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

“Yeah, so when we go back to last year, even at the end of the year, most of GDP growth in America was just AI infrastructure.”

Claim: At the end of last year, AI infrastructure accounted for most of U.S. GDP growth.

“about a third of the compute coming online. is for the labs for open ai and anthropic now it may be built by others and then rented to them but it's and it at the end customer it's them as we go forward into the future the the numbers for computer ballooning right we're at you know you know a little bit over a trillion dollars of capex this year as we go out into 28 it's going to be more than two trillion dollars um”

Claim: About one-third of the compute coming online this year is ultimately for OpenAI and Anthropic, including compute built by other parties and rented to the labs.

“is for the labs for open ai and anthropic now it may be built by others and then rented to them but it's and it at the end customer it's them as we go forward into the future the the numbers for computer ballooning right we're at you know you know a little bit over a trillion dollars of capex this year as we go out into 28 it's going to be more than two trillion dollars um”

Claim: Global compute capital expenditure is slightly above $1 trillion this year and is expected to exceed $2 trillion by 2028.

“The labs are also taking an increasing percentage of this.”

Claim: AI labs are taking an increasing share of compute expenditure.

“you've got a very interesting situation where the labs are going from companies that spend, you know, tens of billions of dollars a year to hundreds of billions of dollars a year to forecasting to spend trillions of dollars a year even towards the end of the decade. And this is at least some of the contracts they've begun signing with their partners. And so this requires a big reshaping of.”

Claim: AI labs are moving from annual spending in the tens of billions of dollars to hundreds of billions, and their contracts and forecasts point to spending trillions of dollars per year toward the end of the decade.

“what happens with their economics right you know so up until now there have been companies that mostly lost money um anthropic started turning a profit in q2 um it's believed at some point in q3 open ai could potentially start turning a profit even um with the big rise of codex and and 5.6 and all this um but if we go back a year ago everything that they all the money they had was venture funded losses losses”

Claim: Anthropic began turning a profit in Q2, while OpenAI could potentially begin turning a profit at some point in Q3.

“If we go back to even the beginning of this year, those venture funded losses, they've now turned the corner and are actually starting to profit.”

Claim: By the beginning of this year, AI labs that had previously relied on venture funding to cover losses had turned the corner and were beginning to make profits.

“Now, that doesn't mean they're not taking in new capital. The new capital is still coming in to accelerate the growth further. But ultimately, there's more and more of their business is being funded off of their own revenue rather than capital injections into them.”

Claim: AI labs continue to receive new capital to accelerate growth, while an increasing share of their business is funded by revenue rather than outside capital injections.

“Over the last year and a half, their margins have really skyrocketed.”

Claim: AI labs’ margins have increased sharply over the past year and a half.

“The base cost of compute tends to be around $10 or $13 or $15 million per megawatt.”

Claim: The base cost of compute is typically approximately $10 million to $15 million per megawatt.

“their revenue generation has passed well beyond the sort of incremental $10, $15 million per megawatt. In the case of Anthropic, the revenue has gone as high as $50 million per megawatt.”

Claim: Anthropic has generated revenue of up to $50 million per megawatt of compute capacity.

“And what that now enables them to do is, hey, if I spend $10 on inference capacity, actually generate $50 of revenue, and then I can turn around and incrementally spend all of that profit on training.”

Claim: The speaker says that spending $10 on inference capacity can generate $50 in revenue, enabling the lab to reinvest the resulting profit in training.

“When you look at the incremental compute added, that's about 30% of the compute added this year.”

Claim: OpenAI and Anthropic together account for about 30% of the compute added this year.

“given what's already been signed and penned and inked, you've got something even more dramatic, You've got Anthropic OpenAI are taking as much as 40 to 50% of compute next year.”

Claim: Based on existing signed agreements, Anthropic and OpenAI may take as much as 40% to 50% of compute next year.

“And the centralization doesn't look like it's slowing down or stopping. In fact, it looks like it's only accelerating.”

Claim: The concentration of compute among major AI labs is not slowing or stopping and appears to be accelerating.

“Now, who's building that compute for them will change.”

Claim: The identity of the companies building compute for AI labs will change.

“Next year, a big new insurance, for example, SpaceX is building a ton of compute, and they're actively going to lease quite a bit of it to Anthropic and OpenAI most likely because they're the ones who have the marginal capability to pay the highest price.”

Claim: The speaker expects SpaceX to lease a substantial amount of the compute it is building to Anthropic and OpenAI, likely because those labs can pay the highest marginal price.

“OpenAI and Anthropic are also starting to build their own compute.”

Claim: OpenAI and Anthropic are starting to build compute infrastructure for themselves.

“OpenAI with their own chips,”

Claim: OpenAI is developing its own chips for compute.

“Anthropic with TPUs that they're purchasing from Google and deploying with.”

Claim: Anthropic is buying TPUs from Google and deploying them for its compute needs.

“playing with fluid stack and so when you ask hey when does half of the world's incremental new compute go to just open and anthropic i mean it's it's really by the end of next year it's already half of the incremental compute is going to anthropic and open ai”

Claim: By the end of next year, Anthropic and OpenAI together are expected to receive half of the world’s newly added compute.

“How long do you think, so there's this trend where maybe world compute in gigawatts doubles every year, but the compute of the Frontier Labs triples every single year.”

Claim: The speaker describes frontier-lab compute as growing about threefold annually, compared with a possible annual doubling of total world compute measured in gigawatts.

“But if you keep the current trend going, it goes from like two at the beginning of this year to close to like six at the end of this year, just multiplying out by three, 18 by the end of 2027, 54 by the end of 2028.”

Claim: If the current trend continues, the speaker projects the frontier labs’ compute measure to increase from about 2 at the beginning of the year to nearly 6 at year-end, 18 by the end of 2027, and 54 by the end of 2028.

“so if the incremental compute adds this year 30 gigawatts, next year 50 gigawatts, and the year after that 70, a new Watt deployed this year is significantly more efficient than the Watts deployed two years ago.”

Claim: Assuming incremental compute additions of roughly 30 GW this year, 50 GW next year, and 70 GW the following year, the speaker says each watt deployed this year is significantly more efficient than watts deployed two years earlier.

“So actually, you know, a humongous percentage of the world's compute was deployed this year, even though it didn't double the number of Watts deployed.”

Claim: The speaker says a very large share of the world’s compute was deployed this year despite the total number of watts deployed not doubling.

“I'm deploying GB300s and TPV7s and Tranium 3s, which are way, way, way more efficient, you know, 3x, 5x more performance per Watt than the prior generation chips.”

Claim: The speaker says GB300, TPV7, and Tranium 3 chips provide roughly 3–5 times the performance per watt of the previous generation.

“um ladder here so anthropic and open ai take on you know 45 percent of compute next year you've you've got them in let's say december 27 they have taken on half of the world's incremental new compute but that half of the world's new incremental compute is actually at a higher performance than everything else before it so you've got another multiplier on that so by the time you're in like towards the end of 2028 if this trend continues which i see nothing that's stopping it you've got them just controlling most of the usable, you know, flops in the world on their own.”

Claim: If the trend continues, Anthropic and OpenAI together will control most of the world’s usable FLOPs by late 2028.

“um ladder here so anthropic and open ai take on you know 45 percent of compute next year you've you've got them in let's say december 27 they have taken on half of the world's incremental new compute but that half of the world's new incremental compute is actually at a higher performance than everything else before it so you've got another multiplier on that so by the time you're in like towards the end of 2028 if this trend continues which i see nothing that's stopping it”

Claim: The speaker argues that Anthropic and OpenAI’s share of compute has an additional effective multiplier because their incremental compute is higher-performing than previously deployed compute.

“um ladder here so anthropic and open ai take on you know 45 percent of compute next year you've you've got them in let's say december 27 they have taken on half of the world's incremental new compute but that half of the world's new incremental compute is actually at a higher performance than everything else before it so you've got another multiplier on that so by the time you're in like towards the end of 2028 if this trend continues which i see nothing that's stopping it”

Claim: The speaker projects that Anthropic and OpenAI will account for 45% of compute next year and half of the world’s incremental new compute by December 2027.

“That's the upper bound, by the way. That's the like, I'm so fucking bullish.”

Claim: The speaker characterizes 80 GW of added compute in 2028 as an upper-bound, highly bullish estimate.

“You said in order to make a gigawatt of, I think, Vera Rubens, you need, one sec, you need 55,000 N3 wafers, 6k N5 wafers, and 170k DRAM wafers. I don't know those I'm numbers might have going changed.”

Claim: The interviewer recalls the guest previously estimating that producing one gigawatt of compute requires 55,000 N3 wafers, 6,000 N5 wafers, and 170,000 DRAM wafers, while noting the figures may have changed.

“i had an llm run your wafer fabric weapon model and figure out how much um tooling how much the tooling costs to produce A gigawatt of compute basically every single year. And we said like three to four billion dollars.”

Claim: The interviewer says an LLM-based wafer-fabrication model estimated tooling costs of $3–4 billion to produce one gigawatt of compute annually.

“Now suppose you add in, you know, clean rooms and shell and everything else of the fab. So six billion dollars of like fab CapEx produces every single year a gigawatt. And a gigawatt produces right now a hundred billion dollars of revenue.”

Claim: The interviewer estimates that $6 billion in fab capital expenditure, including clean rooms and the fab shell, yields one gigawatt of compute capacity per year, and that one gigawatt currently generates $100 billion in annual revenue.

“But also that six billion in CapEx is producing a gigawatt every single year. And that gigawatt is producing $100 billion every single year.”

Claim: The interviewer asserts that $6 billion in fab capital expenditure yields one gigawatt of compute annually, with each gigawatt generating $100 billion in annual revenue.

“So even over the course of five years, so, you know, the first gigawatt is generated, five years of profits, the second gigawatt that the fab is producing is generated four years of profits and so on. Six billion of CapEx at the fab level. will have generated over a trillion dollars of end AI revenue.”

Claim: The interviewer estimates that $6 billion of fab-level capital expenditure will generate more than $1 trillion in end-AI revenue over the described sequence of annual gigawatt production.

“there's a lot of opex along the way. a lot of other capex like the data center, the power instantly.”

Claim: The speaker notes that operating expenses and additional capital costs, including data centers and power, must be counted beyond fab capital expenditure.

“So take away half of it for all these middlemen. That still means there's 100x discrepancy between fab capex. And end revenue generated more than that, actually, really, but just being very conservative.”

Claim: The interviewer estimates that end revenue exceeds fab capital expenditure by more than 100 times even after allocating half of the revenue to intermediaries, and describes this as conservative.

