（预览）SpaceX 的炒作与 Elon 的交易，Nvidia 与新云，Dropbox、Google 问答及 Ferrari Luce 引发的反弹
- SpaceX 拟议的 2万亿美元估值，与 186亿美元营收、49亿美元亏损及增速从35%放缓至33%并不匹配。 引用数据表明，将 xAI、进而 X 纳入后，公司因投入51亿美元进行 AI 研发，从小幅盈利转为巨额亏损；这笔钱投向了一个排名第5的模型，而其整个创始团队最近刚刚离开。Ben Thompson 仍支持 IPO，因为 Elon Musk 一再把信念转化为追逐原本不可能实现的基础设施所需的资本：“他创造运动来为基础设施融资，而不是先建基础设施来承载一场运动。”（"He creates movements to fund infrastructure instead of infrastructure building to a movement."）这笔交易的赌注在于，炒作最终可能“反向证明”（backwards justify）这笔估值。
- Tesla 说明，仅凭传统估值层面的反对意见，不足以分析 Musk 旗下公司。 Model 3 上市约12小时内就有约300,000人支付1,000美元定金；之后 Tesla 增发股票造成稀释，股价却反常上涨。怀疑者在基本面判断上可能完全正确，持有人却仍然发了财：Musk 是“散户与meme文化的大师”。
- SpaceX IPO 为公众投资者提供了真正的风投式机会，也包括血本无归的可能。 Thompson 以 Tesla、Starlink、每周多次着陆的火箭助推器，以及无需方向盘干预就能可靠完成点到点行驶的驾驶辅助系统为证据，认为 Musk 已经赢得了一定可信度；但他强调，仅靠视觉能否实现完全自动驾驶仍无定论，而 Starship 是 S-1 的核心风险。“你想当风投？机会来了。”（"You wanna be a VC? Here’s your chance."）
- 在轨道数据中心成为必需品之前，地面算力仍然更便宜，还有多条替代路径。 搁浅天然气可以用来发电，替代放空燃烧；老旧 Bitcoin 设施已经在改造成 GPU 机房，海上基础设施也可以承载算力。眼下的直接约束是，合规格的涡轮机可能需要7或8年才能拿到。Thompson 也承认，单纯多造涡轮机可能比做太空算力更容易。Andrew 还提到 Anthropic 与 SpaceX 签署的为期3年、金额450亿美元的算力合同，不过任一方提前90天通知即可退出；AI 基础设施生意已经真实存在，预计还会增长。
- 轨道算力最终是一个边际产能命题，而不是认为太空会迅速击穿地面算力成本。 太阳能免费，但冷却一个假设中的卫星规模单元——比如100千瓦——并不容易；国际空间站的功率也只有约70千瓦，而且整个模式取决于 Starship 能否实现规模化运营。地面设施可能继续拥有更高利润率，而 SpaceX 提供新增的边际算力，并试图“靠规模补回来”。
- SpaceX 的发射垄断可能恰恰因为自身成功证明了可重复使用火箭的可行性而逐渐削弱。 Blue Origin 已完成亚轨道助推器着陆，中国也有强烈动机跟进，因此 Thompson 认为，xAI 潜在的模型与工具链网络效应，比单纯的基础设施更持久。如果最终有多家公司都能发射可重复使用火箭并部署轨道算力，即使 SpaceX 失去独占地位，他仍认为 SpaceX 的逻辑已经得到验证。
- 预览将 NVIDIA 的新云经济学留下为一个明确尚未回答的风险。 NVIDIA 的 ACIE 部门覆盖 AI、云、工业和企业客户，服务对象不包括 Google、Microsoft、Amazon 和 Facebook 这几家主要超大规模云厂商，其中也包括 CoreWeave 和 GPU“野猫”玩家。尚未回答的问题是，在供给恢复常态后，这些客户能否存活，同时继续为 NVIDIA 相对于使用更便宜内部 ASIC 的超大规模云厂商所提供的高毛利整套方案买单。
1. Musk 将投机性信念转化为生产性资本
Sharp 读出了 Thompson 刻意挑刺的算术：SpaceX 以 2万亿美元估值对应 186亿美元营收、49亿美元亏损，增速还从35%放缓至33%。引用数据表明，将 xAI、进而 X 纳入后，公司因投入51亿美元进行 AI 研发，从小幅盈利转为巨额亏损；这笔钱投向了一个排名第5的模型，而其整个创始团队当时已不久前离开公司。
Thompson 的框架从一个倒置开始：传统群众运动需要组织多年，才可能形成声势；社交媒体却能在一夜之间召集人群，却没有支撑这些人群的基础设施。Musk 将乔布斯式的“现实扭曲力场”核武化，利用这些人群为承载承诺的基础设施融资。
Model 3 提供了经典案例：约300,000人在大约12小时内支付了1,000美元定金。后来 Tesla 可以增发造成稀释的股票，眼看股价反而上涨，因为投资者把这轮融资理解成实现梦想所需的燃料。
Sharp 的反驳来自实证：他那位做资产管理的朋友连续8年告诉他，Tesla 的基本面说不通；但在信徒得到回报的同时，Tesla 也做到了约100年来没有一家新车企做到的事——这是 Sharp 的表述。不管估值如何，“显然他做的某件事奏效了”。
2. IPO 将风投风险直接卖给散户
