You're Not Buying a Company, You're Buying a Ticket to Mars

“Buffett takes one look and coughs up blood — what is this, there’s nothing here.” — Raymond, on Musk’s one plus one equals five

Guest: Tiannan｜host of the podcast 听懂涨声, on index investing and household asset allocation

On May 11, SpaceX filed its prospectus. A pricing rumored at $1.75 to $1.8 trillion, a company that packs rockets, satellite broadband, 200,000 GPUs and a large model into one shell, and one line written into the industry section that no prospectus in history has ever written: my market is $30 trillion. Of which $23 trillion comes from enterprise applications.

And there’s a queue behind it. Anthropic second, OpenAI probably within two or three months. Three trillion-dollar companies entering one after another and being added to the indices soon after — buy or don’t buy, overweight or underweight, and who do you sell to make room. This summer, American fund managers can forget about a holiday.

This episode is a crossover between Mossfire and the podcast 听懂涨声. Its host, Tiannan, works on index investing and household asset allocation; in his first year out of university he joined Peking University Founder Group and has had a close-range fear of “one company containing a dozen businesses.” Raymond did US equity investment banking and has seen documents like a prospectus from the supply side. The two of them start from the same S-1: one presses on what makes this thing worth two trillion, and the other answers that you’re computing a P/E while they’re selling a ticket.

What follows is the full conversation, edited and condensed.

1. “Space alone isn’t a big enough sea of stars; it takes space plus AI”

Tiannan: People probably have quite a few misconceptions about SpaceX; in their minds it’s rocket launches and Starlink, which doesn’t seem to support the market cap currently projected. What actually is its business — can we start with the basics?

Raymond: People’s impression of it has been shifting for years and is somewhat blurry today, which is normal, like watching a child grow up — this company started in 2002 and has been iterating and evolving for over twenty years. I think the prospectus is its coming-of-age document, and by that document’s own division it’s three parts.

The first is Space, the space business, the one everyone has most clearly in mind — launching rockets, commercial launch. Just over $4 billion of revenue a year, 22%. It is not the largest business.

The real bulk is Connectivity, which is Starlink. Use your own rockets to put satellites up, the satellites form a network, and you install a receiver at home and have broadband. In the desert or the middle of the ocean there was certainly no signal before; install that box and there is. That’s $11.4 billion of revenue, 61%, and the best margins too.

Tiannan: I had one especially striking experience at sea, where the signal in the middle of the ocean was superb. Only later did I learn it wasn’t because a tower had been built nearby, it was Starlink. So it’s effectively the cash cow.

Raymond: Right. Starlink as a business is like your phone plan: seventy dollars or a hundred dollars a month, rolling forward. Users went from 2 million to 4 million to 8 million, and it’s now past 10 million subscribers, minting money every month. EBITDA and revenue growth are both the best of the three.

The rocket-launch part is still in heavy investment, pushing toward Starship and the next generation, burning a lot on R&D. Starlink is subscription; fundamentally it’s a telecom business, space SaaS.

Tiannan: And with no real competitors.

Raymond: No competitors worldwide; its comparable is a company like China Mobile. On the launch side the cadence is already absurd — Falcon 9 alone launched over 40 times this quarter and over 100 times last year. Starship’s cadence is far lower but attracts far more attention, being the most advanced rocket in the world today.

Tiannan: Those two are people’s fixed impression of SpaceX. But last year there was a big change, folding xAI in, and people’s understanding of the company as a whole migrated with it. Can you lay out that timeline, including what actually happened and what the key milestones were?

Raymond: The timeline is indeed a bit complex; it’s a nesting-doll company. Earliest, Musk bought Twitter, which was later renamed X; on top of X he did AI, packaged overall as xAI, which is the Grok model. So xAI is a bundle — the Grok large model, plus Twitter as a social platform, plus Colossus, the 200,000-GPU supercomputing center.

xAI was merged into SpaceX in February 2026, four or five months ago. This merger is particularly interesting: xAI’s revenue last year was only a bit over $3 billion, so after the merger it’s ten-odd percent of the whole company, under twenty. But it brought a new idea — AI. What was a rocket and space company is now a company that burns money, burns a lot of money, but better embodies the future. Space alone isn’t a big enough sea of stars; it takes space plus AI. The imaginative room went up enormously all at once.

