Aman Agarwal

Numbers Mentioned in the Dylan Patel and Dwarkesh Patel Podcast

A point-by-point list of the numbers, estimates, and forecasts discussed by Dylan Patel and Dwarkesh Patel.

Numbers Mentioned in the Dylan Patel and Dwarkesh Patel Podcast

Watch the podcast

Quick unit guide: 1 gigawatt (GW) = 1,000 megawatts (MW). CapEx means money spent building infrastructure. Training improves a model; inference serves users. FLOPs are computer operations.

1. AI infrastructure is becoming a trillion-dollar buildout

  1. At 00:01:03, the conversation puts total AI-related infrastructure CapEx at a little over $1 trillion in 2026.

  2. At 00:01:07, by 2028, the estimate is more than $2 trillion of CapEx.

  3. At 00:01:20, the labs are described as moving from spending tens of billions of dollars per year, to hundreds of billions, and potentially to trillions per year toward the end of the decade.

  4. At 00:00:46, about one-third of new compute coming online in 2026 is said to be for OpenAI and Anthropic. The hardware may be built and owned by another company, but the labs are the end customers renting it.

  5. At 00:04:14, the speakers estimate that OpenAI and Anthropic together may take 40% to 50% of new compute in 2027. Their more specific answer is that half of the world’s incremental new compute could be going to these two labs by the end of 2027.

  6. At 00:05:09, the reason this can become such a large share so quickly is that most compute is new compute. If the total amount of compute is growing rapidly, capturing a large share of what is added each year soon creates a large share of the total.

2. Two labs could control a large share of the world’s compute

  1. At 00:03:50, OpenAI is described as starting 2026 with about 2 gigawatts of compute. Anthropic is described as having less than 2 gigawatts.

  2. At 00:03:58, by the end of 2026, the estimate is that both will be above 5 gigawatts. That is roughly a 3x to 4x increase for each lab as a whole.

  3. At 00:05:26, one simple projection says total world compute doubles every year, while frontier-lab compute triples every year. Starting from 2 gigawatts, that gives a rough sequence of:

    • 6 gigawatts by the end of 2026
    • 18 gigawatts by the end of 2027
    • 54 gigawatts by the end of 2028

    This is a straight-line multiplication exercise, not a measured forecast.

  4. At 00:05:54, the broader world may add roughly 30 gigawatts in 2026, 50 gigawatts in 2027, and 70 gigawatts in 2028. A later estimate at 00:33:45 puts 2029 additions at 90 to 100 gigawatts.

  5. At 00:33:32, the speakers estimate more than 200 gigawatts of global compute by the end of 2028.

  6. At 00:06:16, a new generation of chips can deliver about 3x to 5x more performance per watt than the previous generation. This matters because a new gigawatt is not equivalent to an old gigawatt. The newer one can do much more useful work with the same power.

  7. At 00:13:13, if the current trend continues, OpenAI and Anthropic could each have more than 50 gigawatts by the end of 2028, or 100 gigawatts combined.

  8. At 00:13:41, the more aggressive scenario says the two labs could take 70% to 80% of incremental compute by 2028. Because the hardware would also be newer and more efficient, the speakers argue that the labs might control most of the world’s usable FLOPs, not simply most of its electrical capacity.

3. The economics of a megawatt are changing

  1. At 00:02:32, the base cost of compute is described as roughly $10 million to $15 million per megawatt.

  2. At 00:03:08, Anthropic is said to generate as much as $50 million per megawatt from some of its compute. The simple example is: spend $10 on inference capacity and generate $50 of revenue.

  3. At 00:19:48, the gap between the cost of compute and the revenue a strong model can generate is described as 4x or more in some cases. The implication is that the labs can use the profit from serving users to buy more compute for training.

  4. At 00:25:47, for the end of 2027, one estimate is more than $50 million of revenue per megawatt, with a possible blended figure of $70 million to $80 million if the best models keep improving and can be released.

  5. At 00:17:15, a more aggressive part of the discussion imagines $100 million of revenue per megawatt and says the labs could pay $50 million per megawatt to secure scarce capacity. This is a scenario, not a current price.

  6. At 00:17:57, the speakers also discuss the value of a full AI worker. If one gigawatt could support roughly one million white-collar workers, and each worker created $100,000 of annual value, that would equal $100 billion per gigawatt. With a fully capable general AI system, they imagine the value could be many hundreds of billions per gigawatt.

  7. At 00:18:08, these figures use different units and concepts, so they should not be compared without care. $100 million per megawatt is equivalent to $100 billion per gigawatt when both refer to the same time period, but a compute rental price, annual revenue, and total economic value are not automatically the same thing.

  8. At 00:22:24, the conversation mentions one reported transaction in which SpaceX sold compute to Google at about $40 billion per gigawatt. The forecast at 00:22:31 is that most compute could still transact below $20 billion per gigawatt at the end of the following year because the infrastructure has to be financed before it is built.

  9. At 00:23:02, the speakers also say Anthropic may generate more than $60 billion per gigawatt in revenue. The point is that a lab can value the compute more highly than the company building and renting it.

