Podcast
Semi Doped
Semiconductor and AI infrastructure conversations covering chips, manufacturing, supply chains, and technical strategy.
Original source
NEWS TAKE: China's Optical Ban, Volta's $10B Anthropic Deal, AMD's Earnings
The episode covers a proposed U.S. ban on Chinese optical transceivers, arguing the security case is technically weak and the supply-chain impact could be severe. It also dissects AMD’s earnings and capex surprise, then explains why Volta’s $10B Anthropic compute deal is really a financing innovation for AI infrastructure.
GlobalFoundries Thomas Barber: CPO, Silicon Photonics, 300mm, SiGe, OCI, NRZ
GlobalFoundries’ Thomas Barber argues that silicon photonics is becoming the practical answer to copper’s shrinking reach in AI data centers, especially for scale-up networks inside and across racks. He explains how GF combines 300mm silicon photonics, specialty silicon germanium, and the OCI optics ecosystem to push lower-power, higher-density interconnects.
How Retimers Built an $80B Company: The Story of Astera Labs
Astera Labs built a huge business by solving a mundane but critical problem: keeping high-speed copper links alive inside AI servers with retimers. The episode explains the engineering behind PCIe signal conditioning, then follows Astera from Hopper/H100 wins into switches, cables, and CXL.
News Take: Hyperscaler CDS, SK Hynix Earnings, China's DUV
The episode focuses on three selloff drivers: rising anxiety around hyperscaler AI financing, a sharp but arguably overdone SK Hynix selloff after an earnings miss, and China’s reported progress in immersion DUV lithography. The hosts argue that fundamentals in AI compute and memory remain strong, while market volatility is being amplified by leverage, circular financing fears, and perfection-level expectations.
Datacenter Interconnects: Copper vs. Optics, Nvidia's 78-Layer PCB, Co-Packaged Optics (CPO)
The episode argues that AI datacenters are increasingly limited by interconnects, not compute, and walks through the tradeoffs between copper and optics across scale-up, scale-out, and scale-across networking. It also explains why Nvidia is pushing copper extremely far with a 78-layer PCB and why co-packaged optics remains a promising but hard-to-service “holy grail.”
PicoJool's Al Yuen: The Case for GaAs VCSELs in Scale-Up Interconnects
Al Yuen argues PicoJool can scale AI interconnects with GaAs VCSELs because the supply chain is already built for high-volume data-center production, unlike constrained single-mode/InP alternatives. The episode focuses on error-rate requirements, 200G VCSEL product options, and a roadmap from 1.6T toward 3.2T and 12.8T without requiring new manufacturing breakthroughs.
WEKA's Val Bercovici: KV Cache, DeepSeek V4, HBF, SLC vs QLC NAND, CXL, NVLink, Tokenomics
Val Bercovici argues AI inference is becoming a memory-and-token-economics problem, with KV cache compression, network-attached memory, and NAND tiering reshaping system design. He also makes a blunt case that CXL is losing relevance, HBF will likely need SLC-like endurance buffering, and SaaS economics are being upended by token OPEX.
Micron's Record Profits, Apple's CXMT Plea: AI is Eating All the Memory
Micron’s memory business is in a record-profit supercycle as AI datacenters absorb nearly all available supply, pushing up DRAM/NAND prices across consumer electronics too. The hosts argue this creates a bifurcated market: inelastic AI demand keeps pricing hot, while consumer devices face shrinkflation, repricing, and even controversial sourcing moves like Apple reportedly eyeing CXMT.
Qualcomm's HBC Memory, Alphawave, Modular, and more
Qualcomm’s investor day was presented as a pivot from handset-first communications to a broader business mix led by automotive, IoT, and data center. The hosts focused on Qualcomm’s HBC memory idea, its data-center CPU and networking roadmap, and the long-term edge/robotics opportunity.
Advanced Packaging, TSMC CoWoS, Intel EMIB
The episode argues that advanced packaging has become inseparable from AI chip design, with reticle limits, HBM density, and multi-die stitching now driving architecture. The hosts compare TSMC’s CoWoS variants with Intel’s EMIB and discuss how Google TPU demand could shape the next wave of packaging capacity.
Computex Mania 2026: Optics and Power
Austin and Vik recap their first in-person meeting at Computex, focusing on how the show centered on AI hardware, optics, and data-center power. The episode digs into Marvell’s interconnect strategy, CPO/XPO tradeoffs, micro-LEDs, 800 V rack power, immersion cooling, and Intel’s advanced packaging-led CPU roadmap.