Invest Like the Best with Patrick O'Shaughnessy

Dylan Patel - Inside the Trillion-Dollar AI Buildout - [Invest Like the Best, EP.442]

September 30, 2025

Key Takeaways Copied to clipboard!

  • The massive capital flows underpinning the AI buildout, exemplified by the OpenAI-Nvidia-Oracle deals, are driven by an insatiable, front-loaded demand for compute capacity that requires massive balance sheets to finance. 
  • The value progression of AI models is not linear; the jump from a less capable model (like a six-year-old) to a significantly more capable one (like a 16-year-old) yields a drastic, non-diminishing increase in utility and economic value. 
  • The current phase of AI development is shifting from pure pre-training scaling (using existing internet data) toward post-training reinforcement learning (RL) in simulated environments, which is necessary to teach models complex, iterative reasoning and skills not present in static datasets. 
  • The immense value in AI research and semiconductor manufacturing stems from navigating an impossibly large search space, making the intuition and efficiency of a few key individuals disproportionately valuable, justifying massive compensation packages. 
  • Power dynamics in the AI ecosystem are complex and fluid, involving frenemy relationships between model providers (like Anthropic) and application layers (like Cursor), and a fundamental flow of gross profit dollars from software layers down to the hardware layer (NVIDIA). 
  • The US critically needs the massive GDP acceleration promised by AI to overcome internal social instability and maintain global hegemony against formidable, long-term strategic competitors like China, which is aggressively building an insular semiconductor supply chain. 
  • Geopolitical risk surrounding Taiwan's semiconductor manufacturing capability poses a significant threat to the entire tech ecosystem, potentially outweighing the risk associated with investing in TSMC itself. 
  • The traditional SaaS business model faces a reckoning due to soaring AI-related Cost of Goods Sold (COGS) and persistent high Customer Acquisition Costs (CAC), making it difficult for new software-only businesses to achieve profitability. 
  • Innovation in hard tech, such as battery chemistry (Periodic Labs) and power delivery/networking infrastructure, is critical for unlocking the next major leaps in AI application, like ubiquitous AR/VR interfaces. 

Segments

OpenAI-Nvidia-Oracle Deal Mechanics (Unknown)
Copied to clipboard!
  • Key Takeaway: None
  • Summary: None
AI Scaling Laws and Model Utility
Copied to clipboard!
(00:08:04)
  • Key Takeaway: The utility gain from scaling compute follows a log-log curve where a 10x compute increase might only yield a step-change in capability, analogous to the value difference between a six-year-old and a 16-year-old.
  • Summary: The tech giants believe that continued compute scaling leads to better models, even if the returns appear diminishing on a log-log chart. In software development, current models are already highly valuable, comparable to augmenting staff with 30-year-old senior engineers. However, serving larger, smarter models often results in poor user experience due to inference latency, forcing companies like OpenAI to prioritize cost reduction and serving more users over simply making the model larger.
Inference Demand and Tokenomics
Copied to clipboard!
(00:17:31)
  • Key Takeaway: The rapid doubling of token demand every two months necessitates significant cost decreases in inference compute to maintain user adoption curves for existing model capabilities.
  • Summary: The economics of tokens—or ’tokenomics’—involve balancing compute spend, gross profit, and the value created by the intelligence output. While algorithmic improvements rapidly decrease the cost of serving existing model quality, user experience (latency) is a critical trade-off against model size and cost. OpenAI must balance serving larger, smarter models that users might not adopt due to slowness against serving faster, slightly less capable models to maximize user adoption.
Pre-training vs. Post-training Inning
Copied to clipboard!
(00:29:17)
  • Key Takeaway: Text-based pre-training is in the late innings, but the next major compute consumption phase, reinforcement learning (RL) via environments, is only in the second inning.
  • Summary: While text data for pre-training may be nearing saturation, there is vast untapped potential in scaling multimodality (video, audio, images). The critical bottleneck for future gains is the RL paradigm, where models learn through iterative interaction in specific environments, similar to how humans learn by trial and error. Building these complex, task-specific environments is a significant engineering challenge that will drive future compute usage.
Memory, Context, and Reasoning
Copied to clipboard!
(00:38:37)
  • Key Takeaway: Models currently excel at exact recall within their context window, but struggle with the sparse, compressed reasoning over infinite context that characterizes human memory.
