Gavin Baker owned 15% of Nvidia and 10% of Tesla sub-$2 billion. His biggest regret won't be either. It'll be the price you'll pay for SpaceX after the IPO. "A company like this comes once in your career. Thank God I took it. Plenty of people didn't." His analog: British East India Company. Empire scale, not tech scale. The setup: - 10,000 SpaceX employees could have sold every 6 months for a decade - Almost nobody did - Secondary pricing is a decade underpriced Baker on the Elon-Anthropic partnership: "Four months ago it didn't exist. Then it made sense. Everyone has a price." Incumbents realign around SpaceX gravity when it moves. The IPO doesn't price the company. It prices the last window.

$1.4 trillion in commitments. $13 billion in revenue. Sam Altman just told bears to short the stock. The private-valuation window closes when the S-1 drops. The exchange: Brad Gerstner asked how OpenAI reconciles the gap. Altman: "We're doing more revenue than that. If you want to sell your shares, let me know. It'd be cool to be public so I could tell the loudest voices to short the stock." Translation: revenue higher than the $13B floor, IPO closer than announced, Altman baiting bears — the CEO tell before a public offering. Every OpenAI-adjacent name is repricing: - $MSFT — 49% economic interest - $CRWV — dedicated compute - $ORCL — Stargate partner - $NVDA — largest chip customer - $SFTBY — SoftBank + Stargate When the CEO tells bears to short, smart money is already loaded. Everyone reading this after the S-1 files is buying at a public multiple. Not financial advice!

This is f**king insane. Gavin Baker mentioned Amazon Trainium in a podcast. AI told someone to put 30% into Marvell (MRVL). Two weeks later Jensen Huang publicly named Marvell the next trillion-dollar company. Marvell +20% pre-market. Green all day. The play: - Extract transcripts from Baker, Chamath, Jensen, Amodei - Feed to AI - Ask which specific tickers benefit from the constraints they describe - Wait for the second operator to confirm Alpha isn't Baker saying Trainium is underestimated. Everyone heard that. Alpha is AI translating it into a ticker before the second confirm hits. Two operators on one call = signal. One = noise. Follow — every convergence I catch lands here before it prices in.

Gavin Baker has made 4 major calls this month. Bears haven't caught one. 1. Foundation model survival list: Google, Meta, xAI. Everyone else commoditized. 2. Magnificent 7 goes $12T → $100T. One or two die. Winners take 30-40% each. 3. Anthropic S-1 will break value investors. Tokens aren't subsidized — the majority are profitable across the chain. 4. Orbital data centers aren't impossible. 10,000 SpaceX engineers already solved them. PhDs on X arguing physics haven't done the hours. The through-line: bears are wrong on central facts, not analysis. - If tokens are profitable → "AI is a bubble" collapses - If orbital compute ships → "power constraints kill AI" collapses - If Mag 7 concentrates → index-hugging fails - If foundation models consolidate to 3 → everything else is training data Baker's frame: don't argue theory when operators have already shipped receipts. Four calls. One thesis. Everyone else is a lagging indicator.

Gavin Baker: "The Anthropic S-1 is going to break a lot of people's brains." His frame: "A lot of macro and value investors are confidently making prognostications about AI with the assumption that tokens are subsidized. They're just wrong." The numbers he claims are already true: - Anthropic generating cash. Profitable. - Open-source token providers profitable. - OpenAI generating cash imminently. - "The overwhelming majority of tokens are profitable for everyone in the chain. Everyone." His analogy for the bear thesis: "It's like saying I'm bearish on the world economy because oil is at $500 a barrel. If oil was at $500 a barrel, that'd be a good reason to be bearish. It's just not. The central fact is wrong." Baker's frame: bears aren't wrong about the math. They're wrong about the input to the math. The circular-financing-bonfire crash they're pricing in doesn't happen when the tokens don't need subsidy. The S-1 is the receipt.

