Dylan Patel — Anthropic & OpenAI will have most of the world’s compute by 2028

In one sentence

The AI revolution is transitioning from a venture-funded research phase to a capital-intensive infrastructure war where a handful of labs will control the vast majority of the world's usable compute.

Why read this (and who it's for)

This conversation provides a "macro" view of the AI industry, moving beyond model architecture to the brutal reality of lab economics. It is for investors, researchers, and tech leaders who need to understand the physical and financial constraints (power, chips, and interest rates) that will dictate the pace of AGI. You will be able to forecast the "compute-moat" and understand why the winners of the AI era are being determined by their ability to secure physical infrastructure.

The big idea

Dylan Patel argues that we are entering a period of extreme centralization of compute. Because the marginal cost of running a model is falling while the revenue generated by high-performing models (like GPT-4o or Claude 3.5) is skyrocketing, the "big labs" (OpenAI and Anthropic) are transitioning from loss-making startups to profit-generating titans. This profit allows them to reinvest in more compute, creating a feedback loop where they capture an increasing percentage of all new global compute.

The logic follows a "winner-take-all" trajectory: as the labs become the primary customers for high-end hardware, they gain the leverage to pay the highest prices for scarce resources (like ASML machines and power). By 2028, Patel predicts that these two entities—and their infrastructure partners—will control the majority of the world's usable "flops," effectively creating a private monopoly on the raw materials of intelligence.

Key ideas

The Transition to Profit-Funded Growth

The labs are moving from a regime of "venture-funded losses" to "revenue-funded growth." While they still take in capital, their primary engine for expansion is now the massive revenue generated by their models.

• Why it matters: This changes the "burn rate" logic. Once a lab is profitable, they can spend their own revenue to accelerate growth, making them less dependent on the whims of VC markets and more focused on scaling as fast as the physical world allows.
• In the book: Patel notes that Anthropic began turning a profit in Q2, and OpenAI is expected to follow, with revenue generation now far exceeding the incremental cost of compute.

The "Revenue-per-Megawatt" Multiplier

The value of a single megawatt of compute is increasing exponentially as models become more capable. If a lab can generate $50 million in revenue from a megawatt that costs $10 million to run, they can reinvest that $40 million profit into more training.

• Why it matters: This creates a "compounding" effect on intelligence. The more capable the model, the more revenue it generates, which buys more compute, which makes the next model even more capable.
• In the book: Patel highlights that for Anthropic, revenue has reached as high as $50 million per megawatt, enabling this rapid reinvestment cycle.

The Concentration of Incremental Compute

While the total world compute is growing, the incremental compute (the new stuff being added) is being swallowed by the labs.

• Why it matters: Because compute is growing so fast, "incremental" compute is the market. If the labs capture 40-50% of new compute this year, and that trend continues, they will soon own the vast majority of the world's usable flops.
• In the book: Patel points out that while the world's compute might double annually, the labs are currently tripling their compute every year.

The "Crowding Out" Effect of Capital

The sheer amount of capital required to build 100+ gigawatts of compute will drive up interest rates and "crowd out" other types of lending in the economy.

• Why it matters: This is a systemic economic shift. When the "opportunity cost" of capital becomes the construction of robot factories and data centers, other sectors (like mortgages or traditional business) become more expensive to fund.
• In the book: Patel explains that if the US spends a quarter of its economy on data centers, the interest rates will rise, making it harder for everyone else to borrow money.

The Physical Supply Chain Constraint

The bottleneck for AI is no longer just "research"; it is the physical production of mirrors, wafers, and power.

• Why it matters: Even with infinite money, you cannot build a data center without the physical components (like ASML EUV tools). The "whip" of demand takes time to reach the tail end of the supply chain.
• In the book: Patel discusses the "discrepancy" between the $6 billion in CapEx to produce a gigawatt and the $100 billion in revenue that gigawatt can generate.

Frameworks & tools

The "Lab Economics" Growth Model

To understand the trajectory of an AI company, look at these three metrics:

1. Revenue per Megawatt: How much money is the model generating for every unit of power/compute it consumes?
2. Incremental Capture: What percentage of new compute being added to the world is being used by the lab?
3. CapEx-to-Revenue Ratio: Is the company spending more on infrastructure than it earns? (The "Hyperscaler" model).

The "Crowding Out" Analysis (Macro-Economic Impact)

To predict how AI will affect the broader economy, monitor:

1. Interest Rate Spikes: Are rates rising because of inflation, or because of "Capital Crowding" (massive infrastructure spending)?
2. Tax Revenue Shifts: As labor is automated, how much of the tax base shifts from payroll/income to corporate/data center taxes?
3. Opportunity Cost of Capital: Is the government/private sector prioritizing "AI-pilled" investments over traditional infrastructure?

Memorable examples & stories

The SpaceX/Meta "Optionality"

SpaceX and Meta are unique because they have massive balance sheets and can build compute without needing an immediate, high-margin customer.