2. “On Musk’s side, one plus one equals five”

Tiannan: But there are a few things in between. Musk’s acquisition of Twitter caused a big storm at the time, and there was a whole argument over whether that valuation for Twitter was reasonable. Then when X went all in on AI there was even more doubt — the big three were already running by then, and xAI was weak on both capital and technology, so many people doubted whether you could keep burning money and whether you could catch up. This was plainly a business under continuous doubt; you can’t call it layered packaging, but it really was packed, one layer at a time, into a company that’s going to IPO, and it’s turned instead into something with a lot of valuation potential. Do you think that logic holds?

Raymond: Depends how you look at it. Many people think it’s nesting dolls, but if you take the dolls apart, the structure at each layer is actually fairly logical.

It started with Twitter doing AI. Twitter is where the most enormous corpus in the English-speaking world sits; Musk controls a social platform that large, and he has money and can raise money to buy chips, so evolving toward large models isn’t strange.

With a large model, he needed to buy more chips, so he assembled Colossus, 200,000 cards, extremely expensive. At that point the company’s leverage was high and the cost of borrowing very high. So he said: I also own SpaceX, and Starlink is already profitable and can support the household — why not have the elder sibling look after the cash-starved younger one. Merge the two, the company-level leverage comes down, financing cost comes down, and money spent goes further; and the old telecom-like Starlink business gains more imaginative room.

So at the capital-markets level and the financial-engineering level it makes sense; it isn’t pure nesting dolls. But is every business solidly built? Not necessarily. How to define it — this is a company Buffett would certainly not buy, and yet it has genuinely pushed past the current limits of what humans can do on many dimensions.

Tiannan: Let me offer an aside. In my first year out of university I joined Peking University Founder Group, which at the time had six listed companies and hundreds of subsidiaries; even the founder didn’t know how many businesses the company had, and McKinsey later did a mapping exercise before anyone internally knew there were that many. Fairly frightening. So in China’s capital markets there’s a general fear of conglomerates assembled out of multiple businesses, and historically that type has indeed performed badly. But in the SpaceX case, there are still clear logical links and mutual support between the businesses.

Raymond: In traditional valuation methods, you sum the parts and then apply a conglomerate discount. On Musk’s side there’s no discount; on Musk’s side one plus one equals five — every time you add, you also apply a higher premium. That’s how the market is valuing him this year, and Tesla the same. Whether they’ll still do it next year and the year after, discount or premium, nobody knows.

What does it reflect? That people feel Musk can simply get this done. There are many possibilities inside his company; pay $200 billion for this possibility, $200 billion for that one, and it adds up to over a trillion. Buffett takes one look and coughs up blood — what is this, there’s nothing here.

But I flipped through the first few pages of SpaceX’s prospectus, and there are several illustrations in there, and after looking at them my only reaction was: wow, beautiful. You feel you’re not buying an ordinary company.

Tiannan: What you’re buying is a ticket to Mars. There’s been an interesting change these past couple of years. Five or eight years ago, when Musk told the strong narrative of going to Mars, the capital markets didn’t necessarily buy it a hundred percent; over the last year or two, with AI continuously breaking through commercially across fields, more and more people have started to believe and accept that narrative. So when he paints a blueprint in the prospectus, people don’t find it all that outrageous.

Raymond: Then let’s break the valuation down concretely. There’s an analyst at Morgan Stanley, Adam Jonas, whose framework the media now cite fairly often, and his method is exactly sum-of-the-parts.

Part one, Starlink at $600 billion — a number so large that even its third decimal place could be a listed company. Starlink’s revenue is around $12 billion now, growing forty or fifty percent for two years, so he gives it a P/S in the thirties; Verizon over the same period is around seven or eight. Several times more expensive, because the growth rate is good, and because it’s unlike anyone else: Verizon has to fight AT&T, and who does Starlink fight? No opponent, so the upside is more obvious.