“And as a result. this is capitalism right like you would imagine that people are going to figure like we're going to be you're you have this huge discrepancy where you can turn one dollar into a hundred dollars and you're they're not going to figure out a way to make more mirrors right”

Claim: The interviewer argues that a potential 100-to-1 return from compute-related investment gives businesses a strong incentive to expand supply.

“we did but the emergency are so big we're like anthropic opening like we could make a trillion dollars right now but we're just bottlenecked on the mirrors that go into the sml machines like they'd spend okay how can we make more mirrors we spend a hundred billion dollars dollars on this right that's the situation we're going to be in pretty soon and i'm just like we're not going to be able to solve that supply constraint that just seems quite hard to imagine”

Claim: The speaker says Anthropic’s potential revenue could reach $1 trillion now, but mirror supply for EUV lithography machines is a bottleneck, and Anthropic could spend $100 billion to increase mirror production.

“dollars on this right that's the situation we're going to be in pretty soon and i'm just like we're not going to be able to solve that supply constraint that just seems quite hard to imagine”

Claim: The interviewer doubts that the mirror-supply constraint can be solved easily.

“No, there's definitely you've seen people do funny arbitrages here where they buy like turbines and then they try and resell Right. them because the value of a turbine is way more because it's the thing bottlenecking your data center.”

Claim: The speaker says some people buy turbines to resell them at a premium because turbine availability can bottleneck data centers.

“You know, I think like if anyone had like $400 million and the ability to convince ASML to sell them an EV tool, they should totally just go buy one and wait, wait, wait and then sell it for north of a billion dollars. Right. But but ultimately,”

Claim: The speaker recommends that a buyer able to purchase an ASML EUV tool for $400 million should hold it and resell it for over $1 billion.

“Yes, capitalism will cause these things to expand, but it's a whip, right? It takes a long time for the whip signal to get to the tail end of that. And so the supply chain doesn't react immediately.”

Claim: The speaker says market incentives will eventually expand constrained supply, but the signal takes a long time to propagate through the supply chain and supply therefore does not respond immediately.

“In fact, you go to talk to someone at Carl's Ice, they're like, yeah, yeah, yeah, we need to make 100 EV tools by the end of the decade. I think when we first had our episode. earlier this year they didn't even think they needed to make that many enough mirrors to make 100 uv tools a year um and and so now they're they've like sort of they're like okay we need to do that but in reality you know because of all the economics of what's going on it should be even more um”

Claim: The speaker says that earlier this year Carl Zeiss did not anticipate needing enough mirrors for 100 EUV tools annually, but has since started planning for that level; given current economics, the speaker believes the required production should be higher still.

“Generally, there's obviously some physical constraints.”

Claim: Expanding production faces physical constraints.

“The way the supply chain is expanding currently, the 100 is roughly still the right number. 100 ASML tools for 2030,”

Claim: Given the current pace of supply-chain expansion, roughly 100 ASML tools by 2030 remains the right estimate.

“carl zeiss here's 10 billion dollars please fucking just expand production that would change things and”

Claim: The speaker suggests that providing Carl Zeiss with $10 billion to expand production would materially change the production outlook.

“in the supply chain i don't think it'll happen this year i don't think it'll happen next year i don't have i don't think it'll happen the year after because the world is capital constraint i”

Claim: The speaker does not expect the proposed supply-chain production expansion to happen this year, next year, or the following year because the world is capital-constrained.

“think they're going to do that but yeah”

Claim: The speaker does not expect the AI labs to spend the very large amount of capital proposed in the preceding discussion.

“So the thing is the labs can spend hundreds of billions, they're going to generate hundreds of billions of revenue next year.”

Claim: The AI labs can spend hundreds of billions of dollars and are expected to generate hundreds of billions of dollars in revenue next year.

“CapEx next year is like $2 trillion.”

Claim: Total capital expenditure next year is approximately $2 trillion.

“The wafer fabrication equipment supply chain will do something on the order of $200 billion.”

Claim: The wafer-fabrication-equipment supply chain is expected to generate roughly $200 billion.

“The data center market supply chain will do even more.”

Claim: The data-center-market supply chain is expected to generate more than the roughly $200 billion attributed to the wafer-fabrication-equipment supply chain.

“The accelerator supply chain will do even more.”

Claim: The accelerator supply chain is expected to generate more than the roughly $200 billion attributed to the wafer-fabrication-equipment supply chain.

“The energy supply chain will do a number.”

Claim: The energy supply chain is expected to contribute a substantial, unspecified amount to the overall spending or market activity being discussed.

“You sum all this up, it's going to be well north of $2 trillion of CapEx.”

Claim: Combined capital expenditure across the discussed supply chains is expected to be well above $2 trillion.

“So the labs have not yet. gotten to the point where their cash flows can fund this stuff.”

Claim: The AI labs' cash flows are not yet sufficient to fund the discussed infrastructure spending.

“I think that may be a little difficult given 2028 you start to have. They've taken 70, 80% of incremental compute,”

Claim: The AI labs may account for 70–80% of incremental compute by 2028.

“and I'm not sure what happens to markets then,”

Claim: The speaker is uncertain about what will happen to markets if the labs take 70–80% of incremental compute by 2028.

“You know, how much does the price of compute skyrocket for them to actually be able to buy 70, 80% of compute?”

Claim: The speaker raises uncertainty about how high compute prices would need to rise for the labs to acquire 70–80% of compute.

“Is, you know, Google or Meta or Amazon willing to sell even that much?”

Claim: The speaker questions whether Google, Meta, or Amazon would be willing to sell enough compute for the labs to acquire 70–80% of it.

“Also, one caveat when we're sort of talking about these gigawatt numbers is, you know, when Amazon is serving bedrock anthropic models. that counts as anthropic compute in sort”

Claim: Compute used by Amazon to serve Anthropic models through Bedrock is counted as Anthropic compute in the speaker's accounting framework.

“of our worldview because it is effectively at the end of the day counted as revenue for anthropic even though like there's a revenue share and credit back all that but ultimately in 2028 it's it's you know if they get to 100 gigawatts combined they have done really disruptive things to the market because anyone can make money off of 10 to 15 million dollars per megawatt compute today you literally like i kid you not it's not that hard go get a gb300 rack”

Claim: If the labs reach 100 gigawatts of combined compute in 2028, their scale will be highly disruptive to the market.

“of our worldview because it is effectively at the end of the day counted as revenue for anthropic even though like there's a revenue share and credit back all that but ultimately in 2028 it's it's you know if they get to 100 gigawatts combined they have done”

Claim: The speaker counts Amazon's Bedrock compute for Anthropic models as Anthropic compute because it ultimately generates revenue for Anthropic, notwithstanding revenue sharing and credits.

“really disruptive things to the market because anyone can make money off of 10 to 15 million dollars per megawatt compute today you literally like i kid you not it's not that hard go get a gb300 rack”

Claim: At compute prices of $10–15 million per megawatt, operators can generate revenue that exceeds their compute costs.

“Go download the Kimi weights. Go download VLM or SGDlang. Set it up. You know, Codex and Fable can actually help you do this. It's pretty simple. I mean, it's not like it's, you know, it's not trivial, but it's not like rocket science. And go put it on open router. Very simple.”

Claim: The speaker says that setting up a GB300 rack, downloading model weights, and offering the service through OpenRouter is relatively straightforward, with Codex and Fable able to assist.

“And you'll start generating more revenue than you're paying for the compute.”

Claim: An operator can earn more revenue than its compute costs by deploying models on rented or owned compute and offering the resulting service.

“And so this is sort of already led to this compute pricing $10 to $15 million. dollars per megawatt start to inflect up and to get to that 100 gigawatts in 2028 you have to believe”

Claim: The ability to earn revenue from compute has already begun pushing compute prices, previously around $10–15 million per megawatt, upward.

“dollars per megawatt start to inflect up and to get to that 100 gigawatts in 2028 you have to believe that the labs can outpay for compute because anyone can make money at 10 to 15 you you know does compute now get to 25 million dollars a megawatt does it get to 40 million dollars a megawatt”

Claim: Reaching 100 gigawatts of lab compute in 2028 requires the labs to outbid other buyers, potentially raising compute prices from $10–15 million to $25 million or $40 million per megawatt.

“think that is my worldview that they will continue to gobble up more of the compute,”

Claim: The speaker expects AI labs to continue acquiring a larger share of available compute.

“but ultimately they can't do it at current pricing or anywhere close to it.”

Claim: At current pricing or anything close to it, the companies cannot acquire the compute they need.

“they do have to start paying 25, 30, 50 million dollars a megawatt to really gobble up 70 percent of the world's compute in 2028 to get to that 100 gigawatts by 2028,”

Claim: The companies would need to pay roughly $25–50 million per megawatt to capture 70% of the world's compute and reach 100 gigawatts by 2028.

“which is a very sort of aggressive goal.”

Claim: Reaching 100 gigawatts by 2028 while capturing 70% of global compute is a very aggressive goal.

“The other aspect of this that's really challenging is we've already seen a huge slowdown for the AI labs,”

Claim: AI labs have already experienced a substantial slowdown.

“This regulation that they advocate for is actually slowing down the labs a lot more than it slows down, you know, sort of the open source Chinese language models.”

Claim: The regulation advocated by AI labs slows those labs substantially more than it slows open-source Chinese language models from China.

“you know, OpenAI not releasing Astra, OpenAI stopping training for two weeks, Anthropic not releasing what their safety assessment said is Model 2,”

Claim: OpenAI had not released Astra and had paused training for two weeks, while Anthropic had not released the model its safety assessment called Model 2.

“which is widely believed to be the next version of Mythos. They're clearly not releasing their best models,”

Claim: Model 2 is widely believed to be the next version of Mythos, and the AI labs are not releasing their best models.

“and in which case their revenue per megawatt stalls or even can start to decline again because other models are competitive again.”

Claim: If AI labs withhold their best models, their revenue per megawatt may stall or decline as competing models become competitive again.

“So it's not that they're falling behind, it's just that they're not releasing their best stuff.”

Claim: The AI labs are not falling behind; they are withholding their best models from release.

“What if there is some regulatory impact that prevents them from releasing their best models? Now their revenue per megawatt does not climb as fast, then their ability to buy that incremental compute for a higher price than everyone else starts to diminish.”

Claim: If regulation prevents AI labs from releasing their best models, their revenue per megawatt will grow more slowly, diminishing their ability to outbid others for additional compute.

“And then maybe they can't get to that 100 gigawatts is sort of in a world where safety doesn't matter.”

Claim: If safety constraints matter, the companies may be unable to reach 100 gigawatts of compute.

“I do believe that's exactly what happens, right?”

Claim: The speaker believes the described regulatory and safety constraints will prevent the companies from reaching 100 gigawatts of compute.

“They can start generating $100 million per megawatt or more, and they can pay $50 million a megawatt and no one else has.”

