Gavin Baker | All In Summary

1. Key Themes

AI as the Ultimate Economic Equalizer — But Silicon Valley Is Losing the Narrative

The argument is made that AI democratizes expertise at a civilizational scale, yet the tech community has squandered the opportunity to make that case publicly, leaving a vacuum filled by anti-AI and socialist narratives.

"AI is the greatest economic leveler we'll ever find in our lifetime... it takes that world's knowledge and it allows you to act upon it so that every single man, woman, and child has an equivalent Travis Kalanick as his co-founder, a super founder, this brilliant person that can think through all your problems... And there is no gatekeeping that can prevent you from having them." 00:03:16

"We've done such a poor job in representing it, in bringing it to market, in talking about it. We've let all of our own personal trials and tribulations and insecurities and fights spill out into the open. As a result, Silicon Valley has lost even more credibility with the people at large." 00:03:45

The DSA Takeover of the Democratic Party Is a Structural, Not Cyclical, Phenomenon

The guests argue this is not a fringe protest movement but a deliberate, organized infiltration using the Democratic Party as a vehicle, with a singularly charismatic leader and a methodology borrowed directly from Trump's playbook.

"We're using the Democratic Party as a ballot access vehicle. Not because we share its goals. We build our own organization, get elected under the Democratic label, caucus with Democrats when it's useful, and push our own agenda from the inside. We see the Democratic establishment as an obstacle, not a home." 00:10:27 (quoting DSA co-chair Josh Bloch)

"I think he is one of the most talented politicians I've ever seen in my lifetime... She was talented. She's nothing like Zoram." 00:22:08 — Gavin Baker, comparing AOC to Mamdani

China Is Catching Up to US AI Frontier Models — Fast and Cheaply

GLM 5.2 from Z.ai represents a genuine inflection: a frontier-class open-weight model, trained on indigenous Huawei chips, priced 85% cheaper than GPT 5.5, and freely forkable under MIT license. The US export control and regulatory regime may already be obsolete.

"Z.ai founder told Elon Musk, open weight Fable capabilities will be here sooner than Q1 2027... We do not have months to give away in this race." 00:46:31 — Jason Calacanis, then elaborated by David Sacks

"They're six months behind on the model and they're 24 months behind on the silicon, yet they're only a few months behind in total. So what game are we playing? This is insane. We are going to lose if we keep doing this stuff to ourselves." 00:58:16 — Chamath Palihapitiya

DRAM Is the Single Most Important Bottleneck in AI Infrastructure

HBM DRAM — not GPUs, not power, not networking — is identified as the foundational constraint on AI scaling, with only three companies on Earth capable of producing it at the required quality.

"DRAM is the most important bottleneck... Elon is focusing the TerraFab on memory because he sees it as the most important bottleneck. You know, not lasers, not capacitors, not power supply semiconductors, not NAND flash, not HDDs, DRAM." 01:03:35 — Gavin Baker

"Memory is DRAM is probably going to be 30 to 40% of all hyperscaler CapEx next year." 01:05:49 — Gavin Baker

The Future of AI Is Composable, Not Winner-Take-All Models

Gavin Baker lays out a specific architectural thesis: every enterprise will run a routing layer that sends most queries to fine-tuned open-source models and only the hardest to frontier models, making open-source a complement to — not a replacement of — frontier labs while shifting value to infrastructure.

"You're going to have a router. And every query that somebody comes in, every task that needs to be done at your company, that router is going to send it to, you know, your RL, SFT version of P1.2 or Demetron... at some point in the workflow, a frontier model may or may not come in to kind of check it, add to it." 00:51:01 — Gavin Baker

"To date, frontier tokens are capturing 90% of the economic value. And open source tokens are probably 80% plus of tokens processed. And those ratios may be here to stay. But I just think composable models are the future." 00:49:47 — Gavin Baker

Orbital Compute Has a Concrete Economic Case That Is Improving Relative to Terrestrial

Gavin Baker does the first-principles math showing that as terrestrial data center costs inflate and Starship launch costs deflate, orbital compute will cross over to being cheaper within a few years — not as science fiction but as a capital allocation decision.

