Anthropic Revenue-Compute Gap Signals GPU Price Surge, Patel Says

Patel称Anthropic营收与算力差距预示GPU价格将飙升 Dwarkesh Patel表示，随着AI营收增速超过芯片供应增长，物理硬件瓶颈将推动算力市场价格暴涨十倍。 尽管实验室营收每年增长10倍，但受半导体制造和供应链物理限制的影响，算力容量每年最多只能增长3倍。为了弥合这一差距，算力价格必须上涨、推理利润率必须扩大，或者必须将更多算力从模型训练转移出来。随着AI模型有能力执行人类水平的软件工程等高价值认知任务，每颗GPU的经济回报大幅飙升，从而在奇点前阶段造成了极端的算力短缺。 Anthropic 等顶尖 AI 实验室营收呈现年增 10 倍的指数级爆发，而整体算力供给年增仅约 3 倍；为弥合这一数学鸿沟，算力市场价格飙升、模型推理利润率提高或算力向推理倾斜将成为必然选择。 (0:00) 前沿实验室极力避免将主力算力转移至推理，以防被投资者视为 AGI 研发停滞并沦为普通云服务商；在充分竞争下智能利润率难以长期保持在 90% 以上，因而算力价格大涨将承担主要的宏观调节作用。 (1:34) 若单张 H100 GPU 运行的 AI Agent 达到人类软件工程师能力，其年化租金理论上可超 25 万美元（达现货价 15 倍）；且在阿尔钦-艾伦效应下，极高的算力单价将淘汰低效与低价值应用，使顶尖高效模型享有更高溢价。 (4:06) 受限于摩尔定律放缓、ASML 光刻机产能瓶颈以及台积电先进制程晶圆分配已逼近极限，算力供给缺乏弹性且无法快速扩张，在能够实现自主采矿与芯片制造的奇点到来前，算力稀缺将长期持续。 (6:46)

Dwarkesh Patel argues that an unsustainable mathematical gap between exponential AI lab revenue growth and physical compute expansion will force a dramatic surge in compute prices. Because frontier labs resist pivoting primarily to inference and cannot realistically maintain 90% software margins in a competitive market, compute costs must rise to absorb the value created by increasingly capable models. The Structural Divergence Between Revenue and Compute (0:00) Patel points out a fundamental mathematical mismatch in the frontier AI sector: while leading lab revenues have grown roughly 10x year over year—with Anthropic rising from $9 billion last year toward a projected $100 billion to $150 billion this year—the total compute available to labs expands at only around 3x annually. If current capability growth curves persist, bridging this gap requires some combination of expanding inference margins, dedicating a greater share of compute clusters to inference, or experiencing steep increases in compute prices. All three relief mechanisms are currently visible. Anthropic’s inference margins climbed from 40% in mid-2025 to over 80% for current models, while OpenAI shifted its inference allocation from 25% in 2024 to an estimated 50% or more. Simultaneously, GPU spot prices rebounded by over 40% from their February 2026 trough, indicating that compute capacity is already absorbing pricing pressure. Why Inference Restraint Channels Pressure Into Compute Prices (1:34) Frontier labs actively resist devoting most of their compute capacity to commercial inference. Within these organizations, inference revenue exists primarily to convince investors to fund the next training cluster. Shifting primarily to inference risks signaling that progress toward artificial general intelligence has stalled, effectively turning frontier labs into conventional cloud vendors. Believing that upcoming architectures will render current systems obsolete, labs feel compelled to keep the bulk of their hardware reserved for training and experimentation. Because labs cap inference reallocation, the remaining options to balance revenue and compute are higher margins or higher compute prices. Patel argues that sustained margins above 90% are improbable in a competitive intelligence market. Consequently, compute prices must rise. Frontier developers are already paying substantial premiums for dedicated, secure clusters rather than public spot instances; for example, Google pays SpaceX $900 million monthly for 110,000 GB200 and GB300 GPUs, an hourly rate representing a 2x premium over spot prices. Economic Value Multipliers and the Alchian-Allen Effect (4:06) As model capability increases, the economic yield per unit of compute rises sharply. Patel illustrates this with a software engineering benchmark: if an H100 GPU equivalent can perform at the level of a human software engineer, that chip could