The episode features Gavin Baker, technology investor and founder of Atreides Management, discussing the rapidly evolving AI landscape with host Patrick O'Shaughnessy. Baker brings encyclopedic knowledge of technology markets and decades of experience investing across multiple technology cycles.

The conversation covers the infrastructure war between Nvidia, Google, and emerging players, examining how chip transitions from Hopper to Blackwell are reshaping competitive dynamics and cost structures across the AI ecosystem.

Baker explains the critical importance of three scaling laws - pre-training, reinforcement learning with verified rewards, and test-time compute - and how they multiply together to drive AI progress despite hardware constraints.

The discussion explores frontier model dynamics, the economics of token production, SaaS company strategic failures, semiconductor venture capital, power constraints, and Baker's contrarian thesis on data centers in space as the inevitable future of compute infrastructure.

Staying Current in AI: Process and Information Sources

Baker emphasizes using paid premium tiers ($200/month) rather than free versions: "The free tier is like dealing with a 10-year-old and making conclusions about the 10-year-old's capabilities as an adult" versus paid tiers being "fully-fledged 30-35 year olds"

"AI happens on X" - Baker follows 500-1,000 cutting-edge AI researchers globally who drive all downstream developments, including major public disputes like the PyTorch vs Jax team fight that required lab leaders to intervene

"Everything Andre Karpathy writes, you have to read it three times. Minimum" - Baker treats output from top researchers at OpenAI, Gemini, Anthropic, and XAI (the four leading labs) as essential reading

Baker uses AI itself to process information with minimal friction, including voice-activated assistants and discussing podcast content with AI models to extract insights from technical interviews

Scaling Laws and the Blackwell Transition Crisis

Gemini 3 confirmed pre-training scaling laws remain intact despite no one understanding how or why they work - "kind of like ancient British people's understanding of the sun... perfect measurement" of equinoxes at Stonehenge "but they had no idea how or why"

After XAI achieved 200,000 coherent Hoppers (maximum possible), AI progress should have stalled from mid-2024 through early 2026 waiting for next-generation chips, but reasoning models bridged this gap

Blackwell transition was "by far the most complex product transition we've ever gone through in technology" - going from air-cooled to liquid-cooled, 1,000 to 3,000 pound racks, 30 to 130 kilowatts per rack

Baker analogizes Blackwell deployment to iPhone upgrades requiring complete home infrastructure overhaul: "change all outlets to 220 volt, put in Tesla power wall, generator, solar panels, reinforce the floor"

"Reasoning kind of saved AI" by enabling progress through two new scaling laws (reinforcement learning with verified rewards and test-time compute) while hardware caught up, taking ARC AGI scores from 8% to 95% in three months

Google's Temporary Cost Advantage and TPU Dynamics

Google trained Gemini 3 on TPU v6/v7 (2024-2025 era chips) while competitors waited for Blackwell, creating temporary advantage. Baker compares: "Hopper is like P-51 Mustang, TPU v6/v7 are F-4 Phantoms, Blackwell is F-35"

"AI is the first time in my career as a tech investor that being the low-cost producer has ever mattered" - unlike Apple, Microsoft, or Nvidia's success, token production cost now determines competitive dynamics

Google as low-cost producer has been "sucking the economic oxygen out of the AI ecosystem" by running AI at potentially negative 30% margins, rational strategy given search business subsidizes losses

Google pays Broadcom estimated $15 billion annually (50-55% gross margin on ~$30 billion TPU program) for back-end design and Taiwan Semi management, creating economic pressure to bring program in-house

"Google can go to every person who works in Broadcom Semi, double their comp and make an extra 5 billion" since Broadcom's entire semiconductor opex is ~$5 billion, explaining MediaTek partnership as warning shot

TPU development slowing due to conservative design choices and bifurcated supply chain while GPU development accelerates with annual cadence - "competitive response of Lisa and Jensen to everybody saying we're gonna have our own ASIC"

Blackwell and GB300: Shifting Cost Dynamics

XAI will release first Blackwell model in early 2026 because "according to Jensen, no one builds data centers faster than Elon" and they help Nvidia work out bugs for everyone else

Even after deployment, Blackwell requires 6-9 months to outperform tuned Hopper systems - same pattern occurred when Hopper needed 6-12 months to exceed prior generation A100 performance

GB300 is "drop-in compatible" with GB200 racks, requiring no new power infrastructure or cooling modifications, dramatically simplifying deployment compared to initial Blackwell transition

Companies using GB300 with vertical integration "are going to be the low-cost producer of tokens" - fundamentally changing Google's strategic calculus once they lose cost advantage

When Ruben generation arrives, "the gap is going to expand significantly versus TPUs" while Tranium 3 and 4 from Amazon may be competitive, but most other ASICs will struggle

ASIC Competition and Vertical Integration Challenges

"It takes at least three generations to make a good chip" - TPU wasn't competitive until v3/v4, Amazon's Tranium/Inferentia only becoming viable at generation 3, illustrating learning curve difficulty

Companies making ASICs discover the full stack problem: "What's the NIC going to be? What's the CPU? What's the scale-up switch? Scale-up protocol? Scale-out switch? What kind of optics? What's the software?"