"To stand up a one gigawatt data center, it's $35 billion in semiconductors... And then it's $25 billion of power and cooling equipment. And that is clearly inflationary... When Starship is reusable, it's going to cost $5 billion to put a gigawatt of compute into space... So this is the economics that underpin orbital compute from first principles." 01:11:52 — Gavin Baker

"In three or four years, it's $70 billion versus $40 billion. And that $5, as Starship becomes rapidly reusable, is likely deflationary." 01:12:46 — Gavin Baker

Regulatory Capture Is the Real Game Anthropic Is Playing

Gavin Baker makes the pointed observation that Anthropic's safety-first public posture may be a calculated strategy to erect regulatory moats — and that the rollback of Fable may actually be what Dario Amodei wanted.

"Do you think Dario got exactly what he wanted? It seems to me there's some chance this has been a very calculated strategy to provoke the U.S. government into doing what they just did, and this is what he wants. He has a regulatory moat now." 00:55:16 — Gavin Baker

"We should not reward Dario by giving him exactly what he's always craved, which is some sort of labyrinthine government approval process that does reward regulatory capture." 00:56:05 — David Sacks

Micron's Supply Chain Agreements Represent a Structural Business Model Transformation

Floor-and-ceiling pricing contracts covering 50% of Micron's revenue at margins above prior cycle peaks signal a fundamental re-rating of DRAM from commodity cyclical to quasi-subscription infrastructure.

"They announced that they have these supply chain agreements that have a floor and a ceiling for prices with increasingly large group of large customers... and the floor pricing in these new contracts is ahead of prior cycle peaks from a gross margin perspective. And so this is really, I think, pretty, maybe end up being very transformational for the industry." 01:04:26 — Gavin Baker

The Disaggregation of AI Inference Into Pre-Fill and Decode Creates a New Hardware Arbitrage

The technical split between the memory-capacity-bound pre-fill stage and the memory-bandwidth-bound decode stage is opening a viable path to extend the useful life of legacy H100/A100 GPUs paired with specialized chips like Groq or Cerebras, lowering AI's cost of capital.

"You can put Groq or Cerebras, decode-optimized chips... in front of old NVIDIA GPUs like H100. So you can lift H100s, A100s out of some old data center, put them in one of these megapods... Put a Groq or a Cerebras in front of it, and you can get a very competitive solution... we're going to be using GPUs for seven years, 10 years, 12 years. And that's great because it lowers the cost to finance them, which makes this AI revolution more financeable." 01:24:51 — Gavin Baker

2. Contrarian Perspectives

Social Media Age Bans Are Actually a Censorship Trojan Horse for Adults, Not Child Protection

While the popular argument frames under-16 social media bans as protecting children, Travis Kalanick argues the real mechanism is forcing adult de-anonymization, enabling state censorship of political dissent.

"The real point of banning under 16 is so that you can force adults to identify themselves and de-anonymize themselves. So you can set up a full-scale censorship regime, which they're sort of contemplating in the UK. And what censorship is really about is not about harmful content. It's about content that the people in power don't want you to see." 00:26:55 — Travis Kalanick

"I think if it were not for free speech and X, I think we'd be living in a very different world today that would be a lot worse." 00:28:27 — Gavin Baker

Anthropic's Revenue Trajectory Implies a $3 Trillion Public Market Valuation

This is a number almost no one is saying publicly. Gavin Baker argues Anthropic will end 2025 at well over $100 billion in revenue, with 85% gross margins on inference, implying it would not trade at less than 10x that — putting the valuation in Apple/Nvidia territory.

