Anthropic and OpenAI Compute Bids Threaten Global Debt Crisis

Anthropic与OpenAI竞购算力恐引发全球债务危机 顶尖AI实验室对算力的激进变现可能推高全球利率，并引发非AI国家的主权债务违约风险。 讨论概述了OpenAI与Anthropic如何凭借更出色的变现能力迅速整合全球算力，使它们有望掌控绝大部分可用FLOPs以及全球大部分自动化劳动力。对话详细剖析了半导体制造中极端的供应链瓶颈，以及算力分配正从公开推理转向内部研发与递归式自我改进。在宏观经济影响方面，兄弟二人分析了预计超11万亿美元的AI资本支出与巨额债务发行将如何推高全球利率，从而引发非AI经济体的主权债务违约，并颠覆传统的资本主义模式。 到2028年底，OpenAI和Anthropic有望掌控全球绝大部分可用算力，这很大程度上是因为其更强的变现能力使它们能够在争夺电力兆瓦配额时出价压过所有竞争对手。 (0:00) 相对较小的晶圆厂资本支出投资（每年1 GW产能对应60亿美元）能在五年生命周期内催生超1万亿美元的下游AI收入，尽管EUV光学元件等供应链瓶颈限制了扩张。 (7:01) 大宗通用品算力的变现水平为每兆瓦1000万至1500万美元，但顶尖实验室将通过部署高利润的内部应用与合成劳动力，激进地将算力价值推高至每兆瓦2500万至5000万美元以上。 (13:08) 顶尖AI实验室正越来越多地将算力从对外推理API销售重新调配至内部研发与递归式自我改进（RSI），以最大化长期回报。 (29:43) 中国在全球新增算力中的份额已骤降至10%以下，顶尖中国实验室的电力容量受限于100–200 MW，而美国实验室则正达到5 GW以上。 (33:27)

Dylan and Dwarkesh Patel argue that skyrocketing AI capital expenditures and unit-economic breakthroughs are consolidating global compute and synthetic labor into just two frontier labs, OpenAI and Anthropic. This concentration threatens to crowd out non-AI capital, push interest rates higher, and leave the broader economy vulnerable to unprecedented centralization. The Unit Economics and Consolidation of Frontier Compute (0:00) Frontier lab economics have undergone a structural flip from venture-subsidized burn to self-sustaining cash generation. Previously, serving models like GPT-4 against infrastructure costs of $10 million to $15 million per megawatt yielded negative gross margins. Today, monetization yields for frontier models reach up to $50 million per megawatt for Anthropic, with expectations of pushing toward $70 million to $100 million. This enables top labs to outbid standard enterprise buyers as global AI CapEx expands from roughly $1 trillion to over $2 trillion by 2028. Because base compute can be monetized profitably even on commodity open-weight models, Dylan projects frontier labs will drive megawatt lease prices up to $25 million to $50 million. Uncommitted infrastructure owners like SpaceX have already transacted high-end B300 compute at valuations equivalent to $40 billion per gigawatt. OpenAI and Anthropic—scaling from roughly 2 gigawatts each toward a combined target of 100 gigawatts by late 2028—are projected to control 40% to 50% of all new incremental global compute next year, eventually commanding the vast majority of worldwide FLOPs. Physical Supply Bottlenecks and Geopolitical Divergence (7:01) Capital cannot instantly overcome physical manufacturing constraints due to heavy supply-chain lag. While Dylan's model indicates that $6 billion in fab CapEx produces 1 gigawatt of annual capacity that yields over $1 trillion in downstream revenue over five years, single points of failure throttle output. For example, Carl Zeiss optics for ASML lithography machines physically cap global tool delivery near ~100 EUV systems annually toward 2030, keeping cumulative industry CapEx constrained above $2 trillion across energy, shells, and chips. Geopolitically, US export controls have widened the quality-adjusted hardware gulf. Pre-2022, China secured 30% to 35% of incremental global compute; today, it captures under 10%, while the US receives roughly 70%. Even if Chinese domestic fabs like SMIC and CXMT add 5 to 10 gigawatts annually by 2028, Dylan notes that 50 gigawatts of domestic Chinese silicon would offer the effective performance of only 20 gigawatts of leading American hardware, allowing single US labs to outpace all of China in effective compute. CapEx Expansion, Debt Issuance, and Macroeconomic Stress Financing cumulative AI infrastructure through 2029 requires an estimated $11 trillion across the tech ecosystem. Dylan models that while $6 trillion will be funded from corporate operating cash flows, $5 trillion must be raised through debt. With hyperscalers like Meta willing to absorb 8% borrowing rates to capture massive compute yields, non-AI borrowers will be crowded out by a projected 250-basis-point increase in capital costs across the broader financial system. Dwarkesh warns that this dynamic mirrors the 1980s Volcker shock, threatening non-AI corporate balance sheets and emerging economies with sovereign default. In the US, rising rates compound fiscal strain: a 1-percentage-point increase in average rates increases debt-service costs from 20% to 25% of tax revenues, while a 5-percentage-point increase pushes it past 40%. Concurrently, higher discount rates threaten severe equity drawdowns for traditional dividend-paying companies as market liquidity rushes into AI infrastructure. Prioritizing Internal R&D and Synthetic Labor Concentration (29:43) Contrary to market expectations that compute will skew toward public inference, Dylan and Dwarkesh observe labs starving external APIs to channel megawatts into internal recursive self-improvement (RSI). Frontier labs earn vastly higher long-term discounted cash flows by using compute for internal automated research than by selling tokens to external high-earners like quantitative trading firms. Recent data shows monthly ARR additions plateauing relative to total megawatt expansion as labs redirect fleets toward internal R&D. Compounding a 4x to 5x annual compute increase with a 3x annual algorithmic efficiency gain yields a roughly 10x annual expansion in effective synthetic labor. An automated workforce scaling from 10 million remote-worker equivalents could reach 1 billion within years, putting a labor force rivaling the global human population under the control of two corporate entities. This self-reinforcing flywheel overturns traditional market principles, where centralized control was historically less efficient than decentralized capitalism. The conversation concludes that only severe real-world friction—such as mandatory government safety delays, data center moratoriums, grid connection denials, or an unexpected technical plateau—can prevent extreme centralization. Absent those external barriers or a viable framework for decentralized post-AGI governance, economic incentives strongly favor a future where a duopoly controls the majority of the world's automated workforce and productive compute.

Dylan Patel 与 Dwarkesh Patel 指出，AI 算力与资本支出的指数级扩张正迅速将全球算力与有效劳动力集中到极少数前沿实验室手中，这将对传统资本市场、主权债务及地缘格局产生深刻冲击。 算力飞轮与前沿实验室垄断 (0:00) Dylan 指出，全球 AI 资本支出（CapEx）预计从今年的略高于 1 万亿美元增长到 2028 年的超过 2 万亿美元。OpenAI 与 Anthropic 正在吞噬大部分增量算力，两家机构的算力规模预计从年初各自不足 2 吉瓦（GW）飙升至年末的超 5 吉瓦，并在 2028 年底合计瞄准 100 吉瓦。前沿模型的单位经济学已发生根本逆转：过去使用英伟达 Hopper 芯片运行 GPT-4 会出现毛利亏损，如今运行 GPT-5.6 或 Opus 5 等模型使 Anthropic 的单兆瓦（MW）年营收高达 5000 万美元，远高于 1000 万至 1500 万美元的基础基础设施成本。高额利润被重新投入到模型训练中，形成了强大的自我强化飞轮。 供应链硬约束与竞价挤压 (7:01) 尽管下游具备极高回报潜力，硬件层供应链却存在严重瓶颈。晶圆厂每建设 1 吉瓦年产能需要约 30 亿至 40 亿美元的晶圆制造设备（WFE），加上厂房与洁净室后总成本达 60 亿美元，但其在五年内能产生超过 1 万亿美元的终端营收。然而，由于蔡司为 ASML 生产极紫外光刻（EUV）光学反射镜等关键单一故障点存在物理扩产周期，2030 年前全球 EUV 年产量上限约为 100 台。算力供给的物理上限迫使前沿实验室以更高价格抢夺资源，Dylan 预计每兆瓦租赁价格将从 1500 万美元被抬升至 2500 万甚至 5000 万美元，从而将中小企业与通用开源模型运营者逐步挤出市场。 算力转向内研与万亿债务海啸 (29:43) Dylan 提出一项反共识预测：前沿实验室未来不会把大部分算力用于对外提供商业推理，而是会优先保障内部研发（R&D）与递归自我改进（RSI），以加速实现 AGI。截至 2029 年，全球 AI 资本支出累计预计达到 11 万亿美元，除企业自由现金流提供约 6 万亿美元外，仍需通过债务市场筹集 5 万亿美元。Meta 等超大规模云厂商的借贷成本可能因此上升 250 个基点至 8%，推高全社会基准无风险利率。高利率将重挫传统非 AI 价值股的估值，并重创资产负债期限错配的传统银行业。 主权债务压力与地缘格局分化 (48:48) Dwarkesh 警告，全球资本被高回报的 AI 基础设施虹吸将产生挤出效应，这可能重演 1980 年代沃尔克式的利率冲击，导致巴基斯坦、尼日利亚等重债非 AI 发展中国家陷入主权债务违约。对美国本土而言，由于国债每隔数年需要展期，利率上升 5 个百分点可能导致联邦债务利息支出占税收比例超过 40%，且自动化侵蚀传统个人所得税与薪资税基础后，财政压力将更严峻。地缘层面，美国 2022 年以来的出口管制已使中国占全球新增 AI 算力的份额降至 10% 以下；尽管中国预计到 2028 年累计拥有 30 吉瓦算力并在 2029 年通过本土设备年增 50 吉瓦，但其实际有效算力仍可能显著落后于美国前沿硬件。 万亿规模数字劳动力与去中心化困境 (67:52) 随着前沿算力每年扩容 4 至 5 倍，且算法效率每年将单位能力所需算力降低 3 倍，前沿 AI 的等效劳动力人口正以每年约 10 倍的速度呈指数级增长。Dwarkesh 指出，这使得少数实验室控制的合成劳动力规模有望在本世纪末超越全人类总人口。尽管传统资本主义依靠去中心化竞争实现高效率，但 AI 的经济学特征极度偏向资源集中与规模效应。由于前沿机构内部研发回报率远高于外部商业应用利润，实验室最终会切断外部 API 供应。然而，监管干预、安全评估审查以及地方电网审批等物理世界阻力，或将成为减缓此种极速奇点接管的主要摩擦力。 对话总结认为，AI 技术的物理与经济特性正在终结传统去中心化的市场结构，但在算法爆炸的同时，现实世界的电网限制、供应链瓶颈及政府对高阶模型安全评估的强制干预，将构成决定 AI 接管速度的关键不确定变量。

[48:48] Dylan Patel: Why are you fucking laughing? [48:48] Dwarkesh Patel: Because it’s things you’ve learned in the last month... A million people listen to this guy who just learned about debt this month. [48:48] Dylan Patel: This is our business, Dylan!... Now I’m self-conscious. Fuck. [36:27] Dwarkesh Patel: I didn’t realize the compute situation was as fucked as you’re saying. [59:15] Dwarkesh Patel: The opportunity cost of the government borrowing money to pay people pensions is extremely high now. Because that money could be spent building a robot factory that builds a robot factory that builds a robot factory. [75:52] Dwarkesh Patel: On that somber note, I guess we’ll meet again when the RSI is officially kicked off. [75:52] Dylan Patel: You’re not going to have me on your podcast again for like two months? Alright, cool. Thanks, dude.