Part two is AI. This part is particularly interesting — he’s effectively saying I don’t know how to value it either, but the company folded xAI in at a $250 billion valuation, so I’ll use that number. That’s itself a mildly negative assessment: I don’t know whether you’re expensive or cheap, but you’ve already done the arithmetic yourself.

Tiannan: But a lot has changed at xAI this year.

Raymond: Right. xAI started as Twitter, then did the Grok large model, then had its own enormous supercomputing center, and now rents part of that compute to Anthropic, becoming a bit of a neocloud, a new cloud provider. The price didn’t change; the company did. So periodically you feel that $250 billion isn’t especially outrageous today — after all, 200,000 cards are running, and in America that’s a resource people are desperate for.

Tiannan: Where do those 200,000 cards sit within America’s total compute?

Raymond: The other suppliers are the three clouds, Amazon, Google and Microsoft, and below them smaller ones like CoreWeave and Nebius. There aren’t many in America who genuinely have 200,000 cards and can get them running. There are two hurdles: first you have to afford the cards, and second you have to have somewhere to plug them in, which means power. Power is extremely troublesome in America — you need a proper large facility, energized, connected to the grid with the local government permitting the connection, or your own gas generation; that whole set is an engineering project. So everyone is racing; Microsoft went and bought a nuclear plant outright because it genuinely fears running short of power.

A pile of cards sitting on the floor is worth nothing, nobody wants it in America; a pile of cards energized and running is extremely valuable. So I can’t say whether that $250 billion is high or low; I actually feel it has more substance behind it than a year ago.

Those two together are over $800 billion. Add space launch — everyone has seen Falcon 9 on Xiaohongshu and Douyin — and Starship is still being pushed forward. Add all of these up and you get thirteen or fourteen hundred billion, under 1.4 trillion. How does the rest get to two trillion? The research note has another passage: what an AI data center doing computation up in orbit is worth, and what it’s worth once it’s fully deployed; and the recently acquired Cursor, if in-house model development succeeds and can challenge the top three, what that’s worth. This whole series of options is granted to Musk. But I think the most important thing is still that Musk himself is worth money; that premium is worth a lot, and it’s what ultimately carries it past a trillion.

3. “If I build 5,000 Three Gorges Dams flying overhead”

Raymond: Musk is very good at creating positioning for himself. What he’s doing now is building data centers in space. China’s biggest problem building data centers is cards; power isn’t much of an issue. America is the reverse: cards aren’t restricted, power is severely restricted. So he says: I have cards, so why can’t I send the data center up as satellites? The orbit can be designed to get sunlight around the clock, so the energy problem is solved; and space is cold, so heat dissipation is easy too — running a data center in America means all that liquid cooling equipment, expensive and complex. Doesn’t that sound impressive?

Another example. There’s a line in the prospectus I lay awake thinking about. He has something like a performance wager with the board: if he can build a 100-terawatt compute center in space, he gets an enormous stock grant, possibly trillions of wealth in that single award, with the company’s market cap simultaneously reaching $7.5 trillion. I went and looked up what 100 terawatts means — the Three Gorges is 20 gigawatts, so 1 terawatt equals 50 Three Gorges, and 100 terawatts is 5,000 Three Gorges. What he’s effectively saying in the prospectus is: if I build 5,000 Three Gorges Dams flying overhead, you have to give me trillions.

Tiannan: Too many numbers; this humanities graduate has lost all feeling.

Raymond: Lost all feeling. So if you ask whether he can achieve it in his lifetime, I’m highly doubtful. But you also think that the only person on this planet who could possibly do it seems to be him. Measured against that, is one trillion or two trillion today even a thing? It isn’t.

He’s moved the focus entirely. You stop computing his P/S multiple or P/E, you stop caring about these earthly competitive dynamics and whether he can beat the big three.

Tiannan: None of it matters anymore.

Raymond: I’m on Mars; I’m not like you. It’s extremely clever marketing — I’m in another world, and my promises aren’t on the same dimension as yours. Even if he never manages it in his lifetime.

Tiannan: So do the big Wall Street institutions accept this narrative, this logic, this valuation? Or are there actually not that many institutions paying for the story?