Gavin Baker: "The Magnificent 7 will go from $10-12 trillion to $50-100 trillion. But one or two of them will go extinct." His foundation-model survival list is short: - Google — search-plus-Gemini flywheel - Meta — Llama + ad-revenue self-funding - xAI — X + Tesla data footprint Anthropic, OpenAI, everyone else — Baker treats them as commoditized. His FOMO warning: "Application SaaS is done. If you went all-in on crypto and SaaS in 2021, don't repeat the trade." On AGI: "Elon says 3 years. I say 5 is safer." On regulation: "The leading labs weaponized fear of China. There's going to be no regulation." The math he keeps quoting: $12T → $100T is 8x. But 1-2 die means winners take 30-40% of the total, not 14%. Baker's frame: no regulation + AGI in 5 years + Mag 7 8x = biggest enterprise-value transfer in market history is happening right now.

Gavin Baker: "AI models will go to zero. Data centers are commodity. Energy is commodity. Only two things in the whole stack have real value: data moats and reinforcement learning." His survival list is short: - Google — already multi-trillion, safe - xAI — his call: worth over $1T combined with X - Anthropic — the other closed model that survives Microsoft, OpenAI, Meta — chose open source. Baker treats them as commoditized. What he actually buys: things that increase GPU utilization. Take a GPU from 30% utilized to 60%, you've doubled the output of the AI factory. The math he keeps quoting: GPUs got 50x faster in 4 years. Nvidia sells for $50,000 what Taiwan makes from $700 of sand. Baker's frame: don't rent intelligence. Own the thing it runs on. Full breakdown — in the article below.

Palantir CEO Alex Karp on AI regulation: "How can they tell me not to wear a condom?" His actual argument: hold builders accountable for real harms, but don't preemptively ban capabilities before harms occur. Two regulatory philosophies fighting for the AI industry: - Outcome regulation (courts, liability): common-law tradition, innocent until proven harmful - Intent regulation (approval gates, capability limits): EU precautionary principle, guilty until proven safe Karp is voting outcomes. So is most of Silicon Valley. Brussels is voting intent.

Holy sh*t. The co-inventor of ChatGPT just released the model that makes ChatGPT look like a rotary phone. Diogo Almeida spent 2 years in stealth building "Jev." His claims: 20-200x faster. 40-400x cheaper. Output tokens free. His thesis: chat models optimize for talking. Real AGI needs optimization for deciding. Chat models are fluent and expensive. Jev is decision-optimized and cheap enough to run in loops. If inference cost drops 40-400x, every AI product gets rebuilt from scratch. The guy who built ChatGPT thinks the entire LLM race is optimizing the wrong metric. The next frontier isn't fluency. It's decisions per dollar.

US chip export bans forced China's AI labs to innovate around scarcity. The result: DeepSeek R1 trained for ~$5.6M matches models US labs spend $100M+ on. The efficiency gap is structural: - US labs optimize for compute abundance — throw more GPUs at every problem - Chinese labs optimize for compute scarcity — every architectural improvement matters - Same algorithms, different constraints, different outcomes Kimi K3, Qwen, DeepSeek all release open-source with weights, papers, and training details. US frontier labs release APIs. Anthropic and OpenAI are publicly pushing for AI safety regulation. Critics call it regulatory capture — a framework that grandfathers current frontier labs and locks out both open-source and new entrants. The safety concerns are real. So is the moat the framework builds. The question isn't who's smarter. It's whether US labs win on capability or on regulation. If the answer is regulation, the rest of the world builds against them.

The timing isn't strange — it's strategic. Anthropic just raised at ~$180B. Their essay reads as an AI-safety warning. It also reads as a case for regulation that grandfathers incumbents and locks out startups. The threat is real. So is the moat the safety framework builds.

Anthropic's public framework for how it builds agents internally: Agents → Loops → Graphs → Self-Improving Systems Each layer removes another layer of human input. - Agents: one instruction, one action - Loops: model checks its own output, iterates - Graphs: loops connect, output of one feeds the next - Self-improving: the graph updates its own structure based on results The gap that matters: the Anthropic team ships internally at layer 4. The public Claude API ships at layer 1-2. Most retail users don't leave layer 1. The 3 layers between "prompt engineering" and "self-improving graph" is where the entire moat forms. The people who spend a year at layer 4 while everyone else optimizes prompts at layer 1 aren't smarter. They're just working on a different problem entirely. The tool is the same. The stack you build with it isn't.