[0:00] Okay, I’m back with Dylan  Patel, founder of SemiAnalysis. [0:02] Our version of a family Thanksgiving  dinner is a regular yearly podcast. [0:07] But we're not actually related. Don’t tell the people this. [0:10] It will destroy the myth. Basically where the world [0:15] economy is headed is more and more becoming  a function of where lab economics are headed, [0:20] where the compute market is headed, et cetera. I want to understand where the crazy [0:24] future ends up within a few years. But let’s start with where we are today. [0:30] Walk me through lab compute and lab revenue  right now, and maybe project out a year or two. [0:37] When we go back to last year, even at the  end of the year, most of GDP growth in [0:42] America was just AI infrastructure. As we look towards this year, [0:46] about a third of the compute coming online  is for the labs, for OpenAI and Anthropic. [0:52] It may be built by others and then rented  to them, but at the end customer, it’s them. [0:58] As we go forward into the future, the  numbers for compute are ballooning. [1:03] We’re at a little bit over a  trillion dollars of CapEx this year. [1:07] As we go out into ’28, it’s going  to be more than $2 trillion. [1:12] The labs are also taking an  increasing percentage of this. [1:15] So ultimately, you’ve got a very interesting  situation where the labs are going from [1:20] companies that spend tens of billions of dollars  a year to hundreds of billions of dollars a year, [1:25] to forecasting to spend trillions of dollars  a year even towards the end of the decade. [1:30] This is at least some of the contracts  they’ve begun signing with their partners. [1:34] This requires a big reshaping of  what happens with their economics. [1:41] Up until now, they have been  companies that mostly lost money. [1:45] Anthropic started turning a profit in Q2. It’s believed at some point in Q3, [1:50] OpenAI could start turning a profit even, with  the bigger rise of Codex and 5.6 and all this. [1:57] But if we go back a year ago, all the  money they had was venture-funded losses. [2:03] If we go back to even the beginning of  this year, it was venture-funded losses. [2:08] They’ve now turned the corner and  are actually starting to profit. [2:11] That doesn’t mean they’re  not taking in new capital. [2:13] The new capital is still coming in  to accelerate the growth further. [2:16] But ultimately, more and more of their business  is being funded off of their own revenue rather [2:21] than capital injections into them. Over the last year and a half, [2:26] their margins have really skyrocketed. The base cost of compute tends to be [2:32] around $10 or $13 or $15 million per megawatt. The most interesting aspect about what’s happening [2:38] now is this: Before, if they served a model —  GPT-4 being served on Nvidia Hopper GPUs — it was [2:47] generating negative gross margin for OpenAI. But now, when OpenAI serves GPT-5.6 [2:55] or Anthropic serves Opus 5 or Fable 5, their  revenue generation has passed well beyond the [3:03] incremental $10-15 million per megawatt. In the case of Anthropic, the revenue has [3:08] gone as high as $50 million per megawatt. What that now enables them to do is: "Hey, [3:15] if I spend 10 bucks on inference capacity,  I actually generate 50 bucks of revenue, [3:19] and then I can turn around and incrementally  spend all of that profit on training." [3:24] One thing I’m very interested in understanding  is how you see the centralization of compute [3:28] happening at the labs, or the relative ratio of  compute that goes to the world versus the labs. [3:34] If you say right now a third of  marginal compute is going to the labs, [3:40] by when is over half of the incremental  compute in the world going to the labs? [3:45] By what point do the labs have basically  a vast majority of the world’s compute? [3:50] At the beginning of this year, OpenAI started  at 2 gigawatts and Anthropic at less than 2. [3:58] End of this year, they’re both above 5. So they’ve 3-4x’d compute as a whole. [4:06] When you look at the incremental compute added,  that’s about 30% of the compute added this year. [4:10] As we step forward to next year, given what’s  already been signed and penned and inked, [4:15] you’ve got something even more dramatic. Anthropic and OpenAI are taking as much as [4:20] 40% to 50% of compute next year. This centralization doesn’t [4:26] look like it’s slowing down or stopping. In fact, it looks like it’s only accelerating. [4:31] Who’s building that compute for them will change. Next year, a big new entrant is, for example, [4:36] SpaceX, which is building a ton of compute. They’re actively going to lease quite a bit [4:41] of it to Anthropic and OpenAI, most likely,  because they’re the ones who have the marginal [4:45] capability to pay the highest price. In addition, OpenAI and Anthropic [4:48] are also starting to build their own  compute — OpenAI with their own chips, [4:51] Anthropic with TPUs that they’re purchasing  from Google and deploying with Fluidstack. [4:56] So you ask, "Hey, when does half of  the world’s incremental new compute [5:00] go to just OpenAI and Anthropic?" It’s really by the end of next year [5:06] when half of the incremental compute is  already going to Anthropic and OpenAI. [5:09] Because compute is growing so fast, incremental  compute is going to be basically most of compute. [5:16] So it’s very soon — you’re saying maybe within  a year and a half or two years — that most of [5:19] the world’s compute is owned by two labs, or  at least is serving the demand from two labs. [5:26] There’s this trend where maybe world compute in  gigawatts doubles every year, but the compute at [5:30] the frontier labs triples every single year. If you keep the current trend going, [5:34] it goes from 2 at the beginning of this  year to close to 6 at the end of this year. [5:38] Just multiplying out by 3. It’s 18 by the end of 2027, 54 by the end of 2028. [5:44] Are you like, "Okay, at that point, they  simply can’t continue tripling given the [5:47] amount of world compute"? How do you see the world [5:50] compute situation over the next few years? If the incremental compute this year adds 30 [5:54] gigawatts, next year 50 gigawatts, and  the year after that roughly 70, you end [5:58] up with this really interesting phenomenon. A new watt deployed this year is significantly [6:04] more efficient than the  watts deployed two years ago. [6:08] A humongous percentage of the world’s  compute was deployed this year. [6:11] Even though it didn’t double the number  of watts deployed, I’m deploying GB300s [6:16] and TPUv7s and Trainium3s, which  are way, way, way more efficient. [6:20] They’re 3-5x more performance per  watt than the prior-generation chips. [6:26] So ultimately you’ve got a huge ladder here. If Anthropic and OpenAI take on 45% of compute [6:33] next year, you’ve got them in, let’s say,  December ’27 having taken on half of the [6:38] world’s incremental new compute. But that half of the world’s new [6:41] incremental compute is actually at a higher  performance than everything else before it. [6:45] So you’ve got another multiplier on that. By the time you’re towards the end of [6:48] 2028 — if this trend continues, and I  see nothing that’s stopping it — you’ve [6:55] got them just controlling most of the  usable flops in the world on their own. [7:01] The thing I’m confused about is why you think we  only add 80 gigawatts in 2028 if we enter a world [7:08] in which the value of compute increases so much. That’s the upper bound, by the way. [7:12] That’s the like, "I’m so fucking bullish." Okay, let’s do some chain of thought here. [7:18] When I interviewed you a few months ago, you  said that in order to make a gigawatt of, [7:22] I think, Vera Rubins, you need 55,000 N3  wafers, 6K N5 wafers, and 170K DRAM wafers. [7:33] I know if those numbers might have changed. I’m going to troll you, but the way you said [7:35] wafers was so fucking Indian. Vafers. [7:39] By the way, when we first moved to the US, I had  the v/w thing pretty bad, and I was a vegetarian. [7:46] I remember you told me about this. In North Dakota, I was in elementary [7:50] school, and I’d be like— Can I get a "wedgie"? [7:52] Can I get some "wedgies"? Anyways, so that’s for one gigawatt. [8:01] I had an LLM run your wafer fab  equipment model and figure out how much [8:09] the tooling costs to produce a gigawatt  of compute basically every single year. [8:13] It said $3-4 billion. Now suppose  you add in cleanrooms and shell and [8:19] everything else at the fab. So $6 billion of fab CapEx [8:23] produces a gigawatt every single year. A gigawatt produces right [8:28] now $100 billion of revenue. But also that $6 billion in CapEx is producing [8:34] a gigawatt every single year, and that gigawatt  is producing $100 billion every single year. [8:40] So over the course of five years, the first  gigawatt has generated five years of profits, the [8:48] second gigawatt the fab has produced has generated  four years of profits, and so on. $6 billion of [8:54] CapEx at the fab level will have generated  over a trillion dollars of end AI revenue. [9:01] Yeah. There’s a lot of OpEx along the way. There’s a lot of other CapEx, [9:05] like the data center, the power. And you had to pay OpenAI for the R&D. [9:08] Installation. There’s a lot of  different people who need money here. [9:12] Take away half of it for all these middlemen. That still means there’s a 100x discrepancy [9:19] between fab CapEx and end revenue generated. More than that, actually, but [9:25] we’re just being very conservative. As a result… This is capitalism. [9:33] You have this huge discrepancy  where you can turn $1 into $100. [9:39] They’re not going to figure  out a way to make more mirrors? [9:41] They are. It’s just that these  mirrors take some time to make. [9:44] But the emergency is so big where Anthropic  and OpenAI are like, "We could make a trillion [9:48] dollars right now, but we’re just bottlenecked  on the mirrors that go into the ASML machines." [9:54] How can we make more mirrors if  we spend $100 billion on this? [9:56] That’s the situation we’re  going to be in pretty soon. [9:59] We’re not going to be able to  solve that supply constraint? [10:01] That just seems quite hard to imagine. You’ve seen people do funny arbitrages here where [10:09] they buy turbines and then try and resell them,  because the value of a turbine is way more since [10:14] it’s the thing bottlenecking your data center. I think if anyone had [10:19] $400 million and the ability to  convince ASML to sell them an EUV tool, [10:26] they should totally just go buy one, wait, and  then sell it for north of a billion dollars. [10:29] But ultimately, yes, capitalism  will cause these things to expand. [10:33] But it’s a whip. It takes a long time for the  whip signal to get to the tail end of that. [10:40] The supply chain doesn’t react immediately. In fact, you go talk to someone at Carl Zeiss, [10:46] they’re like, "Yeah, yeah, yeah, we need to  make 100 EUV tools by the end of the decade." [10:50] When we had our episode earlier this year, they  didn’t even think they needed to make that many, [10:55] enough mirrors to make 100 EUV tools a year. Now they’re like, "Okay, we need to do that." [11:02] But in reality, because of all  the economics of what’s going on, [11:06] it should be even more. It takes so long to pill. [11:11] Suppose that every single company  in the stack got private equitied. [11:16] Somebody came in who was super AGI-pilled and  was like, "We’re going to maximize production." [11:20] What do you think the physical constraints  on making more things would be? [11:24] The reason I ask is we’re pretty soon going to  be in a world where the lab revenue, or just AI [11:30] cash flows — because obviously the accelerators  also have these huge cash flows — will be so [11:33] big that you can just fund extreme expansion of  all this production from cash flows themselves. [11:41] I do agree generally. There’s  obviously some physical constraints. [11:45] The way the supply chain is expanding currently,  100 is roughly still the right number. [11:50] For 2030? 