"I think Anthropic is worth $3 trillion today. I think that is roughly where it would probably trade as a public company... They're going to end this year well over $100 billion... And it will be very profitable at that scale because it'll be inference dominated. And people are reporting they have 85% gross margins on inference." 01:28:21 — Gavin Baker

Every IPO Candidate Should Deliberately Under-Price to Avoid Breaking Deal Price — It's a Mechanical Shorting Trigger

Against the prevailing banker advice to maximize IPO price, Gavin Baker argues that breaking deal price activates price-insensitive institutional selling as a matter of rule, which then invites targeted shorting, creating a self-fulfilling death spiral regardless of business quality.

"There are a lot of portfolio managers who, if a stock breaks deal price, they sell it no matter what. They consider it a promise that was broken... shorts, if a stock gets close to deal price, they short it because they want to break deal price. And then they make a quick 10 or 20%... tell your bankers, price this in such a way that we're not going to break deal price in our first nine months as a public company." 01:31:30 — Gavin Baker

The DSA's Ascendance Is Driven by One Person's Charisma, Not Structural Economics

Against the conventional narrative that housing costs, student debt, and inequality are fueling the socialist wave, Gavin Baker argues the specific political success is almost entirely attributable to Zohran Mamdani as an individual talent — which implies it could collapse without him.

"I think he is a singularly talented politician. And he is the reason that the DSA is ascendant. Not their ideas or not dissatisfaction with AI, but it's him... I used to think AOC was by far the most talented Democratic politician. That was the warm-up... She's nothing like Zoram." 00:22:08 — Gavin Baker

Distillation Has Already Largely Leveled the Playing Field Between US and Chinese AI — Export Controls Are Fighting the Last War

Gavin Baker explains that Chinese labs have been systematically harvesting reasoning traces from US frontier model APIs through masked accounts at scale, effectively downloading frontier intelligence at trivial cost — making chip export controls a lagging indicator.

"For sure distillation has happened. There's been an immense amount of distillation... tens of thousands of phones, iPads and computers that are asking the Claude API through masked accounts, very specific questions. And then these reasoning traces are being harvested... those reasoning traces are then fed back into the model during the reinforcement learning process... And that is a way that you can get really, really close to the frontier at a fraction of the cost." 00:47:00 — Gavin Baker

3. Companies Identified

Z.ai (Zhipu AI) Chinese AI lab, maker of GLM models. Released GLM 5.2: 744B parameters, 1M token context, MIT license, highest score ever on Artificial Analysis Intelligence Index for an open-weight model, trained entirely on Huawei Ascend 910b chips, 85% cheaper than GPT 5.5 for comparable performance.

"Z.ai founder told Elon Musk, open weight Fable capabilities will be here sooner than Q1 2027." 00:46:31

"The claim is that this was all done on indigenous chips... GLM 5.2, the inference is optimized for the Huawei chips... they are basically going to package these things up... sell it at a fraction of the cost globally." 00:58:58 — David Sacks

Micron Technology One of only three companies globally that manufacture HBM DRAM. Revenue up 4x year-over-year ($9B to $42B), beat expectations by 16%, Q4 guidance raised to $50B. Stock up 14x since Gavin Baker's 2025 prediction show call.

"They announced that they have these supply chain agreements that have a floor and a ceiling for prices with increasingly large group of large customers. And this covers essentially 50% of their revenue, I think with just four customers. And the floor pricing in these new contracts is ahead of prior cycle peaks from a gross margin perspective." 01:04:26 — Gavin Baker

Cerebras Systems AI chip company focused on inference, recently IPO'd. Described as having a breakout OpenAI contract signed December 2025 whose revenue impact won't appear until Q3 2026 due to wafer-to-server production lead times.

"Let's say they could add 50 megawatts a month in 2027... that means they exit the year at roughly a $9 billion cloud computing run rate. And we're at less than $40 billion of market cap." 01:34:12 — Gavin Baker

Anthropic AI safety company, maker of Claude. Identified as a potential $3 trillion public company based on projected 2025 revenues well over $100 billion and 85% gross margins on inference. Also identified as potentially using safety advocacy to build a regulatory moat.