xAI engineer (ex-Cursor): "I keep Grok 4 in Cursor for the work I'm watching and hand the rest to 20+ Grok agents in a loop. I stopped watching those months ago. Everyone optimizes the prompt. The bill isn't in the prompt." The bill is in the architecture: - How many agents you're running in parallel - How many model calls each makes per task - Whether outputs are cached or regenerated - Which tier handles which subtask Cursor for the work you watch. Agents for the rest. Engineers optimizing prompts save cents. Engineers architecting the graph save thousands. The 10-person billion-dollar companies Altman keeps predicting aren't built on better prompts. They're built on smaller monthly invoices to xAI, OpenAI, and Anthropic.

xAI engineers are increasingly running agent teams instead of writing code manually. The pattern spreading across every major AI lab: - 10-20 specialist agents in a loop-and-graph architecture - Human role shifts from coder to orchestrator - Each agent handles a narrow task; the graph handles coordination At Anthropic, Cursor, xAI, and inside OpenAI, engineers report similar workflows. Different vocabulary. Same shape. The mechanic: 1. Break the codebase into agent-shaped tasks 2. Let each agent iterate autonomously 3. Human reviews graph outputs, not individual commits 4. Reserve the human loop for strategic decisions the graph can't make The trade isn't learning to prompt better. It's learning to build the graph that prompts for you. Engineers who orchestrate agents ship 10x. Engineers who prompt manually ship what they always shipped.

Jensen Huang broke AI into a 5-layer cake: energy, chips, infrastructure, models, applications. The scoreboard: - Energy: "China has twice as much as we have. Our economy is bigger. Makes no sense." - Chips: "We're several generations ahead. Don't be complacent." - Manufacturing: "Anybody who thinks China can't manufacture is missing a big idea." - Frontier models: "We're probably 6 months ahead." - Open source: "Of 1.4 million models, most are open source. China is way ahead there." - Applications: "In China 80% think AI does more good than harm. Here it's the other way around." US leads chips, models, and applications adoption. China leads energy, manufacturing, open source, and public sentiment. Every layer where China leads compounds every layer where they trail. Energy fuels compute. Manufacturing scales chips. Open source captures developers. Jensen isn't warning about China catching up. He's mapping which layers are already gone.

Sam Altman keeps saying the next OpenAI model will make prompting obsolete for most users. Not "reduced." Obsolete. His logic: as models get better at inferring intent, users stop crafting careful instructions and start throwing arbitrary tasks — software, science, whole simulations — at the tool. The transition is invisible from outside. From inside OpenAI, it's already how their teams work. One person does what a team used to spend a month on. Not because the person got smarter. Because the tool stopped needing translation. Prompt engineering was a bridge skill. Bridges get demolished when the other side is finished.

The chief-of-staff agent pattern is quietly becoming standard at every major AI lab. The architecture: one orchestrator agent knows what other agents can do. It routes requests, manages context, handles the coordination overhead a human PM used to eat. The implementation: 1. Base layer — Grok API, Claude API, OpenAI API. Pick your foundation. 2. Chief-of-staff — an agent whose only job is knowing the other agents and delegating. 3. Specialist agents — one for research, one for writing, one for code review, one for outreach. 4. Shared memory — every agent reads the same log so context doesn't fragment. Anthropic ships this internally. So does OpenAI. xAI engineers describe similar setups. Everyone running one prompt at a time is doing manual labor. The teams shipping are running orchestrators. The gap is one architecture decision.

Anthropic just published the 5-step harness build for multi-agent systems. The full stack, condensed: 1. Start with the Claude Agent SDK. Harness handles loops, context, sandboxing. 2. Separate the brain from the hands. Reasoning in one place, tools in sandbox. 60% faster to first token. 3. Run server-side and log every step. Close laptop, keeps running. Crashes resume from log. 4. Make failure cheap. Retry dead sandboxes, replay lost context. 5. Turn yesterday's logs into new memory. Anthropic calls this "dreaming." The harness wakes up smarter than the one you deployed yesterday. Ng described this coming. Cherny confirmed it in production. Jensen and Altman echoed at scale. Now Anthropic published the recipe. Prediction → confirmation → recipe. That's usually the last stage before commoditization.