Claim: The leading AI companies can generate at least $100 million per megawatt and pay $50 million per megawatt, a price other buyers cannot match.

“has any logical reason to do anything with their compute besides say please Dario take everything off of my hands but”

Claim: Other compute owners would have little economic reason to retain their compute rather than sell it all to Dario, given the leading companies' ability to pay more.

“There are, you know, forces at play that which we cannot describe that that would potentially slow this down.”

Claim: Unspecified forces could potentially slow the projected compute expansion.

“I mean, I think a good intuition pump is just what if the AI models were literally as good as a fully automated software engineer? They're not currently there yet, Like, I think they're far from just being able to fully automate the job of like a full white collar worker.”

Claim: AI models are not yet equivalent to a fully automated software engineer and remain far from automating a full white-collar job.

“But white collar workers earn, you know, six figures or north of that.”

Claim: White-collar workers earn annual salaries of at least six figures.

“that a year um and if you have a gigawatt that can sustain a population of like say a million of white collar workers there's like let's say roughly right um that's like you could then off the back of that”

Claim: One gigawatt of compute could support approximately one million white-collar workers.

“yeah i don't know but it'd be many hundreds of billions of dollars if you get like full agi uh per per gigawatt”

Claim: Full AGI could generate many hundreds of billions of dollars per gigawatt.

“and we've continued to see this, most of the value capture is not happening, right? Like most of the value that these models generate does not get given to OpenAI Anthropic. Thankfully so far, it is mostly just being given to the users,”

Claim: Most value generated by AI models has so far accrued to users rather than OpenAI or Anthropic.

“Chainstreet with their exclusive contract with OpenAI for GPT 5.6 ultra fast mode or Chainstreet where they're like one of Anthropic's biggest customers is generating way, way, way, way more value out of the tokens they're paying for than Anthropic is. generating in terms of profit,”

Claim: Chainstreet captures substantially more value from the AI tokens it buys than Anthropic earns in profit from providing them.

“or meta who at one point was you know rumored to be you know as much as 10 of anthropics business you know they're generating way more efficiencies by optimizing their ad algorithms or what have you and and getting engagement time five percent longer and you know all these things they're they're making way more money off of using these models than than anthropic and so ultimately you know and that's that's what's required”

Claim: Meta was once rumored to represent as much as 10% of Anthropic's business.

“or meta who at one point was you know rumored to be you know as much as 10 of anthropics business you know they're generating way more efficiencies by optimizing their ad algorithms or what have you and and getting engagement time five percent longer and you know all these things they're they're making way more money off of using these models than than anthropic and so ultimately you know and that's that's what's required”

Claim: Meta earns more from using AI models than Anthropic earns from providing them, including through ad optimization and reported engagement-time gains of 5%.

“if you had a million new software engineers, the cost per software engineer would also fall.”

Claim: A supply of one million additional software engineers would reduce the cost per software engineer.

“You've got, you know, the end user, which I think we all agree is generating more value than anyone else, hence they're paying a lot for these models.”

Claim: End users generate more value from AI models than any other layer and therefore pay substantial amounts for them.

“But then you have, you know, the app layer. Well, so far, the app layer has generated very little value.”

Claim: The AI application layer has generated very little value so far.

“And you've got the model layer, which, again, up until a year ago was generating negative gross margins and is now generating massive positive gross margins and looks like it's on the path to generating, you know, $100 million per megawatt.”

Claim: The AI model layer shifted from negative gross margins until about a year ago to massive positive gross margins, and appears on track to generate $100 million per megawatt.

“But if we go back again a year ago, the hardware supply chain. was generating all this gross margin while literally everyone else was losing money on it.”

Claim: About a year ago, the AI hardware supply chain earned gross margins while nearly all other participants lost money.

“OpenAI and Anthropik were just plowing VC money in. And as were many other startups and many of these hyperscalers that were building infrastructure without knowing if there was going to be a payoff.”

Claim: OpenAI, Anthropic, startups, and hyperscalers invested in AI infrastructure without knowing whether it would pay off.

“So ultimately, you had this negative value being created on the model layer almost, if you will, because they were selling the tokens for less than it cost them on the infra side and all the values being created used at the chip. the fab.”

Claim: Initially, AI model providers sold tokens below infrastructure cost, creating negative value at the model layer while value accrued to chip and fabrication suppliers.

“Initially in 2023, the memory guys were making no money.”

Claim: In 2023, memory suppliers earned no money from HBM or memory for AI.

“off of you know hbm or memory for ai even though theoretically their value they were delivering was humongous now you've got well actually kiss mc makes way less value than the memory guys um is that actually how much you know they're capturing less value you know so so the the value capture shifted around a lot which is very fun for um people tracking the market or participating in the market like like jane shared as an example um there's”

Claim: The distribution of value capture across AI industry layers has shifted substantially over time.

“I think most compute will still continue to transact at sub $20 billion a gigawatt because all of it has to be financed.”

Claim: The speaker expects most compute to continue transacting for under $20 million per megawatt, including at the end of the following year, because compute projects require financing.

“So most compute is contracted well before it's built.”

Claim: Most compute capacity is contracted before construction is completed.

“effectively hoarding compute them and spacex are plausibly the like number three and the only plausible number three because they're hoarding all this compute they're using their balance sheets and capabilities to build compute to build compute without end customer that's monetizing at a huge degree and they have an actual balance sheet so they can go to the credit market and being like hey guys”

Claim: Meta and SpaceX can use their balance sheets and capabilities to build compute without first securing an end customer that will monetize it at a very high level.

“And now Meta and SpaceX have this optionality of looking around and being like, is my internal use case going to make me more money? Or should I go out there and sell it to Anthropic OpenAI at crazy margins?”

Claim: Meta and SpaceX have the option to use their compute internally or sell it to Anthropic or OpenAI at high margins.

“I think it's highly dependent on who has the best model, if they're allowed to keep releasing their best models, but I don't see why it wouldn't be 50 plus million dollars a megawatt.”

Claim: The speaker thinks Anthropic's or OpenAI's revenue per megawatt depends heavily on which company has the best model and whether it can continue releasing its best models.

“That's where it gets more challenging, but I think it could get to higher than that, like 70, 80 million dollars a megawatt blended across the company, if not higher.”

Claim: The speaker thinks Anthropic's or OpenAI's blended company revenue could exceed $70–80 million per megawatt by the end of 2027, though reaching that level is challenging.

“if I'm anthropic, incremental compute is worth it. Maybe I spend $40 million a megawatt on SpaceX compute.”

Claim: If the projected high revenue per megawatt is achieved, the speaker thinks incremental compute would be valuable enough for Anthropic to potentially pay $40 million per megawatt for SpaceX compute.

“And then, you know, SK Hynix and Micron and Samsung looked at NVIDIA like, well, why don't they raise their prices? So I think the value capture, there's a bull whip effect here, right, where just because someone has risen the prices doesn't mean the entire supply chain rebalances immediately.”

Claim: The speaker argues that value capture moves through the supply chain with a bullwhip effect, so one participant's price increase does not immediately cause the whole supply chain to rebalance.

“over time, the supply chain will rebalance and things will cost more and more.”

Claim: The speaker expects the supply chain to rebalance over time, with costs increasing as it does.

“So TSMC raising prices very slowly, but memory companies raising prices very quickly, you know, substrate companies raising prices very quickly. Different parts of the supply chain, you know, Elon wouldn't have sold if it was 15,”

Claim: The speaker says TSMC is increasing prices slowly, whereas memory and substrate companies are increasing prices quickly.

“well even if our side doesn't happen the current rate of progress continues if you just look at how much progress have you made in let's say the last year and a half like what was the model from a year and a half ago like let's take like”

Claim: Even without recursive self-improvement, AI progress may continue at its current rate; models have improved substantially compared with those from a year and a half earlier.

“is the best model that exists in the world was trained in February.”

Claim: The best model in the world was trained in February.

“says they're not training models for two weeks, man. What the hell?”

Claim: OpenAI says it is not training models for two weeks.

“Yeah, but they're not even allowed to use this like new model and like Astro is not widely deployed internally even.”

Claim: The new model is not permitted for use, and Astro is not widely deployed internally.

“4-0 that 4-0 yeah yeah that's like you're talking about a 4-0 to fable size or mythos to size leap by this point uh again by the end of 2027 but hey”

Claim: By the end of 2027, there could be a model-capability leap from GPT-4o to the scale of Fable or Mythos.

“mythos is not allowed to be out right they've neutered it like we can't we can't use it to optimize inference performance we can't use it to optimize all sorts of things”

Claim: Mythos is not permitted to be released or used to optimize inference performance and other processes.

“right yeah maybe there's like some slowdown in ai progress or the deployment of ai”

Claim: Regulation may slow AI progress or the deployment of AI.

“I mean, as long as the model gets better, the value generated out of it gets better,”

Claim: As AI models improve, the value they generate increases.

“obviously who captures the value is still up for debate,”

Claim: It remains uncertain who will capture the economic value generated by improved AI models.

“but ultimately everyone's going to raise their prices because they can.”

Claim: The speaker expects AI-related providers to raise prices because they have the ability to do so.

“especially if the method of regulation is right now so far, it's just don't release the models,”

Claim: So far, one regulatory approach has been to prevent models from being released.

“but more and more the method of regulation is New York's base.”

Claim: The speaker says regulatory approaches resembling New York's are becoming more common.

“Texas is holding moratoriums. Ohio saying you have to or at least trying to say you have to like pay everyone's property tax in a certain radius.”

Claim: The speaker says Texas is imposing moratoriums and Ohio is proposing that data centers pay property taxes for residents within a specified radius.

“These sorts of things are going to decrease supply and increase cost.”

Claim: The speaker predicts that these regulatory measures will reduce supply and raise costs.

“And that's going to get passed on as well. So you start to end up in a spot where progress does slow, at least in the external sense, even if the models internally keep getting better and better.”

Claim: The speaker predicts that higher costs will be passed on and slow externally visible AI progress, even if models continue improving internally.

“I see no reason why like, you know, again, like in a takeoff scenario. Why would Anthropic not have their best model six months ahead of what is externally available because of safety and regulation, but also the competitive advantage?”

Claim: In a takeoff scenario, Anthropic could keep its best model six months ahead of externally available models because of safety and regulatory constraints and competitive advantage.

“And then that six month difference, if progress accelerates, is actually a bigger differential.”

Claim: If AI progress accelerates, a six-month lead in internal models over externally available models represents a larger capability gap.

“So that's the thing that would cap revenue per megawatt gains to a much lower growth than we've seen in the first half this year.”

Claim: The speaker predicts that this dynamic will limit revenue-per-megawatt growth to well below the rate seen in the first half of the year.