  • Summary: Unlike humans who collapse information densely and sparsely, current transformer models rely on calculating attention across their entire context length, making infinite context prohibitively expensive. The ability for models to write information to external databases and recall it later, mimicking human reliance on notes, must be learned through engineered environments, not just pre-training. This ability to compress and recall information is fundamental to advanced reasoning capabilities.
Bullishness and Timeline Uncertainty
Copied to clipboard!
(00:43:42)
  • Key Takeaway: The speaker is among the most bullish regarding the ultimate upper limit of AI capability but is more cautious than leading researchers regarding the timeline for achieving Artificial General Intelligence (AGI).
  • Summary: Significant economic value will be created by current model capabilities, such as automating software migration, even if AGI is decades away. The speaker believes that researchers like Sam Altman are significantly more bullish on the near-term arrival of AGI (under a thousand days) than he is. The path to true human-level intelligence requires overcoming unknown unknowns, particularly in embodiment and tactile feedback, which are currently far beyond current model capabilities.
Robotics and Data Flywheel
Copied to clipboard!
(00:47:25)
  • Key Takeaway: Robotics value creation requires basic pick-and-place functionality before complex dexterity, and the sector needs its data flywheel to accelerate scaling.
  • Summary: Robotics does not immediately require human-level dexterity like perfectly handling a wine glass to be valuable; simple recognition and placement are sufficient initially. The robotics world is still warming up, far from achieving significant scaling because the necessary data flywheel has not fully engaged. Success in this area depends on iterating through experiments to learn which parameters to adjust.
Talent Wars and Compute Efficiency
Copied to clipboard!
(00:47:49)
  • Key Takeaway: Top AI researchers command massive compensation because small efficiency gains across multi-billion dollar compute fleets yield enormous savings.
  • Summary: The talent war in AI is intense, with top researchers earning huge sums because their insights can save significant compute costs across massive training and inference fleets. Adding more people to complex AI research often slows progress due to the difficulty in coordinating experiments and interpreting noisy data. Companies like Meta have historically wasted compute on failed experiments due to poor leadership or inefficient research processes.
Process Knowledge Acquisition
Copied to clipboard!
(00:50:31)
  • Key Takeaway: Acquiring process knowledge from experts in established manufacturing (like semiconductor fabs) may lead to aggressive ‘aqua-hiring’ globally by leading AI firms.
  • Summary: There is a suggestion that companies should aggressively acquire the process knowledge held by experts in established, complex manufacturing fields, such as semiconductor fabrication. This knowledge extraction is crucial because so much value depends on the nuanced process understanding of a small group of people. This talent competition should extend beyond AI labs to include global experts in critical physical technologies.
Human Capital vs. Capital Goods
Copied to clipboard!
(00:51:13)
  • Key Takeaway: The massive financial incentives in AI are redirecting top human capital away from traditional high-skill fields like specialized medicine and deep science (like Intel’s past challenges).
  • Summary: The high compensation in AI research is drawing smart individuals who might otherwise have pursued fields like specialized medicine or traditional hardware engineering. Historically, industrialization saw human capital decrease relative to capital goods, but AI reverses this by creating immense value through human-like capital. NVIDIA’s success exemplifies the modern dynamic where value accrues to the idea creators (design) rather than the labor exporters (manufacturing).
ML Research and Manufacturing Analogy
Copied to clipboard!
(00:54:16)
  • Key Takeaway: ML research mirrors semiconductor manufacturing as both involve iterating through an impossibly large search space by tuning thousands of knobs based on intuition.
  • Summary: ML research involves tuning thousands of knobs regarding data mixing, architecture, and context length, making exhaustive testing impossible, similar to process optimization in semiconductor manufacturing. Both fields require significant upfront R&D cost (wasted compute or wafer runs) that teaches the necessary intuition for future high-volume economic value. The search space for both chip design and model architecture is too vast to explore systematically.
Power Dynamics in the AI Ecosystem
Copied to clipboard!
(00:56:12)
  • Key Takeaway: Power in the AI ecosystem flows toward those who control the scarce resources, whether it is talent, compute, or customer data, leading to complex ‘frenemy’ relationships.
