Dylan Patel,SemiAnalysis
- 审批已让 AI 算力扩容超出承诺交付日期。 SemiAnalysis 已将多个吉瓦的预期容量从2027年推迟至2028年,部分原定2028年投产的容量也继续延期,因为“根本拿不到许可”。Patel 认为,未来3到4年太空算力都无法带来实质性缓解;即便他谨慎预计本世纪末太空算力可能达到吉瓦级,工程上的不确定性仍然很大。
- 即使底层模型落后,产品体验也能改变 AI 支出的流向。 持续运行的智能体如今占其公司每年1500万美元 AI 支出的45%,其中大部分通过 Perplexity Computer 调用 Claude 和 OpenAI:“说到底只是工具链更好。”他认为开发者切换并不难,但若企业折扣要求将70%的 AI 支出交给某个供应商,也会形成黏性。
- 应用收入增长,并不能说明谁最终留住了利润。 Patel 认为 Sierra、Harvey 等公司已体现出投资回报,同时也承认获客成本可能很高、估值仍有争议。一些应用“纯粹只是转手过账”;Harvey 起伏不定的毛利率,以及他对 Perplexity 能留下多少收入并无把握,都说明光看使用量增长无法证明盈利能力可持续。
- Patel 的5亿美元基金将其基础设施研究延伸至种子轮和 A 轮投资。 这支由他单独担任 GP 的基金覆盖机器人、硅片、光学、制造、网络及相关软件。他划定的边界是能力所及——“我对应用一窍不通”;在他看来,硬件公司通常既需要强创始人,也需要好点子,不能只指望创始人摸索出业务方向。
- 即便长期看多 AI,Citrini 的看空立场也得以保留。 Patel 表示,被收购的研究公司保留了自己的编辑声音和独立身份:“想怎么做都行。”Citrini 的实验性研究方式与 SemiAnalysis 的供应链数据分析形成互补;持有所有权不等于必须认同对方的 AI 投资判断。
- Patel 表示,他不会为了投资宣传牺牲研究公信力。 他区分了 IREN 云业务技术薄弱与其电力资源土地的价值;尽管自己持有 Fluidstack 大额投资,SemiAnalysis 仍披露了其数据中心延期。他的原则是:“我们最爱什么,就批评什么”;基金与研究业务相互独立,咨询信息也受合规约束。
- Patel 预计,尽管当前技术基准表现不佳,AI 仍将在未来一年加速硬件开发。 SemiAnalysis 的 GDPX 测试覆盖芯片设计、材料科学、生物学和网络安全等领域,而模型目前“在大多数基准测试上都很差”。他预计,随着能力提升,制造硬件的公司数量以及真正能把硬件做好的公司数量都会“飙升”。
1. 许可审批正把吉瓦级算力推迟到2027年之后
SemiAnalysis 已将多个吉瓦的预期容量从2027年推迟至2028年,原定2028年上线的“好几个吉瓦”也在继续延期。Patel 给出的原因很明确:这些项目并非受制于施工能力、劳动力或设备,而是“根本拿不到许可”。他补充说,出台禁令或暂停令的大多数辖区,现有或规划中的算力容量都很少。
在德州,Patel 认为电价和用水问题可以应对;工会用工要求则没那么简单,因为当地工会会员稀少,而部分电工和工业管道工的年收入已经达到20万美元以上。官员可以通过水、污水、雨水和空气排放许可阻挠项目。即便设施离网运行、用水很少,也可能因空气排放许可被挡下,因此联邦政府如何落实政策至关重要。
太空算力短期内难以缓解供给压力。对于 Elon Musk 雄心勃勃的2028年时间表,Patel 会押注实际进度要更久;他猜测本世纪内某个时点太空算力或许能达到约1吉瓦,“也许是2030年”。但他也明确表示,等待时间可能长得多:工程难度很高,太空算力无法实质性解决未来3到4年的供给约束。
2. 地方经济影响既解释了反弹,也指向可能的利益交换
Patel 的民调覆盖已有、在建或计划建设100 MW以上设施的县。在已有100 MW数据中心的县,49%的人不知道它就在当地;在计划3年内建成100 MW设施的县,只有40%的人知道项目即将落地。O’Shea 认为,数据中心给了反 AI 人士“一个可以实际去抗议的实体目标”。Patel 认同,对 AI 的反感会蔓延到基础设施建设上。
Patel 提出的用水量对比很醒目,但 O’Shea 立刻提出质疑。Patel 说,Musk 所有数据中心的耗水量还不及3家 In-N-Out;她回应:“这话有个星号。”这个比较将双方的隐含用水都计算在内:餐厅一方包括养牛,数据中心一方则包括芯片、材料和能源。不过 Patel 也承认,设计不佳的设施确实可能大量耗水。
在 Patel 看来,数据中心给社区带来的真实负面影响包括周边噪音、可能推高电价,以及地方政府让出的房产税优惠。建设期会带来短暂就业热潮,但建成后“几乎没人会在数据中心工作”。因此,接入电网、又享受大额税收优惠的设施,留给居民的长期收益可能有限。他还提醒,投资额的 headline 数字中,芯片和设备可能是在州外采购的。
更好的交换方式,是让社区分享收益。Patel 举例称,路易斯安那州一个县获得的税收意外之财,足以向教师发放约2万美元。他还提议补贴项目周边30英里范围内居民的电费,认为与项目规模相比,这笔开支很小。他的实际建议很直接:“补贴一些东西、直接给人发钱,就没问题了。”
3. 更好的产品正在从更强的模型手中争夺支出
O’Shea 问投资者该如何跟踪 AI 领域不断轮换的领先者。Patel 说,估值变化比网上叙事平缓得多:Anthropic 在 Q2 表现突出,OpenAI 在 Q3 领先;如今 Anthropic 凭借 Opus 5.5 和 Fable 5.2 再度“势头正盛”,OpenAI 则推迟模型发布。他起初说 Anthropic 的 Q3 表现很差,随后改口称“还行”。他预计 Google 也会重新追上来。风险在于过度押注单次模型发布。
Patel 认为,Instinct 和 Meta 的 Muse 是“远胜于 OpenAI Dots 的产品”,尽管底层模型更弱。他把问题归咎于限制过多的安全措施,包括 Guardian。O’Shea 反问:如果智能体取消了重要事项,或发出不合适的信息怎么办?Patel 起初回答“那就赌一把”,随后承认这类错误确实不应该发生;不过他认为,竞争对手并没有明显地不断遭遇这类问题。
Patel 自己的支出提供了一个具体例子。Claude Code 在今年大部分时间占据主导之后,短短几个月里,持续运行的智能体在其每年1500万美元 AI 账单中的占比从零升至45%。这笔支出大部分通过 Perplexity Computer,后端仍使用 Claude 和 OpenAI。“说到底只是工具链更好”(“It’s just a better harness”),他说,但也质疑 Perplexity 能否长期留住灵活度高的客户。