"I will be surprised if there are a lot of ASICs other than Tranium and TPU" long-term, and both will eventually run on customer-owned tooling despite current company statements - "economics make it absolutely inevitable"

Amazon has "the best ASIC team at any semiconductor company" with innovations in Graviton CPU, Nitro, and SuperNIC, giving them best chance at ASIC success outside Google

AI Capabilities: Intelligence vs Usefulness Transition

Current frontier models already exceed most human expertise - differences only visible when "you know it really deeply" on specialized topics like PCI Express vs Ethernet protocols for scale-up networking

"We need to shift from getting more intelligent to more useful" - Baker argues ROI S-curve must hand off from intelligence gains to practical usefulness, then to scientific breakthroughs creating new industries

Gemini 3 made Baker a restaurant reservation, "first time it's done something for me" beyond research - if it can book restaurants, "you're not that far from" hotel reservations, airplane tickets, Uber ordering

Building blocks of usefulness include extended context windows to hold "every Slack message and Outlook message and company manual" plus all world knowledge during test-time compute for complex tasks

"50% plus of customer support is already done by AI" in tech-forward companies within the $400 billion customer support industry, with sales following similar trajectory

Karpathy's principle: "With software, anything you can specify, you can automate. With AI, anything you can verify, you can automate" - verification enables reinforcement learning for functions with right/wrong outcomes

ROI Evidence and the Prisoners Dilemma

"The ROI on AI has empirically factually unambiguously been positive" - largest GPU buyers are public companies with audited financials showing ROIC higher than before GPU spending ramped

Revenue growth at major internet companies accelerated from moving recommendation systems and advertising from CPUs to GPUs, generating massive efficiency gains regardless of new AI products

Internal tension at every company: revenue teams "intensely annoyed" at GPUs given to researchers because "it's a very linear equation. If you give me more GPUs, I will drive more revenue"

Companies trapped in prisoners dilemma - "terrified that if they slow down, they're just gone forever" while competitors continue, creating existential risk that overrides economic calculations

"With Blackwell and for sure with Ruben, economics are going to dominate the prisoners dilemma" because the absolute dollar amounts become too large to ignore, forcing ROI-based decisions

"Almost all of them want to live forever" - religious belief in reaching ASI (artificial super intelligence) drives spending beyond rational economic analysis, seeking ultimate return on investment

Bear Cases: Edge AI and Economic Returns

"By far the most plausible and scariest bear case" is edge AI - within three years, phones could run pruned Gemini 5/Grok 4 equivalent at 30-60 tokens/second locally for free

Apple's clear strategy is "we're going to be a distributor of AI" running privacy-safe models on-device, calling cloud "god models" only for complex queries beyond edge capabilities

Critical question: "If 30 to 60 tokens at 115 IQ is good enough" running locally on bigger, bulkier phones with shorter battery life, cloud compute demand could collapse

Other bear cases include scaling laws breaking (now disproven for pre-training through Gemini 3) or low economic returns to ASI limiting continued investment justification

Fortune 500 AI Adoption and Real Productivity Gains

"Fortune 500 companies are always the last to adopt new technology" - first AWS re:Invent in 2013 had every startup on cloud while Fortune 500 took 5+ more years to standardize

"VCs are more broadly bullish on AI than public market investors" because they see real productivity gains - charts show companies today have significantly lower headcount for given revenue versus two years ago

C.H. Robinson (freight forwarder) stock rose 20% on earnings driven by AI - previously quoted 60% of requests in 15-45 minutes, now quotes 100% in seconds using AI

Baker worried about "Blackwell ROI air gap" - three quarters of massive capex spending on training with no inference revenue, potentially causing ROIC decline like Meta experienced

Fortune 500 AI adoption data points becoming critical for navigating potential air gap - companies with "long track record of success" and "internal culture of experimentation" will be early adopters