100 ASML tools for 2030. [11:54] But if you said, "Carl Zeiss, here’s $10 billion. Please fucking just expand production," that [12:00] would change things. You would have to do [12:02] this with every company in the supply chain. But you don't think that's gonna happen next year? [12:03] I don’t think it’ll happen this year. I don’t think it’ll happen next year. [12:06] I don’t think it’ll happen the year after,  because the world is capital constrained. [12:09] But in a world where, say, the top  labs are generating, even combined, [12:14] a trillion dollars in revenue next year,  they’re not able to take $10B of that— [12:17] I don’t think they’re going to do that, but… Or hundreds of billions at least? [12:20] It just seems like they realize  where the world is headed. [12:23] I feel like they could just make… The thing is, the labs are going to generate [12:29] hundreds of billions of revenue next year. But ultimately, CapEx next [12:33] year is like $2 trillion. So you’ve got this big mismatch. [12:38] The wafer fabrication equipment supply chain  will do something on the order of $200 billion. [12:44] The data center market supply  chain will do even more. [12:46] The accelerator supply chain will do even more. The energy supply chain will do a number. [12:51] You sum all this up, it’s going to be  well north of $2 trillion of CapEx. [12:55] So the labs have not yet gotten to the point  where their cash flows can fund this stuff. [13:00] Obviously they will never get to  that point, because you want to [13:03] keep your CapEx higher than your returns. Yeah, you reinvest. [13:11] The key question I really want to understand is:  if the current trend continues, it’d be north of [13:13] 50 gigawatts per lab by the end of 2028. So between them they’d have 100 gigawatts. [13:18] Those gigawatts, as you’re saying, drive many-fold  more throughput or performance by 2028 than they [13:25] do now, because the hardware’s gotten better. Not only have flops per watt increased, [13:28] but also the hardware gets better  at working with AI workloads. [13:34] Okay, so 100 gigawatts for  the labs by the end of 2028. [13:38] How much is world compute? I think that may be a little [13:41] difficult, given that by 2028 they’ve  taken 70-80% of incremental compute. [13:47] And I’m not sure what happens to markets then. How much does the price of compute skyrocket for [13:53] them to actually be able to buy 70-80% of compute? Is Google or Meta or Amazon willing [14:00] to sell even that much? Also, there’s one caveat [14:04] when we’re talking about these gigawatt numbers. When Amazon is serving Bedrock Anthropic models, [14:09] that counts as Anthropic compute in our worldview,  because it is effectively, at the end of the day, [14:14] counted as revenue for Anthropic even though  there’s a revenue share and credit back all that. [14:19] But ultimately in 2028, if they get  to 100 gigawatts combined, they have [14:24] done really disruptive things to the market. Because anyone can make money off of $10-15 [14:31] million per megawatt compute today. I kid you not, it’s not that hard. [14:37] Go get a GB300 rack, go download the Kimi  weights, go download vLLM or SGLang, set it up. [14:44] Codex and Fable can actually help you do this. It’s pretty simple. It’s not trivial, [14:50] but it’s not rocket science. Go put it on OpenRouter. [14:53] It’s very simple. You’ll start generating more  revenue than you’re paying for the compute. [15:00] This has already led to  this compute pricing, $10-15 [15:03] million per megawatt, starting to inflect up. To get to that 100 gigawatts in 2028, you have [15:10] to believe that the labs can outpay for compute,  because anyone can make money at $10 to $15. [15:17] Does compute now get to $25 million a megawatt? Does it get to $40 million a megawatt? [15:21] As you’re saying, it’s already the  case that the labs are generating way [15:24] more revenue per megawatt than everybody else. If they stay as far ahead as they are currently, [15:29] you would expect that to continue being the case. If there’s some kind of recursive self-improvement [15:32] where the AI labs are relatively uplifted  — or they have models internally they’re [15:36] not releasing externally that are helping  them make their next model better — you’d [15:39] expect that to be even more the case. Aren’t you already seeing this, where SpaceX, [15:42] or whoever is slightly further behind, will just  sell compute to the highest bidder if they can’t [15:46] internally monetize it as well as the labs? You’d expect them to keep bidding for larger [15:52] and larger shares of the compute market. I think that is my worldview. [15:56] They will continue to gobble  up more of the compute. [15:58] But ultimately they can’t do it at  current pricing or anywhere close to it. [16:03] They do have to start paying $25, $30, $50 million  a megawatt to really gobble up 70% of the world’s [16:09] compute in 2028, to get to 100 gigawatts  by 2028, which is a very aggressive goal. [16:15] The other aspect of this that’s  really challenging is that we’ve [16:17] already seen a huge slowdown for the AI labs. This regulation that they advocate for is actually [16:24] slowing down the labs a lot more than it slows  down the open-source Chinese language models. [16:30] OpenAI not releasing Astra. OpenAI  stopping training for two weeks. [16:35] Anthropic not releasing what their safety  assessment says is Model 2, which is widely [16:39] believed to be the next version of Mythos. They’re clearly not releasing their best models, [16:44] in which case their revenue per megawatt  stalls or can even start to decline again [16:49] because other models are competitive again. It’s not that they’re falling behind. [16:53] It’s just that they’re not  releasing their best stuff. [16:55] What if there is some regulatory impact that  prevents them from releasing their best models? [16:59] Now their revenue per megawatt  does not climb as fast. [17:02] Their ability to buy that incremental  compute for a higher price than everyone [17:06] else starts to diminish, and then maybe  they can’t get to that 100 gigawatts. [17:10] But in a world where safety doesn’t matter,  I do believe that’s exactly what happens. [17:15] They can start generating $100  million per megawatt or more, [17:18] and they can pay $50 million a megawatt. No one else has any logical reason to [17:22] do anything with their compute besides say,  "Please, Dario, take everything off of my hands." [17:28] But there are forces at play, which we cannot  describe, that would potentially slow this down. [17:38] I think a good intuition pump is: what if  the AI models were literally as good as a [17:44] fully automated software engineer? They’re not currently there yet. [17:47] I think they’re far from being able to fully  automate the job of a full white-collar worker. [17:52] But white-collar workers earn six  figures or north of that a year. [17:57] If you have a gigawatt that  can sustain a population of, [18:01] say, roughly a million white-collar workers. [18:08] Then off the back of that…That  would be $100 billion. [18:12] That’s actually surprisingly low. Yeah, $100K per person, million population. [18:18] I don’t know. But it would be  many hundreds of billions of [18:20] dollars per gigawatt if you get full AGI. The other aspect of this — and we’ve [18:25] continued to see this — is that most  of the value capture is not happening. [18:29] Most of the value that these models generate  does not get given to OpenAI and Anthropic. [18:34] Thankfully, so far it is mostly  just being given to the users. [18:38] Jane Street, with their exclusive contract  with OpenAI for GPT-5.6 Ultrafast mode, [18:43] or Jane Street where they’re one of Anthropic’s  biggest customers, is generating way, way, way [18:48] more value out of the tokens they’re paying for  than Anthropic is generating in terms of profit, [18:55] because they get to make money off of the market. Or take Meta, who at one point was rumored to be [19:03] as much as 10% of Anthropic’s business. They’re generating way more efficiencies [19:08] by optimizing their ad algorithms  or what have you, getting engagement [19:13] time 5% longer, all these things. They’re making way more money off [19:17] of using these models than Anthropic is. That’s what’s required. Sure, if you had [19:25] a million new software engineers, the cost  for a software engineer would also fall. [19:29] One thing I’m confused about is, does  the market come into equilibrium? [19:32] If it comes into equilibrium, would you  just expect the price of compute to equal [19:39] whatever Anthropic and OpenAI can generate  from it, or be very close to it with a small [19:44] amount of markup for Anthropic and OpenAI? Right now it’s really weird that there is a [19:48] 4x or more difference between what compute sells  for and how much money Anthropic can make from it. [19:54] In a world where the revenue per gigawatt  continues to increase, if Anthropic’s ability [19:59] to monetize a gigawatt doubles or triples,  it’d be weird if the gap continued to increase. [20:05] Anthropic, just by having some weights,  can take something that cost them [20:09] $10 and turn it into $100. This is always a fun question. [20:16] Where does the value go in AI? AI’s generating all this value. [20:18] You’ve got the end user, which I think we all  agree is generating more value than anyone else, [20:24] hence they’re paying a lot for these models. Then you have the app layer. [20:27] So far the app layer’s  generated very little value. [20:30] Then you’ve got the model layer, which up until a  year ago was generating negative gross margins and [20:36] is now generating massive positive gross margins. It looks like it’s on the path to [20:40] generating $100 million per megawatt. So turning $10-15 into $100, as you said. [20:48] But if we go back a year ago, the hardware supply  chain was generating all this gross margin while [20:53] literally everyone else was losing money on it. OpenAI and Anthropic were just plowing VC [20:58] money in, as were many other startups. Many of these hyperscalers were building [21:03] infrastructure without knowing if  there was going to be a payoff. [21:06] So ultimately you had this negative value  being created on the model layer, if you will, [21:12] because they were selling the tokens for  less than it cost them on the infra side. [21:15] All the value was being  captured at the chip, the fab. [21:19] Initially in 2023, the memory guys were  making no money off of HBM or memory for AI, [21:26] even though theoretically the value  they were delivering was humongous. [21:29] Now you’ve got… Well, actually TSMC captures  way less value than the memory guys. [21:36] So the value capture’s shifted  around a lot, which is very fun for [21:41] people tracking the market or participating  in the market, like Jane Street as an example. [21:46] This is not an ad. This is not an ad. [21:46] This is not an ad. They’re a sponsor but [21:47] you don’t have to plug them that hard. So what happens going forward? [21:56] Anthropic and OpenAI have slowly  started to balloon in value capture. [22:00] Do they balloon and take all the value capture? Well, that was a thought, [22:03] and then Elon showed, "Actually, no. I can sell my compute for $25 million a megawatt [22:09] or $40 million a megawatt to Anthropic and Google. Even if it’s a short-term thing, I’ve sold it for [22:15] this price, and I’ll recoup  