用户黏性取决于买家类型。Patel 预计,Muse 和 Instinct 这类消费产品比 Claude Code 或 Codex 更容易留住用户。企业合同则可能改变这一点:他见过 Anthropic 提供折扣,条件是客户将70%的 AI 支出投向它。O’Shea 也质疑他对界面的批评,指出多数用户可能只是登录使用,并不会去寻找模型控制选项。
4. 应用普及仍未解决毛利率与估值问题
“大多数初创公司都会失败,”Patel 说,但他也相信其中一些会取得巨大成功。他认为 Instinct 似乎有望长期发展,但并未认可其估值。Sierra 和 Harvey 收入增长很快,在他看来也能带来投资回报,但获客成本可能很高。对于推理服务商的估值倍数,他同时保留两种可能:估值可能过高,也可能由快速增长和良好毛利率支撑。
收入也可能掩盖转手过账的商业模式。Patel 提到 Harvey 的毛利率“一直大起大落”,并坦言不知道 Perplexity 从他的支出中能获得多少毛利,不过他怀疑不会很多。他愿意通过其他服务商调用 Claude,部分取决于经济账是否划算。更好的界面能赢得使用量,但供应商最终能赚到多少,是另一个问题。
5. 5亿美元基金投资方向止于 Patel 的专业边界
Patel 已以单一 GP 身份募集5亿美元,投资基础设施领域的种子轮和 A 轮公司,覆盖机器人、硅片、光学、制造、硬件、网络以及相关软件。他不愿披露单笔投资金额,因为那会“让我失去灵活调整做法的空间”。朋友以及其旗下各项业务约115名员工,都可能为他带来潜在项目推荐。
他对应用投资的克制说得很直白:“我对应用一窍不通。”他或许能帮助应用公司降低基础设施成本,但无法有把握地判断产品执行情况,或公司能否扩展到10亿用户。即便在熟悉的领域,他也说自己一直在学习。他批评投资人动辄对自己并不了解的业务摆出权威姿态。
SemiAnalysis 团队提供了相关的专业深度:成员有在 ASML、Applied Materials、Lam Research、TSMC、Intel 和 Nvidia 的工作经验,也熟悉云基础设施、内核、编译器、模型训练、数据中心和电厂。在硬件领域,他认为优秀创始人通常还需要一个好点子;应用软件则可能给有才华的创始人更多摸索方向的空间。
Patel 将抽象的投资判断与经营公司的现实负担作了对比:解雇员工、可能因此被起诉、谈合同,以及让团队保持积极性。他说这些决定都很难,即便做过几次,“还是只能摸着石头过河”。经营这些业务也意味着牺牲睡眠、健康,以及与所爱之人相处的时间。
6. Citrini 保留看空声音与实验性研究方式
Patel 说,收购 Citrini 的机会在于改善公司的运营,而不改变其研究身份。James 和团队继续沿用自己的方法;研究总监 Kyle 对发布什么内容拥有最大话语权。员工问他是否可以继续看空时,他的回答是:“想怎么做都行”(“Dude, do whatever you want”)。
两家公司的方法不同。SemiAnalysis 跟踪单个设施、芯片、线缆、服务器、半导体设备和内存价格,同时测试硬件、模型和集群。Citrini 则通过直接实验开展研究:为了解 AI 药物设计,他们用 AI 研发了一种蝎毒抗血清。Patel 对结果的描述很克制,只说它“效果还行”,没有把它说成突破。团队还派人前往霍尔木兹海峡,调查石油运输是否会中断。
O’Shea 提到 Citrini 在2月前后发布、曾震动市场的那篇文章。Patel 强调,作者称它为“思想实验”,而非他们必然相信的预测。另有一件事:Citrini 几个月前从看多 AI 转为强烈看空;他的反应是:“这太离谱了。”他自己仍“长期极度看多 AI”,同时接受双方的工作方式不同。
7. 有价值的信息能否卖出去,取决于能否保持独立判断
数据集和咨询业务贡献了 SemiAnalysis 的大部分收入;主要客户包括超大规模云厂商、半导体公司和 AI 实验室。Patel 拒绝广告和付费发表观点。他把 Gartner 的变现模式称为“把戏”,并举例质疑其一份编程智能体排名——不过他承认,自己记不清 IBM 排在 Cursor 还是 Claude Code 前面。他认为,替代做法是靠有用的研究赢得关注。
客观判断会带来商业后果。Patel 说,客户曾因受到批评而威胁取消合作,但客户团队坚持说仍然需要这些数据。出售有利于客户的观点会损害这种需求。他承认:“我觉得自己有时完全判断错了。”但他的原则是发布团队真正相信的内容,而不是让观点迎合客户偏好。
技术批评与资产估值可以并存。SemiAnalysis 曾批评 Iris Energy(IREN)的云集群不可靠,同时认可其电力资源土地的价值。Patel 也披露自己对 Fluidstack 有大额投资,同时公开报告其数据中心项目延期。他认为,持有财务敞口,不意味着分析师必须替公司的每个方面辩护。
基金与研究业务彼此独立;Patel 表示,咨询工作受合规规则约束,公司在合规上投入了大量资金。研究不会替投资组合公司做宣传,不过他的人脉可以帮创始人招聘、对接供应商。他也坚持对 NVIDIA 某些针对新兴云厂商(neocloud)的做法提出反垄断批评,称这是团队基于观察到的证据作出的判断。
8. 严苛的技术基准测试勾勒出下一轮硬件机会
SemiAnalysis 计划测试 Huawei 芯片、Amazon 芯片和 AI 初创公司的芯片在推理性能上的表现。Patel 也关注半导体封装、网络和存储技术。这些工作延续了公司的既有方向:测试硬件实际能做到什么,而不只依赖规格参数或供应链预测。
公司的新基准 GDPX 覆盖网络安全、芯片设计、硬件设计、材料科学、生物学、量化和 RSI。Patel 说,目前“模型在大多数这类基准测试上都很差”,这些测试涉及复杂任务。他兴奋的是性能提升可能为社会带来什么,而不是宣称模型已经掌握了这些领域。
“我认为硬件的发展会疯狂加速。”Patel 预计,未来一年制造硬件的公司数量,以及能把硬件做好的公司数量,都会“飙升”。机器人也是关注方向之一,但他表示,团队的目标是弄懂整条机器人供应链,而不是声称已经具备完整的相关专业能力。
完整逐字稿
本集逐字稿暂无译文,当前显示原文。
You’re building out an empire here.