  • Summary: In the Cursor/Anthropic relationship, while Anthropic captures immediate gross profit dollars from model usage, Cursor captures crucial user interaction data, creating a complex power balance. The Microsoft/OpenAI dynamic is similarly complex, involving IP sharing agreements contingent on achieving AGI, a bar that constantly moves as capabilities improve. NVIDIA currently holds king status by capturing the most gross profit, but they are constrained by their inability to make large acquisitions.
NVIDIA’s Balance Sheet Strategy
Copied to clipboard!
(01:02:09)
  • Key Takeaway: NVIDIA uses its massive cash flow and balance sheet to backstop compute clusters for high-growth, capital-intensive startups like OpenAI, securing future GPU demand.
  • Summary: NVIDIA prefers funding startups that spend the majority of their funding rounds directly on compute, as this guarantees immediate GPU sales. To secure long-term commitments from capital-constrained but high-potential entities like OpenAI, NVIDIA engages in unusual financing, such as backstopping entire clusters. This strategy allows them to win future business even if the customer might otherwise explore alternatives like custom ASICs or TPUs.
Risk in the Middle Layer
Copied to clipboard!
(01:09:05)
  • Key Takeaway: Cloud providers (neo-clouds) face extreme risk by relying on short-term contracts, whereas those securing long-term, balance-sheet-backed commitments (like Nebius with Microsoft) capture superior gross profits.
  • Summary: The neo-cloud business model is either amazing or terrible depending on contract structure; short-term sales yield high initial margins but risk collapse when next-gen chips arrive. Long-term contracts, like Nebius’s $19 billion deal with Microsoft, provide guaranteed, high gross profit dollars because Microsoft’s creditworthiness is seen as superior to that of many AI startups. Companies like Coreweave must balance high-margin but risky contracts (like OpenAI) against secure ones (like Microsoft).
AI Enabling New Capabilities
Copied to clipboard!
(01:13:04)
  • Key Takeaway: AI is already enabling entirely new capabilities, such as accelerating drug discovery and making previously impossible data analysis tasks feasible.
  • Summary: The current phase of AI is moving beyond simply making existing engineering tasks faster (the ‘schemorphic era’) to enabling things previously impossible with deterministic code. AI was instrumental in accelerating the discovery of the COVID vaccine and is being applied to material science and complex optimization problems. Dylan Patel’s own business, tracking global data center buildouts via satellite imagery and regulatory filings, would not be possible without modern AI tools.
The Gigawatt Data Center Boom
Copied to clipboard!
(01:17:16)
  • Key Takeaway: The current AI buildout is forcing the US to relearn how to rapidly construct power infrastructure, leading to supply chain bottlenecks and wage inflation for specialized labor.
  • Summary: Data centers currently consume a small fraction of US power, but the rapid, concentrated demand is straining an electrical grid that has seen little investment for decades. This demand is causing electrician wages to double and forcing creative, sometimes inefficient, power solutions like deploying parallel diesel truck engines. Supply chain constraints are evident everywhere, from transformer coils to gas turbines, forcing companies to invest heavily to expand capacity.
US vs. China AI Geopolitics
Copied to clipboard!
(01:23:47)
  • Key Takeaway: The US needs AI to dramatically accelerate GDP growth to solve internal instability and maintain global leadership, whereas China plays a long game focused on insular supply chain dominance.
  • Summary: Without AI acceleration, the US risks losing global hegemony due to unsustainable debt and internal social fragmentation driven by inequality. China’s strategy involves dumping massive state capital into building an insular semiconductor ecosystem over the long term, unlike the US, which focuses capital on the customer-facing application layer. The geopolitical risk surrounding Taiwan’s chip production poses an existential threat to the US economy if access were lost, highlighting the need for US supply chain security.
US-China Talent and Taiwan Risk
Copied to clipboard!
(01:34:15)
  • Key Takeaway: If the US pushes China too hard on talent acquisition, China possesses the domestic talent pool to build a superior AI compute cluster if Taiwan is lost.
  • Summary: Meta’s alleged poaching of Chinese talent highlights the talent concentration risk. If the US loses access to Taiwan, China could build a larger AI cluster given compute is the primary factor. Geopolitical risk makes investing in TSMC difficult, but avoiding TSMC implies avoiding Apple, Amazon, and Google due to supply chain dependency.