command an annual rental value exceeding $250,000 based on standard tech compensation. This equals a 15x increase over existing H100 spot prices, even before accounting for continuous 24/7 uptime. Citing the lump of labor fallacy, Patel contends that high-skilled AI agents will expand the broader economic pie rather than collapse the marginal value of work. This dynamic is expected to trigger the Alchian-Allen effect across AI infrastructure. When compute becomes expensive, running less capable models that require more tokens to solve a task becomes economically irrational. Frontier labs producing highly efficient models will charge substantial premiums because their systems effectively conserve compute, while lower-value consumer applications and smaller developers will be priced out of scarce hardware. Physical and Supply Bottlenecks Cap Compute Growth (6:46) Unlike commodity markets where high prices eventually spur substitution—as occurred in the Simon-Ehrlich bet—compute supply is bound by rigid, short-term physical constraints. Patel outlines the ceiling that limits compute growth to roughly 3x annually: Moore’s Law scaling contributes roughly a 1.4x factor, fab expansion is restricted to roughly 1.2x through 2030 by ASML’s EUV lithography manufacturing capacity, and wafer reallocation delivers a 1.8x factor. This final driver faces imminent exhaustion. Chipmakers have fueled recent AI hardware expansion by diverting leading-edge silicon wafers away from consumer electronics such as smartphones and personal computers. At TSMC, AI wafer allocation on leading N3 nodes is projected to rise from 60% to 86%, leaving little remaining headroom to reallocate existing fabrication capacity. Patel concludes that while compute will eventually become cheap in a post-singularity environment where autonomous machines handle mining, refining, and fabrication, the transitional period remains severely supply-constrained. During this pre-singularity regime, the massive upfront cost of training models combined with low marginal deployment costs creates extreme economies of scale, threatening to concentrate technical and economic power among a few well-capitalized frontier labs.

Dwarkesh Patel 提出，前沿人工智能实验室每年 10 倍的收入增长预期与算力每年仅约 3 倍的物理扩张上限之间存在严峻的结构性断层。由于实验室极力抗拒将算力主要挪用于推理，加之软件利润率上升存在竞争极限，更智能模型创造的高额经济溢价将作为主要泄压阀，在短期内将算力价格推高数倍甚至十倍以上。 收入爆发与算力扩张的数学脱节 (0:00) 以 Anthropic 为代表的前沿实验室正经历极其迅猛的商业扩张：其年营收已连续三年录得 10 倍增长，去年年底达到 90 亿美元，预计今年底将跃升至 1000 亿至 1500 亿美元区间；若延续此曲线，明年底理论营收需达到 1 万亿美元。然而，支持这一扩张的底层算力整体年复合增长率仅约为 3 倍。Patel 认为，要弥合 10 倍收入增速与 3 倍算力扩容之间的数学鸿沟，逻辑上只能依靠三种机制的单一或组合作用：实验室提升利润率、算力市场价格大幅攀升，或者实验室将更大份额的算力从模型训练切换至推理端。 现实中这三项动态已在同步展开：Anthropic 的推理业务利润率已从 2025 年年中的 40% 快速上升至当前主力模型的 80% 以上；GPU 现货价格已从 2026 年 2 月的周期低谷反弹上涨逾 40%；同时，OpenAI 将其算力分配用于推理的比例已从 2024 年的 25% 攀升至当前的 50% 甚至更高。然而，单纯依赖这些调整机制均面临不可逾越的边界约束。 实验室对推理分流的抵触与算力溢价 (1:34) 前沿人工智能实验室在战略定位上极其排斥将多数算力永久转为推理服务。在顶级研发机构的叙事逻辑中，获取推理收入的核心意图在于向投资人证明商业化前景，从而为下一代超大规模训练筹措资金；若将核心资源倾斜于推理，无异于向市场暗示通用人工智能（AGI）研发遭遇瓶颈，使自身沦为普通云计算服务商。由于业界普遍预期下一代模型将彻底淘汰现有系统，前沿机构抱有极强动机将绝大部分计算资源死守在实验与训练环节。 在推理转移受限的情况下，若期望利润率突破 90% 来平抑缺口，则要求单一顶尖模型必须维持不可动摇的超级垄断，这在竞争激烈的市场中极难维系，使得算力价格飙升成为最核心的泄压通道。此外，前沿实验室因模型权重安全、运营效率和灵活性要求，必须采用专属集群而非标准现货实例。例如 Google 与 Anthropic 租赁 SpaceX 算力资源的协议显示，Google 每月需支付 9 亿美元租用 11 万张 GPU（混配 GB200 与 GB300），其折算的时薪水平比已然处于高位的 GPU 现货市场价格还要高出整整一倍。 劳动力替代价值与阿尔奇安-艾伦效应 (4:06) 随着模型认知水准提升，单位算力能够变现的经济价值呈现非线性暴增。根据 Patel 的测算，假定单张 H100 等效 GPU 运行的 AI 智能体可匹敌人类软件工程师的产出，参考科技行业平均薪酬，该单卡理论年租金价值将超过 25 万美元，相当于当前现货价格的 15 倍，且这尚未计入机器全天候 24 小时不间断运转的额外生产力。针对海量 AI 工程师是否会冲垮劳动力价值的疑问，Patel 引入“劳动力总量谬误”予以反驳，指出如同高技能移民通过拓展专业分工扩大经济总量一样，高技能 AI 同样将支撑起极高的边际劳动与算力价值。 算力昂贵化将引发显著的“阿尔奇安-艾伦效应”（Alchian-Allen effect）。