Thompson 并不是在全面为 Musk 辩护，但这些实体成果确实重要：助推器每周多次在海上回收船上着陆，Starlink 已经部署，Tesla 的驾驶辅助系统也能在无需方向盘干预的情况下可靠完成点到点行驶。仅靠视觉的系统能否实现完全自动驾驶，仍无定论。
Starship 是整个逻辑的承重性前提。SpaceX 在2024年用“筷子”接住了飞船，但约2年后，开发进度仍比预期更慢、更难、成本更高；S-1 的风险因素部分开篇数页都在讨论 Starship，因为更广泛的产能命题取决于它。
支持 IPO 并不等于推荐该股。“你想当风投？机会来了。”Thompson 说——而风投意味着可能血本无归。他可能会买几股，因为“我们可能真的做成”，但反复提醒投资者自行承担风险。
3. 地面上还有大量粗放型算力产能
一位听众提出的挑战很直接：如果电力和选址才是瓶颈，那么海上平台、搁浅能源站点及其他地面替代方案，应该都胜过用完即弃的轨道硬件。Thompson 称这是“完全合理的反驳”，也承认 SpaceX 部分是在用火箭这把锤子，寻找能敲的最大一根钉子。
西德州已有生产商无法外运、只能直接放空燃烧的天然气。把这些天然气转化为 AI 算力，在环保上可能优于白白烧掉；与此同时，前 Bitcoin 挖矿设施正在被拆空、重新装入 GPU，新一代“野猫”式算力玩家正在接手这些场地。
隐形瓶颈是发电设备：把天然气转成电力需要涡轮机，Ben 估计采购周期可能长达7或8年。Siemens、GE，以及可能还有 Mitsubishi Heavy，都掌握有价值的涡轮叶片技术；Bloom Energy 燃料电池等效率更低的替代方案也在推进，因为需求已经存在于当下。
Thompson 甚至替听众把反驳说完：“多造一些涡轮机，难道不比把数据中心建到太空容易吗？”他的回答很坦率：“是的，事实可能就是这样。”
4. 轨道算力只有在边际上胜出
太空有一个根本优势——“免费电力”——不需要像地面那样开采天然气、燃烧天然气再驱动发电机，因此所需的能量转换设备更少。拟议中的设备不是轨道仓库，而是大致 Starlink 卫星规模的单元，只是冷却问题要难得多。
Thompson 以国际空间站约70千瓦的功率作为参照；假设中的算力单元可能达到100千瓦。让一侧背阴、另一侧向阳或许有所帮助，但部署足够大的散热面仍是必须解决的工程问题，而不是已经解决的细节。
Musk 的典型路径是先攻克巨大的前期难题，再落到简单、可重复的终局。如今 Starlink 装上飞机看似理所当然，但背后需要火箭、可重复使用火箭以及整套卫星星座；地面 GPU 站点更容易启动，却仍然面临持续的燃料、涡轮机和维护复杂性。
地面数据中心可能一直更便宜，而且利润极高。轨道算力的逻辑在于，商品化算力的定价取决于新增下一单位的成本：如果需求实际上无限，且只有 SpaceX 能够增加边际算力，它就可以“靠规模补回来”——Thompson 说，这种表述本身就应该让人警惕。
5. 竞争将潜在护城河推向 xAI
SpaceX 的领先地位并不一定永久。Blue Origin 已完成亚轨道助推器着陆，中国也在推进可重复使用发射；一旦先行者证明问题可解，足够大的激励往往就能解决问题。
因此，Thompson 认为继续投资 xAI 有其逻辑。他对 xAI 新团队的暂定判断是，它“本质上就是 Cursor”：一家专注编程、拥有有用数据、企业端增长和巨大算力需求的公司；如果最终模式只是把请求转发给 Anthropic、OpenAI 或其他供应商，就不存在可行的未来。
Sharp 补充说，SpaceX 与 Anthropic 签有金额450亿美元、期限3年的算力合同，不过任一方提前90天通知即可退出。Anthropic 目前确实需要这些产能，这意味着 AI 基础设施生意已经真实存在，即使轨道算力尚未到来，也很可能继续增长。
软件工具链和模型网络效应可能比火箭垄断更持久。Thompson 更大的押注是，10或20年后，智能体推理可能让算力呈现完全不同的形态；如果最终有多家公司都部署可重复使用火箭和太空算力，“这是一个美好的世界，也是我想生活的世界”，因此他并不担心 SpaceX 失去独占地位。
6. NVIDIA 的新云护城河仍是预览片的悬念
Thompson 将 ACIE 定义为 AI、Cloud、Industrial 和 Enterprise，本质上就是 NVIDIA 面向 Google、Microsoft、Amazon 和 Facebook 之外客户的销售。客户包括 CoreWeave、政府、为自用而购买 GPU 的企业，以及将搁浅设施改造成算力中心的“野猫”玩家。
尚未回答的挑战是，ACIE 的强劲表现究竟反映的是稀缺性，还是持久性。新云必须购买 NVIDIA 的高毛利整套方案，而超大规模云厂商可以部署内部 ASIC，避开“NVIDIA 税”；预览在 Thompson 回答加速器供需恢复常态后会发生什么之前戛然而止。
Hello, and welcome to a free preview of Sharp Tech. Hello, and welcome back to another episode of Sharp Tech. I'm Andrew Sharp, and on the other line, Ben Thompson. Ben, how are you doing?