Reasoning Models and the Flywheel Effect

Pre-reasoning AI lacked the flywheel that powered great internet companies: "you made a good product, you got users, those users generated data that could be fed back into the product to make it better"

"Foundation models without unique data and internet scale distribution are the fastest depreciating assets in history" - Baker's modification of Eric Vishria's statement capturing pre-reasoning dynamics

Reasoning fundamentally changed industry dynamics by enabling flywheel to spin - user feedback on similar questions with consistent likes/dislikes creates verifiable rewards fed back into models

"We're very early at this flywheel spinning, like it's hard to do now, but you can see it beginning to spin" - creating increasing returns to scale similar to Netflix, Amazon, Meta, Google

Meta and Microsoft's Frontier Model Failures

Mark Zuckerberg predicted in January 2025 "at some point in 2025, we're going to have the best and most performant AI" - Baker: "I don't know if he's in the top hundred... as wrong as it was possible to be"

Meta's failure despite massive investment, Yann LeCun's departure, and "famous billion dollar bids for AI researchers" proves "what these four companies have done is really hard to do"

Microsoft also failed after buying Inflection AI with expectations their "internal models quickly getting better" to run more of Copilot - neither achieved frontier model status

Amazon bought Adept AI and has Nova models but "I don't think they're in the top 20" - illustrating widespread difficulty of frontier model development

"Wild variations in how well companies run GPUs" - keeping large clusters coherent with high utilization is "really hard" with some achieving 90% uptime versus others at 30%, making competition impossible

The Four Leading Labs and Checkpoint Advantage

OpenAI, Gemini, Anthropic, and XAI are "clearly the four leading labs" - each has more advanced internal checkpoint continuously training next model using current best model

"If you do not have that latest checkpoint, you're behind, you're it's getting really hard to catch up" - creating compounding advantage for leaders with reasoning flywheel accelerating separation

"Chinese open source is a gift from God to Meta" - only way to bootstrap without internal checkpoint is using Chinese models, but this advantage disappearing with Blackwell

AI researchers talk about "taste" - "good intuitive sense for the experiments to perform" - critical because experiments on 50,000 GPUs take days with high opportunity cost of wrong choices

China's Rare Earth Mistake and Geopolitical Leverage

"China's made a terrible mistake" trying to force Chinese open source onto domestic Huawei ASICs instead of allowing Blackwell imports, causing DeepSeek v3.2 to cite compute shortage

DeepSeek's technical paper stated "one of the reasons we struggle to compete with American Frontier Labs is we don't have enough compute" - politically risky but necessary message to Chinese government

Gap between American labs and Chinese open source "going to start to widen" with Blackwell deployment, making it harder for anyone else to catch up using Chinese models as bootstrap

China will realize need for Blackwells "in late 26" but by then "enormous effort underway" with DARPA and DoD programs for rare earth alternatives using enzymes or friendly-country deposits

"Rare earths are going to be solved way faster than anyone thinks" - they're "obviously not that rare, just misnamed" and messy to refine, giving America geopolitical leverage as gap widens

XAI's Competitive Position and OpenAI's Cost Problem

XAI has "dominant share" on OpenRouter processing 1.35 trillion tokens versus Google's 800-900 billion and Anthropic's 700 billion in recent period, though OpenRouter represents ~1% of total API market

XAI will release first Blackwell model and be "first ones probably using Blackwell for inference at scale" - important moment establishing leadership position

OpenAI's Stargate announcement reflects "code red" situation - they "pay a margin to people for compute" and those providers "are not the best at running GPUs," making them high-cost token producer

"$1.4 trillion in spending commitments" signals OpenAI knows they'll need to raise substantial capital, especially if Google maintains strategy of undercutting on price - "pretty fast you go from vibes to code red"

OpenAI will release new model but "will not have fixed their per token cost disadvantage yet relative to both XAI and Google and Anthropic" until vertical integration complete

Anthropic's Efficiency and Nvidia Partnership

"Anthropic is a good company" burning "dramatically less cash than OpenAI and growing faster" - deserves credit for efficient execution with Google and Amazon infrastructure partnerships

Anthropic benefited from "same dynamics that Google has" through access to TPUs and Traniums, enabling low-cost token production without vertical integration investment

Recent $5 billion Nvidia deal signals Dario Amodei "understands these dynamics about Blackwell and Ruben relative to TPU" - strategic shift recognizing coming cost advantage reversal

"Nvidia now goes from having two fighters (XAI and OpenAI) to three fighters" - helps in "Nvidia versus Google battle" by expanding customer base for next-generation chips