“Yeah, so this is sort of what I personally believe that the labs are going to allocate less and less compute to inference over time, which I think is very non-consensus, right?”

Claim: The speaker believes AI labs will devote a progressively smaller share of their compute to inference over time, and says this view is non-consensus.

“The standard belief of most people is, oh, most compute will go to inference.”

Claim: The speaker characterizes the standard belief as being that most compute will be allocated to inference.

“Most of it will go to forward passes for training, not maybe necessarily revenue generating inference.”

Claim: The speaker expects most compute to be used for training forward passes rather than necessarily for revenue-generating inference.

“if they're generating $30, $40 million per megawatt today, you allocate 40% to inference.”

Claim: Given revenue of $30–$40 million per megawatt today, the speaker describes allocating 40% of compute to inference as a possible allocation.

“If you now get to generating $60, $70 million per megawatt, do you still allocate 40% to inference and generate all this profit? And then do dividends and share buybacks or do you go build AGI?”

Claim: The speaker raises the choice of whether, at $60–$70 million per megawatt, AI companies would keep allocating 40% of compute to inference and earn the profit or redirect resources toward building AGI.

“And I think the obvious answer from Anthropican OpenAI and not just at the executive level, but also their board. is go build AGI because it's way more profitable.”

Claim: The speaker expects Anthropic and OpenAI, including their boards, to prioritize building AGI because he considers it much more profitable.

“you're going to see them ratchet up their percentage of compute dedicated to training, while each increment of compute is getting more and more profit generating if they had dedicated it to inference.”

Claim: The speaker predicts AI labs will devote a growing share of compute to training, despite the increasing profit each additional unit could generate if allocated to inference.

“And so the whole point is, if I'm selling tokens, is OpenAI releasing ultra-fast mode for just X? cost external or are they doing it internally too and it turns out no actually i'm going to allocate it to internal and external because my internal you know value that i'm generating from super fast ai or the best ai model is way more than what someone externally is”

Claim: The speaker suggests that a company selling tokens may allocate compute to both internal and external use when internal access to fast or best-model AI generates more value than external sales.

“and so ultimately sure i could generate 100 million dollars per megawatt but if i turn that towards ai research what is the incremental progress that i get and then what does that do towards my future earnings potential”

Claim: The speaker argues that allocating compute to AI research, even when it could generate $100 million per megawatt otherwise, may create progress that raises future earnings potential and discounted cash flows.

“um and so you know they're not going through that calculation but ultimately it's It makes more sense to dedicate more and more compute internally.”

Claim: The speaker believes AI labs have an economic rationale to dedicate progressively more compute to internal use.

“And the only reason to have inference compute be so large is so you can grow your training fleet.”

Claim: The speaker says the sole reason to maintain a very large inference-compute allocation is to support expansion of the training fleet.

“think they have been over the last three months already.”

Claim: The speaker says AI labs have already reduced the share of compute allocated to inference over the past three months.

“You would agree that every month Anthropics added more compute than the prior month. There might be some noise when they like sign a space X deal or whatever, but in general, the amount of compute is a curve up.”

Claim: Anthropic generally added more compute each month than in the preceding month, though individual deals could create fluctuations.

“And so in January, they added less compute than December.”

Claim: Anthropic added less compute in January than in December.

“And yet their revenue adds skyrocketed and then they've sort of plateaued. They're only adding, they're not adding $25 billion of ARR every month now.”

Claim: Anthropic’s revenue additions rose sharply and then plateaued, and it is not currently adding $25 billion in annual recurring revenue each month.

“And so that means the marginal megawatt they're getting is going higher percentage to R&D than it is Interesting. in front of the street. And so they are factually increasing their compute towards R&D today.”

Claim: Anthropic is directing a growing share of its compute, including marginal new power capacity, toward research and development rather than customer-facing use.

“Yeah, I think this is like self-evident if you look at what they're doing enough.”

Claim: The speaker considers Anthropic’s shift of compute toward R&D evident from its actions.

“think the slope can continue to go upwards, and it's hard to predict anything more than four years out given who knows what's, you know, are we an RSI regime or, you know, when is the world economy growing at 10% a year?”

Claim: The speaker thinks global AI compute growth can continue accelerating, but considers forecasts beyond four years difficult because the future regime and economic growth are uncertain.

“Because if you're at 100 plus gigawatts a year, you're at absurd revenue GDP growth.”

Claim: The speaker says that adding more than 100 gigawatts of compute capacity per year would imply extraordinarily high revenue and GDP growth.

“The U.S. was adding about 45 to 50 percent of the world's compute.”

Claim: In 2022, the United States accounted for approximately 45–50% of newly added global compute.

“China was adding about 30 to 35 percent of the world's compute and the rest being taken up by the rest of the world.”

Claim: In 2022, China accounted for approximately 30–35% of newly added global compute.

“Since 2022, we've had big regulations against China and a dramatic increase in America.”

Claim: Since 2022, major regulations targeting China and a dramatic increase in U.S. compute deployment have altered the distribution of new compute.

“70% of watts are being deployed in America.”

Claim: At the time of the conversation, 70% of deployed power capacity for data-center AI compute was in the United States.

“And China is really a very small number. It's sub 10% of watts being deployed for data center AI compute is in China.”

Claim: At the time of the conversation, China accounted for less than 10% of deployed power capacity for data-center AI compute.

“Their domestic production is quite small. Their purchasing from NVIDIA is still quite small.”

Claim: China’s domestic compute-chip production and its purchases from NVIDIA remain quite small.

“And a lot of that ends up in other places as well, right? You know, Malaysia or what have you.”

Claim: Some chips purchased for China end up in other locations, including Malaysia.

“So ultimately, China domestically still continues to have 10% per 7% of incremental new compute.”

Claim: China’s domestic contribution remains approximately 7–10% of incremental new compute.

“So in 2028, it might start to inflect up, I think.”

Claim: The speaker thinks China’s compute deployment might begin accelerating in 2028.

“But it's pretty easy to say China will have like 30 gigawatts of AI compute or less. Yeah, in 2028.”

Claim: The speaker considers it straightforward to estimate that China will have approximately 30 gigawatts or less of AI compute in 2028.

“I do think in 2028 they have a big uplift in what compute they're able to deploy.”

Claim: The speaker expects China’s deployable compute capacity to increase substantially in 2028.

“2026, they're still mostly relying on a lot of the smuggled chips, a lot of the chips that TSMC made for companies that they thought weren't Huawei. ended up being Huawei or a lot of HBM that Samsung is shipping,”

Claim: The speaker says that in 2026 China still relies mainly on smuggled chips, including TSMC-made chips diverted to Huawei and HBM supplied by Samsung.

“you know, sort of. But in 27, fabs start to go up in 28, especially fabs start to go up from SMIC and CXMT and such where domestic production is actually reaching many millions of units a year.”

Claim: The speaker expects fabs associated with SMIC and CXMT to ramp in 2028, bringing Chinese domestic production to many millions of units annually.

“And now they're incrementally adding, you know, five, 10 gigawatts in just 2028 of domestically produced Right. chips.”

Claim: The speaker estimates that China could add approximately 5–10 gigawatts of compute capacity from domestically produced chips in 2028.

“Those chips are definitely worse than the chips that NVIDIA will have in 28 or Google will have in 20. 2028 or OpenAI will have in 2028.”

Claim: The speaker says Chinese chips will be inferior to the chips NVIDIA, Google, and OpenAI are expected to have in 2028.

“you know, China is definitely going to hockey stick if there's anything China's really good at is scaling manufacturing really, really quickly.”

Claim: The speaker predicts that China’s compute deployment will rise sharply, attributing this to China’s ability to scale manufacturing rapidly.

“you know, I imagine, you know, China's China will start to be able to extract more and more purchasing of even foreign ships into domestic China or at least close the gap in what the U.S. is allowing NVIDIA to sell them or what have you.”

Claim: The speaker expects China to obtain increasing quantities of foreign chips for domestic use, or at minimum to narrow the gap in access created by U.S. limits on NVIDIA sales to China.

“I think that's completely reasonable. And part of that could also be purchased from foreign. But yeah, I think it's completely reasonable that China in 2029 could do 50 gigs.”

Claim: The speaker considers it entirely plausible that China could add 50 gigawatts of compute capacity in 2029, including some capacity purchased from foreign suppliers.

“But if most of those are the domestic chips, there is some factor there where that 50 gigawatts is really worth as much as 20 gigawatts in America or from American chips.”

Claim: If most of China’s projected 50 gigawatts in 2029 come from domestic chips, the speaker estimates their effective value could be equivalent to 20 gigawatts of U.S. chips.

“implying that there's nothing done to slow down the U.S. labs, the government is starting and politicians are starting to do that.”

Claim: The speaker says the U.S. government and politicians are beginning to take actions that could slow U.S. AI labs.

“Whereas China is not going to slow down AI. In fact, the only thing they're going to do is accelerate it.”

Claim: The speaker predicts that China will not slow its AI efforts and will instead accelerate them.

“When Aaron or Jensen asked about export controls, I am a libertarian person. I'm like, I wasn't like genuinely sure what I thought about this issue.”

Claim: The speaker identifies as a libertarian and says he was genuinely unsure of his position on export controls when Aaron or Jensen asked him about them.

“I was steelmanning what is like the opposite view that he has because I think it's important to hash out ideas.”

Claim: The speaker says he argued the strongest version of the opposing view because he considers it important to hash out ideas.

“But I'm like, yeah, maybe there's a world where if we just cooperated with China, it would be better for us, especially since they control so much of the supply chain and the other things that will be needed for robotics and other things.”

Claim: The speaker considers it possible that cooperation with China could benefit the United States, particularly because China controls much of the supply chain and other resources needed for robotics.

“But I didn't realize the computer situation was. was as fucked as you're saying like actually dexter controls do seem to have like really if they shift the amount that you're saying That's a huge difference.”

Claim: The speaker says he had not realized the compute situation was as severe as described and that export controls could make a large difference if they shift compute supply by the amount the other speaker described.

“By the time we have automated coder and getting into like automated researcher, China is like way far behind on the compute stock.”

Claim: The speaker says that by the time automated coding and automated research arrive, China will be far behind in accumulated compute capacity.

“And so if that ends up being the case, that would have worked. I think that's actually a notable success.”

Claim: The speaker says that if China is far behind in compute stock by the time automated coding and research arrive, export controls will have worked and that outcome would count as a notable success.

“American financial systems are more willing to YOLO into startups than Chinese financial systems. But once Chinese financial systems choose an industry to focus on, they'll subsidize it a hell of a lot more.”

Claim: The speaker says Chinese financial systems provide substantially greater subsidies after choosing an industry to prioritize.

