Ep. 007 - The 3 Choke Points Killing the AI Boom (Core Research) | Nick Doyle, Nigel Chiang, Konrad Wang, Jordan Nanos

Summary

• The AI infrastructure supply chain is “constrained at almost every layer,” and the market is trading it exactly that way — hyperscaler spenders are down big year-to-date while investors chase “whatever has the biggest shortage.” Konrad Wang’s three choke points: TSMC N3, “where everything starts… sold out through 2027”; InP-based continuous-wave lasers for co-packaged optics; and PCB/CCL substrates, with shortages running all the way down to the drill bits.
• PCB is a bottleneck because AI servers need 30-40+ layer boards versus sub-10 for consumer electronics, and only a handful of suppliers can make the high-grade materials. Ajinomoto is one major ABF film supplier, and perhaps one-to-three vendors (Doosan named) can produce the thick, high-voltage CCL — low yields, long lead times, and much higher premiums than these vendors previously commanded selling into commercial electronics. Taiwanese PCB names Unimicron and Kinsus are up over 100% YTD and “keep hitting the limits” of Taiwan’s 10% daily move cap; a literal drill-bit stock is up 92%.
• Nigel Chiang’s construction call: modular data centers are ~10-15% of builds today and there’s “no reason why that can’t go north of 50%,” because modular solves the scarcest input — field labor. He expects standardization to create single-vendor lock-in for a hyperscaler, with Comfort Systems the name he flags; and in gas power, labor is “probably the most binding” constraint — GE Vernova, Siemens Energy, and Mitsubishi have roughly 100GW of combined global capacity, but can’t divert slots to higher-priced US demand even if they wanted to because there is not enough labor to build the plants. Argan, which Nigel calls the only pure-play gas EPC, rose about 9% on earnings and 30%-plus the next day.
• On memory, Nick Doyle concedes the bear logic — OpenAI/Sam Altman letters of intent and LTAs locking in pricing have “typically been the sign of the top” — but argues this cycle is different because of prepayments. The team thinks prepayment terms will be priced “way higher than everyone expected,” changing memory players’ through-cycle financials, and “that’s not understood” — they are “incrementally even more positive” and would defend the pullbacks, including Micron’s sell-off on “an incredible print.”
• Top signals, per the desk: a rising frequency of Twitter victory laps from “the same retail names,” and unambiguously positive news getting sold — which Micron just demonstrated. The self-aware joke lands too: Jim Cramer calling SemiAnalysis “the gospel” and Dylan Patel asking how to short himself may itself be a sign. Konrad’s comfort check: as long as hyperscalers and neoclouds keep spending, and free-cash-flow/ROI analysis holds up, “diamond hands.” Wega’s aunt is reportedly getting “a lot of alpha” by watching Taiwanese Jim Cramers pitch stocks.
• AI adoption still looks early: Nigel says his former investment-banking colleagues “are not using AI at all,” while Jordan jokes that they may be buying “the Anthropic basket without using any AI” to analyze it. Jordan says the product is “crossing over 25 billion in ARR, like 4× in 3 months or something crazy”; SemiAnalysis estimates vary from roughly 70 people to Jordan’s later count of 79 in the general channel, and the firm spends $5M annualized on Claude with some $10,000 days. Nick’s pitch problem: even computer-savvy friends won’t install Claude Code — “I’m trying to sell you fire insurance to a burning house.”

Deep dive

1. Everything is short — N3, lasers, and substrates

• Nick Doyle’s explanation of Core Research: ten vertical teams each deliver “extremely detailed models,” and the Core Research desk “sits beneath all 10” distilling technical work into actionable insights for hedge funds, long-onlys, and now “even some quant guys.” The context: GPU rental prices are rising — Jordan’s image from last week, “trying to get the last flight out.”
• Konrad Wang’s map of the choke points: “TSMC N3 is where everything starts… it’s sold out through 2027,” with hyperscalers doing internal allocations to manage trade-offs; second, InP-based continuous-wave lasers for co-packaged optics — a major theme this year; third, PCB/CCL substrates, where shortages extend “even to the drill bits.”
• The tape confirms it: “people don’t like the spenders” — hyperscalers down big YTD — while the market buys “whatever has the biggest shortage.”