当算力成本飙升至每小时 20 美元时，使用弱小且低效的模型在经济上将极其荒谬，因为低质模型必须消耗更多 Token 才能勉强达成同等任务；而消耗极少 Token 即可解决复杂任务的顶级模型实质上等效于创造了更多算力，使头部实验室得以索取极高效率溢价。最终，用于生成闲聊休闲内容等低价值场景的算力将被高价挤出市场，资源将强制聚焦于自动化 AI 基础研发等超高回报领域。 算力供给的物理瓶颈与极端刚性 (6:46) 对算力长期短缺的警示常被类比于 1980 年代保罗·埃利希关于资源枯竭的失败赌局，但 Patel 强调算力供给的刚性远甚于大宗矿产开采，完全无法通过市场机制迅速吸收需求冲击。目前 AI 算力每年 3 倍的物理扩容上限由三大不可逾越的乘数因子锁死：摩尔定律带来的硬件性能提升因子仅约为 1.4 倍；晶圆制造产能受制于 ASML 极紫外光刻机（EUV）至 2030 年的交付上限，产能扩张因子仅为 1.2 倍；而依靠挤压智能手机与 PC 份额带来的晶圆重配因子则约为 1.8 倍。 然而，跨行业争夺成熟产能的余量已几乎见顶。在台积电（TSMC）最先进制程 N3 节点上，分配给 AI 芯片的晶圆占比预计将从 60% 飙升至 86%。一旦先进制程晶圆被 AI 完全吃透，未来便再无借道其他下游硬件转移产能的退路。这一绝对物理边界彻底打破了供给弹性神话，注定导致算力资源在几何级需求冲撞下产生剧烈的价格重估。 Patel 总结指出，尽管长远来看进入由自主机器人全流程采矿、提纯硅料与制造芯片的“后奇点”时代后，算力边际成本将彻底跌落至原材料和工具的极低水平，但在当前的“前奇点体制”下，3 倍的算力硬件增速断难匹配日益聪慧的模型所引爆的商业变现势能。由于顶级模型具有巨额一次性研发成本和近乎零边际成本无限分发复制的天然属性，这种强大的规模经济不仅驱动了能力跃升，也必然加剧前沿技术垄断与极端经济权力的失衡集中。

[0:00] Today I want to talk about what the  compute situation for the labs will [0:02] look like over the next few years. For the last three consecutive years, [0:05] Anthropic's revenue has 10x'd year over year, and  it's likely to do so again this year. They ended [0:10] last year with nine billion in revenue. I think  they'll probably end this year with somewhere [0:14] between one hundred billion to one hundred  and fifty billion dollars in revenue. Now, [0:17] for this trend to continue, Anthropic would  need to make one trillion dollars in revenue [0:20] by the end of next year. Of course,  there's no deep reason why this has [0:23] to be true. It's a very wild conclusion, and  it's ultimately a question of AI capabilities. [0:28] Does AI get that useful by the end of next  year? But suppose the trend does continue. I [0:33] want to think through what happens in that world. The other big trend in AI is that lab compute only [0:39] 3x's year over year. For a lab to keep 10x'ing  revenue year over year while compute only 3x's, [0:46] one of the following three things needs to  happen, or some combination of the three: [0:50] One, lab margins have to increase. Two, the  price of compute has to increase. Or three, [0:56] the percentage of compute that labs spend on  inference rather than training has to increase. [1:01] My understanding is that basically all three  of these things are already happening. With [1:05] regards to the margins, Anthropic's inference  margins reportedly went from forty percent in [1:10] the middle of last year to upwards of eighty  percent now for Fable. With regards to compute, [1:14] the spot prices for compute are more than forty  percent higher than they were in the February [1:20] trough that we had earlier this year. And with  regards to the share of compute that goes to [1:24] training versus inference, in 2024, according to  Epoch, OpenAI was spending just a quarter of its [1:29] compute on inference, and that number is likely  closer to fifty percent, if not higher, now. [1:34] Now, labs would prefer not to do this final  thing of increasing the share of compute they [1:37] spend on inference. The way the labs see the  world, the whole point of inference revenue is [1:43] to help convince investors to give you more money  in order to train the next bigger, better model. [1:49] And if you're spending most of your compute  on inference, then you're basically declaring [1:53] that AI progress has stalled and you're just now  in the business of being a cloud provider. This [1:58] is a less compelling business than building AGI,  so the labs do not want to be in this business, [2:03] nor do they think they are in this world. They  think that within a year, they'll have built [2:07] models that make the current ones look extremely  shitty. But they need to invest a lot of their [2:13] compute — the majority of their compute  — into doing the training and experiments [2:16] that are necessary to build the next model. So that leaves only two options for how you [2:22] can get out of this gap between the fact that  lab compute only increases 3x year over year, [2:26] but revenue increases 10x. Either  the lab's margins have to increase [2:30] so that they get the surplus, or the price  of compute has to increase so that everybody [2:35] in the stack below the lab gets the surplus. It's not clear to me which world we end up in. [2:39] Do we end up in a world where we go from 80% for  some of the top models to greater than 90% margins [2:46] if the lab margin effect dominates? That would  require the leading model to be so far ahead of [2:51] the competition, because the nature of margins  — why they exist in a market economy — is that [2:57] the thing you are serving is so much better than  what somebody else could go get and replace you [3:01] on the market. But it's just really wild for me  to consider that the margins for something like [3:05] intelligence will be greater than 90% and  they don't get competed away at that level. [3:09] So that leaves only one other possibility of  this escape valve between these two trends, [3:14] which is that the price of compute has to  increase. As I mentioned, this is already starting [3:18] to happen. And the effect is even stronger when  you look at the tranche of compute that the [3:22] frontier labs actually need to accumulate, because  they can't just go out and buy a spot instance. [3:28] They need to make sure that they get enough scale  to get really good efficiency and flexibility, [3:33] and also that they have the kind of compute that  lends itself to the security they need for their [3:39] own weights and for their customers' information. I think a relevant case study here is to look at [3:44] the compute that Google and Anthropic are renting  from SpaceX. Google, for example, is paying [3:49] nine hundred million dollars a month for a hundred  and ten thousand GPUs that are a blend of GB200s [3:54] and GB300s. The price that Google is paying here  is 2x the spot price per hour for those GPUs. And [4:00] that spot price itself is more than forty percent  higher than it would have been in February. [4:06] I want to emphasize a key conclusion here: as  AI models get smarter, they will be better able [4:11] to monetize the same amount of compute. If a true  human-level software engineer could run on an H100 [4:17] equivalent, then at today's prices for software  engineers, that H100 should rent for over 250K a [4:24] year. That's over 15x the current spot price for  an H100. And this is not even accounting for the [4:30] fact that your AI can work nights and weekends. Of course, you might expect that if we had ten [4:34] million extra software engineers suddenly appear  in the economy, the marginal value of a software [4:40] engineer would decrease, and thus the revenue  that that H100 would be able to generate would [4:45] not be 15x higher than it is right now. But  I actually don't know if this is true. If we [4:50] apply this argument to people instead of AIs, then  this would be the classic lump of labor fallacy. [4:55] For example, economists generally believe that  high-skill immigration does not decrease wages [5:00] in the long run because of how innovation and  specialization increase the value of labor. [5:06] Maybe this labor supply shock will be so big  and so fast that we can't count on this general [5:11] heuristic anymore. But if you believe what  standard economics says, then the marginal [5:15] value of labor, and thus the marginal value  of compute, should stay astonishingly high. [5:20] So let's think about what changes in such a world.  