  10. At 00:29:44, Anthropic is described as potentially having close to 20 gigawatts by the end of 2027. If 10% of its compute equals 2 gigawatts, and each gigawatt could generate a hypothetical $100 billion, shifting that 10% from inference to training would mean giving up a theoretical $200 billion of near-term revenue.

  11. At 00:32:54, the speakers say Anthropic’s revenue growth had at one point been adding about $25 billion of annual recurring revenue per month, but that this had plateaued at 00:33:06. Their interpretation is that additional compute was increasingly going into research and development rather than directly serving users.

  12. At 00:19:03, Meta is described as having represented as much as 10% of Anthropic’s business at one point. The example given is that a 5% increase in engagement time from better ad systems could be worth more to Meta than the profit Anthropic makes from selling the model tokens.

4. Chip supply is a physical bottleneck

  1. At 00:07:18, one estimate for building one gigawatt of advanced AI compute requires approximately:
  • 55,000 N3 wafers
  • 6,000 N5 wafers
  • 170,000 DRAM wafers

N3 and N5 refer to advanced chip manufacturing processes. DRAM is a type of memory used by the systems.

  1. At 00:08:09, the tooling for the wafer fabrication capacity is estimated at $3 billion to $4 billion. Including cleanrooms, the factory shell, and related infrastructure brings the estimate to about $6 billion of fab CapEx.

  2. At 00:08:23, in the simplified example, that $6 billion fab investment produces one gigawatt of compute capacity every year, and one gigawatt generates about $100 billion of annual revenue.

  3. At 00:08:40, over five years, the first gigawatt would generate five years of revenue, the second would generate four years, and so on. On that arithmetic, $6 billion of fab CapEx could be associated with more than $1 trillion of end AI revenue over five years.

  4. At 00:09:12, the speakers then remove roughly half the value for data centers, power, installation, research, and other participants. Even after doing that, they describe a possible 100x gap between fab CapEx and the end revenue generated. In simple terms, the example turns $1 into about $100 across the wider chain.

  5. At 00:10:19, an extreme arbitrage example says that a person with $400 million who could buy an EUV lithography tool might be able to resell it for more than $1 billion, because the tool is a bottleneck. This is a hypothetical example, not investment advice.

  6. At 00:10:46, the supply chain is described as aiming to produce about 100 EUV tools per year by 2030. Carl Zeiss, which makes critical mirrors for the tools, is used as an example of a supplier that may need to expand along with the rest of the chain.

  7. At 00:12:38, the wafer fabrication equipment supply chain is estimated at roughly $200 billion in the following year, while the wider AI infrastructure supply chain would be larger still.

5. China and the United States are adding compute at different rates

  1. At 00:34:37, in 2022, the speakers estimate that the United States was adding 45% to 50% of the world’s new compute, while China was adding 30% to 35%.

  2. At 00:34:58, at the time of the conversation, about 70% of newly deployed watts were said to be in the United States. China was described as accounting for less than 10% of incremental new AI compute.

  3. At 00:35:37, China is estimated to have 30 gigawatts or less of AI compute by 2028.

  4. At 00:36:15, in 2028, China could add 5 to 10 gigawatts of domestically produced chips. The speakers expect those chips to be less capable than the leading Nvidia, Google, or OpenAI systems of that year.

  5. At 00:37:42, by 2029, 50 incremental gigawatts of Chinese compute is described as a reasonable possibility. But the speakers suggest at 00:38:03 that 50 Chinese gigawatts might have the effective performance of only about 20 gigawatts of American chips. This is what quality-adjusting the gigawatt means: counting useful work rather than just power capacity.

  6. At 00:40:17, leading Chinese labs are described as having roughly 100 to 200 megawatts of compute in total, with ByteDance Seed as an outlier. Anthropic, by comparison, is expected to be above 5 gigawatts by the end of 2026 at 00:40:35.

6. Training, research, and inference use the same machines differently

  1. At 00:40:58, a rough current split is described as 60% training and 40% inference.

  2. At 00:41:03, that 60% training figure is then divided into 50% research and 10% development, with the remaining 40% going to inference. In this description, research means trying new ideas, architectures, data mixes, and training methods. Development means running a major model training process.

  3. At 00:41:22, a major pre-training run is described as using less than 200 megawatts at one point in time, over roughly two months. Reinforcement learning uses less compute at one site, although total compute across experiments may be higher.

  4. At 00:41:49, a lab may have multiple gigawatts available but be able to use only 200 megawatts in one coordinated training run. The limiting factors include moving data between sites, coordinating clusters, and producing enough useful reinforcement-learning rollouts.

  5. At 00:42:28, as AI becomes better at coding and research, the speakers expect a larger share of compute to go toward training, continual learning, and internal research. That means a lab may choose not to use every available megawatt to serve paying customers.

7. The buildout creates a financing problem

  1. At 00:42:51, if the world builds 100 gigawatts of compute per year at current prices, the speakers estimate about $5 trillion of CapEx per year for the compute itself.

  2. At 00:42:57, power plants must be built early and may last about 30 years. Data centers may last 15 to 20 years. Because future capacity has to be prepared before it is needed, the total investment could be closer to $7 trillion to $10 trillion per year when power and data centers are included.

