May 20, 2026 · 1h 22m · invest-like-the-best
Watts, Wafers, and the Future of AI Infra | Gavin Baker · Invest Like The Best
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of Invest Like The Best, investor Gavin Baker joins host Patrick O'Shaughnessy to explore the unprecedented expansion of artificial intelligence, analyzing infrastructure bottlenecks, semiconductor physics, capital efficiency, and the geopolitical and economic implications of frontier AI models.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Patrick holds 11% of the talking time here. How this is scored →
Patrick as informed peer 4.8 Guest teaching 6.0 Guest disagreement 0.9 Patrick pushing back 0.9
speaking balance: gold is Patrick, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 19:25 Dismissing orbital compute armchair skeptics
Gavin forcefully dismisses critics of space data centers, arguing that naysayers lack the engineering caliber of SpaceX and fail to grasp the practical physics of orbital hardware.
Hardest push from Patrick ▶ 17:05 Host pushes back on orbital compute cooling and repairPatrick directly confronts Gavin's optimistic orbital compute thesis by highlighting the two major engineering objections: heat radiation in vacuum and hardware repairability.
Biggest teaching moment ▶ 23:00 TSMC wafer supply discipline avoiding historical bubble crashesGavin provides an extensive masterclass connecting 300 years of technology bubbles to TSMC's disciplined capital allocation, demonstrating how wafer supply ceilings prevent a market crash.
Patrick holds their own ▶ 45:02 Patrick concisely defines pre-fill and decode mechanicsPatrick demonstrates sharp technical understanding of chip workload trade-offs by immediately synthesizing pre-fill as context ingestion and decode as token generation.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Patrick as informed peer | Guest teaching | Guest disagreement | Patrick pushing back | Why |
|---|---|---|---|---|---|---|
| Market Drawdowns, DeepSeek Reaction, and Exponential Growth Signals | 4 | 6 | 1 | 0 | Patrick sets the stage by asking Gavin to reflect on the unique volatility of March and April. Gavin delivers an extensive comparison of Anthropic's unprecedented ARR growth against classic SaaS giants and breaks down the market mechanics of DeepSeek. | |
| Geopolitics, Energy Infrastructure, and American Manufacturing Competitiveness | 3 | 7 | 1 | 1 | Patrick prompts with a simple inquiry on why the Strait of Hormuz closure benefits the US. Gavin educates him on relative electricity pricing, natural gas dynamics, and comparative manufacturing advantages. | |
| Multiples, Capital Efficiency, and ROIC: Anthropic vs. OpenAI | 6 | 5 | 1 | 1 | Patrick introduces the valuation comparison between Anthropic and OpenAI using classic sales multiples. Gavin differentiates their capital efficiency, structural ROICs, and token generation dynamics. | |
| Fundraising Strategy and Elon Musk’s Valuation "Superpower" | 5 | 5 | 0 | 1 | Patrick asks why leading AI labs do not raise massive rounds at astronomical valuations. Gavin details the strategic value of underpricing rounds to build long-term investor loyalty, citing Elon Musk's historical playbook. | |
| Mid-Sponsor Break: Ramp, Rogo, and WorkOS | 5 | 6 | 2 | 3 | Following the sponsor break, Patrick frames the core infrastructure bottleneck and directly pushes back on orbital compute feasibility by raising cooling and repair constraints. Gavin counters by reframing space data centers as laser-linked satellite racks. | |
| Wafers, TSMC Discipline, and Bubble Dynamics | 4 | 7 | 1 | 0 | Gavin delivers a detailed lesson on historical infrastructure bubbles from canals to 2000 telecom, explaining how TSMC's wafer supply discipline is the primary barrier preventing an overbuild bubble. | |
| Foundries, TerraFab, and Talent Acquisition | 5 | 6 | 1 | 2 | Patrick presses on fab construction lead times and supplier dynamics. Gavin outlines TerraFab's competitive edge via semiconductor equipment partnerships and aggressive global engineering talent recruitment. | |
| Frontier Models, the Pareto Frontier, and the Bitter Lesson | 6 | 7 | 1 | 1 | Patrick inquires why cheaper open-source models have not eroded frontier model pricing. Gavin breaks down the Pareto frontier of intelligence versus cost and explores potential violations of Sutton's Bitter Lesson. | |
| Harness Engineering, Usage-Based Pricing, and Continual Learning | 5 | 6 | 1 | 0 | Patrick brings up continual learning and memory architectures. Gavin explains the shift from flat subscription tiers to usage-based pricing in enterprise AI harnesses, drawing parallels to cellular telecom billing. | |