We’ve got people in 15 countries. There are about 115 people, with offices in New York, Chicago, San Francisco, Oregon, Singapore, and Korea. We track every data center in the world, and as a firm, we know more about data centers than anyone else. Forty-nine percent of people in counties with a 100-megawatt data center didn’t even know there was a data center in their county.
Oh my God.
My biggest customers are hyperscalers, semiconductor companies, and AI labs.
Dylan Patel, welcome to Sorcery.
Yeah, well, good to be here. It was wizard-themed. Is that—is it Sorcery? I get that it’s like source code or whatever, but—
It’s not source code at all. No.
It’s not based on source code?
It was for sourcing deals. I come from the VC world.
Oh. I was thinking source code, or, you know, that we’re magical.
Magic would be cool. That makes a lot of sense. It was just a play on words.
But your aura has gone down a lot. It was like—
Damn.
It was like source code, wizards, source code—
VC.
—and then VC.
I know.
I think this is—
This is where the hierarchy of things—I really need to go. That’s why I live in L.A. I’m closer to the wizardry, or whatever they call it, all the wizards and, like—
There are wizards in L.A.?
There’s lots of stuff in L.A. There are crystals, Reiki, and—
Yeah, I mean, there are a lot of crystal people, crystal girls, but what do you mean? What’s Reiki? What are wizard people?
Oh, you need to get into this.
Please explain.
It’s all energy healing and the earth—
Oh, yeah, yeah, yeah.
Some of that has actually migrated up north because you guys have taken a liking to mushrooms, psychedelics, and ayahuasca.
Not me. I’ve never done a drug in my life.
Really?
Yeah.
No alcohol?
No, completely.
Holy crap.
Obviously, I drink—
That’s crazy.
I drink—
So this is kind of on theme, though, because we were calling this the Hogwarts library over here.
Mm.
You have a deep, secret library in the back of this office that’s actually quite nice.
Yeah.
Congrats.
Thank you. Thank you.
For today, I was trying to come up with all the different things we could talk about, because we could talk about everything with you. There’s a lot going on, whether it’s IPOs, models, data centers, new regulation, all the research you’re putting out, or even how you’ve been growing SemiAnalysis. It’s probably most relevant to get into the shift in D.C., but also in the U.S., and more regulation. A lot of that comes down to data centers, whether that means pushing them out of states that used to applaud them and now don’t. How are you seeing the stress or pressure from that, and do you see data centers in space as a solution?
I think space is a solution that’s still a few years away, so it’s a lot to say about what happens over the next few years. In terms of regulations, it’s at least a few years away before there are real, high volumes. I know Elon’s trying to do a shitload in 2028, but I’ll take the over on that. At some point this decade, there could be a gigawatt in space. I bet it’s this decade, maybe 2030, if I had to guess.
As far as the space question, it doesn’t solve anything in the next 3 or 4 years, or at least it won’t provide a meaningful amount of capacity. Potentially, space is way more than a few years away, too. The engineering is very hard.
As far as domestic regulations, Governor Hinojosa in Texas has been pretty positive and constructive in her discussions. We’ve tried to talk to people in her administration and her close circle, and we’ve heard what she’s said publicly. She says she’s not going to stop data centers; she’s just going to make sure they don’t raise electricity prices and are reasonable stewards of water. I think most data centers can do that pretty easily, so those things are solvable.
The union labor issue is the other one that’s kind of sketchy, because there’s very little union labor in Texas. How do you mandate union labor when you tell all these electricians, who are making $200,000 or more working on a data center, or industrial plumbers, who are making $200,000 or more at a data center, “Hey, you have to join a union”? That seems like a very unpopular thing. And all the other construction crews, right?
That seems like the biggest issue, at least based on what she’s said publicly, what her administration says, and actually building data centers in Texas. But she has a lot of levers to fight back on any data center. There are water permits, wastewater permits, stormwater permits, water consumption rules, and air permits. She can control and stop those very easily. If she does, a lot of data centers stop. There are data centers that aren’t even connected to the grid and use very little water, but she can still stop them via the air permit.
So it depends on how much of a stickler she is on the union labor point. I imagine it won’t be that crazy because it’s a lot of money and workers coming to Texas. It’s a lot of industry. It’s not the eye-watering numbers you hear about data centers, because most of that comes from outside Texas, like the chips, networking, and a lot of the materials. But it’s still a shitload of people working in the data center industry: tens of thousands of people, and hundreds of thousands in the broader sense.
We’ll see what happens, but we’ve definitely adjusted our timelines. We track every data center in the world and its timeline, and we’ve pushed multiple gigawatts out of 2027 into 2028. In 2028, quite a few gigawatts are getting pushed out. Not because they can’t build them, not because the labor isn’t there, and not because of an equipment shortage, but because permits just aren’t available. I think that’s the biggest risk.
There are a lot of jurisdictions all around the U.S. The vast majority that have banned data centers have no data centers. Most of the jurisdictions that are explicitly banning data centers or running moratoriums don’t have any anyway, or don’t have any meaningful capacity coming online.
We ran a poll in all the counties that already have a 100-megawatt-plus data center, one under construction, or one planned in the next 3 years. Funnily enough, 49% of people in counties with a 100-megawatt data center didn’t even know there was one in their county.
Oh my God.