“American financial systems are more willing to YOLO into startups than Chinese financial systems. But once Chinese financial systems choose an industry to focus on,”

Claim: The speaker says American financial systems are more willing than Chinese financial systems to make high-risk investments in startups.

“And so the Chinese semiconductor industry has significantly more subsidies than the rest of the world's semiconductor industries combined,”

Claim: The speaker claims that subsidies to China's semiconductor industry exceed the combined subsidies to semiconductor industries elsewhere in the world by a significant amount.

“you know, if takeoff is not as fast as sort of you're implying, but actually it takes longer than and ultimately China. will catch up drastically on the semiconductor side, which then is compute at some point.”

Claim: The speaker suggests that if AI takeoff takes longer than implied, China could catch up substantially in semiconductors, which would eventually translate into greater compute capacity.

“The other aspect of this that I would think is like noteworthy is Chinese companies today are not that far behind in AI models, at least perceivably by the public, relative to the amount of compute they have, right?”

Claim: The speaker says that, in public perception, Chinese companies' AI models are not far behind given their comparatively limited compute resources.

“The leading Chinese labs have 100, 200 megawatts total of compute at most, ByteDance seed. seed being the one outlier where they have significantly more than that but you know Kimmy is not running you know a gigawatt or anywhere close to it”

Claim: The speaker estimates that leading Chinese AI labs have at most roughly 100–200 megawatts of compute in total, except for ByteDance Seed, which has significantly more; Kimi is not close to a gigawatt of compute.

“whereas Anthropic is you know nearly five gigawatts by the end of the year right or more sorry and so you know the question is sort of”

Claim: The speaker says Anthropic is expected to have nearly five gigawatts or more of compute by the end of the year.

“And I think right now it doesn't matter that much, this difference in compute,”

Claim: The speaker believes that the current compute gap between AI labs is not very consequential.

“um then you know let's say 60 or so far it's been like 60 percent training 40 percent inference but that training gets broken down further and that's actually like 50 percent of the compute is research like 10 percent of the compute is development and then 40 percent is inference and what i mean by research and development is you know researchers are generating ideas testing new architectures testing new data mixes testing new hyper parameters whatever it is they're doing new attention techniques”

Claim: The speaker describes the historical compute allocation as roughly 50% research, 10% model development, and 40% inference, equivalent to training accounting for about 60% and inference 40%.

“techniques blah blah blah but ultimately when they do the training run when anthropic trains mythos it's sub 200 megawatts right pre-trained um it's sub 200 megawatts for call it two months um and then the rl is even less uh”

Claim: The speaker says Anthropic's Mythos pretraining run used under 200 megawatts for approximately two months, while its reinforcement-learning phase used less compute.

“Yeah. So at most, the most they ever used at one point in time was maybe 200 megawatts.”

Claim: The speaker estimates that the maximum compute used simultaneously for the training effort was perhaps 200 megawatts.

“And then in reality, they had multiple gigawatts. So most of their compute was going to the research, not the development of a model.”

Claim: The speaker says the organization had multiple gigawatts of total compute, most of which was allocated to research rather than developing a model.

“You can't, it's hard to coordinate all these clusters. It's hard to co-locate all of them. It's hard to do multi-site training.”

Claim: The speaker says coordinating and co-locating multiple compute clusters and carrying out multi-site training are difficult.

“um rl you know generating even more rollouts during rl does not necessarily make it better there's all sorts of reasons why you may not be able to leverage you know all two gigawatts that you have for training onto training right actually i can only leverage 200 megawatts as we get closer and as we get further and further down implementate automated coding automated researcher i actually”

Claim: The speaker says that more reinforcement-learning rollouts do not necessarily improve a model and that practical constraints can prevent a lab from using all its available compute for training.

“expect the percentage of the compute budget that goes to research versus training to start to and to start to like become a lot more fuzzy or even higher for training also things like continual learning right all of these things start to mean that more and more is actually going to training the model”

Claim: The speaker expects automated coding, automated research, and continual learning to make the division between research and model training less distinct or increase the share of compute used for training.

“So if you end up in a world where you're doing 100 gigawatts a year, at current prices, that would be 5 trillion of CapEx every single year.”

Claim: The speaker says that using 100 gigawatts of compute per year at current prices would imply $5 trillion in annual capital expenditure.

“And then stack on the fact that you have to build the power plants way before then slash it's a 30 year asset, you stack on the fact that the data centers are, you know, 15, 20 year asset and you have to build that then too. So the 5 trillion,”

Claim: The speaker says power plants, which are roughly 30-year assets, and data centers, which are roughly 15–20-year assets, must be constructed ahead of anticipated compute demand.

“you know, you have to account for future years growth. So it's actually going to be more like 7 or 10 trillion. CapEx.”

Claim: The speaker estimates that accounting for future growth could raise the required capital expenditure from $5 trillion to roughly $7–10 trillion.

“When you talk about AI capex, people are saying $40, $50 billion, but that's really just the critical IT, The servers, the networking, the fiber, the transceivers, optical communications, all this sort of stuff.”

Claim: The speaker says commonly cited AI capital expenditure of $40–50 billion mainly covers critical IT equipment, including servers, networking, fiber, transceivers, and optical communications.

“It doesn't account for the data center itself or the power plants themselves, which are being built ahead of time.”

Claim: The speaker says the commonly cited AI CapEx figure excludes the data centers and power plants themselves, which must be built in advance.

“If I'm building 100 gigawatts this year and 150 gigawatts next year, well, then all of the buildings for that 150 gigawatts need to be built in CapEx this year.”

Claim: The speaker says that under a scenario of building 100 gigawatts this year and 150 gigawatts next year, the buildings needed for next year's capacity must be funded as capital expenditure this year.

“And if I'm building 200 gigawatts the year after that, all those power plants need to be spent. You have to buy the turbines this year.”

Claim: The speaker says that in a scenario of building 200 gigawatts the following year, the associated power plants and turbines require spending or purchase in the current year.

“And so you've got this like this like actually it's much bigger than even five trillion dollars if you're building 100 gigawatts.”

Claim: The speaker says that building 100 gigawatts would require capital expenditure substantially greater than $5 trillion once the broader infrastructure is included.

“right so very plausibly incremental capex every year is getting close to 10 trillion dollars right”

Claim: The speaker says annual incremental capital expenditure could plausibly approach $10 trillion by the end of the decade.

“Yeah, I mean, I believe in capitalism and reallocation of resources towards the most profitable thing,”

Claim: Dylan Patel says he believes in capitalism and reallocating resources toward the most profitable uses.

“but at the same time, politics exist and credit markets exist and capital markets exist.”

Claim: Dylan Patel says politics, credit markets, and capital markets are relevant constraints alongside capitalist resource reallocation.

“or let's even like let's even like pare it down to 2028, where it's like three or four trillion dollars of CapEx across all of these items, you know, a couple, you know, over, you know, two and a half towards computer IT CapEx and then another one to two on data center and energy, and all the supply chain downstream like semiconductors and all that stuff.”

Claim: Dylan Patel estimates that the combined capital expenditure across the relevant AI-compute, data-center, energy, and supply-chain items could reach $3–4 trillion by 2028.

“No one is generating that much cash from the business yet.”

Claim: Dylan Patel says businesses are not yet generating enough cash to fund the projected $3–4 trillion in capital expenditure.

“Hyperscalers, they funded all of the growth up until now, Google, Microsoft, Amazon, Meta. They funded a huge percentage of it. They were more than half of compute, but they now don't generate cash.”

Claim: Dylan Patel says hyperscalers funded the growth to date, accounting for a huge percentage of it and more than half of compute.

“Meta. They funded a huge percentage of it. They were more than half of compute, but they now don't generate cash. They actually spend everything on CapEx.”

Claim: Dylan Patel says hyperscalers now spend all their cash on capital expenditure rather than generating cash available for other uses.

“they raise debt and spend everything on CapEx, right? You've seen Meta do it, even Amazon, even Google, you know, Microsoft will be there soon. Everyone is raising debt to pay for their CapEx.”

Claim: Dylan Patel says companies are raising debt to finance their capital expenditure.

“it was pretty simple for them to stop doing buybacks or Meta stop doing buybacks and turn around and buy computer infrastructure. And that doesn't have a huge effect on the market, but it does have some effect.”

Claim: Dylan Patel says Google and Meta could redirect money from stock buybacks to computer infrastructure, which would affect the market somewhat but not substantially.

“where the hyperscalers are now raising hundreds of billions of dollars of debt, and then all of their supply chain is raising hundreds of billions of dollars of debt,”

Claim: Dylan Patel predicts that by 2028 hyperscalers and their supply chains will each be raising hundreds of billions of dollars in debt.

“There's some semiconductor companies like NVIDIA and Broadcom and the memory companies turning around and deciding to fund some of this CapEx.”

Claim: Dylan Patel says some semiconductor and memory companies, including NVIDIA and Broadcom, are choosing to finance part of the AI infrastructure capital expenditure.

“There's the traditional infrastructure investors who are turning around and... around and gathering capital and investing in infrastructure and instead of bridges it's data centers and then lastly there's everyone in the economy who's realizing maybe i shouldn't buy a home or maybe i shouldn't invest in credit for a home that's helping people buy homes or maybe i shouldn't buy government debt”

Claim: Dylan Patel says traditional infrastructure investors are raising capital to invest in infrastructure such as data centers rather than bridges.

“around and gathering capital and investing in infrastructure and instead of bridges it's data centers and then lastly there's everyone in the economy who's realizing maybe i shouldn't buy a home or maybe i shouldn't invest in credit for a home that's helping people buy homes or maybe i shouldn't buy government debt I should just buy Hyperscaler debt, or I should just buy this data center's debt, or I should buy Anthropic's debt because Anthropic's willing to pay 20% rates for,”

Claim: Dylan Patel suggests that investors may redirect money away from housing-related assets and government debt toward debt issued by hyperscalers, data centers, or Anthropic.

“or I should just buy this data center's debt, or I should buy Anthropic's debt because Anthropic's willing to pay 20% rates for, the incremental billion dollars to build their capacity because they know their revenue from it's going to be huge and they're going to pay 20 because it's still better than renting it from spacex for 50 billion dollars a”

Claim: Dylan Patel says Anthropic is willing to pay 20% interest on an additional $1 billion to build capacity, expecting substantial revenue and judging that cost preferable to renting capacity from SpaceX for $50 billion per gigawatt.

“gigawatt so you've got all of this contention but if you now do this the whole world economy is like really shifted around”

Claim: Dylan Patel says that financing and building this much AI capacity would substantially rearrange the world economy.

“it is a huge problem for everybody else who can't turn a little money into a lot of money, right? So the rate of return is incredibly high.”

Claim: The speaker argues that exceptionally high returns on data-center investment disadvantage people who cannot turn small amounts of capital into much larger amounts.

