Inside the Trillion-Dollar AI Buildout | Dylan Patel Interview

ELI5/TLDR

Dylan Patel (SemiAnalysis) maps out the entire AI compute supply chain, from who pays for the data centers to who captures the gross profit. The short version: OpenAI needs more compute than it can afford, Nvidia is using its balance sheet to backstop deals, and every layer of the stack is playing a high-stakes game of musical chairs with trillion-dollar balance sheets. If model improvements stall, the US economy takes a hit. If they don’t, the value creation dwarfs anything since industrialization.

The Full Story

The OpenAI-Nvidia Deal and the Capex Arms Race

The headline deal is simpler than it looks. OpenAI has insatiable demand for compute but not the balance sheet to fund it. A single gigawatt of data center capacity costs $50 billion to build and $10-15 billion per year to rent over five-year contracts. Sam Altman wants 10+ gigawatts. The math is staggering.

Nvidia’s play: invest $10 billion in OpenAI equity per gigawatt built. Of that $50 billion build cost, roughly $35 billion flows to Nvidia as hardware revenue. At 75% gross margins, about half of Nvidia’s gross profit from the deal goes right back to OpenAI as equity investment. Nvidia keeps the other half. It’s not round-tripping exactly — Nvidia is effectively discounting prices without lowering list price, while accumulating ownership in what could become a $5-10 trillion company.

“It’s about the highest stakes capitalism game of all time.”

Oracle signed a $300 billion deal with OpenAI on a revenue base of maybe $15-20 billion ARR. If the bet works, Oracle clears $100 billion in pure profit. If it doesn’t, they’re raising debt against a counterparty that might go bankrupt. Microsoft, meanwhile, got cold feet in H2 2024, paused data centers, relinquished exclusivity — then recently plugged back in with a $19 billion Nebius deal because the demand didn’t go away.

Scaling Laws, Diminishing Returns, and the Child Labor Analogy

Patel’s core conviction: scaling laws hold. It takes 10x more compute to reach the next tier of model capability, which looks like diminishing returns on a log-log chart. But the value isn’t linear either. A six-year-old versus a 13-year-old — incrementally just seven years, but the economic output is categorically different. A company staffed entirely with high schoolers who get refreshed every six months can dig trenches. A company of 25-to-30-year-olds can build real businesses.

“If I could have an intelligence as smart as a Google senior engineer, that’s $2 trillion of software value.”

For software, we’re already there-ish. Anthropic went from under $1 billion to $7-8 billion revenue, basically all code-related — Claude Code, Cursor, GitHub Copilot. The fastest revenue ramp in history. But the bigger models (Opus, GPT-4.5) are too slow to serve profitably. Anthropic’s revenue comes from Sonnet, not Opus. User experience wins over raw intelligence.

Tokconomics: Why GPT-5 Is the Same Size as GPT-4o

Patel coined “tokconomics” — the economics of token production. Token demand doubles every two months, but hardware capacity doesn’t. Something has to give.

OpenAI’s choice with GPT-5 was telling. They could have gone bigger (they tried with 4.5, couldn’t serve it). Instead, GPT-5 is roughly the same size and cost as 4o. The strategy: serve way more users at current intelligence, push thinking/reasoning modes for those who need more, and ride the cost curve down.

GPT-3-level intelligence is now 2,000x cheaper to serve than at launch. GPT-4-level is 500-600x cheaper. The cost collapses through algorithmic improvement while capability ratchets upward through new training paradigms.

If you could press one magic button: Patel says capacity/cost beats latency. Current latency is good enough for most use cases. The binding constraint is raw serving capacity.

Reinforcement Learning: We’ve Thrown the First Pitch

Text pre-training is in the early innings. The internet’s text has been consumed, but models can still learn faster from it — like the difference between a kid who reads the textbook once and gets 40% versus 100%.

But the real frontier is reinforcement learning, and Patel says we’ve barely started. There are 40-odd Bay Area startups building RL environments — fake Amazons for purchasing tasks, dirty data for cleaning tasks, math puzzles of escalating difficulty. The environments can be anything: medical case diagnosis, spreadsheet navigation, code debugging.

“My brother just had a baby. This baby will literally stick his hand in his mouth… he’s calibrating the senses on his fingers by sticking his hand in his mouth because his tongue is the most sensitive thing.”

Models haven’t done that calibration yet. They can read the entire internet but can’t navigate a spreadsheet. The data just doesn’t exist in pre-training. You have to build environments and let them fail their way forward — exactly how humans learn.

The grokking concept matters here. Models memorize before they generalize. Make a model too big without enough diverse data and it over-memorizes, never reaching the “aha moment.” The challenge isn’t bigger models — it’s generating the right training data.

Memory, Context, and Deep Research

Humans compress information into incredibly sparse representations. We can’t recall exact sentences but get the gist and translate meaning. Models are the opposite: perfect recall within context, terrible at sparse long-term storage.

Deep research was an early answer. The model works for 45 minutes, outputting millions of tokens, using language to compress information, store it externally, compress more, then synthesize across all compressed notes. It’s a workaround for the sparse memory problem, not a solution.

The real research challenge: how do you train models to operate over infinite context? Maybe through databases they write to and reference. Maybe through novel architectures. Nobody knows yet — which is why these companies need millions of GPUs not to build million-GPU models, but to run a bajillion experiments.

Power, Supply Chains, and West Texas Electricians

AI data centers are 3-4% of US power consumption. Not much. The problem is we haven’t built meaningful new power capacity in 40 years. It’s a supply chain bottleneck, not a demand problem.

The solutions are creative and sometimes absurd: one company is parallelizing diesel truck engines because industrial truck engine capacity is huge and nobody’s tapped it. Elon shipped power equipment from Poland because US supply chains couldn’t deliver. GE is doubling turbine production. Transformer (the electrical kind) supply chains are sold out globally.