"I think Anthropic is worth $3 trillion today. I think that is roughly where it would probably trade as a public company." 01:28:21 — Gavin Baker

SpaceX / SpaceX AI Covered extensively for the recent IPO and for Starlink/TerraFab/orbital compute. TerraFab identified as focused specifically on DRAM as the key bottleneck. Orbital compute economics laid out with $5B launch cost per gigawatt once Starship is reusable.

"When Starship is reusable, it's going to cost $5 billion to put a gigawatt of compute into space... it starts to build." 01:12:18 — Gavin Baker

Groq (acquired by NVIDIA) Inference chip company optimized for the decode phase of AI inference. Mentioned as key enabler of disaggregated inference architecture that extends H100/A100 lifespan.

"You can put Groq or Cerebras, decode-optimized chips... in front of old NVIDIA GPUs like H100. So you can lift H100s, A100s out of some old data center, put them in one of these megapods... and you can get a very competitive solution." 01:24:51 — Gavin Baker

Cerebras (inference context) Separately noted as the other available solution for disaggregated decode-optimized inference alongside Groq, enabling competitive inference from legacy hardware.

"Cerebras is the other solution today that is available." 01:24:28 — Gavin Baker

CoreWeave Identified as one of only a handful of companies outside hyperscalers to have successfully brought on more than a gigawatt of compute capacity, cited as a benchmark for Cerebras to aspire to.

"Outside of the hyperscalers, the only companies that have ever brought on more than a gigawatt, I think, are CoreWeave, Crusoe, and SpaceX AI." 01:34:56 — Gavin Baker

Crusoe Energy Named alongside CoreWeave as one of the rare companies to scale past a gigawatt of compute. Also mentioned as working on modularly assembled data centers in shipping container format.

"Crusoe is working on, you know, modularly assembling data centers, you know, kind of like a data center... think of it as like an 18-wheeler... shipping container." 01:23:32 — Gavin Baker

Perplexity Cited for having early on taken Chinese open-source models, forked them, and restored censored content (including Tiananmen Square), demonstrating that political censorship in Chinese models is not a fatal flaw for American commercial use.

"Perplexity did this very early on with Chinese models. They showed that you could sort of put back the content on Tiananmen Square and things like that." 01:00:49 — David Sacks

NVIDIA Discussed for potential strategic response to OpenAI's custom Jalapeño chip (built with Broadcom) — suggesting NVIDIA could enter the foundation model space as a competitive response to customers vertically integrating into chips.

"Don't be surprised if NVIDIA says, you know what, we kind of like the area you're operating in now that you're going to make chips and maybe OpenAI sells those chips to other people. Don't be surprised if NVIDIA starts an OpenAI competitor. You heard it here first on All In." 00:52:40 — Jason Calacanis

Huawei Chinese national semiconductor champion. Its Ascend 910b chips are now claimed to have trained and optimized inference for GLM 5.2, the most capable open-weight model in existence, representing a potential validation of China's chip indigenization strategy.

"Z.ai, they are saying that the GLM-5 family was trained entirely on clusters of Huawei Ascend 910b chips." 00:58:28 — David Sacks

TerraFab Elon Musk's DRAM fabrication venture (in partnership with Intel). Identified as targeting the single most important AI bottleneck. Gavin Baker believes Musk's construction track record suggests it will be stood up faster than conventional fabs.

"Elon is focusing the TerraFab on memory because he sees it as the most important bottleneck... he has a track record of doing what Jensen called impossible, superhuman. And so we'll see." 01:04:03 01:10:52 — Gavin Baker

Vertiv Mentioned as a maker of modular data center systems (alongside Dell), relevant to the Megapod / air-cooled containerized compute opportunity.