Raymond: What I have now is word of mouth from Wall Street friends and may not count — the real vote is what price it opens and closes at on IPO day. The scattered information only tells you two things: first, everyone is extremely focused on this IPO, given a multi-trillion company is about to materialize; second, the doubt is amplified equally. When the company becomes two trillion, the doubting voices amplify to a two-trillion scale as well.

The doubts are a few. First, governance. Beyond super-voting shares he designed a series of thresholds to hold control more firmly, which indirectly damages minority shareholders’ ability to defend their rights.

Second, how you actually count the AI. xAI has now clearly fallen a whole tier behind, not by a little. Is it still a leading model? Probably not. So do you count it as supercompute, as a cloud, as what? And you bought Cursor, so can you get back? The large-model competition, in China as in America, is brutal, with each leading for two or three weeks. Models used to update quarterly or semi-annually; American models are now monthly — Anthropic’s 4.6, 4.7 and 4.8 all happened this year. And xAI has had founding-team departures and all of that. So the question mark is: what is your AI’s position? You’re certainly not leading, so what do you count as, and can you come back? You’re fourth; on what basis do you get a valuation?

Tiannan: Fourth shouldn’t really get a valuation.

Raymond: Fourth shouldn’t get a valuation. Right now I think even Gemini’s valuation deserves a question mark.

Third is Starship. Starship is far larger than Falcon 9, and its payload and the range of what it can deliver are far greater. There should be a line in the prospectus saying that in the second half of 2026 or in 2027 it can enter more definite commercial use, though it’s subject to testing and other factors. Many people will question whether Starship can be commercialized at scale quickly. And this one has a sequence — first you need Starship, Starship has to work well, Starship has to be reusable, Starship has to drive cost down at scale, below $200 a kilogram, before you even get to discuss building data centers in orbit, because a data center means sending far more mass. Then you understand that Starship is fundamentally a freight-hauling business, and if the freight cost isn’t low enough, you can’t send things up at scale.

Fourth, whether an orbital data center is possible at all. Can you have solar around the clock, avoid overheating around the clock, compute fast around the clock, transmit back to Earth around the clock? All question marks right now. Multiply those question marks together and it’s effectively impossible — 0.99 multiplied by itself enough times approaches zero.

And beyond that there are valuation doubts and market-window doubts: whether this market environment can carry a two-trillion company materializing out of nowhere. That has some difficulty.

4. “If you believe him you invest in him; it’s a single-point decision”

Tiannan: On Musk, what people care about is how much control he has in this company. From taking X private, he brought in some investors through share swaps along the way, those investors’ shares swapped into xAI, and this round swapped again into SpaceX, and a lot of people have lost the thread.

Raymond: You put that very well — someone invested in A at the start, A went into B, B into C, C into D, and the original shareholders may not know where they are. But IPO preparation is a long process; banks, auditors and lawyers come in and go through document after document, and how many shares the company has, what was issued along the way, employee holdings exercised or not, all gets mapped out.

The final result: Musk’s economic interest in this company is 40%, and his voting power is over 80%. Absolute control, dual-class structure.

Why design it that way — look at two sets of precedents. On the American side, within the Magnificent Seven only Google and Facebook have dual-class, and in both the founder charged out and said I want super-voting rights, I want 100% control of the company. In America, founders being voted out by the board happens fairly often; Uber is the example. With absolute voting power, one person decides. Chinese precedents are far more numerous; the China ADRs are almost uniformly dual-class with super-voting rights. The early ones like Sina and NetEase may not be, but everything after is.

Tiannan: Especially the cohort that listed after 2014 and 2015. Alibaba failed to list in Hong Kong the first time precisely because Hong Kong didn’t have that structural provision, and later Charles Li filled that gap and a lot changed afterward. If I remember correctly, Xiaomi was the first relatively large company to list in Hong Kong with dual-class, around 2018. The old Hong Kong market believed strongly in protecting retail investors, but hadn’t accounted for one reality: tech companies raise fast, and after five or six rounds the founder may be left with seven or eight percent — Jack Ma may have been left with six or seven.

Raymond: Musk didn’t have that protection at Tesla and was constrained at every turn in managing it. So his thinking was: my next listed company certainly won’t be like that, I’ll control it completely.