One of the things that would happen is that as the [5:24] top labs get better and better at monetizing  compute, and the cost of compute increases, [5:30] it becomes harder for anybody else to  compete against them, because they have [5:34] to bid for this resource against somebody who  is basically able to make better use of it. [5:38] Another thing that will happen — and I think this  is actually the most interesting implication of [5:41] this whole thought exercise — is that if you  can train the best, most efficient model, then [5:46] you'll be able to charge much higher margins than  you can today. This is the Alchian-Allen effect [5:50] in economics, and what it's basically saying is  that if it costs twenty dollars an hour to rent [5:54] an H100, then it would be extremely stupid to use  a weaker, less efficient model, because it's gonna [5:59] burn more tokens on your expensive compute to  get the exact same result. So labs will be able [6:04] to charge a much larger premium if they can train  a model that better economizes this scarce input. [6:11] Basically, if you have a model that can get the  same result by using less compute, then you've, [6:16] in some sense, created more compute, and  the value of compute is gonna increase. [6:21] Another thing that will happen is that a lot of  current popular applications of AI will probably [6:25] get priced out. The reason AI is relatively cheap  right now is that AI just can't do a lot of things [6:30] that top humans can do. But this, at some point,  will no longer be the case. And at that point, [6:35] Google or Anthropic or OpenAI will be  willing to pay more for the tokens to [6:40] automate AI research than you or I will be  willing to pay to make more AI slop talk. [6:46] I'm a bit worried that this kind of analysis  honestly pattern matches a lot onto the ways [6:50] that people in the past have been wrong  about scarcity. I'm thinking, for example, [6:54] of the famous Simon-Ehrlich bet. Paul Ehrlich  was this famous doomer about population growth, [6:59] and he made this bet that a basket of commodities  would increase in price rather than decrease in [7:06] the decade preceding 1990. This is a very famous  bet because it's supposed to illustrate how [7:12] Ehrlich's Malthusian worldview was wrong,  and how he did not anticipate the way in [7:18] which market signals and human ingenuity can  find better ways to economize scarce inputs. [7:24] I'm guessing that the analogy to this bet is  probably wrong. Other analysis has shown that [7:28] if that bet had been made in a different decade,  Ehrlich might well have won. But more generally, [7:33] I think the supply of compute is much less  elastic, much less capable of absorbing large [7:41] demand shocks, and much less capable of being  accommodated by using different substitutes [7:47] than the extraction of different metals is. To illustrate why I think this 3x in compute [7:51] capacity year over year is hard to  budge or potentially even sustain: [7:56] I don't see how any of the three elements that  constitute that 3x can be much accelerated. [8:01] 1.4x of that is coming from Moore's Law. Far  from increasing it, I think it'll be a miracle [8:06] if we can just keep it going for a few more  years. 