| Mid-Sponsor Break: Vanta and Ridgeline | 6 | 6 | 1 | 1 | After the sponsor read, Patrick asks about specialized AI chip startups. Patrick demonstrates domain familiarity by defining pre-fill versus decode, while Gavin uses the iron triangle of tank design to explain hardware trade-offs. | |
| Extending GPU Useful Life, Private Credit, and Hyperscalers | 4 | 7 | 1 | 1 | Patrick prompts Gavin to explain how inference disaggregation impacts private credit. Gavin outlines how pairing specialized chips with older Hopper and Ampere GPUs extends useful lives from 3 years to over a decade. | |
| AI Coding Focus, Data Moats, and the Token Path | 6 | 5 | 0 | 0 | Patrick asks how non-infrastructure AI founders are adapting. Gavin evaluates the strategic focus on coding agents, proprietary vertical data moats, and remaining within the critical token path. | |
| Jensen Huang, Open Source Distillation, and Model Game Theory | 6 | 5 | 1 | 1 | Patrick chimes in with strategic business terminology ('commoditize your complement') as Gavin discusses Nvidia's potential to build frontier models, distillation dynamics, and API release game theory. | |
| Preparing for Advanced Models and Cyber Security Threats | 5 | 5 | 1 | 3 | Patrick presses Gavin beyond defensive cybersecurity measures to ask what human analysts will retain over next-generation models. Gavin responds using The Last Samurai analogy of mastering the machine gun. | |
| Deploying AI Agents and Analyzing Management Incentives | 5 | 7 | 1 | 1 | Patrick prompts Gavin on the breakdown of diversity in investor consensus. Gavin explains massive cross-sectional valuation anomalies between DRAM suppliers, semi-cap equipment, and high-cost commodity laggards. | |
| Big Tech AI Landscape and Startup Engagement Tiers | 4 | 7 | 1 | 0 | Patrick asks for quick takes across major public tech giants. Gavin provides an incisive review of Google's compute base, Meta's talent investments, Amazon's Trainium, and Microsoft's Copilot strategy and startup engagement. | |
| Value Destruction at the Application Layer and GPU Ratios | 4 | 6 | 1 | 0 | Patrick asks about second-order knock-on effects. Gavin discusses net value destruction at the application layer, corporate GPU-per-human ratios, personal security risks, and AI drone warfare in Ukraine. | |
| Pax Americana, Rare Disease Medical Breakthroughs, and AI Equity | 4 | 5 | 0 | 0 | Patrick supplies the phrase 'Pax Americana' as Gavin explores historical parallels, breakthroughs in rare disease drug repurposing via AI agents, and the societal imperative of equitable frontier model access. |
Statements from this episode (59)
Baker: Drawdowns Are Either Flawed Theses or Opportunities for Pent-Up Alpha
“There are drawdowns where you're wrong, a company missed estimates, your hypothesis was invalidated, and you have to take your medicine and you crystallize that loss. And then there are drawdowns or periods of underperformance where you're underperforming beca…”
Baker: Anthropic Added $11 Billion ARR in One Month
“And what I just mean by that is in Anthropik, they added eleven billion dollars of AR. And what is astonishing to me about this is that The SaaS and cloud revolution it created, we'll call it between five and 10 trillion dollars of value. And I would say argua…”
DeepSeek Release Doubled Asian AWS GPU Rental Prices Due to Compute Demand
“And that was really strange because by Deep Seek Monday, it was super clear That this was going to be the most positive thing that had ever happened to compute demand. Prices in the AWS availability zones in Asia had already, like, doubled. You were seeing GPU…”
Closure of Strait of Hormuz Is Net Beneficial for US Economy
“That the Strait of Hormuz being closed is actually relatively awesome for America.”
Baker: US is now the world's largest oil and gas producer and exporter
“The US economy, the United States is now the world's largest producer of oil and gas. And we've become now the world's largest exporter of oil and gas.”
Baker: Tech relative valuations hit 10-year lows during recent drawdown
“I think on a relative basis, tech essentially got as cheap as it's been versus the rest of the market has at any point over the last 10 years.”
Gavin Baker: Anthropic has dramatically lower cost per token than OpenAI
“Anthropic clearly is, has a dramatically lower cost per token than OpenAI.”
Baker: Anthropic burned around 80% less capital than OpenAI
“I think they have, they burned maybe 80% less than OpenAI.”
Baker: Anthropic would do over $100B today with unconstrained compute
“Maybe a true statement is that if anthropic had all the compute, they'd probably be doing
Well north of a hundred billion dollars today.
Maybe a 150.”
Baker: Claude Opus generates 70% fewer tokens per question
“There's an analysis Claude is even on Opus is generating 70% less tokens for the exact same question.”