Filled with regret, Andrew.
Over what?
Well, I was going through the rundown last night, and there was a question about the Ferrari Luce on here. I decided I was going to troll on Twitter and do the guy in the courtroom saying, “I like the Luce.” But I was in the context of Sharp Tech—
Mm-hmm.
—where last week I explained I don't like electric cars.
Right. Not really an EV guy, yeah.
I actually have a much broader take on the Luce that ties into that. But no one on Twitter has that context, so I'm just hanging out on Twitter defending the car everyone hates, and I'm not really defending it. I was actually trying to have this very subtle diss of it.
Okay.
Total failure on Twitter.
I did not—
You can't do—
—pick up on your diss at all. I saw the tweet—
I know, and it's terrible.
—and I was like, “What the hell has Ben doing here?”
I was thinking about Sharp Tech. You can't delete a bad tweet now.
Yeah.
Right? You just—
You gotta own it.
—you leave it because you're getting beaten down by the comments. You gotta own it. So I'm going to explain myself on this podcast. And we actually talked about it on the other thing, so I'm going to double-dip—
Okay, great.
—because I do feel the need to defend myself, but none of that defense is going to make its way to Twitter. This is actually going to be a theme of this podcast: Twitter being annoying.
Mm.
But this one was understandable.
I cannot wait. That's going to be the party in the back. It's going to be mostly male on this episode. We'll try to hit as many questions as possible, and the Luce discussion will be the party in the back, because you've had takes all week long, and I felt like you were pushing back against some of the consensus backlash to the Jony Ive design.
No, the consensus backlash is totally right, but mostly because they don't understand why they're right.
Mm.
And there's a bit of what our friend Spike Eskin has—the what?—the accidentally viral critical consensus—
Yep.
—where it just becomes the thing to have an opinion about something. And the annoying thing about AVCC is, number one, I share Spike's desire to inherently push back against it.
Yeah.
So that definitely fuels the tweet. But the other part where AVCC gets super frustrating is when AVCC is right.
Mm.
But people are just so on the bandwagon that it gets really annoying.
And they're all—
Yeah.
—annoying about it. Okay, well, we'll dive into all of that in about an hour. For now, we will begin with SpaceX. You wrote a big SpaceX article on Wednesday. It was really sort of an Elon, Inc. article, because there was a lot of good Tesla discussion in there as well.
But as to SpaceX and the impending IPO, I'll read from your article: “SpaceX is seeking a $2 trillion valuation on a mere $18.6 billion in revenue, with $4.9 billion in losses last year. Growth actually slowed from 35% to 33%. That slowdown happened despite the addition of xAI, and thus also X, which tipped the company from a small profit to that massive loss thanks to $5.1 billion in AI R&D expense. That R&D, keep in mind, went toward building a model that is in fifth place and whose entire founding team recently left the company. But sure, $26.5 trillion AI opportunity.”
Alongside that note, you also wrote that the SpaceX IPO is an IPO worth supporting. So tell me why. Tell me how you're making sense of all of it.
Well, this is part two of Twitter being annoying. Tae Kim, who I quite enjoy—
Great follow.
He has his own newsletter.
Yep.
Yeah, he's a big NVIDIA guy. He wrote the book The NVIDIA Way, which was a great read. He quoted this paragraph and said, “Oh, I love Ben Thompson. He'll give it straight.”
The reason that paragraph is in there was to say up front that this whole thing, from a financial analyst perspective, doesn't make any sense.
Doesn't fully pencil out.
Right. And so, just in general, when it comes to anything Elon Musk-related, I write about him once every five years.
Mm-hmm.
And the reason is that it's really annoying and impossible. Why do I never cover Tesla earnings? Because Tesla, all along, is fueled by promises and belief, especially of—
Cybercab—
—like retail investors.
Well, but you go—
—coming soon. Imagine how big the market's going to be. Yeah.
Well, so I wrote “It's a Tesla” in 2016 or something like that, when the Model 3 was made available for preorder and 300,000 people put down $1,000 in 12 hours.
