Gavin Baker: Twelve Months of Investing Commentary

• All-In: core frame set - "all of the dollars going into the ground are being used… we're clearly still behind demand." Meta's ad beat and Copilot uptake cited as the first revenue (not just opex) evidence.
• Invested by Aleph: concedes the honest open question - ROI positive so far per the 10-Qs, but the leap to $1-3T of Blackwell-era spend is unproven.
• ILTB EP.451: worries about a "Blackwell ROI air gap" during the training-heavy phase - then is relieved by C.H. Robinson's AI-driven ~20% earnings beat, his first non-tech Fortune 500 proof point.
• BG2: "the math maths" - ~$1.5T of true 2027 capex against ~$300B of AI-lab inference revenue at 50-70% gross margin, justified because per-gigawatt monetization is rising (Anthropic ~$22-23B/GW, Google ~$50B/GW). Altimeter's 55% IRR estimate on Colossus 1 clears any cost of debt.
• Generating Alpha: reframes weak near-term ROI as a temporary "divot" - Blackwell spend is training-heavy until agentic products monetize it.
• X post ("mega bull case"): even if frontier-lab inference margins compress, the margin dollars "get redistributed from the frontier labs to AI infrastructure providers." The bear case for labs is bullish for his book.

"There are no dark GPUs."- recurring, from Dec 2025 onward

"If you can borrow money at 6, 7, 8% and invest in something with a 55% IRR… that math maths."- BG2Pod, June 2026

• Invested by Aleph: GPUs won't be unseated; internal ASIC programs need three generations to compete; chip startups must be "different and hard" (Cerebras, Groq, Next Silicon) - "if you're trying to make a better GPU, good luck."
• a16z (restated Jul 2026): most custom ASIC programs - especially Meta MTIA-class - fail or get cancelled within ~3 years; AMD is a perpetual second source.
• ILTB: Blackwell's air-to-liquid, 30kW→130kW rack transition was "the hardest product transition in semiconductor history" - which explains why TPU-trained Gemini 3 briefly won on cost.
• 13F: money behind the mouth - NVDA calls on 3.5M underlying shares, Astera Labs position up ~25x, Micron added.
• Sohn: "flinty old men and women safeguarding Morris Chang's legacy" - if TSMC doubled capacity, "Nvidia could probably sell $1, $1.5, $2 trillion of chips next year." Watch Intel/Samsung over-expanding or discounting as the bubble tell.
• All-In EP.278: DRAM/HBM displaces power as THE bottleneck - only Micron, SK Hynix, and Samsung can make AI-grade memory; Micron's floor pricing above prior-cycle peak margins signals structural repricing.
• The memory arc: by his own 25-year framework "you should 100% be selling memory right now" - but he stayed long, betting this is the mid-90s true-capacity cycle, the one exception. Validated by Micron's +345.7% YoY revenue.

"Taiwan Semi, if we don't get a bubble, we need to throw a party for them - they will have single-handedly prevented one."- Sohn Conference, May 2026

"TSMC executives… dismissed [Sam Altman] as a podcast bro."- retelling the trillion-dollar fab ask

• All-In: the SMR / Gen-4 nuclear revival is demand-pulled by hyperscalers (price-insensitive, carbon-free preference), not subsidy-driven. Anthropic's 50GW ask ≈ a full year of total US capacity additions.
• ILTB: power - not capital, not chips - becomes the binding constraint; nuclear can't scale in time, so natural gas and solar are the near-term answer. Nuclear stocks flagged as a "rolling bubble."
• Sohn / Watts and Wafers: the watts shortage starts easing 2027-28 via conventional capacity; wafers stay tight longer.
• CNBC + All-In: the orbital endgame - ~$30B/GW in space vs ~$60B (rising toward $70B) terrestrial once Starship is reusable. Terrestrial power and cooling is the inflationary, human-labor half of the bill of materials. Majority of Earth's compute could be in orbit within 10-20 years.
• CNBC: speed-of-energization IS cost - SpaceX/xAI monetize a gigawatt at a premium and bring it online fastest.

