Close friends and family who have had the misfortune of asking what I’m reading know that I organize my leisure reading into quadrimesters. Every four months I pick a subject, figure out what I want to understand by the end of it, and then build a reading roadmap around that question.
Natalia thinks this injects an unnecessary amount of structure and work into something that is supposed to be relaxing. This criticism comes from someone who regularly reads three books at once and abandons two somewhere around chapter six, so yeah, I respectfully reject the premise.
Anyway, here is my spring/summer-ish quadrimester.
My latest four-month rabbit hole has been AI infrastructure. I spend enough time thinking about models and applications, and frankly the pace of change there over the last year has been exhausting. I wanted to move down the stack and understand the physical stuff underneath all that AI voodoo. Semiconductors, memory, accelerators, data centers, the companies that make everything, where the bottlenecks sit, and which countries control the pieces everyone else depends on.
Mostly, I wanted to understand what people actually mean when they casually say things like “we have a compute constraint” or “HBM is the bottleneck.” Cool. But why? What physically has to be manufactured? By whom? Why can’t we just make more of it? And how does some obscure piece of hardware suddenly become a geopolitical issue?
I started broad with Anu Bradford’s Digital Empires and Chris Miller’s Chip War. Bradford gave me the policy frame around the U.S., China, and Europe’s very different approaches to the digital economy. Miller then pulled me down into semiconductors, their history, and the strange global supply chain that developed around them. Chip War is probably still the book I’d recommend to anyone who wants to understand why tiny pieces of silicon now occupy so much space in national-security conversations.
From there I went backward and read Christophe Lécuyer and David Brock’s Makers of the Microchip, which tells the early Fairchild Semiconductor story using old company documents, technical notes, sketches, and other material from the people actually building this stuff. It gets nerdy in places, but I loved that. You get this almost fly-on-the-wall view of people figuring out an industry that barely existed yet.
Then I stopped reading books for about a month and became mildly obsessed with memory.
GPUs get all the attention, but without memory feeding them data, those very expensive processors are basically sitting around waiting. So I worked through SRAM, DRAM, HBM, memory bandwidth, how memory sits alongside GPUs, and why it matters for training and inference. That part of the journey was mostly research papers, SemiAnalysis, RAND and CSIS work, material from Micron, SK Hynix, and Samsung, YouTube, podcasts, and whatever my AI assistant sends me by 6 a.m. on the previous day’s developments in AI infrastructure.
I wasn’t starting from scratch on chips and memory, but my mental model was still mostly pre-AI-accelerator. This was really about updating that map for the AI era and seeing how the whole stack now fits together. From there, I moved into the companies occupying the chokepoints.
Marc Hijink’s Focus – The ASML Way took me inside the Dutch company making the ridiculously complicated lithography machines needed to manufacture advanced chips. Honghong Tinn’s Island Tinkerers broadened the lens to Taiwan and the ecosystem that eventually produced TSMC. And I finished with Tae Kim’s The Nvidia Way, which traces how Jensen Huang and company went from a graphics-chip maker to arguably the company sitting closest to the center of this entire AI moment.
Four months later, I have a much better sense of how the old semiconductor world became the AI stack we have now, how something like DRAM went from feeling like a commodity to becoming part of one of the most strategic bottlenecks in AI, and how all those layers now depend on one another in ways I hadn’t fully appreciated before. ASML relies on specialized suppliers. TSMC relies on ASML. Nvidia relies on TSMC, memory makers, and a manufacturing ecosystem most people never see. Data centers depend on all of them, along with the networking, cooling, storage, and physical infrastructure needed to keep the whole system running. One bottleneck somewhere near the bottom can ripple all the way up to the chatbot sitting on your phone.
That was really what I wanted to understand.
If you only want three recommendations, I’d start with Chip War for the big picture, RAND’s High Bandwidth Memory — What It Is and Why It Matters for why an obscure type of memory suddenly matters geopolitically, and The Nvidia Way for how one very unusual company ended up at the center of this whole moment.
I mean, at this point the only thing missing from this capstone is an invite to dork out at a TSMC fab, Nvidia’s campus, or ASML’s facility.
Just saying…
Roll credits.