All podcasts / Invest Like the Best / Summary

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

2025-09-30 - 117 min - source - Read full transcript

Patrick O'Shaughnessy (host)Dylan Patel

Key insights

The OpenAI-Nvidia-Oracle deals are not simple round-tripping - Nvidia effectively subsidizes OpenAI's compute costs by returning roughly half its gross profit from the deal as an equity stake.

Patel walks through the math: building one gigawatt of capacity costs about $50B, of which roughly $35B goes to Nvidia as hardware revenue at 75% gross margin (about $30B gross profit). Nvidia's $100B equity commitment into OpenAI effectively hands back close to half of that gross profit, so OpenAI gets to pay for a large chunk of its compute in equity while Nvidia still keeps the CAPEX dollars up front and gains ownership in a company that may or may not ever be able to pay its committed contracts.

ai-capex-and-power-dynamics

OpenAI's compute deals are structured as five-year rental commitments costing $10-15B per gigawatt per year, which is why it needs balance-sheet-rich partners to front the capital.

At $10-15B/gigawatt/year over five years, one gigawatt of committed capacity costs $50-75B total. When Sam Altman talks about needing 10+ gigawatts, the cash commitment runs into the hundreds of billions, which OpenAI cannot self-fund - explaining why Oracle, Microsoft, and infrastructure funds are being pulled in as balance-sheet backers rather than OpenAI building everything itself.

ai-capex-and-power-dynamics

Model capability gains are better understood as discrete tiers unlocked by roughly 10x compute jumps than as smooth diminishing returns.

Patel compares tiers to a six-year-old versus a sixteen-year-old: the leap in usable value is drastic even though the underlying compute increase is 'only' an order of magnitude. But this cuts both ways - going from $50B to $500B in spend without algorithmic improvement buys just one such tier, which is exactly what makes the market nervous about whether that next $500B will generate a comparable jump in economic return.

scaling-laws-and-model-progress

GPT-5 held model size roughly flat versus GPT-4o because OpenAI's real constraint was serving capacity, not raw intelligence.

OpenAI's earlier attempt at a much bigger model (GPT-4.5) was smarter but too slow and expensive to serve at scale - user experience collapsed. For GPT-5, OpenAI kept the model close to 4o's size and cost so it could serve far more users and move them up the adoption curve, offloading capability gains into optional 'thinking' (reasoning) modes rather than a bigger base model.

scaling-laws-and-model-progress

Token demand is compounding roughly every two months while hardware supply cannot keep pace, which is the core problem 'tokenomics' is meant to describe.

Patel frames tokenomics as the interaction between compute available, the intelligence tier being served, and the resulting cost/value per token: a fixed gigawatt of capacity can serve a huge volume of a cheap/bad model, a moderate volume of a good model, or a tiny volume of an excellent model. Because a given intelligence tier's serving cost falls with algorithmic progress (GPT-3-level quality is now roughly 2,000x cheaper to serve), demand keeps outrunning the hardware buildout even without model quality improving.

tokenomics-and-inference-economics

If forced to pick one lever to unlock more AI value, Patel says capacity/cost beats latency, because current response speed is already good enough for most use cases.

He argues that if inference latency were 10x lower, OpenAI could have shipped a 10x bigger GPT-5 at the same perceived speed - but that would only recreate the same capacity bottleneck at a higher cost tier. He personally defaults to Claude Sonnet over the smarter but slower Opus specifically because of this latency-versus-capability trade-off.

tokenomics-and-inference-economics

The next wave of model improvement is coming from reinforcement-learning 'environments' - synthetic tasks that teach skills the internet's text corpus never captured.

Examples include a fake Amazon storefront for purchasing-decision practice, messy spreadsheets for data-cleaning practice, and graded medical cases. Roughly 40 startups are now building these environments, and Patel credits this shift (not raw pretraining scale) for the rapid jump in model math ability between Q4 of last year and Q2 of this year - much of it models learning to write code (e.g. Python) to solve problems rather than memorizing answers.

scaling-laws-and-model-progress

Power dynamics between application-layer companies and foundation model labs run in both directions - it is not simply the model provider that holds all the leverage.

Using Cursor and Anthropic as the case study: Cursor sends most of its gross profit to Anthropic for model access, but Cursor owns the user relationship, the usage data, and can switch to a different model provider (or train its own embedding/autocomplete models) at will. Patel calls the resulting relationship 'frenemies' - similar to the ambiguous OpenAI/Microsoft dynamic.

talent-and-power-in-ai

Neocloud business models diverge sharply based on contract duration and counterparty balance-sheet quality.

Short-term GPU rental contracts look extremely profitable in year one but collapse in value once the next Nvidia generation ships at 10x the speed for 3x the cost, forcing prices down. Long-term contracts with strong counterparties are the durable model - Patel cites Nebius's roughly $19B Microsoft-backed deal (at least $6B of gross profit) versus CoreWeave's riskier OpenAI-backed contracts, since 'the market literally believes Microsoft will pay its obligations before the US government.'

ai-capex-and-power-dynamics

Patel believes the US needs AI-driven GDP growth to avoid a slow-motion social and economic unraveling, because the alternative is fighting over a fixed pie.

