Hynix are each building two new fabs in the southwest of the country, part of an eight hundred trillion won national semiconductor ecosystem project. At current exchange rates, that's roughly five hundred eighteen billion dollars in committed infrastructure.
Advanced packaging is becoming its own independent capex story. Samsung Electro-Mechanics is committing fifteen trillion won, roughly eleven point four billion dollars, to FC-BGA and MLCC expansion across Busan and Sejong.
Two things to watch in the near term. TSMC reports second quarter earnings on July sixteenth.
The through-line across all of this is the same. The infrastructure layer of AI is hitting physical limits that capital alone can't immediately fix.
Chapter summary auto-generated from the verified script. Listen to the full episode for the complete content.