"There are companies, you know, like Dell makes these racks, companies like Vertiv makes these modules." 01:18:08 — Chamath Palihapitiya

Bittensor Decentralized compute network mentioned as a live example of permissionless distributed inference, with validators ensuring contributors deliver what they claim.

"The magic of Bittensor is there are validators that make sure you're putting in what you say you're putting into the network." 01:26:07 — Jason Calacanis

Targon Mentioned as a permissionless marketplace where anyone can contribute H200s at ~$3-4/hour, representing the emerging distributed compute economy.

"Targon, which is just straight up people are putting and you can rent H200s by the hour for three bucks, four bucks, and it's permissionless." 01:26:07 — Jason Calacanis

Atreides Management Gavin Baker's fund. Called the Micron/HBM trade correctly in the 2025 prediction show; Micron up 14x from that call.

"In our 2025 prediction show, he gave a call on HBM makers like Micron as the best performing asset since that time. Micron up 14x." 01:02:21 — Jason Calacanis

Adams (Travis Kalanick's new company) Travis Kalanick's stealth startup. Described only obliquely. Gavin Baker notes it is generating significant investor interest.

"Your super awesome startup, Adams, which, man, loads of people have been calling me to say how, like, they've been in the... investor, your dog is hunting with investors, Travis." 00:50:42 — Gavin Baker

4. People Identified

Gavin Baker Managing Partner, Atreides Management. Called the Micron/HBM trade (14x return), laid out the orbital compute economics from first principles, identified the composable model architecture thesis, questioned whether Anthropic's safety posture is strategic regulatory capture, and valued Anthropic at $3 trillion.

"Memory is DRAM is probably going to be 30 to 40% of all hyperscaler CapEx next year... these stocks still trade cross-sectionally cheap relative to the rest of AI." 01:05:49

Andrej Karpathy Former Tesla/OpenAI AI researcher. Cited for coining the concept of the "Council of LLMs" — the thesis that enterprises will run multiple frontier models simultaneously rather than committing to one.

"You're going to every enterprise... you're going to have what Andrej Karpathy called the Council of LLMs. You're going to have Grok. You're going to have Anthropic. You're going to have OpenAI. Google." 00:48:22 — Gavin Baker

Brad Gershner Investor (Altimeter Capital). Cited for the concept of AI needing to renegotiate "the social contract" as the costs of building AI infrastructure escalate and become visible to society.

"This may give us as a society time to adapt, to adapt, you know, what our friend Brad Gershner calls the social contract." 01:06:39 — Gavin Baker

Dario Amodei CEO, Anthropic. Identified as potentially engineering the US regulatory environment through strategic safety advocacy — seeking an FAA-equivalent for AI models — while also having his own most capable model rolled back in the process.

"Dario posted a blog just a few weeks ago saying he wants an FAA for AI. They wanted government approval, a government approval process for AI models. And so in a sense, they've gotten exactly what they wanted." 00:55:39 — David Sacks

Zohran Mamdani NYC Mayor and DSA political figure. Identified by multiple guests as the most talented politician they have seen in their lifetimes — more skilled than AOC or Bernie Sanders — and the primary driver behind the DSA's political success, independent of the movement's ideology.

"I think he is a singularly talented politician. And he is the reason that the DSA is ascendant. Not their ideas or not dissatisfaction with AI, but it's him." 00:22:08 — Gavin Baker

Elon Musk Referenced extensively for SpaceX IPO, TerraFab/DRAM focus, orbital compute, Megapod/Tesla Powerwall distributed compute speculation, and T-Mobile acquisition rumor.

"Elon has done more to decarbonize the planet than every activist combined." 00:19:50 — Gavin Baker

Travis Kalanick Founder of Uber and current founder of Adams. Discussed the composable model architecture with Gavin, described distributed compute via restaurant kitchens, and engaged on political economy themes.