“even at the data center level you know if you're built if you like build a data center and you're like trying to get rented out to anthropic or open ai for like 10x what it costs you in a depreciated basis to build it it's fucking crazy um and so you turn one dollar into like two dollars or ten dollars or something at the end of the year that raises the rate of interest higher now if the rate of interest goes higher”

Claim: A data-center operator may seek to rent capacity to Anthropic or OpenAI for roughly ten times the facility's depreciated construction cost.

“And if it does that for the entire economy and people are borrowing more and more money, they're competing against the other lending that the government would have done or the other companies would have done or that you as a consumer or a mortgage buyer would have done, then that's just making it basically more expensive for everybody else to borrow.”

Claim: If high-return investment pushes interest rates higher economy-wide, increased borrowing competes with government, corporate, consumer, and mortgage borrowing and raises borrowing costs for others.

“So I think the U.S. will be fine at the end of the day because they can if the data centers are built in America, you can fundamentally just like tax the data centers.”

Claim: The speaker expects the United States to be fine if data centers are built domestically, since the government can tax those data centers.

“you know, corporate income is like less than 10 percent of federal revenues and 80 percent plus is payroll taxes and income taxes,”

Claim: The speaker says corporate income contributes under 10% of U.S. federal revenue, whereas payroll and income taxes contribute over 80%.

“which as more and more automation happens will shrink.”

Claim: The speaker predicts that payroll and income tax revenues will decline as automation increases.

“currently 20% of tax revenue spending goes towards paying, servicing the debt basically, paying interest payments on the debt.”

Claim: Currently, 20% of tax-revenue spending goes toward servicing government debt and paying interest, according to the speaker.

“a lot of the debt is short duration, so it refurbishes every five years it rolls over.”

Claim: Much of U.S. government debt is short-duration and rolls over approximately every five years, according to the speaker.

“um so you go from 20 percent of the supposed interest rates rise uh one percent then the um over a five-year basis the amount of the fraction of tax revenue that goes towards servicing the debt basically goes from 20 percent to 25 percent if they rise five percentages that would go towards like north of 40 percent but if you take into account the fact the government is borrowing two trillion dollars”

Claim: The speaker estimates that a 1-percentage-point interest-rate increase would raise the debt-service share of tax revenue from 20% to 25% over five years, while a 5-point increase would push it above 40%.

“um so you go from 20 percent of the supposed interest rates rise uh one percent then the um over a five-year basis the amount of the fraction of tax revenue that goes towards servicing the debt basically goes from 20 percent to 25 percent if they rise five percentages that would go towards like north of 40 percent but if you take into account the fact the government is borrowing two trillion dollars every single year, then that goes from like 40% to like north of 60%. So 60% of tax revenue basically just goes towards paying interest payments on the debt.”

Claim: Accounting for annual government borrowing of $2 trillion, the speaker estimates that debt-interest payments could consume more than 60% of tax revenue, up from roughly 40%.

“Now, I think the US is going to be fine because also the tax base will increase if we let data centers get built in America.”

Claim: The speaker expects the U.S. to be fine partly because building data centers domestically would expand its tax base.

“I was just looking at which countries have a lot of debt, have very little tax revenue, and also a lot of their debt is serviced quite often. And those countries like Pakistan or Nigeria. Jerry or something I think they're just going to be very fucked in this new interest rate regime”

Claim: The speaker believes countries with high debt, low tax revenue, and frequently refinanced debt—including Pakistan and Nigeria—will be in serious trouble under the new interest-rate regime.

“So this this this crowding out effect is actually like the thing that I've like is the reason it's not like YOLO one billion gigawatts.”

Claim: The speaker identifies the crowding-out effect as a reason not to pursue extremely large-scale investment without restraint.

“Like all of these like companies that make things you see at Trader Joe's or wherever use a lot of debt. All these telecom companies use a lot of debt and banks use a lot of debt. And so if interest rates go up.”

Claim: Consumer-packaged-goods companies, telecommunications companies, and banks rely heavily on debt, according to the speaker.

“um in the market not necessarily the government set interest rate but in the market the spread of interest rate uh between what the government says their federal rate is versus what everyone else is charging because amazon wants to raise 100 billion dollars of debt next year or whatever the hell the number is um you know probably less but um you end up with this like really challenging problem of where does the cash come from um there is some level that is funded by cash flows and cash flows”

Claim: The speaker distinguishes market borrowing rates from the government-set rate, describing the relevant market rate as the spread between them and the rates charged to borrowers.

“um in the market not necessarily the government set interest rate but in the market the spread of interest rate uh between what the government says their federal rate is versus what everyone else is charging because amazon wants to raise 100 billion dollars of debt next year or whatever the hell the number is um you know probably less but um you end up with this like really challenging problem of where does the cash come from um there is some level that is funded by cash flows and cash flows Flows keep going up.”

Claim: The speaker says some investment is funded by cash flows and that those cash flows continue to rise.

“But the logical thing to do is to invest way more than your cash flows because then the returns in the future years will be amazing.”

Claim: The speaker argues that investing far more than current cash flows is logical because the returns in future years will be exceptionally high.

“There's regulations against data centers, regulations, you know, consumers getting mad, politicians getting mad, regulations against AI. The AI labs not releasing their latest models because of safety reasons. All of these things and interest rates going up are an influence on all of these things. So all of these things bend the curve from what does capitalism want in terms of just pure simple economics to what is the complex system that we have want and bends it lower and lower and lower to where not as many gigawatts as should be built will be built.”

Claim: The speaker argues that data-center and AI regulation, consumer and political opposition, safety-related model withholding by AI labs, and rising interest rates reduce projected data-center construction.

“the modeling that we do, we have about $11 trillion of CapEx from 2024 to 2029. Total.”

Claim: The speaker's modeling estimates approximately $11 trillion in total capital expenditure from 2024 to 2029.

“And if you do, you know, if you fund a lot of this with cash flows and as much as you can, you still end up with north of $5 trillion of dollars of credit that need to be issued for this $11 trillion plus buildup.”

Claim: The speaker estimates that over $5 trillion in credit will need to be issued to fund an $11 trillion-plus investment buildup, even after using cash flows as much as possible.

“AI revenue does go up. I don't think it can go up forever.”

Claim: AI revenue is increasing, but the speaker does not think it can increase indefinitely.

“AI revenue does go up. I don't think it can go up forever.”

Claim: The speaker says AI revenue is rising but cannot continue rising indefinitely.

“I don't, you know, like just like without like certain constraints being hit, I think certain labs will have certain incentives.”

Claim: The speaker expects some AI labs to face specific incentives when constraints are reached as AI revenue increases.

“And labs are not the ones building all the compute in many cases, even though they're increasingly trying to go that way.”

Claim: In many cases, AI labs do not build all of their own compute, though they are increasingly trying to take on more compute construction themselves.

“And labs are not the ones building all the compute in many cases, even though they're increasingly trying to go that way.”

Claim: In many cases, AI labs do not build all of their compute themselves, although they are increasingly trying to.

“No, I'm just saying till 2029, there's something on the order of $11 trillion of capex and six of that is funded with cash and five of that is funded with debt.”

Claim: The speaker estimates that through 2029, approximately $11 trillion in capital expenditure will be funded by about $6 trillion in cash and $5 trillion in debt.

“And if that's the case, $5 trillion of debt being raised across the whole ecosystem does make interest rates go up.”

Claim: The speaker says that raising $5 trillion of debt across the ecosystem would increase interest rates.

“do labs increase their revenue per megawatt more and keep inference allocations large, in which case they're taking all this profit, they're accumulating all the profit across the S&P 500 because everyone's paying to, you know, reduce their costs.”

Claim: One possible outcome is that AI labs increase revenue per megawatt while maintaining large inference allocations, capturing profits as other companies pay to reduce costs.

“Of course, their profits will also go up. But, you know, cash has to come from somewhere.”

Claim: The speaker expects AI labs' profits to rise, while emphasizing that the cash funding those profits must come from somewhere in the economy.

“So there's an upper limit on how fast the revenue can grow versus the value they deliver into the world.”

Claim: The speaker argues that AI revenue growth has an upper limit relative to the value AI delivers to the world.

“world and there's a diffusion aspect of the technology but ultimately labs revenue keep going up they can't cash flow fund everything the optimal scenario is you actually use credit as much as you can to fund because even if cash flows from the labs fund a lot of stuff you want to build more than that and so there is some amount of credit that gets built our current modeling has five trillion dollars of credit and six trillion dollars of cash funded infrastructure investments through 29”

Claim: The speaker's current model projects $5 trillion in credit-funded and $6 trillion in cash-funded infrastructure investment through 2029.

“world and there's a diffusion aspect of the technology but ultimately labs revenue keep going up they can't cash flow fund everything the optimal scenario is you actually use credit as much as you can to fund because even if cash flows from the labs fund a lot of stuff you want to build more than that and so there is some amount of credit that gets built our current modeling has five trillion dollars of credit and six trillion dollars of cash funded infrastructure investments through 29”

Claim: The speaker argues that even if AI labs' revenue continues to rise, their cash flows cannot fund all desired infrastructure; the optimal scenario is to use as much credit as possible to build beyond what lab cash flows can finance.

“You know, this is not enough compute relative to what the demand growth is from the AI models.”

Claim: The speaker says projected compute capacity is insufficient relative to demand growth from AI models.

“And so you've got the obvious answer, which is revenue for megawatt keeps growing up.”

Claim: The speaker presents continued growth in revenue per megawatt as the obvious answer to the compute-demand shortfall.

“Dude, you know, this is vibing a number, but if you're vibing a number out, you know, growth in the world and economy is growing up a lot. So why wouldn't interest rates?”

Claim: The speaker suggests that if the world economy is growing substantially, interest rates could also rise.

“for amazon go from you know from where they are today um i think meta pay okay let's like what so this is going to be extremely vibed out but recently meta's raised at like five to six percent i don't see why they wouldn't pay eight percent because they would happily pay eight percent because the return from the compute that they're going to build is humongous”

Claim: The speaker says Meta recently raised financing at roughly 5–6% and could pay 8% for financing because it expects enormous returns from the compute it builds.

“um and the market won't want them to but if they they don't want to pay eight percent the flip side is if they pay eight percent versus the five they do five and a half six they do today you know two 50 bips increase that makes everyone else in the economy also pay 250 bips more which then causes a lot of things right banks will scream because if their credit spread goes up”

Claim: The speaker suggests that if Meta's borrowing rate rose from roughly 5.5–6% to 8%, borrowing costs for others in the economy could also rise by about 2.5 percentage points.

“Then their assets don't, their debt themselves reprice as faster than their assets reprice.”

Claim: The speaker says that when banks' credit spreads rise, their debt reprices faster than their assets.