Data Centers in Space: First Principles Analysis

"The most important thing that's going to happen in the world in the next 3 to 4 years is data centers in space" with "profound implications for everyone building a power plant or data center on planet Earth"

Solar energy in space provides 6x more irradiance than Earth - sun is 30% more intense, satellites can stay in sun 24 hours by positioning, and no atmospheric interference

"Because you're in the sun 24 hours a day, you don't need a battery" - eliminating giant percentage of cost that ground-based solar requires for night operation

"Cooling is free. You just put a radiator on the dark side of the satellite" approaching absolute zero - "majority of the mass and weight" in current racks is cooling infrastructure that disappears

Laser communication through absolute vacuum is faster than fiber optic cables - "you actually have a faster and more coherent network than in a data center on Earth" for linking satellite racks

Direct-to-cell capability demonstrated by Starlink enables phone-to-satellite-to-phone path, eliminating cell tower, base station, metro aggregation, and data center fiber - "much better lower-cost user experience"

"In every way, data centers in space from a first principles perspective are superior to data centers on earth" - power, cooling, networking, and latency all favor space deployment

Space Infrastructure Requirements and Convergence

Primary friction is launch availability: "We need a lot of Starships" - only economically viable launch system, though Blue Origin recently landed booster offering potential alternative

Elon Musk stated "Tesla, SpaceX and XAI are converging" - XAI provides intelligence for Optimus robots, Tesla provides vision and manufacturing, SpaceX provides space data center infrastructure

"Each one is kind of creating competitive advantage for the other" - XAI has built-in Optimus customer relationship (subject to intense vetting as intercompany agreement), space infrastructure advantage, and customer base through Tesla/SpaceX

Training in space will "take a long time just because it's so big" but inference deployments feasible sooner, with training eventually following as satellite networks scale

Power Constraints and Natural Gas Solutions

"Having watts as a constraint is really good for the most advanced compute players" - when power limits deployment, TCO becomes irrelevant and only tokens-per-watt matters, giving best technology unlimited pricing power

"We just can't build nuclear fast enough in America" - NEPA regulations mean "a rare ant that we could move and it could be in a better environment can totally delay construction of a nuclear power plant"

"The solutions are natural gas and solar" - AI data centers for training can locate anywhere, enabling deployment in natural gas basins like Abilene with abundant fracking supply

Caterpillar announced 75% capacity increase for turbine manufacturing over next few years after initial reluctance - "the system on the power side is beginning to respond" to AI demand

"Humans need to come first. We need to have a human-centric view of the world" - Baker's perspective on balancing environmental regulations with infrastructure needs

SaaS Companies' Strategic Failure with AI

"Application SaaS companies are making the exact same mistake that brick-and-mortar retailers did with e-commerce" - refusing to accept lower margins despite customer demand for new technology

Brick-and-mortar retailers saw Amazon losing money and concluded e-commerce would be low-margin because "customers pay to transport themselves to store and transport goods home" versus delivery costs

"Good AI company might have gross margins of 40%" versus SaaS 70-90% margins, but AI companies generate cash earlier through dramatically lower headcount, not high margins

"If you are trying to preserve an 80% gross margin structure, you are guaranteeing that you will not succeed at AI. Absolute guarantee" - AI natives running at 35-40% will win customer relationships

Cloud transition provided existence proof - Adobe and Microsoft both saw margin compression moving to SaaS but investors accepted it as long as gross profit dollars grew and margins eventually improved

"There is room for someone to be a new kind of activist" telling SaaS companies: "Here are my AI revenues, here are my AI gross margins, it's real AI because it's low gross margins"

Salesforce, ServiceNow, HubSpot, GitLab, Atlassian "all of them could run this" playbook - make agents for core customer functions (CRM → customer communication agents) at 10-20% margins accessing existing data

"This is a life or death decision. And essentially everyone except Microsoft is failing it" - other companies' agents are "accessing your systems, pulling data into their system, and then you will eventually be turned off"

Semiconductor Venture Renaissance and Ecosystem

"Your average semiconductor venture founder is like 50 years old" - Nvidia's success and market cap "singlehandedly ignited semiconductor venture" by showing outcomes possible in data center markets

Best architects leaving stable high-paying jobs at public companies to start ventures - "maybe he's the head of networking at a big public company, he's making a lot of money and has a good life" but sees opportunity

"Silicon Valley stopped being Silicon Valley long ago" - Baker's firm "maybe has done more semiconductor deals in the last seven years than the top 10 VCs combined"

Ecosystem critical because "there are thousands of parts in a Blackwell rack" and Nvidia only makes 200-300 of them - "they couldn't go to this one-year cadence if everything else was not keeping up"