It’s pretty fucking hilarious, because how do you not know this? We did this poll with David Shore, who collaborated with our team. One of the other findings was that only 40% of people knew a data center was even coming when a 100-megawatt data center was planned in the next 3 years.
I think it’s a function of this: even if you look at all the exit polls, data centers weren’t one of the top issues for the election. It was inflation, the war in Iran, and really pertinent things in people’s lives. It’s pretty clear that most people don’t realize there are data centers coming to their area or region, or that there are data centers already there.
They just have this nebulous view that data centers are bad because of stuff on the internet—both reasonable stuff and stuff from bad sites, as well as completely unreasonable, fake astroturfed stuff. Climate people need something to focus on, and there are also external actors who want to push anti-data-center narratives. Most people don’t know about the data centers that are coming, and the places that have them haven’t banned them yet. Texas is an example. In Ohio, some counties have banned them, but not the state as a whole.
I think it’ll become more and more pertinent, but these companies can easily do things to make it work. We put out a post saying you could just subsidize the power for everyone within a 30-mile radius, and it costs almost nothing. You can subsidize things and just give people money, and it’ll be fine.
So where does the negative narrative come from? I’ve heard amazing stories, too. Dina Powell was speaking about this on stage at the All-In Summit, and she brought, I think, a city council leader. They can now fund all the education in their area because of the benefits and revenue the data center brings to their county.
Yeah. There’s a county in Louisiana where they paid out all their teachers like $20K.
Yeah, I think that probably was it.
There’s a rule that all the property tax you collect has to be credited back to government employees, so they’re like, “Oh, shit. Okay.” If done improperly, a data center can be loud if your house is right next to it. If done improperly, data centers can use a lot of water. But if done properly, they use very little water. All of Elon Musk’s data centers use less water than 3 In-N-Outs.
Really?
Yeah. You have to count the supply chain of cows, but yes. If you count the supply chain of cows—
Oh my God.
—3 In-N-Outs is—
You’ve got an asterisk.
No, I think it’s fair, right? It’s like, what’s the embedded water usage? We did the same for the data center: all the embedded water usage of the materials in there, including the fabs building the chips, and all the embedded water usage of the power being generated from the turbines, and of getting all the oil and gas. Even if you compare the embedded water usage of the data center with the embedded water usage of In-N-Out, 3 In-N-Outs is equal to all of the data center capacity Elon Musk has.
That’s just because agriculture wastes a lot of water, and no one cares about that. But if a data center intakes that water and it evaporates for cooling, people freak out. There are obviously ways to use water in data centers that are very inefficient, but no one does that now.
Yeah.
Likewise, there’s a lot of concern about data centers not returning much to the community if done incorrectly. If you give them a property tax break and then they’re grid-connected, it’s possible they could cause grid prices to increase. They might not give much property tax back, and almost no one works in a data center. It’s great while they’re being constructed, but afterwards it’s like, who cares? So there’s this big boom and then a bust.
Most people aren’t paying attention to that. I think most people just don’t like them because they’re a little loud. If you’re close to them, sure. But if you’re not close to them, there’s no noise, which is the whole point, right? With that survey, 49% of people don’t know there’s a 100-megawatt data center in the county. It’s not like it’s some small one.
Well, do you think it also leans into the huge negative aura and narrative around AI? Data centers are the physical thing you can actually protest—
Yeah.
—versus software.
I agree. I think there are also a lot of incorrect numbers. One of the books on the New York Times bestseller list overestimated the water usage of data centers by something like 1,000 times. Then they revised it, but no one pays attention. They just cite this.
No one... Yeah.
But yeah, I think people dislike AI more than they dislike data centers, but data centers are an easy way to point at it.
So let’s talk about hating AI. We have some awesome, huge IPOs coming out. SpaceX had a big one earlier this year, and it’s now an AI company. We have Anthropic coming up. If this doesn’t land around that time, it’ll be before or after, who knows? And then we have OpenAI. We had Cerebras earlier this year. There’s a lot happening in the AI world, and I recently had Apoorva Agarwal on the pod. We talked about the four seasons of AI—
Altimeter?
He has this whole thing on the four seasons of AI. Google will be in vogue, then Anthropic, then OpenAI, and then—what was the fourth one? SpaceX. How do you deal with juggling who’s top dog at the moment?
The thing is, the market doesn’t actually shift the valuations that much. People talk about this company doing well, and that company doing well, but the market doesn’t go all over the place. It kind of smooths the curve.
Obviously, Anthropic crushed it in Q2, and then OpenAI crushed it in Q3, while Anthropic did pretty poorly. They did okay in Q3. Now they’re back to crushing it because of Opus 5.5 and Fable 5.2. OpenAI keeps delaying model releases, so maybe... And then Google will get back in.
These are competitive markets, and people like us, who watch this stuff and are terminally online, over-index on any individual data point. That’s just human nature.
Are we still watching model releases? Are we watching applications now? It seems like Meta is winning with the release of... It’s not called Jolly. What is it called?
Muse?
Muse.
Do you name your Muse Jolly?
No, that’s what it’s called. I’m just obsessed with the little—
The bear is... The, like, the—
Caricature.
The thingy is called Jolly?
Yeah.
Really? Like, the white thing?
Yeah. Mm-hmm.
Oh, I didn’t realize that.
I don’t know if it’s a bear, either.
It’s a teddy bear. It looks like a teddy bear, right?
Okay. Yeah.
I don’t know. It’s not a bear. It’s a teddy something.
It’s something.
Yeah.
It’s something fluffy.
Yeah.
But—
The fluffy thing. Got it.
The fluffy thing, that is Muse. How do you handle Muse coming into the scene and AI applications starting to pick up some momentum?
Yeah, I think applications are definitely important to watch. It’s pretty clear that Instinct and Muse, despite being worse models than OpenAI’s models, are way better products than OpenAI Dots.
I think some of OpenAI’s focus on safety restricts them. They have this thing called Guardian, and it stops the model from doing certain things in Codex and Dots, often overzealously. So Dots is just a shitty product compared to Muse and Instinct. Manus and Muse will run with what you want them to do.
There’s something to be said about being safe, but good products need to be good products first. Safety is important, but I don’t know. If it texts someone incorrectly or books an appointment incorrectly, fine, whatever.
What if it cancels something that’s important or says something not good?
