Qualcomm - Broadcom

Episode 48 covers Broadcom's $103 billion bid for Qualcomm (about $130 billion including assumed debt) — potentially the largest technology acquisition of all time, and at the time of recording still a rejected offer. The hosts joke that both companies' names sound like modem standards ("the battle of the comms... 1x versus 2x... reminds me of 3Com"), and that one of the episode's themes is "what's old is new again — we're back in 2001." Notably, neither acquirer nor target is in Silicon Valley: Broadcom is Singapore-based (pending its relocation to Delaware) and Qualcomm is in San Diego — ironic, since semiconductors are the "silicon" of Silicon Valley.

A quick primer: both are fabless semiconductor companies — they design chips but don't own fabs, the fabrication facilities that manufacture them. With the exceptions of Intel, Samsung, and TSMC, most chip designers no longer manufacture, because fabs are so expensive to build. Both companies design chips and work with contract manufacturing partners, making everything from processors to wireless radios — component makers for phones, computers, servers, cars, toasters, everything ("IoT explosion"). The hosts also plug their Slack community (over a thousand strong) and ask for Apple Podcasts reviews.

David's stage-setting: a semiconductor renaissance is underway, sparked a couple of years earlier by the machine-learning boom. NVIDIA — long just a maker of graphics cards for PCs and consoles — has been on a crazy tear as GPUs became central to machine learning: from $20 a share in 2015 (or $26 in January 2016, per the hosts' independently-run math) to $212 — a 10x-plus in two years, in the public markets. It turns out graphics cards are good for more than graphics: the linear algebra and matrix transforms behind rendering happen to be the core of machine learning. Nervana, a startup making GPU/ML-dedicated chips, was acquired by Intel for $350 million — in many ways the start of a renaissance of actual semiconductor startups, which hadn't happened in a long time. Just that week, Sequoia led a $50 million round in GraphCore (UK), which had already raised $60 million — $110 million raised in a bit over a year to compete in a now-crowded market of startups making specialized machine-learning chips to challenge NVIDIA.

What enabled that innovation? Beyond market demand, the consolidation of fabs made sense: as transistors-per-wafer grew and process nodes shrank, manufacturing became more and more expensive — a fab costs $10 billion plus — so only a few companies (TSMC, Samsung, Intel) still manufacture. Suddenly there's a platform where you don't need the capital to manufacture: you can start a smaller company that just does designs, licensing them or working with the manufacturers. Ben likens it to Jeff Bezos's point that Amazon wouldn't be possible without the electric grid, UPS, the internet, and the phone companies — platforms built by predecessors that you can innovate on top of.