Mm-hmm.
That remains pretty incredible for a car. It spoke to the power of Elon Musk and the brand.
Yeah.
I wrote another one a few years later, “Mistakes and Memes,” about this flip. You go back to the March on Washington, a very famous civil rights march, which entailed years of preparation, both in terms of seeding the political ground and organizationally. That specific march, I think, was 2 years in the planning.
Mm.
And it culminated in this incredible moment, one of the most famous speeches in history: Martin Luther King's “I Have a Dream” speech. That was how mass movements had to be built. You built them up, and they culminated in this massive spectacle on the National Mall.
I didn't know that.
On the internet, it doesn't work that way. On the internet, you can spin up masses overnight.
Mm-hmm.
I think I wrote “Mistakes and Memes” in the context of the whole GameStop sort of thing, but there have been other things, like the color revolutions, and there was a sense of, Why is it that we get these huge movements and nothing happens? It's because this whole process has been inverted. You can spin up a mass movement on social media—
Mm-hmm.
—but if you didn't put in the years of building the infrastructure around that movement, then it just sort of sputters out and everyone forgets about it.
Mm-hmm.
However—and this is where I tied Musk into that article—what makes him unique—there are a lot of things that make Elon Musk unique, and unique can be good and it can be bad. But no question, he's unique.
Yeah.
There's a bit of a Jobsian reality distortion field, but Musk has nuclearized that sort of capability, where you leverage the masses to achieve the means by which you build the structure that spun up the masses in the first place.
Hmm.
The Tesla stock has always been the best story of this. I have not written about Tesla earnings because any sort of breakdown—even once they achieved success and, honestly, the stock went to the moon—still doesn't make sense if you actually sketch out their profits and losses and expected value.
Not super rational.
People on Wall Street have been writing that TSLA stock does not make sense for ages.
For 10 years, yeah.
And they look dumb, though, because the stock went up. They got their shorts washed when the Model 3 worked and the Model Y worked, and the stock went to the moon. But just because it went to the moon, it's like, okay, if it had stayed where it was, then it actually made sense—but it didn't. It got escalated.
The crazy thing about the Tesla stock that always amazed me, and I'm sure this is the thinking for SpaceX going public, is that they would do a stock issuance to get more money. If you issue more stock, you're diluting the stock that's in the market. The fundamental value of the company—if the company's worth $1,000 and you have 10 shares worth $100, if you issue 10 more shares, nothing has changed about the quality of the company. So the shares should go down to $50 each.
Less valuable. Yeah, you mentioned this last week.
Tesla would issue shares and the stock price would go up. And so it's like they—
Just pile in, man. You know, go for a ride with Elon.
See where you end up.
Right. And everyone’s like, “Oh, yes, he’s investing more money in the company. We’re going to achieve our dreams, achieve our goals.” It’s easy to sit around and mock this and look at these rubes being led astray by Elon, and you’re seeing a lot of it with the SpaceX thing, right? “I can’t believe these idiots are going to invest in this company with these numbers.” That’s why this paragraph was there, in a mocking tone, to reflect that reality. But what Musk is doing and has done again and again is sell people on a dream such that they will put money into the dream, and that’s the means by which he goes and pursues the dream.
Mm-hmm.
He’s flipped over the process. He creates movements to fund infrastructure instead of infrastructure building to a movement. It’s a complete inversion of the way things used to work. He’s the master of retail in that way, the master of the internet, of memes.
Mm-hmm.
SpaceX is the new meme stock. That’s just going to be the case. And so the question then, if you accept that this is what happens—it seems irrational, but this is what happens—is there a path—
To build an infrastructure—
—in terms of what he’s promising—
—that’s actually going to make a ton of money and justify any sort of valuation in this neighborhood?
Right. Backwards justify. So you look at all the idiots who are buying unbelievably expensive Tesla stock and doubling down when they were issuing more. They all made a lot of money.
Yeah.
Right? At the end of the day, all the people who were pointing out that was dumb did get rich—
And by the way—
—the people who held the stock did.
I will just note that my best friend is a money manager. He manages my wife’s money. I’ve mentioned in the past that my wife has Tesla stock. For literally 8 years, my friend Ed has been coming to me and saying, “The fundamentals just don’t make sense with Tesla. You need to get out.” Alice will always say no, and the fundamentals still don’t really make sense with Tesla when you look at the multiple and the valuation and the Cybercab stuff. It is definitely a real possibility.
On the SpaceX front, there are real things that they have working in their favor. They got a $45 billion contract for compute with Anthropic over the next 3 years. Either side can walk away from that deal with 90 days’ notice, which is pretty interesting, but Anthropic definitely needs the compute right now. The AI infrastructure business alone is real and should grow for them. But what I love—
Right. And I should have acknowledged this: I do like xAI. I think the way I’m interpreting it is xAI’s new team is basically Cursor.
Mm-hmm.