Yeah, let me roll the dice.
Oh, boy. Okay.
I don’t know. No, I think obviously that’s incorrect. But you’re not hearing about Muse and Instinct doing this all the time, so I think OpenAI is just on the overzealous end.
Products are very important. Throughout the year, our spend has seesawed all over the place in terms of what we’re using, with the majority being Claude Code for most of the year. But over the last few months, we’ve gone from zero persistent agents to 45% of our $15 million-a-year AI spend being persistent agents.
The vast majority of that 45% isn’t through Anthropic’s Claude Tags or OpenAI’s Slack agent or Dots or anything like that. It’s through Perplexity Computer, which is still using Claude and OpenAI on the back end. It’s just a better harness.
It’s very interesting. I don’t know if there’s much of a sticking point for us long term. We’re a small, nimble organization that moves fast, but it says something that products are really important. Instinct doesn’t have the best model. They’re primarily using open models, and yet OpenAI is terrible compared to them.
I think product is very important. Products will be stickier in general. Muse is probably going to be a lot stickier, and Instinct is going to be a lot stickier. People are going to index to that use case. I don’t think Claude Code and Codex are going to be that sticky because developers can switch easily. Enterprise contracts will be sticky, though, and there are contracts floating around where Anthropic’s like, “Oh, if 70% of your AI spend is with us, we’ll give you this discount.”
And if it’s not, you don’t get that discount. I’ve seen them offer that to enterprise customers. Something like that is very sticky, so we’ll see what kind of stickiness happens.
I do wonder how much of this is because Anthropic and OpenAI are just bad at product. Have you seen the OpenAI chat app? It’s so dumb. You open Chat, and there’s a Chat section and a Work section. In each one, you can use a slider to choose what model you want. On the Chat side, you can choose an instant response, Pro, high, extra high, or a medium amount of reasoning effort. On the Work side, it selects the models, but there’s no Pro option. It’s like, what the fuck is this?
I think most people just log in and start using it. They have no idea they can toggle between models. Genuinely, I think probably 90% of people don’t know.
I agree, but why do you have all these options, then?
I don’t know.
Yeah, it’s bad product. I was like, this is a billion-user app.
It’s a fair point, because Tesla will at least take things off the production line. Why don’t the models do that?
Yeah. Meta is good at product, and Muse is actually intuitive. I was able to beta test Muse. Alex texted me, like, “Hey, try this new product out.” It was the week Instinct was popping off. He was like, “We’ve been working on this for a while. Hey, try it out.” So I tried it, and I was like, “Fuck, this is pretty good.” It was clear and had a good UI. That’s not the case with OpenAI and Anthropic, so I don’t know what they’re doing.
It’s this cycle of hiring a Meta VP who wasn’t actually that important at Meta, creating a bunch of garbage, and hiring a new Meta VP. I don’t know what’s going on.
So, as the market shifts into applications, do you remain bullish? I know your sentiment changed a little bit. What’s your temperature check?
Sentiment of what?
The ROI. Is it there, given the amount of spend it takes to grow these companies, apart from the big leader?
Most startups will fail. That’s just a fact of life. But some of them will be successful as hell. It seems like Instinct has some staying power. They’ve gotten a lot of users. I don’t know if it’s worth the valuation they’ve got, but they’re kicking ass, and we’ll see. We have a legitimate shot at having some staying power.
Companies like Sierra and Harvey are growing revenue really fast. Maybe their customer acquisition cost is really high, but they’re solid, and they’re in there. Those companies have ROI. There are other companies that may not. Are the inference providers worth the valuation multiples they have? Potentially not. Or maybe they are, because they’re growing revenue so fast and their margins are good.
I think most startups will fail, and some will succeed. I wouldn’t want to rain on anyone’s parade. I have my own thoughts on each company, but most of these companies are doing badass shit. The market dynamics are such that everyone has to pivot and move really fast. Some application companies will work. Some of them are just pure pass-through, though. I don’t know if you saw Harvey’s blogs; they showed that their margins have seesawed all over the place. Or if you’ve seen, “Yes, 45% of my company’s AI spend now flows through Perplexity,” I don’t know what Perplexity’s margin is on that.
Yeah.
I don’t think it’s that high. If it were a more expensive way of consuming Claude than Claude itself, I probably wouldn’t switch off of it.
1. A $500 Million Hardware Fund
This is great. You’re really getting into the investor seat, which leads us into our next topic: your starting a VC fund—or you started one.
Yeah, I already started it.
So you can’t talk about it too much, but why—
Well, I can. I just don’t want to.
You can? Okay. Then you don’t want to.
I just don’t want to.
Okay. Well, then we’re going to talk about it.
Yeah, sure.
So why extend out? You’re building out an empire here. You acquired Citrini, and you started a VC fund. What’s the goal here? How are you building out some analysis to extend?
I think there are many ways to do it. One of the things I want to do professionally and in life is have fun. I want to make money. I want to know everything, which is obviously a crazy and impossible goal. I want to spend time with those I love, like my family. Those are the basic goals.
From there, what does that mean? If I’m trying to know everything, that makes sense as a research business, as an analyst business. And I want to influence the people and things that matter. That makes sense if you’re an analyst in a research firm. You’re influencing people based on your research, opinions, and the facts you provide. You’re closest to knowing everything, although it’s obviously impossible to know everything.
There are many ways to monetize that. There’s selling research, and there’s consulting. That’s the bulk of our revenue: the datasets we sell to people or the consulting projects we work on with hyperscalers, infrastructure funds, or whoever. We work on these one-on-one consulting projects, and that work is firewalled away from the research.
Another way is investing when you see something awesome super early. We started an early-stage venture fund. I raised $500 million as a solo GP. We’re focusing on seed and Series A. That’s about all I’ll say. It’s in the infrastructure space, all the way from robotics to silicon, optics, manufacturing, and all these aspects of hardware and networking. There are many different places we’re investing, but it’s all early-stage.
We’ll never use the platform I built to pump the companies I’ve invested in. But there are many ways to help them. I have 115 people who work for me across SemiAnalysis and Citrini, and now the third company is the fund, which is separate and firewalled away. There are many ways to help people. If a company needs help, maybe we sell them research, maybe we do consulting, or, if they’re super early-stage, we can invest and give them some help pro bono.