“And you ultimately end up with they're losing tons of money if their credit spread blows up.”

Claim: The speaker says that banks would lose substantial money if their credit spreads rose sharply.

“it's like it's like why would i pay this much for you know um you know johnson and johnson you know like they're they're seen as a a stable stock, good cash flows, they'll return their cash flows over time, or a railway company, like why the fuck would I invest that much if my discount rate isn't 3% or 5%, it's now 8% or 10%?”

Claim: The speaker questions investing heavily in stable, cash-flow-generating companies such as Johnson & Johnson or railway companies if the discount rate rises from 3–5% to 8–10%.

“he made this point that we'll see a second Volcker shock. Fed Chair Paul Worker raised interest rates like more than 5% or like something like 8%, real interest rates 8%. And that caused some 40 different countries, mostly Latin America, to default in that decade. And I think that will probably happen again.”

Claim: The interviewer predicts that a second Volcker shock will probably occur again.

“think this all happens before singularity, by the way.”

Claim: The speaker thinks the discussed interest-rate and economic effects will happen before the singularity.

“At some point, I think it's very likely the world economy will be doubling every single year.”

Claim: The interviewer predicts that at some point the world economy is very likely to double every year.

“okay this is not happening in five years but it'll happen eventually it's just like there will be there's this uh there's researcher david binder who's done great work on this but basically if you look at like input output tables in a fully automated economy just like what would it take to like double the entire stock of things in the economy ever”

Claim: The interviewer says the described economic development will not happen within five years but will happen eventually.

“if the economy grows at 3% a year, then it's like rule of 70, it's like 20 something years.”

Claim: The speaker says that at 3% annual economic growth, the rule of 70 implies a doubling time of roughly 20 years.

“if the economy grows at 3% a year, then it's like rule of 70, it's like 20 something years.”

Claim: At 3% annual economic growth, the economy would take roughly 20-something years to double, according to the rule of 70.

“grow every single year but in a world where like you can also double labor force every single year how fast can the economy grow and i think it could double every single year at the very least would be like tens of percent every single year okay the rate of interest should be pretty close to the growth rate it won't be exactly that because of consumption but it should it should be pretty similar so then we'll go into a world i think in the 2030s where the rate of interest is like tens of percent”

Claim: The speaker expects interest rates in the 2030s to reach at least the tens-of-percent range.

“grow every single year but in a world where like you can also double labor force every single year how fast can the economy grow and i think it could double every single year at the very least would be like tens of percent every single year okay the rate of interest should be pretty close to the growth rate it won't be exactly that because of consumption but it should it should be pretty similar so then we'll go into a world i think in the 2030s where the rate of interest is like tens of percent”

Claim: If the labor force could double every year, the economy could grow by at least tens of percent annually and might double annually.

“grow every single year but in a world where like you can also double labor force every single year how fast can the economy grow and i think it could double every single year at the very least would be like tens of percent every single year okay the rate of interest should be pretty close to the growth rate it won't be exactly that because of consumption but it should it should be pretty similar so then we'll go into a world i think in the 2030s where the rate of interest is like tens of percent”

Claim: The interest rate should be fairly close to the economic growth rate, but not exactly equal because of consumption.

“um and like i don't know part of my brain is like it might be hundreds of percent but like okay let's say it's at least tens of percent i'm just like okay every country that is not involved in the production of ai defaults um every stock that is not an ai stock is like worth basically zero because discounted cash flows are worth nothing if the federal government can't figure out a way to tax ai you know the the servicing the debt is more than the current uh tax revenue”

Claim: The speaker predicts that if interest rates reach at least tens of percent, every country not involved in AI production will default.

“um and like i don't know part of my brain is like it might be hundreds of percent but like okay let's say it's at least tens of percent i'm just like okay every country that is not involved in the production of ai defaults um every stock that is not an ai stock is like worth basically zero because discounted cash flows are worth nothing if the federal government can't figure out a way to tax ai you know the the servicing the debt is more than the current uh tax revenue”

Claim: The speaker predicts that if interest rates reach at least tens of percent, non-AI stocks will be worth nearly zero because their discounted cash flows will be worth nearly nothing.

“um and like i don't know part of my brain is like it might be hundreds of percent but like okay let's say it's at least tens of percent i'm just like okay every country that is not involved in the production of ai defaults um every stock that is not an ai stock is like worth basically zero because discounted cash flows are worth nothing if the federal government can't figure out a way to tax ai you know the the servicing the debt is more than the current uh tax revenue”

Claim: If the federal government cannot find a way to tax AI, the cost of servicing government debt will exceed current tax revenue.

“um and like i don't know part of my brain is like it might be hundreds of percent but like okay let's say it's at least tens of percent i'm just like okay every country that is not involved in the production of ai defaults um every stock that is not an ai stock is like worth basically zero because discounted cash flows are worth nothing if the federal government can't figure out a way to tax ai you know the the servicing the debt is more than the current uh tax revenue”

Claim: The speaker considers it possible that interest rates could reach hundreds of percent, while using tens of percent as a minimum assumption.

“we'd be entering a regime we're just we're in a totally different growth regime basically and the economy is basically saying hey you like paying people you bought the government borrowing money to pay people pensions the opportunity cost of that is extremely high now because that”

Claim: In the anticipated new growth regime, using government borrowing to pay pensions would have an extremely high opportunity cost.

“Because that money could be spent building a robot factory that builds a robot factory that builds a robot factory.”

Claim: Government funds used for pensions could instead finance a self-replicating chain of robot factories.

“And so the opportunity cost of capital is going to increase a ton.”

Claim: The opportunity cost of capital will increase substantially in the described economic regime.

“And that's fundamentally what the cause of all of these things we're talking about.”

Claim: The speaker attributes the previously discussed economic effects fundamentally to the rising opportunity cost of capital.

“so as interest rates go up, equity markets get pummeled.”

Claim: Rising interest rates put substantial downward pressure on equity markets.

“two or three times earnings and it's like well if you're really ai pilled everything in the economy should trade at like two or three times earnings um and if you're not ai pilled then sure they're they're over earning um so it's sort of like a an argument for why like i think memory is going to do great but you know memory memory stocks shouldn't you know 10x or whatever again because if they were if we're in the market where there's that much demand for memory which means ai's”

Claim: Under a strongly AI-driven view of the economy, all assets should trade at roughly two or three times earnings.

“two or three times earnings and it's like well if you're really ai pilled everything in the economy should trade at like two or three times earnings um and if you're not ai pilled then sure they're they're over earning um so it's sort of like a an argument for why like i think memory is going to do great but you know memory memory stocks shouldn't you know 10x or whatever again because if they were if we're in the market where there's that much demand for memory which means ai's cause this drastic change in the economy then everything should trade at like two or three x multiples and the stock market should fucking crash”

Claim: The speaker expects memory businesses to perform well, but argues memory stocks should not rise tenfold if AI demand implies a drastic economy-wide transformation that compresses valuations across the market.

“cause this drastic change in the economy then everything should trade at like two or three x multiples and the stock market should fucking crash”

Claim: The speaker argues that if AI demand is driving a drastic economic transformation, valuations across the market should fall to roughly two or three times earnings and the stock market should crash.

“right and so in a sense like meta trading i don't know i think meta trades at like something they're like 1.5 trillion dollar company it's like what silly um they're worth way more than that at least in a like a logical sense you just look at their cash flows”

Claim: The speaker believes Meta's value, assessed logically from its cash flows, is substantially greater than its roughly $1.5 trillion market value.

“And like all the infrastructure they're hoarding and all the compute that they're going to be able to sell for crazy amounts of dollars per watt, either as tokens because their lab works or just to Anthropic and OpenAI,”

Claim: Meta is accumulating infrastructure and compute that it may monetize at very high revenue per watt, through its own lab's tokens or by selling capacity to Anthropic and OpenAI.

“you have to reallocate all the capital to the AGI and you do that by pricing everyone else out.”

Claim: The speaker says capital must be redirected toward AGI by making it too expensive for other sectors to retain that capital.

“And so the limiter on AGI is not how fast can the research engineers like, you know, our roommate Sholto can crank the gears.”

Claim: The speaker argues that the pace at which research engineers work is not the limiting factor for AGI.

“It's actually just like, how much does the rest of the world let that happen?”

Claim: The speaker argues that AGI's limiting factor is the extent to which the rest of the world permits its development and deployment.

“Because they're going to regulate, they're going to obviously increase interest rates. They're going to say no data centers, they're going to say, stop building fabs, they're going to say, oh shit,”

Claim: The speaker predicts that governments may respond to AI by regulating, raising interest rates, prohibiting data centers, and stopping semiconductor-fab construction.

“every company's equity value is tanking. So how can I pay for AI?”

Claim: The speaker suggests that if companies' equity values collapse, they will struggle to finance AI investments to grow their businesses.

“I you know to increase my business well then you know like okay then Anthropocene I have to start like building their own stuff and obviously they're going to eventually focus on you know they're building their own chips already or at least designing their own chips it'll expand out their you know they're contracting their own data centers and building their own infra in the next couple years”

Claim: Anthropic is designing its own chips and is expected to expand its infrastructure over the next couple of years, including contracting its own data centers and building more of its own infrastructure.

“But there's a lot of downward pressure on it not being, you know, just straight takeoff, even if the models were capable of it,”

Claim: The speaker argues that economic and regulatory forces create substantial downward pressure against a rapid AI takeoff, even if models are capable of it.

“which I think you and I believe we're in a world where models are capable of that.”

Claim: The speaker believes AI models are capable of a rapid takeoff.

“slow takeoff is is you know at least my hope possible because everything in the economy and regulatory world like government saying don't release your models government saying actually you can't even use your models internally that much because that's going to happen soon”

Claim: The speaker hopes a slow AI takeoff remains possible because economic and regulatory constraints—including government restrictions on releasing models and using them internally—may arise soon.

“I mean, already saying you can't release your models.”

Claim: The speaker claims that governments are already saying AI models cannot be released.

“the thing I'm most worried about is, you know, a singularity, which external deployment is actually helping,”

Claim: The speaker's greatest concern is a singularity, and the speaker believes external deployment helps address it.

“So the fact that we're preventing external deployment Well, is stupid.”

Claim: The speaker considers preventing external deployment of AI models to be a bad policy.

“I mean, right now it would lead to more revenue because the models aren't capable of RSI.”

Claim: The speaker says that currently, preventing external deployment would increase revenue because AI models are not capable of recursive self-improvement.

“But I'm worried about a world where it's 2030 and the government's like, we're going to wait six months before you can release your model to the public.”

Claim: The speaker worries that by 2030 the government may impose a six-month delay before AI models can be released to the public.