Baker: Anthropic has strongly positive gross margins on inference
“Anthropic I'm sure is at very positive gross margins on inference today.”
Baker: Anthropic will probably generate cash this year
“Anthropic probably starts generating cash this year if they are not already generating cash, which I think is probably the case.”
Baker: SpaceX valuation compounded in the low 30% range annually for a decade
“SpaceX compounded it, you know, low 30% per year for whatever that was, a decade.”
Baker: Anthropic could raise capital at at least a 100% premium
“But could Anthropic raise money? At probably a, at least a 100% premium. To this rumored latest mark? Of course.”
Orbital Compute Will Solve AI Power Shortages Starting in 2027
“The Watts shortage will probably begin to alleviate 27, 28, and then I think orbital compute will really solve that.”
Baker: SpaceX operates 98% to 99% of all satellites in orbit
“SpaceX operates the world's largest satellite fleet, which is like 98 or 99% of all satellites in orbit.”
Baker: AI training will remain terrestrial while orbital compute handles inference
“Inference, I think, is very sensible for orbital compute. Training will be done on Earth for a long time.”
Baker: Jensen Huang has never had a formal contract with TSMC
“Jensen has never had a contract with Taiwan Simi. They do business on what seems fair in handshakes.”
Baker: AI infrastructure build-out is overwhelmingly funded by cash flows, not debt
“One thing I'm really happy about, really good about the current build out is it's still overwhelmingly funded out of operating cash flows, which is a really important fundamental difference versus the year 2000.”
Baker: Every GPU runs at 100% utilization versus 99% unutilized fiber in 2000
“Every GPU is running at a hundred percent utilization when 99% of fiber was unutilized.”
Baker: TSMC wafer supply constraints will prevent an AI bubble
“I have been optimistic that this fundamental shortage of wafers, Which really today is controlled by Taiwan Simi, will prevent one.”
Baker: NVIDIA could sell $2T-$3T in GPUs in 2026-2027 if TSMC expanded enough
“If Taiwan Simi did what Jensen wanted, I think NVIDIA could sell two trillion dollars of GPUs. In 20, in 26 or 27. Maybe two and a half trillion. Maybe three trillion.”
Baker: Intel or Samsung will break discipline in the semiconductor foundry market
“One of Intel and Samsung, they're not going to stay disciplined. They will break, and then at some level that will force everyone else to break.”
SpaceX and Tesla's TerraFab Domestic Semiconductor Joint Venture Will Succeed
“The Turfab, it's a SpaceX, I believe Tesla's involved as well joint venture to build the world's largest fab here in America, and I'm, I think they're going to be successful.”
Baker: Musk built a data center in 122 days versus the industry's three years
“Everybody else has taken three years to build a data center. He built one in a 122 days.”
Gavin Baker: Frontier AI models capture overwhelming majority of model-layer returns
“All the economic returns to AI at the model layer, not all of them, but an overwhelming amount of them have been at the frontier, which is surprising to me”
Baker: Anthropic and OpenAI surpassed Google on the AI intelligence-cost Pareto frontier
“Google dominated that nine months ago. At every point on the Pareto frontier, OpenAI, XAI, and Anthropic were inside of them. Now, the Pareto Frontier is dominated by Anthropic, OpenAI, and then Grok 4.3 is on the Pareto Frontier.”
Baker: Google is likely subsidizing Gemini 3.1 out of pride
“And if I were to bet, I would bet that they were subsidizing that out of pride.”
Baker: Violating Sutton's Bitter Lesson is the biggest risk to AI
“A violation of Richard Sutton's bitter lesson is for sure the biggest risk to this trade. To all of AI.”
Baker: Flat-rate subscriptions deliver a lobotomized version of AI
“And if you're on that 250 or 300 or 280 dollar a month plan or whatever it is,
You were getting severely rate limited.
You were getting a lobotomized version of the AI.”
Baker: OpenAI and Anthropic will probably exceed $200B ARR this year
“So I think the shift to usage based pricing is probably why you will see OpenAI and Anthropic exceed well over two hundred billion dollars in ARR this year.”
Baker: AWS Trainium is doing best among chips trying to beat GPUs
“And you have TPU, Tranium, and AMD, Which are all you know, essentially trying to be a better GPU, and today I think probably Tranium is doing the best.”
Baker: Switched scale-up networks are needed for economical MoE inference
“The Tranium III needs to ramp into production because it has a switch scale up network, which you really need to economically inference MOE models.”
Baker: 1% AI chip market share will equal $100B in value
“Like my rule of thumb is one percent market share is going to be worth a hundred billion.”