There are a variety of ways to help people. The key thing is that I want to know everything, have fun, make money, and have influence—and obviously spend a lot of time with those I love.
As you think about that, where do you find the advantages and conflicts? Obviously, there are so many advantages to having you as an investor, and there are some conflicts with that, too. Most investors do have their own thing or their own company. You mention that you have a lot of influence and knowledge, and it’s clear that you’re compiling that together for the fund. Why did you choose the hardware side of it?
There’s some software stuff, too. There’s software for hardware, software for infrastructure, software infrastructure.
I don’t know shit about applications, so why would I invest in applications? I don’t know how to scale a fucking app to 1 billion active users. Why would I be useful? I could be useful to, let’s say, an Instinct, right?
Mm.
Help them think about infrastructure and cost or whatever. I met the founder recently, and we talked about some of this stuff, but at the end of the day, I don’t know shit about the product, and I don’t know how to scale users. How can I? I don’t think I could even make the determination that he’s going to succeed or not. He seems smart. They seem like good people, but how am I supposed to invest in that?
Whereas in my company, SemiAnalysis, I have people who have worked at ASML, Applied Materials, Lam Research, and all these companies making equipment for semiconductors. I have people who did R&D at TSMC and Intel. I have people who’ve designed chips and worked on kernels at NVIDIA, and people who’ve done infrastructure at Google, Microsoft, and Amazon. I have people who’ve worked on training models, people who’ve worked on compilers, people who’ve built data centers, and people who’ve built power plants. I have people who’ve done all of these things in my company, and I can tap into that expertise.
I think we, as a firm, know more about data centers than anyone else. I think we know more about where and how power is being generated for these data centers behind the meter. I think we know more about chips and all the aspects of chips and software and evals. We’ve started testing eval models with GDPX. There are all these different things I think I have expertise in as a company, or that we’re trying to become experts in. We’re trying to sell data, and we’re trying to do consulting, but it’s not application stuff. It’s models, infrastructure, and hardware.
There are aspirational places, right? Like robotics. No one really knows the whole robotics supply chain, but we’re trying to be experts in it. Hopefully, we’ll be the experts in it, and then it would make sense, right? I don’t understand investing in shit you don’t know anything about. I think one of the most toxic traits of investors, whether in San Francisco or New York, is that they think they know everything about something that—
They don’t know anything about—
—they don’t know shit about. Yeah. I don’t know shit about a lot of things.
Mm.
I think I know some about the things I’m investing in, and I think I know more than the vast majority of people. But even in places I’m investing, I’m learning stuff all the time. If I didn’t know anything, I’d be so lost. How do you invest like that?
Obviously, you invest in founders, and then you invest in ideas, and a great company is both of those things. I think in hardware, it’s harder to just have a great founder who figures things out. You kind of have to have a great founder and a great idea. Whereas in a lot of the application software world, you can just invest in a great founder, and they’ll figure it out.
So what does early stage mean to you, though? What’s the check size you’re going at? Because it’s such a sliding scale.
I’ll keep that confidential.
Confidential?
Yeah. Why would I tell people what I’m doing? That removes optionality for me to do something different.
Maybe someone wants your money.
If they’re awesome, they’ll find me, or I’ll find them, right? Hopefully.
I guess that’s a fair point. Are you doing a lot of outbound sourcing, or is it kind of coming to you? Do you get it through your cabal?
Cabal’s a strong word.
I don’t really know if I’m even using that word right.
I think a cabal is a group of people who are conspiring and, you know, evil. It has some evil connotation to me.
What’s the non-evil version of that?
I don’t know.
Your friends.
Yeah, my friends—
Okay.
—and team.
Besides your friends.
Friends, team, you know. There are 115 people who work for me. If they see something interesting, they’ll refer it.
Oh, you’re gonna get inundated with pitches now.
Yeah, but most people’s pitches are boring and wrong. I don’t know. I think that’s the broad view that I have.
Yeah. VC life.
VC life. I don’t know. But anyway, I think there are a lot of good ideas out there. There’s a lot of interesting work and a lot of innovators, so we’ll find them.
Fair.
The check sizes—I don’t want to talk about the specific details.
Okay. We’ll just leave that one up to the imagination. On a $500 million fund. Well, congratulations on that. I know you don’t like to talk about it, so I appreciate you sharing. In terms of Citrini, can we talk about that? What was the thought process there? It seems like you guys kind of conflict sometimes in how you view the world.
Yes. They’re bearish on AI. That’s okay. If they want to be bearish on AI, that’s fine. Look, James is awesome, and the whole Citrini team is awesome. They’ve captured the market’s imagination. They’ve been right a shitload of times, and their way of doing research is very different from the way SemiAnalysis’s research is done.
Let’s operationalize it. Let’s help them be better. Let’s improve that. I think there are a lot of things they can do better, from the way I see it, while at the same time not losing their ideal of who they are in research. I think it just made sense to join forces with them and help them with the things we can help them with.
It is a separate company. It remains a separate company. Kyle, who’s their director of research, has the most say on what gets published and what doesn’t. I think at one point when we acquired the company, they were like, “Oh, are we allowed to be bearish?” I’m like, “Dude, do whatever you want.” I’m not here to change your voice. I’m here to help you run this as a better company, right? Make it a better company.
I think the team has been pretty positive on that. They’ve been nailing a lot of calls. They got bearish on AI months ago, and I was like, “That’s crazy.”
I mean, they’ve kind of always been a little bearish on something, but I remember when the Citrini paper broke out, probably in February, right? It shook the markets, and then that was a whole topic. That was just one thing, and it had ripple effects, but—
I think some of the things they said have come true. They said it was a thought exercise, right? They were like, “It’s not what we believe is going to happen.” They’ve been super bullish on AI, but then they flipped and were super negative on AI a few months ago.
I’m still super bullish on AI in the long term, but our jobs are very different, right? SemiAnalysis is tracking the exact infrastructure: every data center, every chip. We’re testing the chips. We’re testing the models on bio, security, cyber, and chip design, and all these different aspects. We’re testing clusters, and we’re tracking every bit of the supply chain: how many cables are shipping, how many servers are shipping, how many GPUs are shipping, how many tools are shipping for the fabs, and what the price of memory is. SemiAnalysis is trying to answer all these questions on a very technical and numerical basis.