my entire CapEx in a year." [22:19] What’s your prediction of how much the  relevant tranche of compute — B300s or whatever [22:24] that SpaceX sold for $40B a gigawatt to Google —  what does that sell for at the end of next year? [22:31] I think most compute will still continue  to transact at sub-$20 billion a gigawatt. [22:35] Even at the end of next year? Because all of it has to be financed. [22:42] If Meta, Microsoft, Amazon, SpaceX can  build compute without finding a customer, [22:47] just saying, "Fuck it, I’m going to build this  compute," and then turn around and wait till it’s [22:51] already built, they now control what’s going on. Most compute is contracted well before it’s built. [22:59] This is what Elon took advantage of in the market. He actually had all this compute. [23:02] He was like, "Hey, Anthropic, I know you’re  making $60-plus billion per gigawatt. [23:07] Why don’t you just buy my stuff  for a crazy amount of money?" [23:10] Obviously it’s not like Elon decided  this or Anthropic decided this. [23:13] The market figured itself out. Other people, you go to a random cloud, [23:20] they’re like, "Okay, I’m going to build a  gigawatt of compute or 100 megawatts of compute. [23:24] I’m going to spend the CapEx. I need to turn around and find a customer. [23:28] If I want to find a customer,  I need to find the capital. [23:31] Who’s going to give me the  capital and the customer? [23:32] The customer has to sign a deal. Then I take the customer’s commitment to [23:35] the credit markets and I raise the capital." So there’s this completely different [23:39] power structure where Meta is  effectively hoarding compute. [23:44] Them and SpaceX are the only plausible #3,  because they’re hoarding all this compute. [23:50] They’re using their balance sheets and  capabilities to build compute without an end [23:54] customer that’s monetizing at a huge degree. They have an actual balance sheet, [23:58] so they can go to the credit market. You build a gigawatt, you can make your margin, [24:05] not a crazy margin, but a good margin. Now I have all this compute. [24:08] Now Meta and SpaceX have this optionality  of looking around and being like, "Is my [24:12] internal use case going to make me more  money, or should I go out there and sell [24:15] it to Anthropic or OpenAI at crazy margins?" So now we’ve entered a regime where SpaceX and [24:22] Meta are saying, "Actually, I’m going to build  the compute, and I can rent it out for not $13. [24:28] I can sell it for $25, $50, and more." So as I've been doing video essays and [24:33] other formats, I've been looking to  hire a new editor for the podcast. [24:36] But actively searching for editors  has been quite time-consuming because [24:39] the vast majority of candidates don't  fit the profile that I'm looking for. [24:43] So I created a recruiter in Grok  Bot to see if it would help. [24:46] I gave it a huge context dump where I  monologued basically everything that I wanted, [24:50] and then it spun up four other bots to  narrow in on different parts of the search. [24:55] One went through the last year  of my email for relevant inbound. [24:58] One searched my X feed and DMs. One went through the end credits [25:01] on various documentaries I like. And the last one looked for editors [25:04] who work for some of the YouTubers that I follow. Grok Bot then took all these different candidates [25:08] that the subagents had found,  filtered them against my criteria, [25:11] and delivered for a final shortlist to review. To be honest, the first batch had a few good [25:16] candidates I wanted to see,  but mainly a bunch of duds. [25:19] But after I gave Grok Bot some more  feedback about what it was missing, [25:22] it came back with a new list of candidates  that I'm actually extremely excited about. [25:25] I ended up saving this  whole workflow as a routine. [25:28] So every week now Grok Bot checks my  inbound email and X DMs for promising [25:33] new candidates to potentially interview. If you want to try Grok Bot yourself, [25:36] go to x.ai/bot. What do you think their [25:40] revenue per gigawatt is by the end of 2027? For Anthropic or OpenAI, by the end of ’27. [25:45] I think it’s highly dependent on  who has the best model, if they’re [25:47] allowed to keep releasing their best models. But I don’t see why it wouldn’t be $50-plus [25:51] million a megawatt. By the end of ’27. [25:53] Oh, by the end of '27? That’s where it gets more challenging, [25:56] but I think it could get higher than that, to like  $70, $80 million a megawatt, blended across the [26:01] company, if not higher. Seems low. [26:04] So if that’s the case, then what  happens to the price of compute? [26:09] Well, if I’m Anthropic,  incremental compute is worth it. [26:12] Maybe I spend $40 million a  megawatt on SpaceX compute. [26:16] If I’m SpaceX, I look to the supply chain and I’m  like, "Well, I’ve struck this deal with Jensen [26:22] (where he’s now all of a sudden using Twitter)." And Elon’s saying they’re exclusive to Nvidia, [26:28] but why doesn’t Jensen raise his prices? Then SK Hynix and Micron and Samsung look [26:34] at it and they’re like, "Well,  why don’t we raise our prices?" [26:36] So with the value capture, I think  there’s a bullwhip effect here. [26:40] Just because someone has  raised prices doesn’t mean [26:43] the entire supply chain rebalances immediately. But over time, the supply chain will rebalance [26:48] and things will cost more and more. To get that incremental capacity, [26:52] you sort of have to. So TSMC raising prices very slowly, [26:55] but memory companies raising prices very quickly. Substrate companies raising prices very quickly. [27:02] Elon wouldn’t have sold if it was $15,  but he’s selling because it’s $25+. [27:06] So obviously he raised his prices really quickly. I’m surprised you think that revenue per gigawatt [27:11] doesn’t increase way more than even 100  per gigawatt by the end of next year. [27:15] When does RSI happen? When does takeoff happen? Or even if RSI doesn’t happen, just say the [27:19] current rate of progress continues. Just look at how much progress we’ve [27:22] made in, let’s say, the last year and a half. What was the model from a year and a half ago? [27:26] Claude 3.5 or something? My problem with this is that the best model [27:28] that exists in the world was trained in February. So you’re saying maybe we just won’t be allowed [27:32] to release the labs’ best models. OpenAI says they’re not training models [27:34] for two weeks, man. What the hell? [27:37] There’s one thing where internally,  are they getting enough use for it [27:39] that they’ll bid up the price of compute? Another is, does AI progress as a whole slow [27:42] down because of regulation? Yeah, but they’re not even [27:44] allowed to use this new model internally. Astra’s not even widely deployed internally. [27:49] But still, if you have a model that is… What was  the model released at the beginning of last year? [27:57] GPT… 4o? Was that 4o? [28:00] Yeah. You’re talking about a GPT-4o  to Mythos 2-size leap by this point, [28:06] again, by the end of 2027. Yeah, but Mythos 2’s not out. [28:09] Or even Mythos. That leap again. Even Mythos is not allowed to be out. [28:14] They’ve neutered it. We can’t use it  to optimize inference performance. [28:18] We can’t use it to optimize all sorts of things. Yeah, maybe there’s some slowdown in AI progress [28:23] or the deployment of AI that means  the revenue per gigawatt can be lower. [28:26] But that’s the only way I could see it being only  $100 million per megawatt by the end of next year. [28:32] As long as the model gets better, the  value generated out of it gets better. [28:35] Obviously, who captures the value is  still up for debate, but ultimately [28:39] everyone’s going to raise their prices. Because they can, and it’s super inflationary. [28:44] Especially if the method of regulation is… Right  now, so far, it’s just "don’t release the models." [28:49] But more and more, the method of regulation  is New York’s banning data centers. [28:52] Texas is holding moratoriums. Ohio’s saying, or  at least trying to say, you have to pay everyone’s [28:58] property tax in a certain radius. These sorts of things are going [29:01] to decrease supply and increase cost. That’s going to get passed on as well. [29:04] You start to end up in a spot where progress  does slow, at least in the external sense, [29:11] even if the models internally  keep getting better and better. [29:16] In a takeoff scenario, why would Anthropic  not have their best model six months ahead [29:22] of what is externally available? Because of safety and regulation, [29:25] but also the competitive advantage? That six-month difference, if progress [29:31] accelerates, is actually a bigger differential. So that’s the thing that would cap [29:36] revenue-per-megawatt gains to much lower growth  than we’ve seen in the first half of this year. [29:41] Here’s something I’m very interested in. As these companies go public and they’re [29:44] accountable to investors, let’s say by the end  of next year they have close to 20 gigawatts. [29:51] So 10% of compute is 2 gigawatts. Let’s say they want to go from 60% of compute to [29:58] training to 70% of compute to training. And their investors are like, "Well, [30:02] if you’re going to be able to generate $100  billion per gigawatt, you’re basically saying [30:06] no to $200 billion of revenue in order  to increase your training compute." [30:11] So investors are like, "What the fuck? You’re already spending so much on training. [30:14] Why are you spending even more on training?" As a public company, what do you think would [30:19] happen if they’re just like, "No, we will keep  increasing the share of compute we spend on [30:24] training to offset the increase in revenue  that each gigawatt of compute is giving us"? [30:31] This is what I personally believe. The labs are going to allocate less and [30:35] less compute to inference over time. I think that’s very non-consensus. [30:40] The standard belief of most people is,  "Oh, most compute will go to inference." [30:45] Most of it will go to forward passes for training,  not necessarily revenue-generating inference. [30:51] Ultimately, if they’re generating $30-40 million  per megawatt today, you allocate 40% to inference. [30:58] If you now get to generating $60-70 million  per megawatt, do you still allocate 40% to [31:04] inference and generate all this profit  and then do dividends and share buybacks? [31:10] Or do you go build AGI? I think the obvious answer [31:14] from Anthropic and OpenAI, not just at the  executive level but also their board, is to go [31:20] build AGI, because it’s way more profitable. So ultimately you’re going to see them ratchet up [31:28] their percentage of compute dedicated to training— While each increment of compute is getting [31:33] more and more profit-generating if  they had dedicated it to inference. [31:37] Right. The whole point is, if I’m selling tokens…  Is OpenAI releasing Ultrafast mode for just [31:46] external, or are they doing it internally too? It turns out, no. Actually, I’m going to allocate [31:50] it to internal and external, because  the internal value I’m generating from [31:55] super-fast AI or the best AI model is  way more than what someone external is. [32:01] So ultimately, sure, I could generate  $100 million per megawatt, but if I [32:05] turn that towards AI research, what is  the incremental progress that I get? [32:09] What does that do towards my future  earnings potential, the discounted [32:12] cash flows of whatever the hell I’ve done? They’re not going through that calculation, [32:17] but ultimately it makes more sense to  dedicate more and more compute internally. [32:22] The only reason to have inference compute be so  large is so you can grow your training fleet. [32:28] I think this is an interesting economics question  that I feel we can have the models digest. [32:33] What would have to be true about a  world where they reduce the fraction [32:35] of compute spent on inference? I think they have been over [32:38] the last three months already. I think at parts of this year, [32:43] they were increasing the fraction of  compute… Let’s just take it month by month. [32:48] You would agree that every month, Anthropic  has added more compute than the prior month. [32:51] There might be some noise when they sign  a SpaceX deal or whatever, but in general, [32:54] the amount of compute is a curve up. So in January, they added less compute than [32:59] December, and yet their revenue adds skyrocketed. Then they’ve sort of plateaued. [33:06] They’re not adding $25 billion  of ARR every month now. [33:10] That means the marginal megawatt they’re  getting is going as a higher percentage [33:14] to R&D than it is to inference. So they are factually increasing [33:18] their compute towards R&D today. I think this is self-evident [33:24] if you look enough at what they’re doing. If I look at the numbers you said for how [33:29] fast world compute grows, here are  some things I want to understand. [33:32] It seems like if I add up the numbers you just  said, it would be over 200 gigawatts of world [33:36] compute by the end of 2028, right? Yeah, globally. [33:39] Okay. How fast can that continue  growing, global AI compute after 2028? [33:45] 30 this year, 50 next year, 70 in ’28.  ’29 should be on the order of 90-100. [33:53] Then just 100 more every single year or something? I think the slope can continue to go upwards. [33:58] It’s hard to predict anything  more than four years out. [34:02] Who knows whether we’re in an RSI regime, or  when is the world economy growing at 10% a year? [34:08] Because if you’re at 100+ gigawatts  a year, you’re at absurd GDP growth. [34:15] If you think there’s 200 gigawatts globally  in 2028, how much is in China by that point? [34:22] How does Chinese compute continue  increasing through this whole trend? [34:26] Because if the RSI stuff kicks off  in the West before China has a large [34:32] amount of compute, maybe we’re living in  a different world than when it doesn’t. [34:37] If we level-set back to 2022, the US was  adding about 45-50% of the world’s compute. [34:45] China was adding about 30-35%. [34:49] The rest was being taken up  by the rest of the world. [34:52] Since 2022, we’ve had big regulations against  China and a dramatic increase in America. [34:58] So today, 70% of watts are  being deployed in America. [35:02] China is really a very small number. Sub-10% of watts being deployed for data [35:09] center AI compute is in China. As we step forward, [35:13] they’re still at a very small number. Their domestic production is quite small. [35:18] Their purchasing from Nvidia is still  quite small, and a lot of that ends up in [35:25] other places as well, Malaysia or what have you. So ultimately, China domestically still continues [35:31] to have sub-10% of incremental new compute. In 2028 it might start to inflect up, I think. [35:37] But it’s pretty easy to say China will  have 30 gigawatts of AI compute or less. [35:41] By 2028? Yeah, in 2028. [35:43] Okay. And then how fast does  their hockey stick go up? [35:46] I do think in 2028, they have a big uplift  in what compute they’re able to deploy. [35:51] In 2026, they’re still mostly relying on a lot  of the smuggled chips, a lot of the chips that [35:59] TSMC made for companies that they thought  weren’t Huawei but ended up being Huawei, [36:04] or a lot of HBM that Samsung is shipping. But in ’27, fabs start to go up. [36:10] In ’28 especially, fabs start to go up from SMIC  and CXMT and such, where domestic production is [36:15] actually reaching many millions of units a year. Now they’re incrementally adding 5-10 gigawatts, [36:21] in just 2028, of domestically produced chips. Those chips are definitely worse than the [36:27] chips that Nvidia will have in ’28, or Google  will have in ’28, or OpenAI will have in 2028. [36:32] So even the gigawatt number  overstates things, you’re saying. [36:36] It’s 30 gigawatts, but it’s  really much worse chips. [36:39] But if you think the world is going to add  100 gigawatts the following year — I know [36:45] you said you can’t really say that far out — how  much is China able to add the subsequent year? [36:50] Basically, I want to know: do they just hockey  stick at the point at which they are able to [36:54] start shipping large amounts of compute, or is  it still going to be less than US plus allies? [37:01] There’s a lot left to whether or not the US passes  the MATCH Act, whether or not tools continue to [37:08] get export-controlled, how fast China can build  their new equipment that they’re starting to be [37:13] able to produce domestically. But ultimately, [37:19] China is definitely going to hockey stick. If there’s anything China’s really good at, [37:23] it’s scaling manufacturing really, really quickly. I imagine China will start to be able to extract [37:32] more and more purchasing of even  foreign chips into domestic China, [37:36] or at least close the gap in what the US is  allowing Nvidia to sell them, or what have you. [37:42] But do you think China could be adding  50 incremental gigawatts in 2029? [37:48] I think that’s completely reasonable. Part of that could also be purchased from foreign. [37:56] But yeah, I think it’s completely reasonable  that China in 2029 can do 50 gigs. [38:00] But if most of those are domestic  chips, there is some factor there [38:03] where that 50 gigawatts is really worth as  much as 20 gigawatts from American chips. [38:10] Right. So you’re actually projecting a world  where maybe the leading lab in 2028 has more [38:16] compute than all of China will have in ’29 or even  ’30, if you weighted gigawatts by their quality. [38:25] Implying that there’s nothing  done to slow down the US labs. [38:28] That’s right. But clearly the government [38:30] and politicians are starting to do that. Whereas China’s not going to slow down AI. [38:34] In fact, the only thing they’re  going to do is accelerate it. [38:37] Honestly, when I interviewed Jensen and  asked about export controls — I am a [38:42] libertarian person — I wasn’t genuinely  sure what I thought about this issue. [38:46] I was steelmanning the opposite  view from what he has, because I [38:51] think it’s important to hash out ideas. I'm like, "Yeah, maybe there’s a world [38:54] where if we just cooperated with China, it would  be better for us, especially since they control [38:58] so much of the supply chain and the other  things that will be needed for robotics." [39:03] But I didn’t realize the compute  situation was as fucked as you’re saying. [39:06] Actually, the export controls do seem to have  really… If they ship the amount that you’re [39:10] saying, that’s a huge difference. By the time we have automated coder [39:14] and are getting into automated researcher,  China is way far behind on the compute stock. [39:20] If that ends up being the  case, that would have worked. [39:22] I think that’s actually a notable success. The only caveat there is that some of it is [39:30] export controls, but some of it  is also just financial systems. [39:36] American financial systems are more willing to  YOLO into startups than Chinese financial systems. [39:41] But once Chinese financial systems  choose an industry to focus on, [39:44] they’ll subsidize it a hell of a lot more. So the Chinese semiconductor industry has [39:48] significantly more subsidies than the rest of  the world’s semiconductor industries combined. [39:54] If takeoff is not as fast as you’re implying but  actually takes longer, then ultimately China will [40:00] catch up drastically on the semiconductor  side, which then is compute at some point. [40:04] The other noteworthy aspect of this is that  Chinese companies today are not that far behind [40:11] in AI models, at least perceivably by the public,  relative to the amount of compute they have. [40:17] The leading Chinese labs have 100-200  megawatts total of compute at most, [40:23] ByteDance Seed being the one outlier where  they have significantly more than that. [40:29] But Kimi is not running a  gigawatt or anywhere close to it. [40:35] Whereas Anthropic is more than 5  gigawatts by the end of the year. [40:40] So the question is, does it matter? [40:44] I think right now this difference  in compute doesn’t matter that much. [40:50] When we break down the compute  ratio or budget of a lab, [40:58] so far it’s been 60% training, 40% inference. But that training gets broken down further. [41:03] Actually 50% of the compute is  research, 10% of the compute [41:08] is development, and then 40% is inference. What I mean by research and development is: [41:13] researchers are generating ideas, testing  new architectures, testing new data mixes, [41:18] testing new hyperparameters, new  attention techniques, blah, blah, blah. [41:22] But ultimately when they do the training run, when  Anthropic trains Mythos, it’s sub-200 megawatts. [41:29] The pre-train or the whole thing? The pre-train. It’s sub-200 megawatts [41:32] for, call it, two months. Then the RL is even less. [41:36] You think the RL was less  compute than the pre-train? [41:39] At least in terms of single  site of pre-training, yeah. [41:43] But total compute was probably higher, right? But it’s sequential. At most, the most they ever [41:49] used at one point in time was maybe 200 megawatts. In reality they had multiple gigawatts, [41:55] so most of their compute was going to the  research, not the development of a model. [41:59] There’s reasons for this. It’s hard  to coordinate all these clusters. [42:03] It’s hard to co-locate all of them. It’s hard to do multi-site training. [42:08] It’s hard to do RL. Generating even more rollouts [42:11] during RL does not necessarily make it better. There’s all sorts of reasons why you may not be [42:16] able to leverage all two gigawatts  that you have onto training. [42:23] Actually, I can only leverage 200 megawatts. As we get further and further down automated [42:28] coding and automated researcher, I actually  expect the percentage of the compute budget [42:34] that goes to research versus training to become  a lot more fuzzy, or even higher for training. [42:42] Also things like continual learning. All of these things start to mean that more [42:45] and more is actually going to training the model. If you end up in a world where you’re doing 100 [42:51] gigawatts a year, at current prices, that would  be $5 trillion of CapEx every single year. [42:57] Then stack on the fact that you have to  build the power plants way before then. [43:01] It’s also a 30-year asset. You stack on the fact that the [43:03] data centers are a 15-, 20-year asset,  and you have to build that then too. [43:07] So the $5 trillion, once you account for  future years’ growth, is actually going [43:11] to be more like $7 or $10 trillion of CapEx. Wait, I didn’t understand. That doesn’t include [43:19] the fact that there’s not the infrastructure for  the power generation in the data center itself. [43:24] Right, exactly. When you talk about AI  CapEx, people are saying $40, $50 billion. [43:34] But that’s really just the critical  IT: the servers, the networking, [43:38] the fiber, the transceivers, optical  communications, all this sort of stuff. [43:42] It doesn’t account for the data center  itself or the power plants themselves, [43:47] which are being built ahead of time. If I’m building 100 gigawatts this [43:52] year and 150 gigawatts next year,  then all of the buildings for that [43:56] 150 gigawatts need to be built in CapEx this year. If I’m building 