Now Playing
TSMC COUPE Goes Commercial, Cloud GPU Controls & Inference Economics Shift
Audio is available on Spreaker — see link below.
TSMC's co-packaged optics platform moves from roadmap to production, reshaping AI data centre interconnects — while new export rules target cloud GPU access and custom inference silicon challenges Nvidia's dominance. Six stories every investor and engineer needs today.
Audio is available on Spreaker — see link below.
TSMC has moved co-packaged optics off the roadmap slide and into commercial production. That's the signal worth leading with this morning.
The practical implications for interconnect margins are real, and the risks are equally real. TSMC has identified three manufacturing constraints: fiber-array supply, wafer-level testing yield, and packaging assembly.
The interconnect story connects directly to memory. At the same conference, the Memory Executive Summit reframed DRAM and HBM as architecture decisions, not commodity purchases.
Elsewhere, the Commerce Department is drafting a cloud-access rule expected to circulate by September. The target is remote access to Nvidia GPUs in third-country data centers, specifically Thailand and Singapore.
On physical enforcement, Taiwan prosecutors have indicted nine people in a B three hundred server diversion scheme routed through Indonesia and Hong Kong. Seventy-four of one hundred thirty servers reached China.
The broader theme threading through this week is inference economics. Hot Chips revealed that OpenAI's Jalapeño chip outperformed H one hundred, H two hundred, and GB three hundred on cost per token for inference in independent testing.
The signals worth tracking from here are narrow. On CPO: fiber-array supply ramp and wafer-level yield numbers from ASE and TSMC's packaging partners.
Chapter summary auto-generated from the verified script. Listen to the full episode for the complete content.