2. PCB: 40-layer boards and a handful of suppliers

• Konrad’s primer: AI server PCBs run “30, 40 plus layers” versus sub-10 for commercial electronics, requiring specialized fiberglass grades such as T-glass, low-Dk glass, and L-glass, with “much lower yield” and longer lead times, plus more precise — and more numerous — drills. The supplier base is tiny: Ajinomoto is one major ABF film provider, and only “maybe one or two, maybe three suppliers” (Doosan named) can make the thick, high-voltage CCL — hence much higher premiums than these vendors previously commanded when supplying commercial electronics.
• Nick’s framing of why this suddenly matters: “It just didn’t matter a couple years ago. It all has to do with the end of Moore’s Law” and the move to GPU-based compute demanding new materials.
• The price action, per Konrad: “this year is the year of networking. Last year was probably the year of ASIC accelerators” — Lumentum +100% YTD, AAOI +200%, Unimicron and Kinsus up over 100% and repeatedly pinned at Taiwan’s 10% daily limit, Elite Material (CCL) +70%, and Nigel’s favorite story — literal drill-bit companies, one up 92%.

3. Time-to-market: modular construction and the labor choke

• Nigel Chiang’s thesis: modular is ~10-15% of data centers today with “no reason why that can’t go north of 50%,” because it moves the scarcest, most expensive input — aggregated field labor — into a factory. He expects standardization to create lock-in: “what’s the point of having two guys doing it” once one design is set, making Comfort Systems his name to watch.
• Behind-the-meter power and modular are “different expressions of the same thing, which is time to market… The moment you bring GPUs online, revenue flows” — install plants on site rather than wait three years for grid interconnection.
• The unappreciated angle: labor is “probably the most binding in gas power plants.” GE Vernova, Siemens Energy, and Mitsubishi have roughly 100GW of combined global capacity, but “even if they wanted to, there’s not enough labor to build these power plants.” Argan, which Nigel calls the only pure-play gas EPC and which the team first highlighted in November, rose about 9% on earnings then 30%-plus the next day — “evidence of higher pricing power… coming through into their P&L.”

4. Memory: the LTA top signal versus the prepayment counter

• Nick grants the bear case its logic: OpenAI letters of intent and long-term agreements that lock pricing have “typically been the sign of the top” — once buyers say “just give us a price and we’ll lock it in for three years,” upside is capped and “the stock prices don’t work anymore.”
• His counter, hedged as team view: “we think that this time there’ll be a really big prepayment aspect that completely changes the financials of the memory players through cycle,” with prepayment pricing “way higher than everyone expected — and that’s not understood.” He’d defend the pullbacks, is “incrementally even more positive,” and notes that Ray and Mats That Matter were early in calling out that capacity is not arriving as quickly as people think — even ‘27-into-‘26 CapEx pull-ins don’t fix supply.
• On top signals generally, Nigel offers two: rising “victory-lapping” screenshots from “the same retail names” on Twitter, and stocks trading down on “unambiguously positive” news — which Nick says Micron just did, “selling off an incredible print.” Nick’s joke candidates: “We just saw Dylan go on CNBC, and we just saw TBPN get bought.” Plus Wega’s aunt in Taiwan getting “a lot of alpha” by watching Taiwanese Jim Cramers pitch stocks.

5. Claude Code: fire insurance for a burning house

• Nick’s adoption puzzle: even “very computer-savvy, computer-programming-type friends” won’t clear the install hurdle — “I’m trying to sell you fire insurance to a burning house.” His own conversion came only because “my boss is pushing me to do it”: after Doug told him that his plain-Claude IRR calculator was wrong and “not what I’m talking about,” Nick went back in, asked it to manipulate data and produce outputs, and his “jaw dropped” — a moment even bigger than ChatGPT in 2023.
• Jordan says the product is “crossing over 25 billion in ARR, like 4× in 3 months or something crazy,” alongside Super Bowl ads and New York Times coverage. Konrad’s version: GenAI “turned from reactive to proactive” — he’s building a personal RAG, running scraping bots, and “Claude for Excel is just so good.” Konrad estimated about 70 people at the firm; Jordan later counted 79 in the general channel. SemiAnalysis spends $5M annualized on Claude, with some $10,000 days.
• Nigel’s confirmation and kicker: his ex-IB colleagues “are not using AI at all” — Jordan’s gag is that they may be buying “the Anthropic basket without using any AI” — which Nigel reads bullishly as early innings. Using Claude for Excel, “if someone tells me that this is AGI… I wouldn’t have a strong pushback,” and this is on current-generation hardware, before the Blackwell- and Rubin-vintage models come in.
• Closing hedge from Konrad: expect volatility from macro and “the war”; the desk needs to tie supply-demand to CapEx sustainability via free-cash-flow ROI work. Until then: “Diamond hands.”