Favorite AI Bear Perspectives
Copied to clipboard!
(01:34:57)
  • Key Takeaway: AI researcher ‘gods’ like Yann LeCun are considered AI bears because they argue current autoregressive pre-training methods alone will not achieve AGI.
  • Summary: Dylan Patel respects AI researcher bears but believes they are wrong about the ultimate outcome, though he agrees that pure web-scale pre-training is insufficient for AGI. Some investors are bearish on the current spending but invest ruthlessly based on market perception, such as buying Oracle ahead of OpenAI earnings.
Startup Focus: Hard Tech
Copied to clipboard!
(01:36:42)
  • Key Takeaway: Significant value creation lies in accelerating hard tech bottlenecks, exemplified by Periodic Labs applying RL to battery chemistry optimization.
  • Summary: Periodic Labs, founded by former OpenAI and Google personnel, is using reinforcement learning principles on real-world physical testing cycles to improve battery chemistry efficiency. Breakthroughs in hard tech like batteries are gated by slow physical iteration cycles compared to digital RL flywheels. Advancing battery technology is seen as crucial for enabling high-powered, face-worn computing devices.
Hardware Bottlenecks and Innovation
Copied to clipboard!
(01:39:17)
  • Key Takeaway: The software supporting semiconductor manufacturing tools is often outdated, presenting an opportunity for acceleration, while direct accelerator competition against NVIDIA is capital-intensive and unlikely to succeed.
  • Summary: The software stack within semiconductor fabrication is significantly less advanced than the hardware tools themselves, suggesting areas for AI-driven improvement. Competing directly with NVIDIA in the accelerator space is viewed as too capital-intensive without a revolutionary leap. Innovation is needed in power delivery (transformers) and chip-to-chip networking to handle massive context length memory demands.
World Models and Manufacturing Data
Copied to clipboard!
(01:42:30)
  • Key Takeaway: Breaking the slow experimentation cycle in hardware manufacturing requires cultural change and better world models that simulate physics accurately, rather than relying solely on classical methods.
  • Summary: Internal data sharing barriers at companies like Intel slow down the crucial experiment-analyze-refine cycle necessary for lithography improvements. World models can simulate complex physical realities, such as molecules or chemical reactions, offering an AI-driven alternative to slow, expensive physical testing. This simulation capability is applicable across robotics, chemistry, and materials science.
Speed Round: Company Views
Copied to clipboard!
(01:44:35)
  • Key Takeaway: Anthropic is currently viewed as executing better on the software monetization side than OpenAI, while Google is becoming bullish due to aggressive awakening across hardware, models, and infrastructure.
  • Summary: Meta is positioned uniquely to own the next human-computer interface paradigm shift (voice/intent-based interaction) because it controls the full stack: hardware (glasses), models, serving capacity, and recommendation systems. AMD is loved as an underdog despite being ‘mid’ currently, while XAI faces a capital raising challenge unless its business model evolves beyond current monetization strategies. Google is seen as well-positioned to capture both consumer and professional AI markets.
Reckoning for Software Economics
Copied to clipboard!
(01:52:28)
  • Key Takeaway: The high COGS associated with AI tokens and persistent high CAC will likely prevent many new AI SaaS businesses from achieving the critical mass needed to amortize costs, unlike the traditional SaaS model.
  • Summary: The economics of traditional SaaS, characterized by flat R&D and low COGS, are fundamentally broken in the AI era due to soaring token costs. In markets like China, low software development costs historically suppressed the SaaS model, suggesting AI development cost reduction could similarly favor building over renting. Platform owners, like YouTube, are better positioned than application builders within existing platforms because they control the content generation and distribution layer.
Closing Reflection on Kindness
Copied to clipboard!
(01:57:50)
  • Key Takeaway: The kindest thing anyone has done for Dylan Patel is his brother unconditionally loving and correcting him despite his self-admitted tendency toward being inconsiderate and task-unoriented.
  • Summary: Dylan Patel acknowledges wrestling with being overly focused in the moment, which often leads to being inconsiderate of others’ feelings or forgetting tasks. His brother consistently provides unconditional love and course correction when he acts like an ‘asshole.’ This unconditional support helps compensate for his personal weaknesses in task orientation and social consideration.