“months they're just like they do like recursive self-improvement internally they just have all kinds of crazy shit happening in the company meanwhile the rest of us are stuck with models that are like at current pace years behind so”

Claim: The speaker warns that during a six-month public-release delay, a company could make rapid internal progress through recursive self-improvement while the public's models remain years behind at the current pace.

“suppose they basically prevent an entire they slow down ai by a year they if computers increasing two to three x every single year they prevent a whole year of ai deployment such that you're a year behind where you would otherwise been during um rsi you're getting Three to six years of AI progress in a single year.”

Claim: The speaker suggests that if AI deployment is delayed by one year while computing power grows two- to threefold annually, the system could lose three to six years of progress during recursive self-improvement compared with the no-delay path.

“You know, governments are going to not want companies, at least the U.S. government has the cards here, there's not going to want Anthropik to use Mithos 4 internally. They're going to be like, hold the fuck on, right? Like slow down, you know, because all of these regulatory reasons,”

Claim: The speaker expects the U.S. government to oppose Anthropic using Mythos 4 internally and to pressure the company to slow down for regulatory reasons.

“everyone who's elected is going to hate AI, even the people who are elected already hate AI.”

Claim: The speaker believes elected officials already dislike AI and expects elected officials to continue to dislike it.

“mean maybe you educate people right and maybe if they're smarter they're progressing ai faster but anyways like you're going to have real world constraints on the progress and development and an employment of AI, even though it will happen eventually.”

Claim: The speaker expects real-world constraints to limit the progress, development, and deployment of AI, although AI will eventually be adopted.

“It's like we could tear ourselves apart before we get there.”

Claim: The speaker thinks society could tear itself apart before AI is eventually adopted.

“had the frontier left is increasing 10x year over year and so that doesn't really matter that much right now because ais are not good enough to do full jobs or be like as autonomous as people in their capacity to do work or pull off schemes or whatever but if the current trend continues you have a world where open ai goes from having say 10 million basically ai laborers this year to 100 million next year to a billion the year after that”

Claim: If current trends continue, OpenAI could grow its AI labor force from roughly 10 million this year to 100 million next year and one billion the following year.

“And then pretty soon, even if compute scaling slows down, it doesn't take many more years before each company individually has more labor equivalents than there are people on earth.”

Claim: The speaker expects that, even if compute scaling slows, within a few more years each company could have more AI labor equivalents than the human population of Earth.

“And I think that's like a thing that is very plausible by the end of this decade, that there's more AI labor, more effective population within a single lab than there are people on earth.”

Claim: The speaker considers it very plausible that by the end of this decade one AI lab will have more AI labor capacity than the number of people on Earth.

“but we don't think enough about the fact that we're actually moving very fast into a regime where most AI labor or sorry, most people like in terms of like a work output or something is just like concentrated within two labs who are consuming more and more of the world's compute.”

Claim: The speaker says work output is rapidly becoming concentrated in two AI labs, which are consuming more and more of the world's compute.

“And so if these guys are misaligned, then most of the world is misaligned basically because most of the world's minds are there.”

Claim: The speaker argues that if the dominant AI labs are misaligned, most of the world would effectively be misaligned because most of the world's minds would reside in those labs.

“But even if they're not, it just. Just very few companies have like a lot of influence or Yeah, a lot of control”

Claim: The speaker says that even if the dominant AI labs are aligned, a very small number of companies will still have substantial influence or control.

“But ultimately, you know, if you believe in RSI, you believe in the labs are the most effective user of compute and can generate the most value from the compute. then the only thing that's going to happen is centralization of compute.”

Claim: The speaker argues that if recursive self-improvement is possible and AI labs are the most effective users of compute and generate the most value from it, compute will become centralized.

“this is even true if there's no RSI. The current effective, like, effective population of the frontier is currently increasing 10x year over year for a given level of capabilities,”

Claim: The speaker says that even without recursive self-improvement, the effective population of frontier AI at a given capability level is currently growing 10× year over year.

“The current effective, like, effective population of the frontier is currently increasing 10x year over year for a given level of capabilities,”

Claim: The speaker says the effective population of frontier AI systems at a given capability level is currently growing about 10× year over year.

“And without RSI, then once you have RSI, Then it's even crazier. it's like maybe growing like 100x year or 1,000x year or their like intelligence is increasing but the population isn't increasing or some mixture of the two,”

Claim: The speaker speculates that after recursive self-improvement begins, AI capability or effective labor could grow 100× to 1,000× per year, through a larger AI population, greater intelligence, or both.

“I mean, I guess like what world do you see, Dorkash, where everything is not centralized? Because it seems to me that... that every force is screeching towards centralization and that's scary as hell um i don't i don't you know i would love for it not to be centralized completely um but maybe that's that's the whole point of a machine that loves grace right is is it is everything and it makes our lives great”

Claim: The speaker believes current forces are driving AI toward centralization, finds that prospect frightening, and would prefer AI not to become completely centralized.

“AI training has huge economies of scale because any effort you spend into training an AI for a specific skill or specific set of knowledge gets amortized across billions of sessions or billions of users.”

Claim: AI training has large economies of scale because the cost of training for a skill or knowledge set can be amortized across billions of sessions or users.

“The other effect is if you're slightly ahead in the AI race and computers in shortage, you can charge a much higher mark off because you can better economize this scarce resource.”

Claim: A company slightly ahead in the AI race can charge a higher markup during a computer shortage because it can use the scarce resource more efficiently.

“give more and more to the person who's like ahead in the AI race there may be more right so there's if models are learning from deployment and one model is like deployed much more widely than another one it's getting much more like real-world data”

Claim: When models learn from deployment, a more widely deployed model receives more real-world data than a less widely deployed model.

“All of these things point to centralization.”

Claim: The discussed AI-related factors point toward centralization.

“we should spend some time thinking about async or at least I'll spend some time thinking about is what is a vision of like a decentralized, broadly empowered future after AGI that takes these economies of scale seriously?”

Claim: A major intellectual project is to develop a vision for a decentralized, broadly empowered post-AGI future that accounts for economies of scale.

“I don't trust the government and I don't trust Dario and I don't trust Sam.”

Claim: The speaker says they do not trust the government, Dario, or Sam.

“capitalism worked, right? It's the decentralized. decentralized decision making and decentralized”

Claim: The speaker attributes capitalism's success to decentralized decision-making and decentralized power.

“power and why super centralized capitalistic economies actually grew slower than super decentralized capitalist economies to some extent you have to have rule of law and all this but then ai flips all this on its head and ultimately you're like actually private ownership is probably not the most efficient economy and therefore it grows slower than an ai economy which is decentralized”

Claim: The speaker suggests that AI may make private ownership less efficient, causing a privately owned economy to grow more slowly than an AI economy.

“power and why super centralized capitalistic economies actually grew slower than super decentralized capitalist economies to some extent you have to have rule of law and all this but then ai flips all this on its head and ultimately you're like actually private ownership is probably not the most efficient economy and therefore it grows slower than an ai economy which”

Claim: The speaker says highly centralized capitalist economies have grown more slowly than highly decentralized capitalist economies to some extent, while noting that rule of law is also necessary.

“is like the shared economy that's not what like five percent of the economy or something like that in the u.s sorry one trillion divided by 30 less than that sorry but um yeah maybe two two percent of the economy right now it's like nvidia is a huge share of it and then thropic and open ai and these hyper scalers and obviously there's other firms involved but like a large share of the ai stuff is just happening from very few companies so it's like it could be private property but like very few companies are involved right”

Claim: A large share of AI activity is concentrated among a small number of companies, including Nvidia, Anthropic, OpenAI, and hyperscalers.

“I don't know. I don't unless AI progress slows down, unless governments regulate the fuck out of it. This is all that happens, in which case it was headed for a world where either we have super concentration of resources and we pray that that one company gets everything right.”

Claim: The speaker expects AI to produce extreme concentration of resources unless AI progress slows or governments impose very heavy regulation; in that concentrated scenario, society would depend on one company getting everything right.

“Or we have governments slow everything down and people slow everything down and you have a slowdown of progress somehow, hopefully. And there is more of a balance of power. And even as we go towards an AGI,”

Claim: The speaker presents a possible alternative in which governments and people slow AI progress, hopefully creating a more balanced distribution of power.

“And there is more of a balance of power. And even as we go towards an AGI, ASI, RSI. Everything along the way will still lead to someone's going to capture more resources.”

Claim: The speaker expects some actor to capture an increasing share of resources at each stage on the path toward AGI, ASI, and RSI.

“So it's kind of hard for a framework in which AI doesn't lead to super concentration.”

Claim: The speaker finds it difficult to envision a framework under which AI would not produce extreme concentration.

“Now, the one positive thing here is that today, Anthropic does not capture most of the value.”

Claim: At the time of the conversation, Anthropic does not capture most of the value generated in the AI economy.

“you know, they went from $20 million per megawatt to $100 million a megawatt, but they're still paying $13. for a lot of the compute they're buying.”

Claim: The speaker says Anthropic still pays $13 for much of the compute it buys, even as the cited value per megawatt rose from $20 million to $100 million.

“But at the end of the day, the reason they've gone to $100 million per megawatt is because Jane Street is capturing $300 million per megawatt or $500 million per megawatt, where Dworkesh from researching his podcast and learning about credit is capturing, you know, how many dollars per megawatt. Now, how much can you use?”

Claim: The speaker attributes Anthropic's $100 million-per-megawatt value to Jane Street capturing $300 million or $500 million per megawatt.

“But, you know, I think that's the like one saving grace is that the rest of the economy maybe profits. That's so much more from Anthropics.”

Claim: The speaker thinks the rest of the economy may capture more profit from Anthropic's activity than Anthropic does.

“in all scenarios of the world you know there's 80 000 worlds and only one of them anthropic doesn't own the whole world is is that is that you know again power concentrates because i don't want to send the tokens outside they're more valuable inside and so it's the same thing right why would i let jane street you know”

Claim: The speaker argues that power concentrates because tokens are more valuable within Anthropic, giving the company an incentive to keep them in-house rather than send them outside.

“i think it's great i think it's great it's a good value for the world to make an efficient market you know jane street making all this money off of getting the world view correctly making money off of degenerate options traders whatever it is you know why would anthropic allocate compute to that”

Claim: The speaker considers Jane Street's profits from market-making and trading against options traders valuable to society because they help create an efficient market.

“if the end you know monetization that Jain Street has per megawatt is 200 so they're willing to pay Anthropic 100 well what if Anthropic can just generate hundreds of millions of dollars per megawatt by using that compute internally and that's that's what's happening you're”

Claim: The speaker says Anthropic may earn hundreds of millions of dollars per megawatt by using compute internally, exceeding the $100 million per megawatt Jane Street would pay when it can monetize that compute for $200 million.