1.2x is coming from building new fabs. [8:12] This process is ultimately gonna be bottlenecked  up to 2030 and potentially even beyond by just [8:17] building new ASML EUV machines. Dylan, when he was  on the podcast a few months ago, talked about this [8:23] in great detail. And 1.8x comes from the fact that  AI is absorbing a lot of wafer allocation that was [8:30] previously going to smartphones and PCs. This is  probably gonna hit a wall by the end of next year, [8:36] when at the leading edge N3 nodes at TSMC, AI  will have gone from 60% to 86%. At some point, [8:42] you have just absorbed all leading-edge wafer  capacity for AI, and you can't keep increasing [8:46] this number. So I don't know how we even continue  to do 3x compute scaling year over year for the [8:53] next few years, much less go beyond that. At the end of the month, I go through the [8:57] time-honored tradition of closing my  books. I start by opening Mercury, [9:00] which is my banking platform, to make sure that  all my transactions are properly categorized. [9:05] Auto-categorization rules handle the predictable  stuff pretty well. But I'm constantly working [9:09] with new contractors—tutors, researchers,  and videographers—and I'm also trying new [9:14] tools. Manually categorizing all of  these transactions would add a couple [9:18] of hours of overhead every single month. So instead of going through them one by one, [9:22] I have Command, Mercury's built-in AI,  take a stab at all of them at once. [9:26] Command proposes a category for each transaction  and provides its rationale. I just review, [9:31] fix anything that's off, and approve. Once all this work is done in Mercury, [9:34] it syncs everything with QuickBooks. And  Command's judgment calls are genuinely [9:38] good. It does the obvious things, like looking  at the vendor, but it also investigates who on [9:42] my team made the purchase and looks at notes and  memos to build up as much context as possible. [9:47] This is just one of the ways you can use Command  to automate the back end of your business. [9:51] To learn more, go to mercury.com/command. Mercury is a fintech company, not an [9:57] FDIC-insured bank. Banking services are provided  through Choice Financial Group and Column N.A., [10:00] Members FDIC. AI-generated responses and  suggested actions may vary and are not guaranteed. [10:05] Now, I wanna clarify that at some point in  the future, compute will get cheap again. At [10:10] some point, we'll just have robots that can  convert shores of silica sand and mines of [10:15] copper into new computer chips, and then the  price of compute is basically the raw inputs [10:20] and the tools required to do this processing. I'm  just talking about this current pre-singularity [10:24] regime where AI compute merely 3x's year  over year, which is not enough to offset [10:29] how much more valuable AI is becoming over time. By the way, the fact that Anthropic's revenue has [10:34] been 10x'ing year over year, whereas their  compute has only been 3x'ing year over year, [10:39] I think illustrates how strong the economies of  scale are in the model business. And logically, [10:43] this makes sense. When you train a model,  you just have to spend this one-time cost [10:47] to learn all these different skills that then  get to be shared across all your users. This is [10:52] very unlike human labor, where each instance has  to be retrained from scratch. I wish we didn't [10:58] live in a world with such strong economies of  scale for intelligence, because I'm worried [11:02] about power concentration, but it seems we do. Okay, this was a narration of a blog post that I [11:08] also released on my website at dwarkesh.com. Check  it out for other posts or to be notified when I [11:13] release a post in the future. Otherwise,  I'll see you for the next full episode.