• The Lesson: This allows them to "hoard" compute, creating a massive power imbalance where they can choose to sell to the labs at "crazy margins" or use it for their own internal dominance.

The "Mirror" Bottleneck

The demand for AI is so high that the bottleneck is literally the mirrors used in ASML machines.

• The Lesson: Capitalism will eventually force the expansion of the supply chain, but it is a "slow whip." The physical world cannot move as fast as the digital demand.

The "White-Collar Population" Calculation

If a gigawatt of compute can sustain a population of one million white-collar workers earning $100k/year, that's a $100 billion revenue stream.

• The Lesson: This illustrates the sheer scale of value capture possible once AGI can automate high-value human labor.

Key quotes

"The labs are going from companies that spend tens of billions of dollars a year to hundreds of billions of dollars a year, to forecasting to spend trillions of dollars a year."

• Why it matters: This highlights the sheer scale of the industrial revolution happening in the data center.

"A new watt deployed this year is significantly more efficient than the watts deployed two years ago... they’re 3-5x more performance per watt than the prior-generation chips."

• Why it matters: It explains why the "doubling" of compute isn't just a numbers game; it's a quality/efficiency multiplier.

"The opportunity cost of capital is going to increase a ton... the rest of the world... is going to be priced out."

• Why it matters: This identifies the "winner-take-all" mechanism of the AI era: it isn't just about who has the best code, but who can afford to outbid the rest of the world for the resources needed to run it.

How to apply it

This week

• Audit your "Compute" exposure: Look at your portfolio or business. How much of your value is tied to "labor" (which is being automated) vs. "capital" (which is being concentrated)?
• Monitor the "Interest Rate" signal: Watch for news on data center construction and power grid expansion. Understand that these are the "foundations" of the AI economy.

Ongoing

• Think in "Megawatts": When evaluating AI companies, stop looking only at "Model Accuracy." Start asking: "How much revenue can they generate per megawatt of compute?"
• Watch the "Concentration" of Capex: Follow the spending of Meta, Microsoft, and SpaceX. They are the "landlords" of the AI era.

Common pitfalls

• The "Research" Fallacy: Thinking that the winner will be the one with the smartest research paper. Patel argues that because of the "CapEx" and "Power" constraints, the winner may simply be the one who can build the most infrastructure.
• Ignoring the "Crowding Out" Effect: Assuming that AI growth happens in a vacuum. It actually competes for the same dollars and interest rates as your mortgage, your retirement fund, and other industrial projects.
• Underestimating the "Margin" of the Labs: Assuming the labs are just "selling" a product. They are actually capturing a massive "markup" on scarce resources (compute) because they are the only ones with the demand to pay for it.

Connects to

• The "Resource Curse" (Economics): How access to a critical resource (Compute) creates a monopoly that can lead to extreme centralization.
• The "Great Acceleration" (History): Comparing the current AI-driven infrastructure build-out to the industrial and electrical revolutions of the 19th and 20th centuries.

If you remember only three things

• The Lab Monopoly: OpenAI and Anthropic are on track to control the majority of the world's usable compute by 2028.
• The Revenue Loop: High-performing models generate massive revenue per megawatt, which funds the purchase of even more compute.
• The Macro Effect: This massive concentration of capital will drive up interest rates and "crowd out" other parts of the economy, fundamentally changing the global financial landscape.

Internalize it

• Reflection questions
  • If you were a "landlord" of compute, how would you decide who gets to rent your gigawatts?
  • How does the "crowding out" of capital affect your specific industry or career path?
  • If the "opportunity cost" of capital is now building robot factories, what is the "opportunity cost" of not being AI-pilled?
• Flashcards Q: What is the "Revenue-per-Megawatt" metric? A: It is the amount of revenue a lab generates from a unit of compute power; it is the key to the "reinvestment loop" where profit from inference funds further training. Q: Why is "Incremental Compute" the most important metric? A: Because compute is growing so fast, the "new" compute added each year represents the vast majority of the total market. Q: What is the "Crowding Out" effect? A: The phenomenon where massive AI infrastructure spending drives up interest rates, making it more expensive for everyone else in the economy to borrow money. Q: Who are the "Landlords" of the AI era? A: Meta and SpaceX, because they have the balance sheets to build compute without needing an immediate customer, allowing them to hoard resources. Q: What is the "Whip" effect in the supply chain? A: The delay between the "signal" of high demand and the physical expansion of the supply chain (e.g., making more mirrors for EUV machines).

Fidelity note

Confidence is high. The summary is strictly grounded in the transcript provided. The "Crowding Out" and "Interest Rate" sections are based on the host's and guest's specific financial monologue regarding debt, tax revenue, and the 20% vs 40% debt servicing figures. One nuance to verify: The specific "gigawatt" numbers (e.g., 100GW by 2028) are "vibed" (estimated) by the guest and should be treated as projections, not hard data.