  3. At 00:44:10, by the end of 2030, annual incremental CapEx could approach $10 trillion, which the speakers compare with roughly one-tenth of the world economy. If most of this happened in the United States, they say it could represent roughly one-quarter to one-third of the US economy at today’s scale.

  4. At 00:44:59, for 2028, one rough estimate is $3 trillion to $4 trillion of CapEx across the system. More than $2.5 trillion would go to IT infrastructure, with another $1 trillion to $2 trillion for data centers, energy, semiconductors, and downstream suppliers. These are rounded figures and the ranges overlap.

  5. At 00:54:34, from 2024 to 2029, the model discussed in the interview uses about $11 trillion of total CapEx. Roughly $6 trillion would be funded with cash and $5 trillion with credit.

  6. At 00:57:06, the speakers use Meta’s borrowing costs as an example. Meta had recently raised debt at about 5% to 6%, and the estimate is that it might willingly pay 8% to build more compute. That would be a 250-basis-point increase, or 2.5 percentage points, over a 5.5% midpoint.

  7. At 00:46:17, the reason a company might accept a high interest rate is that it expects the compute to generate an even higher return. The wider problem is that if large technology companies borrow heavily, they compete with governments, other businesses, and households for the same pool of capital.

8. Higher rates could spread the cost through the economy

  1. At 00:50:20, in the United States, corporate income taxes are described as contributing less than 10% of federal revenue, while payroll and individual income taxes contribute more than 80%.

  2. At 00:50:32, about 20% of tax revenue is described as currently going toward servicing government debt. Much of that debt is short-duration and rolls over in roughly five years.

  3. At 00:51:34, in the speakers’ example, a 1 percentage-point increase in interest rates could push the share of tax revenue used for debt service from 20% to 25% over five years.

  4. At 00:51:46, a 5 percentage-point increase could push it above 40%. If the government continues borrowing about $2 trillion per year, the figure could rise above 60% in their more extreme calculation.

  5. At 00:58:51, the historical comparison is the early 1980s. Paul Volcker raised interest rates by more than 5 percentage points, reaching about 8% real interest in the description given. The speakers say roughly 40 countries defaulted during that decade.

  6. At 00:59:27, if an economy grows at 3% per year, the rule of 70 suggests it takes a little more than 20 years to double. In a fully automated economy, the speakers imagine the effective labor force could double every year. They say that could produce growth of at least tens of percent per year, and possibly much more.

  7. At 01:00:16, in that extreme scenario, interest rates in the 2030s could reach the tens of percent because the return on productive AI investment would be so high. This is one of the most speculative claims in the conversation.

  8. At 01:01:33, one market implication discussed is that many companies could trade at only 2x or 3x earnings if investors demand a much higher return. Meta is described as being worth roughly $1.5 trillion at the time of recording, with the argument that its future compute and cash flows could justify a higher value.

9. The labor equivalent could grow faster than the human workforce

  1. At 01:08:12, frontier compute, measured in FLOPs, is described as growing 4x to 5x per year.

  2. At 01:08:18, at the same time, the amount of compute needed to reach a given level of capability is described as falling by about 3x per year. Combining those two effects gives roughly 10x growth per year in the effective AI population at the frontier.

  3. At 01:08:41, the illustrative sequence is:

  • 10 million AI laborers this year
  • 100 million the next year
  • 1 billion the year after that

“AI laborers” here does not mean actual people. It means a rough comparison between the work a collection of AI systems could perform and the work performed by human workers.

  1. At 01:08:57, the speakers say it is plausible that, by the end of the decade, a single lab could contain more AI labor equivalent than the number of people on Earth.

  2. At 01:10:45, if recursive self-improvement, or RSI, begins, the speakers discuss even faster possibilities of 100x or 1,000x growth per year in effective intelligence or labor capacity.

  3. At 01:04:20, another hypothetical is that a six-month delay in releasing a model could correspond to 100x more capability if progress had entered a rapid takeoff phase. They also suggest at 01:04:53 that 3 to 6 years of AI progress could occur in one year during RSI.

10. The main issue is concentration, not only speed

  1. At 01:11:24, the speakers argue that AI training has strong economies of scale because an improvement can be spread across billions of sessions or users.

  2. At 01:11:41, a company that is slightly ahead can also charge a higher markup when compute is scarce, because it can produce more value from each unit of hardware.

  3. At 01:12:00, those two forces, combined with learning from deployment and the possibility that better models help create the next better models, point toward more compute and economic value being concentrated in a small number of labs.

  4. At 01:14:49, the conversation’s final positive observation is that the labs do not currently capture all the value. Jane Street is given as an example of a customer that may generate $300 million to $500 million per megawatt from its use of models, while Anthropic may capture only part of that value through its pricing.

  5. At 01:16:20, the counterargument is that if a lab can generate hundreds of millions of dollars per megawatt internally, it may eventually prefer to use the compute for its own research rather than sell it to another company.

Generated by OpenAI Codex using GPT-5.6 Luna with xhigh reasoning effort.