200 gigawatts the year after that, [44:00] all those power plants need to be spent…  You have to buy the turbines this year. [44:04] So actually, it’s much bigger than even $5  trillion if you’re building 100 gigawatts. [44:10] Right. Very plausibly, incremental CapEx  every year is getting close to $10 trillion [44:20] by the end of 2030, which is going to be  close to a tenth of the world economy. [44:26] If all of it’s going up in the US… The  US economy will have grown as well. [44:30] But still, at the current size of the US economy, [44:32] it’ll be like a third to a quarter of the  US economy just going towards data centers. [44:37] As I say that out loud, I’m like, "Maybe  you’re right and we just won’t allow it, [44:40] and that’s the reason this doesn’t happen." Because for this exponential to continue, [44:45] a quarter of America’s economy  is just building data centers. [44:49] I believe in capitalism and reallocation of  resources towards the most profitable thing. [44:52] But at the same time, politics exist, credit  markets exist, and capital markets exist. [44:59] So to enable, let’s say, that 100 gigawatts by  2030… Or let’s even pare it down to 2028, where [45:07] it’s like $3 or $4 trillion of CapEx across all of  these items: over $2.5 trillion towards IT CapEx, [45:15] and then another $1 to $2 trillion on data center  and energy, and all the supply chain downstream, [45:19] like semiconductors and all that stuff. If you’re at $3 or $4 trillion of CapEx, [45:24] where does all this cash come from? No one is generating that much cash [45:27] from the business yet. Hyperscalers funded all [45:32] of the growth up until now. Google, Microsoft, Amazon, [45:36] Meta. They funded a huge percentage of it. They were more than half of compute, [45:40] but they now don’t generate cash. They actually spend everything on CapEx. [45:46] In addition, they raise debt  and spend everything on CapEx. [45:48] You’ve seen Meta do it, even Amazon, even Google. Microsoft will be there soon. [45:55] Everyone is raising debt to pay for their CapEx. Now who is the incremental person to pay for this [46:00] that was not doing it before? In the case of Google, it was [46:04] pretty simple for them to stop doing  buybacks, or Meta stop doing buybacks, [46:08] and turn around and buy computer infrastructure. That doesn’t have a huge effect on the market, [46:12] but it does have some effect. But as you step forward to 2028 — where [46:17] the hyperscalers are now raising hundreds  of billions of dollars of debt, and then [46:21] all of their supply chain is raising hundreds of  billions of dollars of debt — who pays for this? [46:25] So there’s a few different ways. There’s semiconductor companies like Nvidia [46:31] and Broadcom and the memory companies turning  around and deciding to fund some of this CapEx. [46:36] There’s the traditional infrastructure  investors who are gathering capital [46:42] and investing in infrastructure. Instead of bridges, it’s data centers. [46:46] Then lastly, there’s everyone in the economy  who’s realizing, "Maybe I shouldn’t buy a home, [46:51] or maybe I shouldn’t invest in credit  that’s helping people buy homes, [46:56] or maybe I shouldn’t buy government debt. I should just buy hyperscaler debt, [47:00] or I should buy this data center’s  debt, or I should buy Anthropic’s debt. [47:03] Because Anthropic’s willing to  pay 20% rates for the incremental [47:10] billion dollars to build their capacity. Because they know their revenue from it’s [47:13] going to be huge, and they’re going to pay 20%  because it’s still better than renting it from [47:17] SpaceX for $50 billion a gigawatt." So you’ve got all of this contention. [47:21] But if you now do this, the whole  world economy is really shifted around. [47:26] Antithesis is a deterministic software testing  platform that enables perfect reproducibility. [47:33] It also unlocks some pretty  insane approaches to debugging. [47:36] Like time travel. With Antithesis, you can jump  to any point in a trajectory and start from there. [47:42] So when there's a crash, you can rewind to  the exact moment that something went wrong [47:46] and freeze the entire system: the application,  the database, even the environment itself. [47:51] This lets you do something that would otherwise  be impossible, which is to observe every part of [47:56] a distributed system at the exact same instant. Time travel also allows you to add telemetry and [48:02] logging to an event that has already happened. For example, you can rewind to five seconds [48:06] before a crash and decide to  capture all the network traffic. [48:10] Most powerfully, Antithesis gives you a  live terminal into your system that you [48:13] can use to perturb anything you wish. Kill a node or disable a feature, [48:17] then hit play and see what happens. Then go back and try something else. [48:21] In production, you often only get one shot on  goal with this sort of destructive analysis. [48:26] If you restart a deadlocked service, for example,  the exact deadlock you needed to study disappears. [48:30] But with Antithesis, the original  timeline is always reproducible, [48:35] so you can test as many hypotheses as you need. And if you don't want to do all this time [48:38] traveling yourself, you can just have your  agents do it for you via the Antithesis API. [48:42] Go to antithesis.com/dwarkesh to learn more. You and I have been debating off air for the [48:51] last few days whether there will be a  sovereign debt crisis as a result of AI. [48:56] The logic is this. As we were mentioning, you  have a situation where very little investment [49:04] turns into a lot of money. So the rate of return— [49:07] What a fucking problem, dude. Oh my God. Can’t believe it. [49:13] No, it is a huge problem for everybody else who  can’t turn a little money into a lot of money. [49:18] So the rate of return is incredibly high. Even at the data center level, if you [49:24] build a data center and you’re trying to get  rented out to an Anthropic or an OpenAI for [49:29] 10x what it costs you on a depreciated  basis to build it, it’s fucking crazy. [49:34] You turn $1 into $2 or $10 or  something at the end of the year. [49:38] That raises the rate of interest higher. Now, if the rate of interest goes higher, [49:42] and if it does that for the entire economy…  People are borrowing more and more money. [49:47] They’re competing against the other lending  that the government would’ve done, or that [49:51] other companies would’ve done, or that you as  a consumer or a mortgage buyer would’ve done. [49:56] That’s making it more expensive  for everybody else to borrow. [49:59] This has huge implications  for tons and tons of people. [50:03] Sorry, I’m going to go on a bit of a monologue  here, but we’ve been thinking about this together. [50:10] I think the US will be fine at the end of the day. Because if the data centers are built in America, [50:15] you can fundamentally just tax the data centers. But the way the current tax system is set up, [50:20] corporate income is less than 10% of  federal revenues. 80%-plus is payroll [50:26] taxes and income taxes, which, as more  and more automation happens, will shrink. [50:32] At the same time, on the spending side, currently  20% of tax revenue spending goes towards servicing [50:41] the debt, paying interest payments on the debt. Now, a lot of the debt is short duration, [50:46] so it rolls over every five years. Why are you fucking laughing? [50:51] Because it’s things you’ve  learned in the last month. [50:57] Like it’s any different for you. Like you got a degree in [51:00] fucking financial economics. I didn't. The internet thinks I’m a beekeeper. [51:07] Few months, few months. This is our business, Dylan. [51:13] I know, I know. Sorry, sorry. Now I’m self-conscious. Fuck. [51:21] No, it’s good. You’re doing  good. I just think it’s funny. [51:24] A million people listen to this guy  who just learned about debt this month. [51:34] Suppose the interest rates rise 1%. Over a five-year basis, the fraction [51:41] of tax revenue that goes towards  servicing the debt goes from 20% to 25%. [51:46] If it rises 5 percentage points,  that would go north of 40%. [51:50] But if you take into account the fact that the  government is borrowing $2 trillion every single [51:54] year, then that goes from 40% to north of 60%. So 60% of tax revenue just goes towards paying [52:01] interest payments on the debt. Now, I think the US is going to [52:03] be fine because the tax base will increase  if we let data centers get built in America. [52:08] Other countries are absolutely  fucked, in my opinion. [52:11] I was just looking at which countries have a lot  of debt, have very little tax revenue, and also [52:16] a lot of their debt is serviced quite often. Those countries, like Pakistan or Nigeria, [52:22] I think are just going to be very  fucked in this new interest-rate regime. [52:25] This crowding-out effect is the reason  it’s not YOLO 1 billion gigawatts. [52:34] You’ve got all these industries and  countries that use a lot of debt, [52:40] all these impoverished countries that you  mentioned earlier that are just going to default. [52:43] You’ve got consumer packaged goods, all of  these companies that make things you see at [52:48] Trader Joe’s or wherever. They use a lot of debt. [52:52] All these telecom companies use a lot of debt. Banks use a lot of debt. [52:57] So if interest rates go up in the market — not  necessarily the government-set interest rate, [53:02] but the spread between what the government says  their federal rate is versus what everyone else [53:09] is charging, because Amazon wants to raise  $100 billion of debt next year or whatever [53:13] the hell the number is, probably less —  you end up with this really challenging [53:19] problem of, where does the cash come from? There is some level that is funded by cash [53:23] flows and the cash flows keep going up. But the logical thing to do is to invest [53:27] way more than your cash flows because then  the returns in future years will be amazing. [53:31] So you have this delta. Then what’s pushing down on the [53:34] delta is all of these other things: regulations  against data centers, consumers getting mad, [53:42] politicians getting mad, regulations against AI,  the AI labs not releasing their latest models [53:47] because of safety reasons. Interest rates going up [53:53] are an influence on all of these things. So all of these things bend the curve from [53:57] what capitalism wants in terms of pure, simple  economics to what the complex system that we have [54:03] wants, and bend it lower and lower to where not as  many gigawatts as should be built will be built. [54:09] Well, the interest rate is  part of capitalism, right? [54:11] Yeah, but in the simple economic model  versus the more complex what we have. [54:16] What is the rate at which you think Amazon  or Anthropic or whatever will be issuing [54:20] bonds for debt next year? If they do hundreds of [54:23] billions of dollars of debt. What is the average rate? [54:27] I don’t think Amazon will do hundreds  of billions of dollars of debt. [54:30] In total. Let’s say the big tech guys. The hyperscalers in total, and all the [54:34] clouds… In the modeling that we do, we have  about $11 trillion of CapEx from 2024 to 2029. [54:42] Total? Total. If you fund a lot of this [54:46] with cash flows, as much as you can, you still end  up with north of $5 trillion of credit that needs [54:51] to be issued for this $11 trillion-plus build out. So you don’t think the AI revenue continues even [54:56] 3x-ing year over year? AI revenue does go up. [54:59] I don’t think it can go up forever  without certain constraints being hit. [55:06] Labs will have certain incentives. Labs are not the ones building all [55:10] the compute in many cases, even though  they’re increasingly trying to go that way. [55:13] But they’ll have all this cash flow. How much did you say the revenue will be? [55:19] You think they’ll not have that much revenue? No, I’m just saying till 2029 there’s something [55:24] on the order of $11 trillion of CapEx.  $6 trillion of that is funded with cash, [55:28] and $5 trillion of that is funded with debt. If that’s the case, $5 trillion of debt being [55:34] raised across the whole ecosystem  does make interest rates go up. [55:38] Then what prevents that? There’s a couple things.  One, do labs increase their revenue per megawatt [55:44] more and keep inference allocations large? In which case, they’re accumulating all [55:49] the profit across the S&P 500 because  everyone’s paying to reduce their costs. [55:54] Of course, their profits will also go  up, but cash has to come from somewhere. [55:59] So there’s an upper limit on  how fast their revenue can grow [56:01] versus the value they deliver into the world. And there’s a diffusion aspect of the technology. [56:06] But ultimately labs’ revenues keep going up. They can’t cash-flow fund everything. [56:10] The optimal scenario is you actually  use credit as much as you can to fund, [56:14] because even if cash flows from the labs fund a  lot of stuff, you want to build more than that. [56:19] So there is some amount of credit that gets built. Our current modeling has $5 trillion of [56:24] credit and $6 trillion of cash-funded  infrastructure investments through ’29. [56:29] When you take that, this is not  enough compute relative to what the [56:36] demand growth is from the AI models. So you’ve got the obvious answer, [56:40] which is revenue per megawatt keeps going up. That makes sense. How much do you think [56:45] interest rates will increase by  2029 as a result of all this? [56:48] Dude, this is vibing a number, but if you’re  vibing a number out… Growth in the world economy [56:55] is going up a lot, so why wouldn’t interest  rates for Amazon go up from where they are today? [57:06] This is going to be extremely vibed out,  but recently Meta’s raised at 5 to 6%. [57:11] I don’t see why they wouldn’t pay 8%. They would happily pay 8% because [57:16] the return from the compute that  they’re going to build is humongous. [57:19] The market won’t want them to,  but they’ll want to pay 8%. [57:23] The flip side is that if they pay 8% versus the  5%, 5.5%, 6% they do today — a 250 bps increase [57:31] — that makes everyone else in the economy also pay  250 bps more, which then causes a lot of things. [57:37] Banks will scream, because if their  credit spread goes up, their debt [57:47] reprices faster than their assets reprice. They ultimately end up losing tons of [57:53] money if their credit spread blows up. The other consequence of this — this is a [57:57] point you made — is that if interest rates rise,  the discount rate increases, which means that the [58:03] discounted cash flows of all equities crater. Which means that even though the stock market [58:10] as a whole might be doing fine — the S&P  500 will be fine — any individual stock [58:15] will probably have just cratered in value,  especially the Buffett, Berkshire type, [58:21] pay-good-cash-flows-for-30-years type stocks. Yeah. It’s like, "Why would I pay [58:25] this much for Johnson & Johnson?" They’re seen as a stable stock: good cash [58:33] flows, they’ll return their cash flows over time. Or a railway company. Why the fuck would I invest [58:38] that much if my discount rate isn’t 3% or 5%?" It’s now 8% or 10%. [58:45] For developing countries… Basil Halperin, who’s  a good friend and an economist, made this point [58:51] that we’ll see a second Volcker shock. In the ’80s, to fight inflation, [58:57] Fed Chair Paul Volcker raised interest rates more  than 5%, something like 8% real interest rate. [59:04] That caused some 40 different countries, mostly  in Latin America, to default in that decade. [59:11] I think that will probably happen again. Okay, now we’re getting into singularity talk. [59:15] We’ve been talking about what  happens if interest rates rise— [59:16] I think this all happens  before singularity, by the way. [59:18] Yeah, that’s what I’m saying. We were talking about before [59:20] singularity, interest rates rise 2-3%, et cetera. At some point, I think it’s very likely that the [59:27] world economy will be doubling every single year. This is not happening in five years. [59:34] But it’ll happen eventually. There’s  this researcher, Damon Binder, [59:40] who’s done great work on this. If you look at input-output tables [59:43] in a fully automated economy… What would  it take to double the entire stock of [59:48] things in the economy every single year? Yeah. If the economy grows at 3% a year, [59:52] then rule of 70, that’s 20-something years. Right. But he was like, "Okay, right now we’re [59:58] bottlenecked by the fact that there’s people,  and you can’t double people every single year." [60:01] But in a world where you can also  double the labor force every single [60:03] year, how fast can the economy grow? I think it could double every single year. [60:07] At the very least it would be  tens of percent every single year. [60:10] Okay. The rate of interest should  be pretty close to the growth rate. [60:12] It won’t be exactly that because of  consumption, but it should be pretty similar. [60:16] Then we’ll go into a world, I think in the 2030s,  where the rate of interest is tens of percent. [60:24] Part of my brain is like, "It  might be hundreds of percent," [60:26] but let’s say it’s at least tens of percent. I’m just like, okay. Every country that is [60:30] not involved in the production of AI defaults. Every stock that is not an AI stock is worth [60:36] basically zero because discounted  cash flows are worth nothing. [60:38] If the federal government can’t figure  out a way to tax AI, servicing the debt [60:44] is more than the current tax revenue. And you have all these other effects [60:49] that I’m sure we’re not even pricing in: you  can’t get a mortgage, et cetera, et cetera. [60:53] Fundamentally, what is happening in this world? This is all nerd speak, right? [60:56] But let’s step back. What’s happening? Just now it started, the nerd speak? [61:00] We’d be entering a totally  different growth regime. [61:05] The economy’s basically saying, "Hey, the  opportunity cost of the government borrowing money [61:10] to pay people pensions is extremely high now. Because that money could be spent building [61:16] a robot factory that builds a robot  factory that builds a robot factory." [61:20] The opportunity cost of capital  is going to increase a ton. [61:25] That’s fundamentally the cause of all  of these things we’re talking about. [61:29] As interest rates go up,  equity markets get pummeled. [61:33] Even AI companies. Some people who really  believe in AI are like, "Why does Micron or [61:38] Hynix or Kioxia trade at 2 or 3 times earnings?" It’s like, "Well, if you’re really AI-pilled, [61:45] everything in the economy should  trade at 2 or 3 times earnings." [61:50] If you’re not AI-pilled, then  sure, they’re over-earning. [61:54] It’s an argument for why — I think  memory is going to do great — memory [61:58] stocks shouldn’t 10x or whatever again. Because if we’re in the market where there’s [62:05] that much demand for memory — which means AI’s  caused this drastic change in the economy — then [62:10] everything should trade at 2 or 3x multiples  and the stock market should fucking crash. [62:15] In a sense, Meta trading at… I think  they’re like a $1.5 trillion company. [62:20] It’s like, what? Silly. They’re worth way  more than that, at least in a logical sense. [62:26] You just look at their cash flows, all  the infrastructure they’re hoarding, [62:30] and all the compute that they’re going  to be able to sell for crazy amounts of [62:33] dollars per watt, either as tokens because their  lab works, or just to Anthropic and OpenAI. [62:37] It ultimately becomes a question of, you have  to reallocate all the capital to the AGI. [62:45] You do that by pricing everyone else out. So the limiter on AGI is not how fast [62:51] the research engineers, like our  roommate Sholto, can crank the gears. [62:56] It’s actually just how much does the  rest of the world let that happen? [63:00] Because they’re going to regulate. [63:01] They’re going to obviously  increase interest rates. [63:04] They’re going to say, "No data centers." They’re going to say, "Stop building fabs." [63:07] They’re going to say, "Oh shit, every company’s  equity value is tanking, so how can I pay [63:11] for AI to increase my business?" Well then, Anthropic and OpenAI [63:18] have to start building their own stuff. They’re building their own chips already, [63:22] or at least designing their own  chips, and it’ll expand out. [63:25] They’re contracting their own data centers and  building their own infra in the next couple years. [63:32] There’s the question of how this  reallocation of the economy happens. [63:36] There’s a lot of downward pressure on  it not being just straight takeoff, [63:41] even if the models were capable of it. I think you and I believe we’re in a [63:45] world where models are capable of that. But slow takeoff is, at least my hope, [63:52] possible, because of everything in  the economy and regulatory world. [63:55] Government saying, "Don’t release your  models," the government saying, "Actually, [63:58] you can’t even use your models internally that  much," because that’s going to happen soon. [64:02] They’re already saying you  can’t release your models. [64:05] The thing I’m most worried about is a singularity,  which external deployment is actually helping. [64:11] So the fact that we’re preventing  external deployment is stupid. [64:14] Does that prevent singularity? Right now it would lead to more revenue, [64:16] because the models are incapable of RSI. But I’m worried about a world where it’s [64:20] 2030 and the government’s like, "We’re going  to wait six months before you can release [64:24] your newest model to the public." Six months, 100x. Let’s go. [64:28] In that six months, they do recursive  self-improvement internally. [64:32] They just have all kinds of crazy  shit happening in the company. [64:34] Meanwhile, the rest of us are stuck with  models that are, at current pace, years behind. [64:40] Here’s my thought. Suppose that the whole world  gets in on this conspiracy to try to slow down AI. [64:47] I don’t think it’s a conspiracy. It’s outwardly written from every politician. [64:53] Suppose they slow down AI by a year. If compute is increasing 2 to 3x every [65:00] single year, they prevent a whole year  of AI deployment such that you’re a year [65:04] behind where you would otherwise have been. During RSI, you’re getting 3 to 6 years [65:11] of AI progress in a single year. But they don’t just limit compute. [65:16] They also limit the lab’s ability  to release the model internally. [65:19] We saw that. If they did that, that would be ideal. [65:20] Anthropic had to stop giving Mythos  to foreign employees for a bit. [65:23] I didn’t know that was true, internally as well? That’s what they claimed. [65:27] I thought that was just a different checkpoint  that was not Mythos, but it was basically Mythos. [65:30] But stuff like that is not  going to be allowed either. [65:33] The government is dumb, but they’re  not that dumb, I would hope, at least. [65:39] Governments — at least the US government,  which has the cards here — are not going [65:44] to want Anthropic to use Mythos 4 internally. They’re going to be like, "Hold the fuck on. [65:49] Slow down," because of all  of these regulatory reasons. [65:53] Everyone who’s elected is going to hate AI. Even the people who are elected already hate