"Not even Nvidia can do it alone. AMD can't do it alone. Google can't do it alone" - need transceiver makers, wire makers, backplane makers, laser makers all accelerating together

"First time where every level of the stack that I look at, the most important competitors are public and private" - creates unique investment landscape across entire AI infrastructure

Memory Constraints and DRAM Cycle Risk

"Wave of innovation in memory which is really exciting to see because memory is such a gating factor" - innovations like KV cache offload enabling longer context windows critical for AI progress

Potential natural governor: "First true DRAM cycle since the late '90s" before Taiwan Semi smoothed supply and DRAM became oligopoly with modest price swings

Historical DRAM cycles saw prices increase 10x during shortages when "DRAM wafer valued like 5-carat diamond" - recent cycles only see 30-50% increases or prices just stop declining

"If it starts to go up by X's instead of percentages, that's a whole different game" - could dramatically slow AI deployment if memory becomes severe constraint

AI-Native Entrepreneurs and Productivity

"These young CEOs, they're just so impressive in all ways and they get more polished faster" because they're constantly consulting AI: "How should I pitch this investor? I'm meeting with Patrick O'Shaughnessy"

AI provides expert guidance on difficult situations: "I have this difficult HR situation, how would you handle it?" and "We're struggling to sell our product, what changes would you make?" - "it's really good at all of that today"

"VCs are seeing massive AI productivity in all their companies" because companies are "full of these 23, 24 or even younger AI natives" who use AI fluently for all business functions

"Impressive young people come in and they're where I was as an investor in my early 30s and they're 22" - knowledge accessibility through podcasts and internet plus AI tutoring accelerating development

Rolling Bubbles: Nuclear, Quantum, and Hype Cycles

Since 2020, "series of rolling bubbles" - EV startups (not Tesla) went down 99%, meme stocks like GameStop, now "nuclear and quantum" experiencing speculative excess

Fusion and SMR (small modular reactors) would be "transformative technology" but "none of the public ways you can invest in this are really good expressions of this theme or likely to succeed"

Quantum leaders are "Google, IBM, and Honeywell Quantum" - public quantum companies "are not the leaders" making them poor investment vehicles despite exciting theme

"Quantum supremacy is very misunderstood" - doesn't mean quantum computers better at everything, just "you can do some calculations that classical computers cannot do. That's it. That's going to be really useful" but limited scope

Baker fascinated that "for the last two years, whatever AI needs to keep growing and advancing, it gets" - public opinion on nuclear "changed so fast" exactly when AI needed power solutions

Taiwan Semi's Cautious Mistake and Intel's Opportunity

"Taiwan Semi is not expanding capacity as fast as their customers want" and "in the process of making a mistake" - so paranoid about overbuilding they're constraining AI progress

Taiwan Semi executives "met with Sam Altman and laughed and said he's a podcast bro, has no idea what he's talking about" - skepticism about AI demand leading to conservative capacity decisions

Intel has "all these empty fabs" and new CEO Lip-Bu Tan is "really good executive" reaping benefits of Patrick Gelsinger's strategy - "eventually given shortages those fabs are going to be filled"

"It's shameful that the Intel board fired [Gelsinger] when they did" - Baker believes Gelsinger put Intel on "only strategy that could result in success" before being removed

Taiwan Semi and power constraints act as "natural governors" preventing AI overbuild - "if Taiwan Semi opens up at the same time data centers in space relieve all power constraints... you get an overbuild really fast"

Baker's Path to Investing: History, Climbing, and Competitive Drive

"Investing is the search for truth" - finding truth first that others haven't seen generates alpha, requiring "most thorough knowledge possible of history" intersected with "most accurate understanding of current events"

Early fascination with history - "earliest thing I can remember is being interested in history... Phoenicians, Egyptians, Greeks, Romans and pyramids" leading to intense focus on current events as "applied history"

Original plan: "ski bum in winters, work on river in summers, climb in shoulder seasons, try to be wildlife photographer and write the next great American novel" after leaving college

Parents' only request: "Get one professional internship, just one, we don't care what it is" - led to Donaldson Lufkin Jenrette internship mailing research reports to clients

"Day three I went to bookstore" and consumed Peter Lynch, Warren Buffett letters (twice), Market Wizards, taught himself accounting from Why Stocks Go Up and Down

"This is the only thing I've been vaguely competitive at. I'd love to be good at something else. I'm just not" - picked last for sports teams, not good at skiing despite fortune on lessons, never beat chess players in Cambridge park