He points to unsustainable debt, slowing productivity growth, and rising social instability (partly from visible income inequality amplified by social media and algorithmically fragmented culture) as reasons the US cannot simply hold its current trajectory. His summary framing: 'if we don't accelerate, we die' - not because AGI is required, but because AI-driven productivity gains are what let the economy keep growing the pie rather than dividing a shrinking one.

us-china-ai-race

China's AI strategy is fundamentally different from the US's - it prioritizes an insular, self-sufficient supply chain over a single largest compute cluster.

China has spent an estimated $400-500B over a decade building its own chip ecosystem (through SOEs, tax policy, land grants, and state venture funds) even though its chips and memory still lag the US/Taiwan/Korea by several years. Patel argues this mirrors China's earlier playbooks in EVs, steel, solar, and rare earths: a willingness to subsidize a strategic industry for a decade-plus until it wins the global market, rather than optimizing near-term ROI.

us-china-ai-race

AI-driven software cost collapse threatens the classic SaaS business model by adding a large new COGS line while eroding the moat of custom software development.

Traditional SaaS profitability depended on flat R&D and low COGS scaling against high customer acquisition cost, which pays off once a company hits critical mass. AI breaks both sides at once: falling development costs make it easier for customers to build competing tools themselves (as China's historically cheap developer labor already discouraged SaaS adoption there), while AI inference adds a heavy, hard-to-amortize COGS burden - meaning many software-only businesses may never reach SaaS-style escape velocity, while platforms with existing scale and distribution (e.g. YouTube) become relatively more advantaged.

tokenomics-and-inference-economics

Media referenced

K-pop Demon Hunters - movie - Patel's friend brought it up as an example of drama; he says the real-world AI power struggles between labs and hyperscalers are a better soap opera than this.

Companies

OpenAI - Central case study for the episode - insatiable compute demand, the Nvidia and Oracle deals, the Microsoft relationship, and the bet that frontier intelligence justifies hundreds of billions in committed spend.
Nvidia - Discussed as the industry's dominant gross-profit capturer; its $100B equity commitment to OpenAI is unpacked as effective price subsidization rather than simple round-tripping.
Oracle - Signed a $300B compute deal with OpenAI and is taking on CAPEX and debt risk betting that OpenAI's revenue will scale to cover it.
Microsoft - Pulled back from being OpenAI's exclusive compute provider in 2024, then re-engaged; holds a 49% profit-cap structure and IP-sharing terms with OpenAI that Patel calls hard to fully parse.
Anthropic - Patel says he's more bullish on Anthropic than OpenAI because its revenue (driven by Claude Code and coding use cases) is growing faster and is more concentrated on the $2 trillion software market.
Meta - Discussed for its superintelligence lab talent wars, its new display glasses, and Patel's view that it has the closest-to-complete stack (hardware, models, serving, recommendation know-how) to own the next computing interface.
Google - Patel says he flipped from bearish to bullish given TPU sales, aggressive infra investment, and its position to capture both consumer (Android, YouTube) and enterprise AI markets.
Amazon - Cited as an example of a company that has wanted to migrate mainframes to cloud for two decades and still hasn't fully done it - illustrating how slow enterprise migration is even with obvious ROI.
Apple - Referenced alongside Nvidia as a company that has exported manufacturing labor to Asia while keeping nearly all the value/profit domestically.
TSMC - Discussed as inseparable from US tech geopolitical risk - if Taiwan falls, the entire US AI and consumer electronics stack loses its chip supply.
AMD - Patel calls them 'pretty mid' as a business today but says he has a personal soft spot from AMD being his first multi-bagger stock trade.
xAI - Discussed as compute-rich (building the largest single data center) but at risk of running out of capital without a real business model beyond Grok's companion product.
CoreWeave - Cited as a neocloud whose early lenders should have taken equity instead of debt; it lost Microsoft as a customer then found new demand from Google and OpenAI.
Nebius - Signed a ~$19B GPU deal with Microsoft after Microsoft briefly paused and then restarted data center buildout, illustrating the value of long-term, well-backed compute contracts.
Cursor - Used as the case study for the ambiguous power dynamics between AI application layer companies and the foundation model providers (Anthropic) they depend on.
SemiAnalysis - Dylan Patel's own firm; he describes how AI-driven vibe coding let a 3-person team build an image-recognition/satellite-tracking data center business that would have taken 50-100 people before.
Periodic Labs - Patel's most recent investment - an RL-for-real-world-chemistry startup applying the environments paradigm to physical experiments like battery chemistry instead of purely digital tasks.
ByteDance - Cited as possibly the second- or third-largest GPU buyer in the world after OpenAI and Meta, illustrating China's serious (if less publicized) compute scale.
DeepSeek - Referenced as an example of algorithmic efficiency gains (5-600x cheaper serving cost) that don't require bigger models.

