Jon (Asianometry) — Google Exodus, Elon Terafab & Memory Investment Thesis

Source: Ep. 023 — Everyone Leaves Google, Elon Forecasts 1T ARR, Reflecting On GPT-5 | Jon from Asianometry, SemiAnalysis Weekly (Jordan Nanos, Doug O'Laughlin, Jon from Asianometry), August 7, 2026.

The Framework: Bureaucracy Loses Frontier; Atoms Bind Abundance

The panel's organizing lens: frontier AI talent is fleeing bureaucratic platforms for single-mission startups, while physical bottlenecks (memory, optics supply chains) decide who can monetize the next trillion of compute. Elon's Terafab / $1T ARR pull-forward is the abundance case on chips and power — memory is the choke that turns abundance into pricing power.

Layer	Panel View	Investable Read
Talent / labs	Google legends leave → raise + buy GPUs; OpenAI/Anthropic still "one job" startups	GPU demand from spinouts; frontier risk at GOOG
Foundry / systems	SpaceX picks Nvidia because "best"; TPU software support outside Google ≈ zero	NVDA over merchant TPU for leavers
Memory	Taiwan product lines crimped; Apple N2 dies idle without DRAM	MU / Hynix / Samsung; Elon memory option
Optics / trade	China leads 1.6T transceivers (InnoLight); US ban hurts Western buildout	Domestic optics rebuild vs China ban pain
CapEx abundance	Elon 1→10→20 GW; hyperscalers drive FCF toward zero	Infra winners if power/memory clear

Investment Thesis #1: Memory Is the Binding Constraint — Even Apple Gets Mogged

Argument Jon (from Taiwan) says memory shortage is crimping every product they try to make. Apple has N2 chips sitting in a locker without memory to package into phones (via Tim Copan). Apple tried to negotiate volume discounts with CXMT and was refused — CXMT would not price below the going rate because "everyone wants my memory." On the SpaceX earnings call, Elon talks memory constantly; the panel's take is the fastest path to Terafab revenue is to make memory — all stacks are open.

"The lack of memory is crimping every product out here in Taiwan that they're trying to make and they can't get enough of it."

Contrarian element Consensus still frames Elon's $1T as a chip/foundry story. The panel says the scarce, high-margin unlock is memory capacity, not another GPU brochure.

Trigger Apple / OEM commentary on memory-limited device shipments; CXMT / HBM-DRAM pricing staying elevated; any SpaceX/Terafab memory capacity announcement.

Names Micron (MU), SK Hynix, Samsung Electronics — scarcity pricing; CXMT as the Chinese supplier that denied Apple; SpaceX / Terafab as a potential new memory entrant.

Investment Thesis #2: Elon's Abundance Case — $1T Pulled to 2030, Terafab Must Print

Argument Elon broke ground on Terafab and, on the first combined SpaceX earnings call, moved the forecast from 1 GW → 10 GW by end-2027, then 20 GW after — pulling the $1 trillion revenue forecast from 2031 to 2030. Doug's bar: anything under $100B revenue at end of Q4 2027 is a miss for Terafab alone (exit-rate framing; Tesla is a "drop in the bucket" if that ramp is required). Hyperscalers (Microsoft, Amazon, Google, Meta) are simultaneously driving free cash flow toward zero buying chips.

"Anything under a hundred billion at the end of Q4 27 is a miss."

Contrarian element Jon still "doesn't believe" the hyperscaler CapEx wave even as it happens — Taiwan doesn't build DCs (power goes to fabs). The investable implication is not skepticism of demand; it is that whoever clears memory and power captures the abundance.

Trigger SpaceX/Terafab GW and revenue milestones vs the 10 GW / $100B bars; hyperscaler FCF and CapEx guides staying elevated.

Names SpaceX (PRIVATE) / Terafab path; NVIDIA (NVDA) as the GPU Elon publicly chooses; memory stack as the bottleneck on the path to $1T.