Tiannan: Do you think that’s good or bad for the company?

Raymond: For investors I think it’s good. The core reason you invest in this company is that you believe in Musk the person; if you don’t believe in him you wouldn’t invest at all. Suppose there were no dual-class; then you’d have to think one layer further: Musk wants to do this thing, will the board support him? Your investment decision would instead require second-guessing every decision-maker in the company, the board members, what the other shareholders think — that’s complex. Dual-class makes the judgement simple: if you believe him you invest in him. It’s a very single-point decision.

Of course, believing he can pull it off and believing he should be worth two trillion today are two different things. You can believe he can pull it off and still think he should be $500 billion, or should be $5 trillion. He’s separated those two decisions; you no longer have to worry about other considerations inside the company.

There’s another layer. Many people think the rich corporate governance of an older company is a protection; I don’t think that’s necessarily so. The real protection for investors and retail holders is a company with better prospects, faster growth and a clearer sense of what it’s doing. There are far too many cases of a board taking all the money.

Tiannan: China has plenty of examples of the reverse. In companies with no controlling person, every executive stakes out a private plot inside; there’s no boss in any real sense, and therefore no long-term strategy in any real sense. Real estate is more extreme: the listed entity and the off-balance-sheet debt can’t be reconciled, and the balance sheet you see may be on a completely different order from the true leverage; and by the time you have warlords everywhere, insider rent-seeking and gambling with company money all come out. We joke about it as a poor temple with rich monks.

Raymond: Whether a boss having complete control is good or bad is genuinely case by case. There are many companies whose bosses I don’t much trust, where I’d think it’s better for him to be constrained, better to have a board and shareholders supervising him. But Musk has enough credit in my mind, I believe in him enough, I’d give him greater tolerance, and I’m willing to think his having full control of this company is a good thing.

5. “This is a very busy summer”

Tiannan: You used to do US equity investment banking. What stages does a company go through in America from private to a US IPO, and what key milestones can outside investors observe?

Raymond: Take an example: ByteDance isn’t listed today, and you’ll see WeChat accounts writing this and writing that, all of which may be unreliable, and the company won’t come out and verify. At that stage there’s no first-hand authoritative information.

Which is why the prospectus moment is critical. SpaceX’s prospectus came out on May 11, a large legal document that hundreds of people have reviewed, where every word is a potential legal consequence. Before that, the boasting you did could sit in those videos; after that day every sentence you say is a public company speaking — say something wrong and cause a loss to my investment, and in US markets I can bring a class action, with lawyers who specialize in exactly that business.

Before that day you move in the dark: find bankers, find auditors, sort out a pile of financials, and Chinese companies also need particular regulatory opinions. After that day you move in the open: meet analysts, meet investors, roadshow through New York, London and Hong Kong, meet hundreds or thousands of institutional investors, buy my stock, buy my stock; if enough people want to buy, you price it, and having priced it you ring the bell, and after that you report every quarter. Many bosses will say listing is their first day; but for many companies listing is the last day — the best of the business may already be behind them.

Tiannan: So how much pressure comes from these companies crowding together in the second half of this year?

Raymond: This year SpaceX is first, then Anthropic second, and OpenAI should be very soon, possibly all within the next two or three months. If you’re an investor at a large American tech fund, you’re busy right now; you may not get a summer holiday.

Because once these three trillion-dollar companies list they’ll be added to the indices quickly. Do you buy or not? You’re a fund manager: against the benchmark do you overweight or underweight, do you buy fully, not buy, or hold judgement for one period? Once it enters the index, who gets squeezed out? Do you sell, sell whom, sell how much, how do you rebalance? If it rises after listing everyone’s happy; if it falls, how do you defend? That whole series of questions is enormous. For American fund managers, this is a very busy summer.

Tiannan: A-share index rules have clear windows, some semi-annual, some quarterly. Does Nasdaq include on day one?

Raymond: Very good question. Nasdaq has actually changed rules for large companies so they can enter indices faster, possibly within a month or two. Think about a trillion-dollar company: the moment it enters it goes straight into the index’s top ten, and several companies certainly get squeezed out behind it.