Baker: LLM pre-fill is capacity-bound while decode is bandwidth-constrained
“And that is fundamentally a memory capacity bound problem. Decode is the process of generating new tokens, and that is memory bandwidth constraint.”
Baker: Cerebras hardware can theoretically run any model size
“A Cerebrus machine can theoretically run any size model.”
Baker: AI chip startups will easily raise capital after Cerebras IPO
“Everybody's going to get funded after the Cerebris IPO.”
Baker: Private credit SaaS loans need further markdowns
“Well, just, you know, private credit they're in pain from these SaaS loans. And however much they're marked down, they probably need to be marked down more.”
Baker: Disaggregated inference will extend GPU lifespans to 10-15 years
“And the disaggregation of inference means that I think these GPUs are going to have 10 or 15 year lives.”
Baker: AI profits accrue to infrastructure and models, not applications
“Today in that, in Jensen's five layer cake of AI. The profits. They're accruing to energy. They're accruing to data centers. They're accruing to chips. They're accruing to models. Not really accruing to the applications.”
Baker: Coding may be the shortest path to ASI and useful AI
“Bitter lesson adjacent is the fact that coding might be the shortest path to ASI and useful AI. Because if you're really good at coding, you can write yourself code to do anything.”
Baker: Lower frontier token returns will trigger an AI application explosion
“Now on the other hand, if the returns to these frontier tokens relative to other tokens come down, there's going to be an explosion in value creation at the application layer.”
Nvidia Could Build Near-Frontier AI Models Whenever Jensen Huang Wants
“I have a belief that whenever he wants, Jensen can probably get pretty close to the frontier. Patrick O'Shaughnessy: With his own model? Gavin Baker: With his own model.”
Baker: Open-Source AI Frontier Consists of Chinese Models Using Stolen US Tokens
“Open source frontier, which today consists of, you know, Chinese models with stolen American tokens.”
Baker: Alternative Foundries Would Improve Rapidly if NVIDIA or AMD Used Them
“The reality is like if a company like Nvidia or AMD were to ever really, really use one of these other foundries, that foundry would get better really quickly.”
Gavin Baker: Families and companies must establish offline safe words against AI deepfakes
“Everybody needs to have a say for it. Everybody needs to go Leave your digital devices behind. Literally go to the ocean and have a family say for it or a company say for it. And it can't be one that can be like socially engineered.”
Baker: Tracking proxy compensation changes is ideal work for AI
“Going through the proxy, pulling the PSU thing, looking at how it's changed versus all the proxies because there's signal in that. And that's very labor intensive and that's so good for an AI.”
Baker: Semi-cap equipment trades at 40x earnings vs DRAM at single digits
“You have semi-cap equipment companies trading at 40 times next quarter's annualized earnings, and DRAM companies trading at mid-single digit.”
Baker: Highest-cost suppliers rise most in commodity bull markets
“In a real bull market for a commodity. The commodity suppliers with the highest costs go up the most, because it's the most beneficial to them. They go from on the verge of bankruptcy to gushing cash.”
Baker: Astera Labs cannot be a copper loser because its primary product is a switch
“Astera, their biggest product is going to be a switch. You use both copper and optics to connect switches to accelerators. And so definitionally, if you're a switch company or an accelerator company, you cannot be a copper loser because you're going to be on t…”
Baker: Google holds the largest installed base of compute
“They have the biggest installed base of compute.”
Baker: Amazon retail will see P&L efficiencies from robotics within 18 months
“You're going to see like real P&L efficiencies from robotics, Over the next 18 months in their retail business.”
Microsoft Stock Would Be $800 if GPUs Were Exclusively for AI Labs
“Microsoft probably be an 800 dollar stock today if they were using their GPUs to serve open AI Solely open AI and anthropics capacity instead of using them for their own products.”
Baker: Amazon and Nvidia lead big tech in startup engagement by a mile
“The two companies who are the most deeply engaged with startups are Amazon and Nvidia by a mile. Then there's a really intense engagement with Google, their next most intense.”
Baker: AI Has Net Destroyed Trillions in Application-Layer Value
“AI has net destroyed. Even if you count cursor cognition, the most successful AI natives, value has been, trillions of dollars of value has been destroyed by AI at the application layer.”
Baker: Top Value-Creating Companies Have Highest GPU-to-Human Ratios
“The companies that are doing the best today, that are seeing kind of their values increase the most, that are creating economic value, Are the companies with the highest ratio, highest effective ratio of utilized GPUs per human.”
Baker: Ukraine Has the Best Battlefield AI Outside US and Israel
“I think the reason Ukraine is really winning is they have the best battlefield AI outside of probably America and Israel.”