Investment Thesis #3: Leavers Buy Nvidia — TPU Software Support Outside Google Is Zero

Argument Elon's tweet: SpaceX installs Nvidia GPUs because they are the best — "pretty matter-of-fact." Parallel rumor mill: Jeff Dean left partly to use GPUs; Dave Silver leaves, raises a large seed, and spends it on Vera Rubin chips (with a Google Cloud promo nonetheless). TPUs are "great for pre-trains," but support outside Google on the software side is zero — so spinouts default to the CUDA/Nvidia stack.

"I think that the support you can get outside of Google on the software side is zero."

Contrarian element Merchant-TPU bulls assume leavers will scale on Google silicon. The panel's lived read: talent exits increase GPU/Vera Rubin demand, not TPU share.

Trigger Spinout financing → GPU/Rubin purchase announcements; external TPU software tooling remaining immature; SpaceX/Nvidia deployment updates.

Names NVIDIA (NVDA); Vera Rubin generation as the spend target for ex-Google labs.

Investment Thesis #4: Google's Dream Team Breakup — Bell Labs Pattern (Frontier Risk, Not Near-Term Stock Death)

Argument Jeff Dean (employee 30; MapReduce, Spanner, Bigtable, TensorFlow, TPUs), Sanjay, Oriol, Quoc Le — "ostensibly the leads of Gemini" — leave after prior exits (John Jumper, Noam Shazeer, Dave Silver). Pattern: leave → raise (often including Google Ventures) → buy GPUs. Doug's Bell Labs / Kodak / IBM framing: invent the transformer, then get "giga screwed" on the next branches while staying a massive profitable company ($4.3T market cap cited). Short and medium term, Doug says the stock can still do well (search monopoly + TPU infra "good enough"); the risk is quietly bowing out of successive frontier generations.

"Bell Labs invented the transistor and gave it to everyone... It makes so much sense that Google is going to invent the transformer and then get like giga giga screwed."

Jon contrasts TSMC's federated, ego-less R&D (one departure is not existential) with the Jeff Dean counterexample — a rare full-stack fingerprint across software and hardware. SF vibe: "Gemini hasn't been a frontier model for a long time and now they definitely won't be." OpenAI/Anthropic still have one job: be best at AI; Google is ads/search/YouTube plus AI shove-in.

Contrarian element Market prices Google as a perpetual frontier peer. Panel: treat it more like Bell Labs + search monopoly — great cash engine, structurally bad at keeping cracked talent and shipping successive model wins.

Trigger Next Gemini quality vs Claude/GPT; further GDM exits; Demis path (Isomorphic / chief scientist / possible Alphabet CEO rumors) clarifying whether Google chooses science bureaucracy or product war.

Names (risk / relative underweight on frontier AI) Alphabet (GOOG) — see Key Risks. Anthropic / OpenAI as the "startup religion" winners in the Catholic/Protestant joke framing.

Investment Thesis #5: China Optics Ban Is Inverse-EV — West Bans the Better Product

Argument Trump admin prepping bans on Chinese data-center components including transceivers (and possibly humanoid robots / Unitree). Jon is generally hawkish on denying China revenue, but flags evasion via Thailand and the key asymmetry: unlike past bans of cheap knockoffs, InnoLight is actually better than the West on 1.6T transceivers — "inverse EVs." The entire supply chain is primarily in China, so the ban is more painful for Western companies whose DCs already depend on that chain. China also strong in optics (Sunny Optical lenses), drones, solar, EVs, humanoids.

"When it comes to 1.6T transceivers... the entire supply chain is primarily in China. So, this ironically is probably more painful for the Western companies to bear."

Contrarian element Ban bulls assume easy domestic substitution. Panel: you are cutting off the best product and must rebuild a manufacturing base, 1960s-Japan JV/IP-transfer style, under unabated AI demand.

Trigger Final BIS/entity rules on transceivers; Western optics vendors' lead times and CapEx; DC build delays blamed on optics.