Which I think is quite interesting. Cursor needs compute. They need to build their own models. It’s not a viable business if their ultimate model is throwing out to Anthropic or OpenAI or whatever it might be. But they also have lots of interesting data. They have a very good enterprise business that’s still growing.
Mm-hmm.
And if it’s important to be really good at coding, it’s hard to imagine a better place to start. So I don’t even hate where xAI is in that regard. Almost certainly, that deal was structured because you can’t do a $50 billion acquisition with your S-1 already filed.
Yeah.
So I don’t hate where that’s going either, and I probably should’ve given more credence to it. But I wanted to write about space data centers, and I wanted to acknowledge that. I think this framework is actually important.
Mm-hmm.
It’s important to understand the Elon framework, which is not conducive to your friend Ed.
Exactly.
For obvious reasons.
Well, but it has been—I mean, look, the odds of starting a new car company 15 or 20 years ago, whenever Tesla started, and actually succeeding—literally, it had not been done in 100 years.
It had been 100 years.
Yeah.
Yeah, exactly.
And so clearly something he’s doing is working, and I just appreciate any company that’s IPOing. I feel like I came to Sharp Tech in the post-IPO era, and half the companies we talk about are private companies, so I’m excited to open up the balance sheets and have some fun, first of all.
Yeah. Well, this is part of why the IPO is worth supporting. Number 1, just in general—look, this is not a defense of Elon Musk, up and down, left and right.
Yeah.
But there is a bit where both sides go too far. So, on one hand, again, this is why I come back to that tweet: I just had tons of Elon bros in my mentions—
Like Will Ferrone calling into your mentions. Great.
—going off on me, who clearly didn’t read the article, right? They were just attacking that one screenshot, and they’re listing all these things. Whatever, I do engage.
Your article was pretty positive, all things considered, on what’s possible for SpaceX.
I know, but it widely got interpreted on Twitter as very negative, which is funny. I was worried about coming across as way too much of a believer, and so I needed to shield myself and be like, “Look, this is all kind of ridiculous; it doesn’t make sense.” Of course, that’s the part that totally got anchored on. This is why I don’t write about Elon. There’s no winning.
Yeah.
But to the Elon point, look, Tesla exists. The self-driving is amazing.
Mm-hmm.
Is it ever going to get to fully autonomous with vision only? We’ve litigated that. It’s still an open question. As it is, it goes door to door and you don’t have to touch the steering wheel, and it does it reliably.
Yeah.
It’s amazing. Everyone needs to have experienced it at least once. You owe it to yourself just to appreciate how good it is. So that exists. Rocket boosters do land on ships routinely.
Yeah.
It happens every week, multiple times a week. It’s not even a news story. And—
Starlink exists.
Starship, we’ll see. It’s taking a long time.
Mm-hmm.
They’re spending a lot of money on this. It’s definitely, I think, been much more difficult than they think, and I’ve made the case for iteration: getting to the limit, finding exactly where it is—stuff should blow up. There’s a reason the risk factors in the S-1 start immediately with Starship. It’s multiple pages long, because all this stuff depends on getting that sort of capacity into space. In 2024, they caught that thing with chopsticks: “Oh, they’re almost there, they’re going to get it.”
Yeah.
Right? It’s been 2 years, I should say. So it’s a real risk factor. But, number 1, Musk has earned it, and Musk IPOs are by and large good for society. I think that’s been shown, number 1. Number 2, their ambitions are so large and the commensurate risk factors are so high. People talk about, “Oh, it’s not fair that retail investors don’t get a chance to get into startups, and all they do is catch stuff at the end.” It’s like, well, you want to be a VC? Here’s your chance.
Yeah.
And, by the way, being a VC means you might lose all your money. But that’s great. I think it’s great that that opportunity exists. Buyer beware. This is not stock advice. I don’t invest in stocks directly. If I did, probably not for me because I—
Well, it looks pretty overvalued right now, as IPOs go, but who knows?
I don’t know. Maybe I’d buy a few shares just because, hey, we might pull it off. But I’m glad it exists.
Yeah.
And—
I agree.
So there you go.
Well, what I love about SpaceX is it’s not a software company. They’re building real infrastructure that actually matters and could change the way we live in a variety of ways. They’ve already changed national security in some very positive ways, delivering internet in new ways, building out data centers faster than anyone on Earth, at least in this country. Maybe going to Mars, I don’t know. I don’t know if we’ll get to Mars, but I support the ambitions in the physical world. I think that’s pretty cool. So we’ll see what it turns into.
Speaking of the physical world, though, Justin says, “Before we go to space for data centers, aren’t there uglier but more plausible terrestrial or near-terrestrial escape hatches? For example, modular data centers attached to offshore oil platforms or other stranded energy sites powered by gas that might otherwise be flared, with existing industrial maintenance infrastructure, fiber/backhaul, and the ocean as a heat sink.” Lot going on there.