Electrician wages have doubled for mobile workers willing to relocate to data center sites. West Texas is having a fracking-era talent moment. Air permit regulations create bizarre constraints — you might fail environmental rules if you run backup diesel generators more than eight hours a month.

China: The Existential Backdrop

Patel’s most striking claim: without AI, the US probably falls behind China as world hegemon by end of decade. The math on debt, consumption, social instability, and industrial decline doesn’t work without dramatic GDP acceleration.

China has been playing the long game for decades — steel, rare earths, solar panels, semiconductors. They’ve dumped $400-500 billion into semiconductor self-sufficiency. ByteDance is the second or third largest GPU user globally. DeepSeek engineers are world-class but paid a fraction of Silicon Valley rates.

China builds faster. Period. They can construct physical infrastructure at a pace the US can’t match. They don’t have the best chips or memory, but they have the most power and construction capacity. Elon moves fast by American standards; he’s slow by Chinese standards.

The geopolitical risk isn’t just about Taiwan. If China blockades or destabilizes Taiwan, the US can’t make refrigerators, cars, or data centers. People say they won’t invest in TSMC because of geopolitical risk — but if Taiwan falls, Apple and Amazon are equally screwed.

The Speed Round

Anthropic over OpenAI — Anthropic’s revenue is accelerating faster because they’re focused on the $2 trillion software market. OpenAI is split across consumer, science, agents, everything.

Meta has the cards — Full stack from hardware (glasses with screens) to models to serving capacity to recommendation systems to capital. The next human-computer interface paradigm is voice-to-AI, and Meta might be the only company that can deliver end-to-end.

Google is waking up — Bearish two years ago, bullish now. Selling TPUs externally, models getting competitive, aggressive infrastructure spend. Has both consumer (YouTube, Android, Search) and professional channels.

XAI is in danger — Biggest individual data center, but no business model beyond anime chatbots. Elon can fund one more round. After that, needs revenue.

Oracle’s bet is binary — In most worlds where OpenAI pays Oracle $300 billion, OpenAI is a $5-10 trillion company. Oracle shareholders are implicitly long OpenAI’s success.

SaaS Is in Trouble

The sleeper insight. AI fundamentally breaks the SaaS model by spiking cost of goods sold (inference costs), maintaining high customer acquisition costs, and cratering the cost for competitors to build equivalent software. China never had a big SaaS market because software development was 10x cheaper — developers earned a fifth of US wages and were arguably twice as good. AI is doing the same thing to every market globally. Many software businesses won’t hit escape velocity.

Google has an edge here: lowest cost of goods sold per token because they own the vertical stack (TPUs). He who controls the platform wins.

Key Takeaways

A gigawatt of AI data center capacity costs ~$50B to build, ~$10-15B/year to rent, and 5-year contracts put $50-75B on the line per gigawatt
Nvidia’s OpenAI deal works out to roughly half of Nvidia’s gross profit going back as equity investment — a disguised discount without lowering list prices
Token demand doubles every two months; hardware capacity does not — algorithmic efficiency must close the gap
GPT-3 intelligence is now 2,000x cheaper to serve than at launch; GPT-4 is 500-600x cheaper
Anthropic’s revenue ($7-8B) is the fastest ramp in history, nearly all from code tools, and most of it flows from Sonnet, not Opus
GPT-5 was deliberately kept the same size as 4o — the strategic choice was more users over smarter model
The capacity/cost bottleneck matters more than latency — current speeds are adequate for most applications
Reinforcement learning is in the first pitch of the first inning — 40+ startups building training environments, but barely scratching the surface of what’s possible
Models grok: they memorize before they generalize, and over-parameterization without diverse data makes them worse, not better
AI data centers are only 3-4% of US power, but we haven’t built meaningful new power capacity in 40 years
China has dumped $400-500B into semiconductor self-sufficiency; ByteDance is the 2nd or 3rd largest GPU consumer globally
Without AI-driven GDP acceleration, Patel argues the US loses hegemonic status to China within the decade
SaaS business model breaks under AI: COGS spike (inference), competitors can now build instead of buy, customer acquisition costs stay high
Google has the lowest token COGS in the industry because of its vertical TPU stack
Meta is the only company with the full stack for the next computing interface: hardware (smart glasses), models, serving capacity, recommendation systems, and capital

Claude’s Take

This is one of the more information-dense conversations about AI infrastructure I’ve encountered. Patel brings something unusual to the table: he’s a supply chain analyst who actually understands the ML research landscape, which means he can connect the cost of a transformer coil to the economics of token production to the geopolitical implications for US-China competition. Most people can do one of those. He does all three.

The child labor analogy is crude but effective. The core insight — that each tier of model capability is categorically more valuable despite costing an order of magnitude more compute — is the central bull case for AI capex. He’s right that “diminishing returns” is a misleading frame when the value function is step-wise, not linear.

Where I’d push back: the “US falls without AI” thesis feels a bit dramatic. The US has structural advantages (reserve currency, geographic isolation, immigration, institutional depth) that don’t disappear in a decade. But the directional argument — that AI is the best shot at growing the pie rather than fighting over it — is solid.

The SaaS insight at the end deserves more airtime than it got. If inference costs remain high and development costs crater, the entire software industry reprices. That’s not speculative — it’s already happening in China, and Patel is essentially saying AI will do to the US SaaS market what cheap Chinese developers did to China’s.

Score: 8/10. Deep domain knowledge, concrete numbers, good frameworks. Loses a point for some rambling sections and the anime chatbot monetization tangent. Gains it back for being one of the few people who can credibly talk about diesel truck engines and scaling laws in the same breath.