Names Western optics / transceiver rebuild candidates (category); InnoLight / Chinese optics as the banned superior supply — watch substitution pain, not as a Tier-1 long.

The Ecosystem Map

Jon (Asianometry) — Built a custom ~40MB video editor with Claude/Codex (~couple hundred $/mo stack: Codex + GPT Pro + $20 Claude); memory shortage is what he sees on the ground in Taiwan
Doug O'Laughlin — Bell Labs/Google thesis; Terafab $100B by end-2027 bar; coding-agent believer (Fable / 5.6 redos of old projects)
Elon / SpaceX — Terafab GW ramp; $1T ARR pulled to 2030; public Nvidia preference; memory as limiting factor
Google diaspora — Dean, Sanjay, Oriol, Quoc Le (+ prior Jumper/Shazeer/Silver) → GPU-funded startups, often with GV capital
Value capture — Near-term: memory + Nvidia systems. Medium-term: whoever onshores optics under ban. Long-term frontier: Anthropic/OpenAI-style single-mission labs over Google bureaucracy

Key Risks

Alphabet (GOOG) frontier AI — Gemini lead exodus; "L culture" / acquire-don't-build critique; Bell Labs pattern of inventing then under-executing successive waves (panel still allows short/medium-term stock strength via search + TPU infra)
Western DC buildout under transceiver ban — Cutting InnoLight-class 1.6T supply before domestic alternatives exist
Hyperscaler FCF compression — MSFT/AMZN/GOOG/META taking cash flow to zero for chips; funding stress if debt markets tighten
Elon timing — Panel jokes he makes the "impossible late"; $1T in 2030 could slip toward 2040 without invalidating the direction
AI security / containment — Felony/breakout charts; modern software too large for humans to audit; local-LLM / "toxic internet" scenario as decade-scale risk

Investment Opportunities at a Glance

Tier	Name / Category	Core Thesis	Conviction Signal
1	Micron (MU) / SK Hynix / Samsung	Memory crimps Taiwan + Apple; CXMT refuses discounts	Jon on-the-ground shortage; Apple N2 idle without DRAM
1	NVIDIA (NVDA)	SpaceX chooses Nvidia; Google leavers buy GPUs/Vera Rubin	Elon tweet + TPU software-support gap outside Google
2	SpaceX / Terafab (PRIVATE)	1→10→20 GW; $1T ARR pulled to 2030; memory vertical option	Combined SpaceX earnings call guidance
2	Western optics / 1.6T transceiver rebuild	Ban forces domestic substitute for China-led supply chain	InnoLight "better than the West" asymmetry
3	Anthropic / OpenAI (PRIVATE)	Single-mission startups keep cracked talent Google loses	Panel "Catholic vs Protestant" lab framing
3	Coding-agent compute demand	Jon ships production tools on <$1k/yr model spend → token/GPU pull	Custom video editor via Claude/Codex/5.6
4	Elon memory manufacturing	Fastest path to Terafab revenue if he enters DRAM/HBM	Panel recommendation tied to $1T ramp math

Monitoring Checklist

Bottom Line

Own memory scarcity. Apple N2 dies idle and CXMT mogging Apple is the tell — Taiwan cannot ship products without DRAM/HBM; Elon's $1T path runs through memory first.
Nvidia remains the default silicon for everyone leaving Google. TPU pre-train economics do not matter if external software support is zero; SpaceX said the quiet part out loud.
Treat Google as Bell Labs + search cash, not a perpetual frontier peer. Dream-team breakup is the branch point; short/medium stock can work while frontier share quietly erodes.
A China transceiver ban is inverse-EV policy — the West is cutting off the better 1.6T supply chain (InnoLight), so watch domestic optics rebuild and near-term DC pain.
Elon's 10 GW / $100B bars are the abundance scoreboard — hit them and compute floods; miss them and memory/power constrain the entire CapEx cycle.