Obviously, marine environments are brutal, and maintenance would still be a nightmare, but that seems dramatically more tractable than launching disposable rack satellites and accepting orbital hardware depreciation. My question is, if the core sticking point is terrestrial power and zoning opposition, why shouldn't we expect the industry to exhaust every conceivable alternative before orbital data centers pencil out? I'm very pro-space industry, and I'd love billionaires to subsidize orbital industrial capability. I'm just struggling to see why Earth-side ugly solutions don't dominate space compute economically for a long time. What do you think, Ben?
So this is totally valid pushback.
Mm-hmm.
So, in fact, a lot of the pushback we're going to go through is valid, and for SpaceX specifically, there's a bit of, “You have a hammer; what's the biggest nail I can find?”
Yeah.
Right? So there was something on Techmeme today—there's an article from The Information—about Apple touting how its own chips give it an advantage in AI. Did Apple develop its chips to give it an advantage in AI?
Hmm.
Or is it casting about and retrofitting its advantages to tell a story about how it has an advantage in AI? Now, that's definitely true, and it also might be true that they do have an advantage in AI—
Mm-hmm.
Because of their chips. All these things can be true at the same time. This definitely applies to SpaceX. SpaceX was not created to put data centers in space.
Yeah.
Right? And so there is a certain retrofitting of mission, and there's a bit where AI is arguably rescuing the SpaceX economic case—
Mm-hmm.
By virtue of delivering this theoretical market that is so large that they didn't plan for, but it happened to come along at the right time.
$26 trillion opportunity right there.
Right.
Yeah.
SpaceX is making the argument for data centers in space because they're a space company.
Mm-hmm.
Right? So, totally valid response. Now, in pushing back on Justin, there are a few reasons why I am fairly optimistic about data centers in space.
Yeah.
Number one, I absolutely think there will be things like this, maximizing everywhere we can find energy. And, by the way, in the meantime, there's still a lot of places to build data centers, particularly in places like West Texas. To his point, there's the bit about how they actually have more natural gas than they know what to do with.
Mm-hmm.
Because it's a byproduct of extracting oil, and yet they literally just set it on fire.
Yeah.
Right? Because they can't get rid of it. It's actually a net positive. It is carbon positive to the Earth to build a data center instead of flaring it up. A lot of existing data centers that have been retrofitted for AI were Bitcoin-mining operations for the same idea: Look, we can just monetize what is actually excess energy.
Mm-hmm.
And then everyone's like, “Oh, wait. I know what's better than Bitcoin: AI. Let's make tokens,” and tear it all out and put stuff in there. There's this whole ecosystem of rando old-school—what was the name for people who find oil wells? Wildcatters, or—
Yeah, wildcatters.
Wildcatter-type people who would just find oil wells and stuff. That's going on with AI and finding places to put GPUs.
Okay.
So, absolutely, absolutely a thing. And so, yeah, data centers on oil platforms, sure. Why not? You do need—there's lots of stuff that goes into that. You have to convert that natural gas into power.
Yeah.
You need turbines. Guess how long it takes to buy a turbine right now?
Probably quite some time. I'm sure—
Like, 7 or 8 years or something like that.
Most of that supply chain is rooted in China.
Yeah. Well, no, actually, a lot of it is not, believe it or not.
Okay. That's good news.
In part because turbine-blade technology is one of the West's great secrets. There's Siemens, there's GE, there's—
Oh, that's right.
It might be Mitsubishi Heavy Industries, the third one.
Yeah.
But—
One of those companies had a big espionage scandal 3 or 4 years ago because that's a nut China has yet to crack.
Yep, turbine-blade technology.
Mm-hmm.
And so there's been a big push toward much less-efficient turbines, temporary ones, or generators, like Bloom Energy, which is doing fuel cells. Those are not nearly as efficient as a proper generator, but you can't get a proper generator, and there's demand right now.
Mm-hmm.
Now, all these things can, and I think will, be solved over time. So it's not dispositive that this means we're fundamentally constrained in how many turbines we have.
Yeah.
We can make more. Wouldn't it be easier to make more turbines than to have data centers in space? Let me just write Justin's next email for him: Yes, that might be the case. So the thing about data centers in space is that there are some fundamental advantages.
Okay.
The fundamental advantage is pretty straightforward, which is free power, right?
Hmm.
And so you do get power, and it's actually relatively simple, all things considered. It's not capturing gas, setting it on fire, boiling water, and spinning blades. If you think about it, the actual power generation is kind of nuts in how it works. And this gets back to the bit we talked about last time. We're not talking about putting a warehouse in space. It's putting a bit of compute that's about the size of what Starlink satellites are today.
Mm-hmm.
Now, you need more cooling. Just to put it in context, I think the International Space Station's 70 kilowatts, and it was very difficult to figure out how to deploy the radiators to cool that. You're going to have to figure that out if you're going for, let's say, 100 kilowatts.
Yeah.