"I think the energy implications of AI are probably bigger than the Inflation Reduction Act and certainly will be longer-lasting."- All-In, August 2025

"It costs $30 billion to put a gigawatt of compute in space relative to $60 billion here on Earth."- CNBC, June 2026

• All-In: GPT-5 read as OpenAI's first non-decisive release - lead possibly eroding.
• Aleph / a16z: full reframe - "GPT-5 is a smaller model… designed to be more economical. Any reference to GPT-5 and scaling laws is crazy." Reasoning models are the flywheel; a four-lab oligopoly (OpenAI, Anthropic, xAI, Google).
• ILTB: Gemini 3 confirms pre-training scaling laws after an ~18-month gap - to him the single most thesis-relevant fact of the year. Low-cost-producer doctrine introduced: OpenAI structurally high-cost (pays a margin for compute → "code red," $1.4T of commitments); Anthropic burning ~80% less cash to similar scale.
• Sohn / ILTB: usage-based pricing shift ("all you can eat" → "pay by the drink") is "wildly bullish"; takes the over on OpenAI+Anthropic $200B combined ARR within 12-18 months.
• BG2: declares the open-source-catches-up thesis "decisively wrong" so far - frontier may take ~90% of economic value on ~20% of token volume.
• CNBC: Kimi K2/K3 don't scare him (they need 2-3x the tokens per equivalent answer). His "symphony" framework: frontier conductor plus cheap open-source instrumentalists fine-tuned on proprietary data. Harness and product lock-in (he cites Claude Code) matter as much as the model. Biggest tail risk: a "bitter lesson" violation. Best-articulated bear case: free, good-enough edge AI on devices.

"If you are a frontier model without access to unique, valuable data and internet-scale distribution, you're the fastest depreciating asset in history."- a16z, October 2025

"There is an enormous return to being at the frontier of intelligence."- recurring, 2026

• All-In: the Apple broadside - ~$700B of buybacks vs no AI capability is a false dichotomy; "it almost feels like they're trying to lose."
• a16z: Google is Nvidia's most serious competitor and possibly already the biggest AI company by usage; the ASIC-graveyard call; skeptical of AI browsers against Chrome's 5B users.
• ILTB: Google's TPU cost advantage is a temporary window that GB300 closes. Anthropic's TPU/Trainium deals are a real cost edge. Meta/Microsoft/Amazon lab catch-up efforts underperform because they lack the reasoning flywheel.
• Sohn / ILTB / Huxiu: window closed on schedule - Google's Pareto dominance "erased by conservative TPU v8 design choices." Trainium is "by far" the most underestimated chip for 2026: only Nvidia and Amazon have functioning switched scale-up networks for MoE inference. Satya credited for the "courageous" call to keep Azure GPUs for Microsoft's own products.
• All-In: Anthropic "worth roughly $3 trillion today," ending 2026 well over $100B revenue at ~85% inference gross margins.
• CNBC: holds NVDA; holds neither OpenAI nor Anthropic (the SpaceX-Anthropic détente came too late). xAI's Cursor acquisition ($60B, ~$3B ARR) "looking very smart."

"Incompetence is indistinguishable from malice."- on Apple, All-In, August 2025

"Trainium is going to be to 2026… what TPUs were to 2025."- Sohn, May 2026

• a16z: the Cisco-150-180x-vs-Nvidia-~40x framing debuts.
• ILTB: names the froth precisely - nuclear and quantum are "rolling bubbles" where the public tickers aren't the real leaders.
• Watts and Wafers: flags late-cycle warnings inside his own bull case - semicap equipment at ~40x forward vs DRAM at mid-single-digits; the lowest-quality suppliers outperforming; a Mauboussin-style "diversity breakdown" beginning.
• Motley Fool coverage: expressed as a relative-value pair - long cheap memory/compute (MU, SNDK, NVDA) against triple-digit-P/E optical names (Lumentum, Coherent).
• Generating Alpha / BI: "anyone who says we're in a valuation bubble is just not paying attention." Burry's bearishness is "a godsend" that keeps the market honest - "watts, wafers, and scars of the internet bubble keep a true bubble from happening."

"We don't want a bubble. Bubbles are terrible… The aftermath of them is even worse."- Generating Alpha, July 2026

• Aleph: SpaceX is "probably the highest risk-adjusted return available in the markets."
• Publicly backs Musk's ~$1T Tesla pay package against proxy-advisor consensus - a conviction signal on the AI/robotics optionality thesis.
• Q4-2025 13F: NVDA calls (3.5M underlying), ALAB up ~25x, MU added, GitLab halved on AI coding disruption, Unity and Wix added.
• Capital Allocators: the meta-framework - "investment hypotheses, not theses"; "you either panic early or double down late"; drawdown behavior, not hit rate, is the edge.
• 13F coverage: QQQ puts the largest line - the barbell of AI-bottleneck longs vs index hedge.
• BG2: trims core positions from "large" to "medium/small" on elevated prices, macro concerns, and a needed consolidation.
• After the July shakeout (MU/WDC/SNDK/MRVL down 24%+ in a month), flips back to offense: "Risk/reward seems attractive again. Lots of cheap stocks with durable competitive advantages that are going to crush numbers for the next 6-12 quarters."
• On SpaceX's post-IPO drawdown (roughly -$1T from peak): "I'm not really thinking about it."