Citrini, I think, tries to be technical. They wanted to know what was going on with AI drug design, so they went and developed a scorpion antivenom with AI, and it works okay. It works okay. It’s pretty surprising, but they’re super deep into—They wanted to know if oil was going to stop being shipped, so they sent someone to the Strait of Hormuz. They’re going to keep doing crazy shit like this.
It just made a lot of sense. It’s not that I’m going to make them become part of SemiAnalysis. It’s a separate company.
Mm-hmm.
They’re awesome. There’s a lot they can do better, and we want to help them do better.
I’m really curious, because as you build out multiple businesses, who’s a business leader that you look up to? Are you going off the playbook of someone else, or how have you learned—
I think there are a lot of very interesting businesses across history and humanity, but ultimately, no one has really done the mix of things that we’re doing. I think it is uncharted waters.
Media is obviously being completely rewritten. Information services businesses are being rewritten. Everyone recognizes that Gartner, IDC, and those kinds of companies just suck and are useless, and yet information is more valuable than it’s ever been. What does a new kind of information firm look like?
In media and attention, the old guard is, in many ways, bad. I think there are still awesome reporters at places like The Wall Street Journal, Bloomberg, or The New York Times, but those brands have definitely taken a huge hit and fallen off in quality a lot. There’s the creator economy, and there are people who are just big on YouTube or some other platform—Twitter, X, Instagram, Substack. Attention is important. How do you monetize attention? How do you not just attention-hack, but also provide useful things?
Through the SemiAnalysis newsletter, the Citrini newsletter, and their social media, there are a lot of valuable things you can put out there. What’s the correct way to monetize attention? If you look at Gartner, the way they monetize it is that they’re a racket. They’ll be like, “Pay us, or we’re going to rank you poorly.” It’s not that explicit, but you could look at their AI software agent index, and it’s like, yeah, they put IBM’s coding agent—whatever the fuck it is—above, I think, Cursor or Claude Code. I don’t remember which one.
I don’t know.
But they put it above them.
I could be sleeping on IBM.
They have the Magic Quadrant, right? How the fuck is IBM’s coding agent better? Whatever. They’ve done this for a lot of folks. Obviously, the correct way is not to do that kind of bullshit, because then you lose your credibility and no one cares what you say.
One thing we will never do is take paid partnerships. We’ve never done ads, and we’ve never taken payment for an opinion. Our biggest customers are the ones we hate on the most. My biggest customers are hyperscalers, major semiconductor companies, and AI labs, and I don’t think there’s one that we only say positive things about.
We fall in and out of love with Anthropic, and I’ll say that clearly. We’ll say positive and negative things, and the same goes for any of these companies. I think it’s important for people to care what you say. You have to be objective and say what you truly believe. That’s sometimes antithetical to what people want. We’ve had people try to pull their contract because they were mad at what we said, but then their team was like, “Well, no, actually, we still want their data.”
It’s interesting: you have this attention, and you have this business where you’re selling data and consulting services. People only care about your data and consulting services because you’re an objective, truth-seeking player. You could monetize attention by selling an opinion that’s paid, but if you do that, you lose the whole thing. You always want to give your own opinion. You can be wrong sometimes. I think I’ve been totally wrong sometimes, but my whole team and I will always only say what we believe, even if we’re wrong.
That’s a little bit different. We’re a lot more cavalier. We’ve said things like, “We think some of Nvidia’s practices are anticompetitive,” with regard to how they treat certain neoclouds. We’re not going to shy away from that, even if they tell us it’s false. We’ve seen plenty of data points that make us think otherwise.
We’ve also had opinions like, “OpenAI’s product sucks in many cases,” or, “Anthropic’s product teams suck in many cases.” What is going on? Why is Perplexity my primary persistent agent supplier? Yes, they’re doing a lot of things right, but OpenAI and Anthropic are just fucking up. And yet, at the same time, we can say, “Wow, these are the best companies in the world, too.”
As long as people believe our opinion is true to our word, there will always be people who cast shadows and doubts. They’re like, “Oh, you’re just saying that because someone paid you to.” It’s also important to separate opinion. Sometimes we’ll talk about something technical publicly, and then our research will say something completely different.
For example, we’ve said Iris Energy, IREN, has terrible technology on the cloud services side. Their clusters can be unreliable. They don’t have all these other things in ClusterMAX. At the same time, we’ve said, “Oh, but they have powered land, and that’s worth a lot.” We’ve said both of these things. Or vice versa: I have a large investment in Fluidstack, and yet we’ve still said publicly that their data centers are delayed. You don’t have to compromise your morals or beliefs.
I think a lot of VCs especially think they have to be the tail that’s wagged by the dog that is their companies, and actually, it doesn’t matter. Good founders will recognize, “I’m going to give you honest feedback, and we’re going to tell you what we think is true. If you disagree with us, that’s fine. There’s no animosity here.” In fact, I think we criticize the things we love the most, at least as analysts. Just because you’re invested in something or have a financial interest doesn’t mean you have to be the tail that’s wagged by the dog.
So is that the biggest difference you’re going to make as an investor?
Oh, yeah. I think one thing is that we’ll never push our companies that we’ve invested in publicly, because those are separate companies.
Mm-hmm.
The fund is a completely separate entity and company from SemiAnalysis and Citrini. They will never say anything to pump a company that’s on the venture side. But at the same time, we have all these connections through those businesses. I just happen to know people in every layer of the supply chain, and I can help you get introduced to them or hired, or whatever. So I think there’s differentiation.
It’s the same with information I have under a consulting contract: there’s compliance. You spend a lot of money on compliance. You can’t just willy-nilly do stuff. Even if I’ve invested in a company through my fund, or if that company is a big customer of SemiAnalysis’ data or consulting, that doesn’t mean that, if we’re testing something empirically or tracking some supply chain number, we’re not going to just say what we think. Or if we think technically the direction is wrong, we’re not going to say that we think that. I think that’s the key point.
2. The Hard Parts Of Growth
I do think you’ve built a really unique business. You’ve built a phenomenal business, and you did it in a quite short amount of time. You’ve ridden an incredible market, and it seems like you’re creating some durability with the fund and other things, and with how you’re planning out research and pricing things.