Positioning it so that one side's always in shade and one side's always in sunlight will solve some of these issues. But the long and short of it is that there's an elegance and simplicity to this that I think is underappreciated—
Well, and that's—
This—
My question is, in terms of where they're—
By the way—
Driving to, would it be a cheaper solution, in addition to more availability in space? Is that the goal—the SpaceX value proposition that they're trying to offer?
No, I don't think so. Or—
Yeah.
The bit is that Elon, at his best, ends up with very elegant solutions that took astronomical amounts of investment up front to get in place.
Mm-hmm.
Starlink is an example of this. Starlink on airplanes is so obvious. But to get to the state where, of course, every airline is going to sign up for it, they had to invent rockets, they had to invent rockets that land, they had to do Starlink. They had to do all this.
Yeah.
There's a complexity in Justin's solution that is an ongoing operational complexity. You have to continually figure out what you're going to do with these GPUs, getting the natural gas, powering the turbines, maintaining the turbines, and all that. It's complex on an ongoing basis, even if the path to getting there is much easier.
Mm-hmm.
What Musk usually does is the path to getting there is hilariously difficult—
A nightmare.
And seems stupid.
That no one sane would ever try, yes.
But if you get there, it's super elegant, super reproducible, and super scalable.
Mm-hmm.
The idea that we talked about last week, too, is that space is so much bigger than people realize. And if this is a bet on the demand for compute being effectively infinite—
Yeah.
Where these racks in space would be is providing marginal compute, which just means you need one more bit of compute. Where are you going to go? Of course, you're still going to have the data centers on Earth, which, by the way, are probably going to be extremely profitable because they're cheaper.
Mm-hmm.
But the marginal—if compute becomes a commodity, the price of the commodity is set by whoever adds one more bit of the commodity.
Mm-hmm.
What's their cost structure? On one hand, it's the most expensive one that has the worst margins that adds that marginal compute. But if demand is large enough, then that marginal piece of compute—if you're the only one that can add marginal compute—you can make it up in volume.
Yeah.
That is literally the answer here. There’s a make-it-up-in-volume aspect, which, by the way, should always make you worried and suspicious, and feels very bubbly. “We’re gonna make it up in volume” has always been the excuse to do things that make no sense. But that’s the case: it’s a make-it-up-in-volume situation.
At some point, the demand for compute becomes so large, and the constraints on doing it on Earth become so difficult, and SpaceX gets so good at this, and actually, Starship works—
Right.
—they can deploy at large scale, that this very elegant solution, where we basically move all our compute to space, you look back—
And if it works—
—it’s like, of course, this is how you do it.
And it’s streamlined, yeah. Because you can envision a scenario where we get 10 years or 15 years down the line, and sure, you can set up an offshore data center on an old oil rig that you bring natural gas in for, and you generate your compute—
No, no, the natural gas is from the ocean, to be clear. You’re not bringing in the natural gas.
Okay.
But yes. That’s Justin’s point: the energy is there, but it’s still one of those things you look at, and if the elegant solution works, it’s like, why aren’t we gonna do it that way? Just launch more satellites. It’s very easy.
And my point would be, you do that, and then you get 6 months down the line and say, “Oh, actually, we need more compute.” And then a year down the line, “We need more compute.”
Well, you’re just on this treadmill forever, right? And then—
Exactly.
Yeah.
So if there’s an alternate path where they’re offering marginal compute in perpetuity, and SpaceX—
Yeah.
—has cracked that nut—
Exactly, yes.
—then it’s much easier to offer it, even if it’s more expensive. So—
Right.
—we’ll see whether we get there.
Because the implementation cost of the alternatives gets very, very high. No, that’s exactly right. Now, again, everything I just described to you may sound stupid and impossible, and you might totally be right. That’s why this is an unbelievably risky investment. This is why the laws exist.
Now, should the laws be changed to make it easier for retail to get into private companies? That’s certainly a debate that could be had, particularly as the world has shifted to larger and larger companies staying private longer and longer. This isn’t Microsoft going public at, like, $100 million or whatever it was.
Mm-hmm.
You could ride it all the way up. That’s definitely a discussion to be had, but the reason we have protections against retail investors getting into private companies is because you’re likely to lose your money.
Yeah.
And SpaceX is like, “What if we create a company so audacious and large that we could give the public the opportunity to lose all their money without any regulations whatsoever?” Here we are. So it’s—
Here we are, and we’ll see where we go. Again, maybe we’re going to Mars 10 or 15 years from now. Adam says, last SpaceX question:
“How durable is the SpaceX rocket monopoly? At present, it seems that SpaceX has 2 incredibly valuable moats: a monopoly on space delivery, commercial rocket contracts, and a monopoly on low Earth orbit internet, Starlink. You can currently project out a few years, as Ben’s done, and see how it’s possible that they’ll have a third moat in the future: racks in space.