"The name is called Adams. It's not super secret. It's super awesome." 00:50:38

5. Operating Insights

Price Your IPO to Never Break Deal — It Triggers a Mechanical Short Cascade

The conventional wisdom is to maximize IPO price. Gavin Baker's counter-advice is specific and actionable: breaking deal price triggers price-insensitive institutional selling by rule, which shorts then weaponize to force a further decline regardless of fundamentals.

"Tell your bankers, price this in such a way that we're not going to break deal price in our first nine months as a public company. And that's what I always advise everyone to do." 01:32:23 — Gavin Baker

"If a stock breaks deal price, they sell no matter what. They consider it a promise that was broken... shorts, if a stock gets close to deal price, they short it because they want to break deal price. And then they make a quick 10 or 20%." 01:31:30 — Gavin Baker

Sequence Your Public Market Storytelling Around Lead Times, Not Milestones

When Cerebras reported a "slow" quarter after a transformational OpenAI contract, they failed to communicate the 7-month production pipeline between contract signing and revenue recognition. The operating insight is to pre-educate public market investors on your specific supply chain lead times before they interpret a quiet quarter as weakness.

"What I would have said is we signed this transformational contract with OpenAI in December... it takes four months to make the chip. Then it takes us two months to turn that chip into a server. And then if we're lucky, it takes us a month to energize that chip... the first time you're going to see the impact of this OpenAI deal at the earliest is probably around Labor Day." 01:33:15 — Gavin Baker

For Enterprise AI, Build a Router First — It Unlocks Both Cost Efficiency and Capability

Rather than committing to a single frontier model provider, the practical playbook for any company building on AI is to implement a routing layer that directs 80-85% of queries to a fine-tuned open-source model and escalates only the hardest tasks to frontier models. This simultaneously reduces per-token costs and creates a proprietary intelligence layer.

"Every query that somebody comes in, every task that needs to be done at your company, that router is going to send it to your RL, SFT version... at some point in the workflow, a frontier model may or may not come in to kind of check it, add to it. And that's what I mean by a composable model." 00:51:01 — Gavin Baker

Distributed Training Requires Physical Co-location; Distributed Inference Does Not

A critical operational distinction: if you are building or considering building distributed compute for training workloads, physical separation between GPU clusters — even two kilometers over fiber — kills efficiency by orders of magnitude. For inference workloads, geographic distribution is viable and increasingly commercially interesting.

"Even if it's on your own fiber connected, like, two kilometers away, you're screwed. It's not a thing... But everything that cuts against training works for inference." 01:22:40 01:22:47 — Travis Kalanick and Gavin Baker

6. Overlooked Insights

CXMT Going Public in China Is the Release Valve for Consumer DRAM — But Only Consumer DRAM

Everyone on the podcast focused on the Micron blowout and the AI DRAM shortage. But a single sentence identifies the actual mechanism that will eventually relieve consumer electronics price pressure without touching the AI DRAM market: China's CXMT is IPO'ing and will flood the market with commodity-grade DRAM. This bifurcates the DRAM market permanently — a structural short opportunity in consumer electronics components and a structural long in AI-grade HBM that most analysts are treating as one market.

"CXMT is going public in China. They are going to, they may be the cure for Apple's ills. They will flood the market with, to some degree, cheap consumer grade DRAM. But for the DRAM you need in these AI servers, there are three companies that can make it." 01:05:19 — Gavin Baker

A Startup Is Deploying 4-GPU Inference Units Into Homes With Locked Hardware and Neighbor-Scale Inference — Right Now

Gavin Baker mentions in a single sentence that a startup already exists that is installing locked 4-GPU units at residential properties, offering homeowners a power discount in exchange for contributing neighborhood-scale inference compute. This is the actual early-stage version of the Powerwall-plus-GPU distributed compute thesis being discussed theoretically by the whole group — and it is already operational.

"There is actually a startup that is trying to put four GPU units with kind of a battery on people's houses and give them a discount on their power. And then you can do inference for that neighborhood, you know, from those four GPUs. And it's like lock sealed so nobody can get in." 01:23:05 — Gavin Baker