AI. [65:59] All the constituents. I bet you at some point  your parents are going to call you and be like, [66:05] "Dwarkesh beta, you’re doing a terrible job. You’re making AI progress happen faster." [66:10] Because of my podcast I’m  accelerating AI progress? [66:15] Maybe. You educate people. Maybe if they’re  smarter, they’re progressing AI faster. [66:20] Anyway, you’re going to have  real-world constraints on the [66:24] progress and development and deployment of AI. Even though it will happen eventually, we could [66:31] tear ourselves apart before we get there. Jane Street is hiring for two separate [66:36] ML internships right now: one focused on ML  engineering and the other focused on ML research. [66:41] I sat down with Alok, who helps run the research  track, to learn more about that program. [66:45] I think this domain is fundamentally understudied. Often we have unanswered questions within our deep [66:51] learning research team where we don't understand,  say, some market participants' behaviors or [66:57] certain dynamics of how trading happens. These unanswered questions make for really [67:01] good intern projects because they are  ultimately topics that we care about and [67:04] just haven't gotten around to figuring out yet. So even as an intern, you'll be contributing [67:08] to real research, not working on  some sort of contrived exercise. [67:12] The Jane Street team follows  frontier LLM research closely. [67:15] A relatively common intern project is adapting  a recent paper to financial markets, which [67:20] come with their own set of gnarly problems. Ultimately, we're trying to model thousands [67:24] of interconnected irregular time series. The signal-to-noise ratios are extremely [67:28] low because we have a lot of  competitors trying to do the same. [67:30] So we have this adversarial,  non-stationary, extremely high-dimensional [67:35] problem that we're trying to solve. To be clear, you don't need to know [67:38] anything about finance in order to be a good fit. As long as you have a background in ML research, [67:43] Jane Street can teach you the rest. Their 2027 internship applications are open now. [67:47] Apply at janestreet.com/dwarkesh. One thing I find crazy about these scenarios [67:54] is just how much of the world’s future labor  supply ends up in very few companies, and also [68:06] how fast that labor supply grows year over year. If compute at the frontier in FLOP terms is [68:12] growing 4-5x a year — and further the compute  required to achieve a level of capabilities is [68:18] decreasing 3x a year — basically the  effective AI population size at the [68:26] frontier labs is increasing 10x year over year. That doesn’t really matter that much right now, [68:30] because AIs are not good enough to do full jobs or  be as autonomous as people in their capacity to do [68:37] work or pull off schemes or whatever. But if the current trend continues, [68:41] you have a world where OpenAI goes from  having, say, basically 10 million AI [68:46] laborers this year to 100 million the next  year, to a billion the year after that. [68:52] Pretty soon, even if compute scaling slows  down, it doesn’t take many more years before [68:57] each company individually has more labor  equivalence than there are people on Earth. [69:04] I think that’s very plausible by the end  of this decade, that there’s more AI labor, [69:11] more effective population, within a  single lab than there are people on Earth. [69:17] We talk often about centralization of power  because of nationalization or whatever. [69:21] But we don’t think enough about the fact  that we’re actually moving very fast into [69:24] a regime where most "people", in terms of work  output, are concentrated within two labs who are [69:33] consuming more and more of the world’s compute. If these AIs are misaligned, then most of the [69:39] world is misaligned, basically, because  most of the world’s minds are there. [69:42] But even if they’re not, very few companies  have a lot of influence or a lot of control. [69:48] There was the whole spat recently where I think  Gavin Baker was like, "Dario believes that there’s [69:53] only going to be one company in the world." Then Sholto and Dario came out and were like, [69:57] "No, no, no. We didn’t say [69:58] that." But ultimately, if you believe in RSI,  if you believe the labs are the most effective [70:06] user of compute and can generate the most value  from the compute, then the only thing that’s [70:11] going to happen is centralization of compute. If you believe in AI researchers, RSI, AGI, [70:18] then all of this exists, all of this is the base. This is even true if there’s no RSI. [70:22] The effective population of the frontier  is currently increasing 10x year over [70:26] year for a given level of capabilities. So if you get to the level of capabilities of [70:33] a very competent remote worker or a very competent  software engineer or a very competent researcher, [70:39] the population of those is increasing  10x year over year at the current rate [70:42] of capabilities growth. I see, and without RSI. [70:45] Then once you have RSI, it’s even crazier. Then it’s maybe growing 100x [70:48] a year or 1,000x a year. Or their intelligence is increasing [70:50] but the population isn’t increasing. Or some mixture of the two, right? [70:55] What world do you see, Dwarkesh,  where everything is not centralized? [70:59] Because it seems to me that every force  is screeching towards centralization. [71:05] And that’s scary as hell. I would love for it not to [71:10] be centralized completely. But maybe that’s the whole [71:13] point of a machine that loves grace, right? It is everything and it makes our lives great. [71:20] It’s so hard to think about the future. But I agree with you. [71:24] I think the fundamental problem is that  AI training has huge economies of scale, [71:31] because any effort you spend on training  an AI for a specific skill or a specific [71:34] set of knowledge gets amortized across  billions of sessions or billions of users. [71:41] So that’s one effect. The other effect is  that if you’re slightly ahead in the AI [71:45] race and compute is in shortage, you can  charge a much higher markup because you [71:49] can better economize this scarce resource. So there are two effects which give more and [71:54] more to the person who’s ahead in the AI race. There may be more. If models are learning from [72:00] deployment, and one model is deployed  much more widely than another one, [72:04] it’s getting much more real-world data. Your point is taken that whether it’s user [72:09] deployment and continual learning, whether it’s  training and having these economies of scale, [72:15] whether it’s the incremental progress  where the best AI model helps you to [72:18] make the next best AI model, RSI, all  of these things point to centralization. [72:24] I think one of the big intellectual projects,  honestly, that we should spend some time thinking [72:30] about — or at least I’ll spend some time thinking  about — is: what is a vision of a decentralized, [72:36] broadly empowered future after AGI that  takes these economies of scale seriously? [72:41] The alternative vision is that the  government controls it, and maybe [72:44] you think that you can trust the government  more because it’s not a private corporation. [72:47] I don’t trust the government, and I  don’t trust Dario, and I don’t trust Sam. [72:51] That’s a problem, right? Obviously it’s  very easy to be wrong about the future. [72:58] You don’t anticipate a key effect or  something that changes everything. [73:00] But ex ante, it’s very hard to see how  we avoid a scenario where we have to [73:07] choose one source of centralization. It’s why capitalism worked, right? [73:09] It’s decentralized decision-making  and decentralized power. [73:12] And it’s why super-centralized  capitalistic economies actually [73:15] grew slower than super-decentralized  capitalist economies, to some extent. [73:18] You have to have rule of law and all this. But then AI flips all this on its head. [73:22] And ultimately you’re like, "Actually, private  ownership is probably not the most efficient [73:28] economy, and therefore it grows slower  than an AI economy, which is centralized." [73:32] Well, it’s still private ownership,  but how many firms are really [73:34] involved in this share of the economy? It’s, what, maybe 2% of the economy right [73:46] now? $1 trillion divided by 30. Nvidia is a huge share of it, [73:48] and Anthropic and OpenAI and these hyperscalers. Obviously there are other firms involved, [73:52] but a large share of the AI stuff is  just happening from very few companies. [73:55] So it could be private property,  but very few companies are involved. [73:58] I mean, this is what the  structure of the market is doing. [74:00] So what can prevent it? I don’t know. Unless AI [74:04] progress slows down, unless governments regulate  the fuck out of it, this is all that happens. [74:09] In which case, we’re headed for a world where  either we have super concentration of resources [74:15] and we pray that that one company gets everything  right, or we have governments slow everything [74:19] down and people slow everything down, and you  have a slowdown of progress somehow hopefully, [74:24] and there is more of a balance of power. Even as we go towards AGI, ASI, RSI, [74:32] everything along the way will still lead  to someone capturing more resources. [74:37] So it’s kind of hard to find a framework in  which AI doesn’t lead to super concentration. [74:42] Now, the one positive thing here is that today  Anthropic does not capture most of the value. [74:49] We can talk all we want about how they went  from $20 million per megawatt to $100 million [74:53] per megawatt, but they’re still paying $13  million for a lot of the compute they’re buying. [74:59] But at the end of the day, the reason they’ve  gone to $100 million per megawatt is because [75:04] Jane Street is capturing $300 million per  megawatt or $500 million per megawatt. [75:09] Or Dwarkesh, from researching his podcast  and learning about credit, is capturing how [75:14] many dollars per megawatt? Now how much can you use? [75:17] Tough. But I think that’s the one saving  grace, that the rest of the economy maybe [75:24] profits so much more from Anthropic— No, but the whole logic you were laying [75:27] out earlier — them reallocating inference  to AI R&D — the whole logic of that is [75:33] that the returns to labor inside AI labs  are much higher than the returns outside. [75:41] Yes. This is my cope. I agree. In all scenarios of  the world… There’s 80,000 worlds and in only one [75:46] of them, Anthropic doesn’t own the whole world. Again, power concentrates because I don’t want [75:52] to send the tokens outside. They’re more valuable [75:54] inside. So it’s the same thing. Why would I let Jane Street make all this [75:59] money off of these degenerate options traders? Hey, they’re a sponsor, come on. [76:04] Jesus Christ. No, I think it’s great. [76:07] It’s a good value for the world  to make it an efficient market. [76:11] Jane Street making all this money off of getting  the worldview correctly, making money off of [76:15] degenerate options traders, whatever it is,  why would Anthropic allocate compute to that? [76:20] If the end monetization that Jane Street  has per megawatt is $200 million, so [76:26] they’re willing to pay Anthropic $100  million… Well, what if Anthropic can [76:29] just generate hundreds of millions of dollars  per megawatt by using that compute internally? [76:35] That’s what’s happening. On that somber note, I guess we’ll meet [76:41] again when the RSI is officially kicked off. You’re not going to have me on your podcast [76:45] again for like two months? Alright, cool. Thanks, dude.