Tiannan: A lot of older investors will think back to the 2000 dot-com bubble: first the enormous volume of M&A, second some companies at absurdly high market caps, with the AOL–Time Warner merger as the marker event, and a lot of people carry PTSD about it. A-shares have even more examples; the wave of super-large central-SOE mergers in 2014 and 2015 was the same. Will these mega IPOs cause problems for US market liquidity or sentiment?

Raymond: Let me split your question in two. First, whether SpaceX, OpenAI and this sequence of listings are a marker of a US market top — US markets are indeed high, all investors are saying it’s too high, too high, and yet their bodies are honest and they keep buying. Second, from a liquidity standpoint, whether it can technically absorb it.

On the first. I don’t think a mega IPO is necessarily a precursor to the market cracking. Everyone cites 2000, because that was genuinely a big year. But pull up the IPO leaderboard: Saudi Aramco listed in 2019, do you remember? You may not. Alibaba listed in 2014, the world’s largest IPO at the time, over $20 billion. Agricultural Bank of China and AIA were both 2010, that batch after the financial crisis. Further back, ICBC listed in 2006 at over $20 billion, an extraordinary size for the time. Would you call 2006 a bubble? Looking back, it doesn’t seem so. So there’s no direct link; it isn’t that raising a lot of money means they cleverly escaped the top.

But conversely, it does confirm something: SpaceX’s achievements in space over twenty years, and everyone’s achievements in AI over two or three years, may have been recognized once, provisionally, at this point in time, and so this company exists. Does that necessarily imply a bubble? I don’t think so. But is it possible that after listing it performs badly and breaks issue, triggering a chain reaction, and at a moment when sentiment is relatively fragile gives it a push down? That’s possible.

On the second question, look at the data. US markets trade roughly $500 billion to $1 trillion a day, and these three raising $75 billion combined — call it a lot, it is somewhat; call it small, it seems absorbable, especially since it isn’t all on the same day. And this isn’t news, it isn’t zero-expectation; fund managers who want to buy have the money ready, those who should sell have sold, and retail is the same. It isn’t that money gets suddenly pulled from elsewhere on listing day.

Ultimately nobody knows how it performs; that’s the market’s decision, and there’s geopolitics, oil prices, the Fed and a pile of other signals. But I think the next two or three quarters of results, and how people react to those results, may matter more. If it performs badly, we’re discussing a two-trillion company falling, and the drag on the market then is greater.

And because of SpaceX’s listing this year, several dozen companies up and down its supply chain are listing alongside, and the commercial space theme in A-shares had a hot run too. If SpaceX performs badly, none of the others can possibly do well.

Tiannan: Fortunately A-shares have already fallen back. I remember attending a private-market conference at the end of 2020 where the EV and autonomous driving companies were all euphoric, purely because Tesla’s share price was good — high public valuations make private fundraising easier, secondaries sell better, and raising from LPs is easier too. When a benchmark leader IPOs and public valuations are high, you can’t quite call it every dog and chicken ascending, but every company along that chain benefits. Once SpaceX actually lists, it will help the valuations of many companies along the supply chain, including Chinese private-market companies benchmarked against it.

Raymond: Very obviously. 2014 was exactly that kind of watershed — JD listed in April and Alibaba in September, a defining moment. Before that, people’s impression of Chinese tech stocks wasn’t strong, and what I personally worked on were companies nobody has heard of, PPS and Shanda Literature types. After that year, the world put its focus on Chinese tech, a large batch of VCs was suddenly founded, raising money in America was very easy, and everyone was saying I’m all in on China Tech — the past decade of Chinese tech and internet development and its talent pipeline may all trace back to that year. I believe that looking back three years from now, this year may also be a very important one: a big year for space, and a big year for AI. Where it leads afterward, nobody knows.

Tiannan: Suppose SpaceX has already listed. What developments in fundamentals or information flow would affect the broad judgement on this company’s share price and valuation?