“This, of course, assumes a lot, and Ben makes that point in his article. For me, however, the most crucial assumption is that SpaceX will be the only game in town for space delivery going forward. Now, there’s no doubt that SpaceX has a huge lead in reusable commercial rockets, but is it safe to think that that will stay the case for a decade, 2 decades, longer?
“In other words, if racks in space become as important as Ben thinks they might be, there would seem to be enormous incentives—far greater than any that exist for satellite delivery and internet satellite constellations—for investment in second-mover R&D, whether from Blue Origin, other startups, or even nation-states.”
What do you think, Ben?
Yeah.
I’ve wondered the same thing about the SpaceX rockets.
No, it’s a very relevant question. Blue Origin has landed boosters—done suborbital, I believe—but they’ve done it.
Mm-hmm.
China is obviously working on this. And to Adam’s point, if the economic incentives get large enough, problems get solved, particularly if it’s been demonstrated that they are solvable.
Yeah.
It’s way easier to invent the second booster that lands precisely because you know it’s doable.
Mm-hmm.
Just like it’s gonna be easier, in some respects, for China to invent—
Turbines.
—Starship, right, because they know it exists.
Yeah.
And a lot of the equipment they need, they can take it apart, look at it, and try to figure it out. And so that’s absolutely a fair point.
This is maybe an argument for why continued investment in xAI is a good thing. If you get there first, and your model and your tooling accrue software-type network effects, in part because you were there and you had way more capacity than everybody else, that provides more durability than just having infrastructure. I think that’s probably the point there.
Actually, it is important to get something defensible in the long run. The way I look at it, though—not coming here as an Elon Musk fanboy or a SpaceX stock advocate—if we get to the state where multiple companies are making reusable rockets—
Mm-hmm.
—and there are multiple companies putting compute in space—
The thesis was correct for SpaceX.
That’s right.
Yeah.
It’s a wonderful world. It’s a world I want to live in. So I’m actually not worried about it. But—
Fair enough. That makes sense.
This is why I had to give myself permission in this article. I’m not being a stock analyst here. I think the—this, you know, I wanted to tie it into the agentic inference idea, this idea that there’s just gonna be a lot of compute, that everything we think about how compute has to look today—
Mm.
—I’m not sure is exactly what compute’s gonna look like in 10 or 20 years.
Yeah.
And part of that is certainly the thesis here. But yes, it’s a fair point, Adam, and I don’t care.
Fair enough. All right. Well, shifting gears, let’s run through some mailbag questions. Aaron says:
“Hi, guys. In Ben’s Tuesday post on NVIDIA, he discusses the AI stack and makes an interesting distinction between where the benefits accrue in the short term because of bottlenecks, and in the long run because of moats. The distinction tees up how NVIDIA’s new reporting segmentation, which was likely done to show just how much of their sales were going to the ACIE customer base, where they have a much stronger moat than they do—
Yeah, is there—
—with hyperscalers.”
Is there a definitive way to pronounce this—
So I saw ACIE in this email, and I immediately thought of Acie Law. In honor of Acie Law, former Texas A&M great, who had some great battles with Kevin Durant about 15 or 20 years ago now, I’m going with ACIE. I’m probably getting that wrong, though.
Yeah, so just to be clear, ACIE is A-C-I-E, which is AI, Cloud, Industrial, and Enterprise. Basically, it’s all of NVIDIA’s sales to non-hyperscalers.
Right.
So exclude Google, exclude Microsoft, exclude Amazon, exclude Facebook. Is that all of them? It’s hard to keep straight.
So we’re talking about CoreWeave here?
It’s not hard to keep straight. There aren’t very many of them. No, CoreWeave is in the ACIE category.
Right.
All these wildcatters that are finding old Bitcoin data centers and putting GPUs in there, those are in the ACIE category. Companies buying GPUs for themselves are in the ACIE category, and governments buying—
Yeah.
—anyone but the big guys is ACIE. We’ll go with ACIE.
Acie Law.
A.C. Earl or Acie Law? Yeah, I thought of both—
A.C. Earl.
—A.C. Earl and Acie Law.
Another good one.
Yes.
Remember some guys?
That’s right.
Isn’t the ACIE customer segment itself, though, currently benefiting significantly from bottlenecks? Aren’t neoclouds largely only thriving because there’s a huge supply shortage? And ultimately, when there’s a balance between supply and demand, won’t the very fact that these guys need to buy an entire NVIDIA bundle, paying NVIDIA’s huge gross margins, by definition mean that they will lose significant share to lower-cost hyperscalers who have their own ASICs and can avoid the NVIDIA tax? What do you think about that possibility? All right, and that is the end of the free preview. If you'd like to hear more from Ben and I, there are links to subscribe in the show notes, or you can also go to sharptech.fm. Either option will get you access to a personalized feed that has all the shows we do every week, plus lots more great content from Stratechery and the Stratechery Plus bundle. Check it out, and if you've got feedback, please email us at email@sharptech.fm.