Techniques and frameworks

Tokenomics - Patel's framework for reasoning about AI economics: given a fixed amount of compute (a gigawatt), how much intelligence and how many tokens can you serve, and what is the resulting gross profit and value created.
Scaling laws (log-log compute vs. quality) - The 10x-compute-per-capability-tier framework Patel uses to explain why moving from $50B to $500B in spend only buys one step up in model quality absent algorithmic gains.
Reinforcement learning environments - Synthetic task environments (fake e-commerce sites, data-cleaning exercises, math puzzles, medical case grading) used to teach models skills that pure internet-text pretraining cannot, now being built by roughly 40 startups.
Grokking - The phenomenon where a model memorizes data before it generalizes; Patel uses it to argue that making models bigger without better data just produces more memorization, not more understanding.
Aggregation theory - Applied to explain why OpenAI and similar consumer AI platforms can absorb near-term losses (like YouTube did) while building toward monetizing a massive user base later.
World models - Models trained to simulate physical or chemical reality (not just video) so that other AI systems - robots, chemistry, materials science - can learn via simulated experience rather than only real-world trial and error.

Summary

Dylan Patel, founder and CEO of SemiAnalysis, returns to Invest Like the Best for a wide-ranging tour of the physical and financial mechanics underlying the AI buildout. The conversation opens with the OpenAI-Nvidia-Oracle circular deals, which Patel argues are widely misunderstood as simple "round-tripping." He walks through the actual cash flows - Nvidia's $100B equity commitment effectively hands back roughly half its gross profit from the deal, subsidizing OpenAI's compute costs without Nvidia visibly cutting its prices - and explains why OpenAI's five-year, $10-15B-per-gigawatt-per-year compute commitments require balance-sheet-rich partners like Oracle and Microsoft to front the capital OpenAI cannot self-fund.

From there the discussion moves into what Patel calls "tokenomics": the relationship between available compute, the intelligence tier being served, and the resulting cost and value per token. He explains why GPT-5 held model size roughly flat compared to GPT-4o rather than scaling up - OpenAI's real bottleneck was serving capacity and latency, not raw model quality, so it optimized for broader access and added optional reasoning modes instead of shipping an unservably large model. This leads into a detailed explanation of why reinforcement-learning "environments" (synthetic tasks like fake e-commerce sites, data-cleaning exercises, and graded medical cases) are now the primary lever for model improvement, since the open internet's text has been largely exhausted as a pretraining source. Patel estimates the industry is in the earliest innings of this environments-driven RL paradigm.

The episode's middle section covers power dynamics across the AI stack: the ambiguous relationship between application-layer companies like Cursor and their underlying model providers like Anthropic; the divergent economics of "neocloud" GPU rental businesses depending on contract length and counterparty quality (with Nebius's Microsoft-backed deal held up as the durable model versus riskier OpenAI-backed contracts); and a candid speed-round of Patel's views on OpenAI, Anthropic, AMD, xAI, Oracle, Meta, and Google, in which he says he is now more bullish on Anthropic than OpenAI given its concentration on the $2 trillion software market, and has flipped from bearish to bullish on Google.

A substantial section addresses the physical constraints of the buildout - power. Patel notes data centers still represent only a few percent of US power consumption, but that the country has essentially forgotten how to build power infrastructure after four decades of underinvestment, leading to strained supply chains for turbines, transformers, and skilled electricians (whose wages have doubled), and improvised solutions like paralleled diesel truck engines for emergency generation. This bleeds into a substantial geopolitical section on US-China competition: Patel argues the US needs AI-driven GDP growth to avoid a genuine social and economic unraveling, while China is playing a longer, more insular game of supply-chain self-sufficiency that has already worked in EVs, steel, solar, and rare earths - and that a Taiwan contingency would instantly collapse the US AI and consumer electronics stack given its total dependence on TSMC-made chips.

The episode closes with Patel's argument that AI-driven cost collapse threatens the traditional SaaS business model itself - falling software development costs make in-house building more viable for customers, while AI inference adds a heavy new cost-of-goods-sold line that many software companies will struggle to amortize, favoring incumbents with existing scale and distribution. He ends on a personal note, naming his brother as the person who has most kept him accountable and grounded despite what he describes as his own difficulty with task orientation and consideration for others.

Notable Quotes

"It's about the highest stakes, like, capitalism game of all time." - Dylan Patel

"I have like a very pessimistic view that if we don't accelerate, we die." - Dylan Patel

"The talent war should actually be, it shouldn't be meta and open AI. It should be like the U.S." - Dylan Patel

"Tech is the most deflationary thing in the world ever, right? In terms of quality of life, it gets cheaper way faster than the revenues go up. But the revenues still go up." - Dylan Patel