But along this journey, I can’t imagine it’s all been that great. What have been the hardest lessons you’ve had to learn, or the most challenging moments?
Running a company is very hard. You lose sleep, so I don’t know, that’s fun. But it’s a lot of fun. You miss a lot of things. You have to choose what you’re going to work on, so you’re going to miss a lot of things, whether it’s in some people’s lives, your own life, sleep, or health. Obviously, that stuff gets sacrificed sometimes.
It’s really tough when you have to make tough decisions. When you’re running a business that does a lot of revenue, you have tons of tough decisions, whether it’s negotiating with a customer or team members, fighting between team members, or—
Kittens.
Yeah, we had a kitten in the office. Something really cute. But I think there are tough decisions to be made all the time. I think that’s another difference: a lot of investors have never made tough decisions.
Do I invest in this company or not? It’s like, eh. Have you fired people? That fucking sucks. Have you been sued by someone you fired? That fucking sucks. These kinds of things are annoying, and you have to learn how to deal with them and know the playbook. How do you negotiate with a customer or a supplier? How do you hire someone? How do you convince someone to work for you? How do you keep everyone motivated? All these things—I don’t know. I actually don’t know the answer to these things, right? You just fly blind.
Mm.
Maybe I’ve done it a few times, but I still haven’t done it much.
But you’ve built it out to how many offices and how many employees?
We’ve got people in 15 countries. There are about 115 people. Offices in New York, Chicago, San Francisco, Oregon, Singapore, and Korea. But we’ve got people in a bunch of other countries.
You speak all those languages?
Fuck no. I speak English, Gujarati, some Hindi, and Spanish. I’d say I can get by in 2 of those languages. I’m good. I’m decent. I can converse, but not well. No, I don’t know that many. I wish I did. I know enough Chinese that I can make people laugh.
Really?
Yeah.
Is it because your Chinese is so bad, or…?
Well, that’s exactly it.
Okay.
It’s bad, and I only know phrases that are insane.
Really?
Yeah.
Should I dare to ask which ones you know?
No, I mean, it’s up to people to ask me, right? I think—
Okay.
There are some old Chinese proverbs, and I think there are some things you can say based on the body language of what’s happening. Even if you don’t understand what the fuck is going on, you just say this phrase. An easy example is, you hear some people, and there’s a certain intonation when they’re speaking, and you say a phrase that means kind of like, “What? Why?” People will be like, “What the fuck?” It kind of makes sense, and they’ll think you know Chinese. You don’t know Chinese. But there’s obviously much more complicated stuff.
Okay. So if I wanted to—
You can talk about cows playing piano—
Oh.
—and how you can’t share an umbrella with a fat person. There are all sorts of things like this that you can say that are really insane.
So if I wanted to show that I can hang, what should I say?
Don’t ask me. I don’t know shit.
But you just said—
I lived in—
Those work.
I lived in Taiwan for a while. I don’t know. It’s context. If you learn it, then I’m giving you that, and that alpha’s decayed. It’s leaked.
All right.
It’s over.
Whatever, man.
It’s over. You’ve got to come up with your own wizardly spells.
You’re right. It all comes back to sorcery. It’s a good full-circle moment. Okay, I think we’re just about at time, if not more. But, Dylan Patel, I have to ask you: What are you most looking forward to in the next 12 months?
I think there’s a ton. There’s just a lot of stuff that we’re launching and doing. There’s testing of various chips, including Huawei chips, Amazon’s chips, and various AI startups’ chips for inference performance. The new model releases, I think, are sick. We launched a new benchmark called GDPX. It tests across many domains: cyber, chip design, hardware design, materials science, bio, quant, and RSI.
I’m excited to see the models get good at them, because the models suck at most of these benchmarks. They’re very complicated tasks, and as the models get better at them, what does that unlock for society? I think we’re going to see tremendous progress in technology, especially hardware. I think hardware’s going to accelerate like crazy. In a year, the number of companies making things in the hardware realm is going to skyrocket, and the number of companies that can do it well is going to skyrocket. It’s going to be really fun to see all the innovation in society from that.
Robotics, obviously. There are tons of niche semiconductor technologies I’m really excited for, especially in packaging, networking, and memory. So, yeah, there are a lot of things I’m excited about.
Cool. Is your favorite chip still Takis? Is that what you said?
Takis.
Takis? Takis?
Takis. T-A-K-I-S.
I don’t.
Takis. The blue ones. Or purple ones. I don’t remember what color.
They’re purple?
They’re blue or purple. I’m partially colorblind.
Really?
Yeah. They say it’s red-green color blindness, but I can tell the difference between red and green, and I don’t fucking understand that. It’s a really interesting one, right? It lives on the X chromosome, the type that I have. I can see most colors. It’s fine.
So this has been clinically confirmed?
Yeah, yeah. It lives on the X chromosome.
Okay.
It can only express itself in men. It goes from the maternal grandfather to his daughter. His daughter always has it. If you draw a Punnett square, that daughter will pass it to 50% of her sons because she has 2 X chromosomes.
Wow.
If a woman passes it to her daughter, it can skip a generation too, right? But it lives on the X chromosome, and it generally comes from your maternal grandfather. It’s an interesting thing. I see color a little bit worse, but I see way better in the dark. I think it’s rods versus cones. The light detectors in your eyes: your rods are better at telling whether it’s dark or light, and cones are better at seeing a spectrum of color. I can see every color around us. It’s just my color skills are a little worse.
What color is this?
It’s red.
What color is this?
It’s blue.
Well, then you can’t see purple? Is that what it is?
No, I see purple too. I just see purple and blue as similar.
Oh, okay.
But I’m also just kind of dumb.
No, I don’t think so.
No, I think so. I think, stylistically, right?
You’re stylistically dumb?
Maybe.
Maybe less aware.
Less aware. There we go.
There you go.
Stylistically less aware.
There you go. That sounds much more sophisticated.
Hmm. Anyways, I think it’s a fun thing to be able to see better in the dark.
How did we start talking about this? I don’t know. All right, Dylan Patel, thank you so much for your market update.
Yeah, for sure.
So—
Thanks for having me on Wizards of Waverly Place.
I know. I’m going to steal this room from now on. I think it’s very fitting.
Awesome. Awesome.