Raymond: My first reaction is still to watch the Fed. Many tech companies are now at a fairly extreme place, and at that point fundamentals of course matter, but the Fed’s draining or adding liquidity may matter more. After all, there’s a new Fed chair, and Kevin Warsh is a distinctive person — what he says to Trump may be one thing, what he says at a Congressional hearing another, and his first press conference maybe a third. How he treats market liquidity going forward is ultimately a question of the cost of money. And his statements are in turn constrained by the US-Iran conflict, oil prices, CPI and non-farm payrolls — whether employment has been affected by AI. Warsh has said repeatedly that he thinks America’s employment environment is deflationary, because AI makes society more deflationary, so cutting rates is fine. Whether AI is cutting enough jobs is a factor he’ll be watching.

Second, watch whether the other AI companies do well; that’s the overall effect of market sentiment. If SpaceX, Anthropic and OpenAI all list well, with a clear wealth effect, plus impeccable results from the harder hardware companies, they’ll pull the index up on their own. But if some company performs badly — say Nvidia comes out and misses, and semis have a problem; or an AI end customer like Anthropic says this quarter fell short — then the dominoes start falling.

Conversely, I think investors and Musk’s fans have the highest tolerance for Starship launches; it isn’t the kind of stock that falls on a failed launch. Hard to say for certain, given there’s no precedent for this kind of listing. But my sense is their tolerance on AI will be lower.

6. “What replaces you is an invoice from Anthropic”

Tiannan: Are the top American companies still maintaining the overall scale and pace of borrowing to fund AI hardware capex? I saw Google apparently did another $80 billion recently. Is this process still accelerating?

Raymond: Still accelerating. Q1 2026 is very clear: Anthropic single-handedly set the whole of America alight. The logic chain is this: Anthropic started doing coding, doing the token business, enterprises found it helped employee productivity, and they started handing out unlimited tokens, giving and giving. Uber’s CFO said recently that they used up their full-year IT budget in four months — four months, and the whole year’s token money was gone, because employees genuinely used it well. I’m a heavy, heavy Anthropic user myself.

That immediately opened up everyone’s ceiling on the B2B market and on token demand from white-collar work. Once it opened, people found revenue forecasts accelerating past expectations quarter after quarter, with extremely, extremely fast revenue growth. So people think: if you’ve grown this much this year, then what about next year, and the year after? It hasn’t done many things yet; some models it has produced have been validated but not released. The growth expectations are extremely strong.

Those expectations for token usage convert into buying cards, buying data centers, buying memory, buying all kinds of components, buying gas turbines — because you need generation. Everything is rising, and the most fundamental demand is that people want to use tokens.

Tiannan: So in the end everyone focuses on one metric: whether Anthropic’s ARR each month or quarter is sustainable and continues to beat, and nothing else matters.

Raymond: It’s an extremely important metric; I feel there’s nothing more important right now. Talking to investors, attention to OpenAI’s ARR seems far lower — I’m being a bit absolute saying that. But look at how fast Codex has risen these past two or three months; that story isn’t necessarily settled.

Tiannan: Each leads for three to five months. I’ve also seen another sign lately: scientists who originally worked in drug discovery or other fields are increasingly moving bodily into AI companies. A superb brain plus unlimited tokens, cutting into one concrete setting, may in three to five months find the next market like coding — and once a market appears, the size gets rewritten again.

Raymond: I have to show you that chart. There’s a section in SpaceX’s prospectus called the industry section, and I got a different quantity of information out of it. Musk was certainly involved in a section like that — the bankers will ask: Mr Musk, what market are we in? You have to define which market you’re in, and then define how large that market can grow.

Let me read a few numbers. SpaceX says its total addressable market is $30 trillion. That number is too large; people hearing it may already have lost all feeling. But $30 trillion is the largest market size any company has ever written into a prospectus in history. He broke it down: the space portion $370 billion, the Starlink portion $1.6 trillion. Subtract that $370 billion and that $1.6 trillion from $30 trillion, and the remaining $26.5 trillion is all AI. He’s telling you, in a very particular way, that the AI part is the bigger sea of stars. And within that $26 trillion, $23 trillion comes from enterprise applications.

Tiannan: That’s an important observation. At AI’s current stage, what people focus on is more compute and the hardware side, and on the application side many investors haven’t formed a consensus about market size. China’s last generation of super-large internet companies were all consumer, and applying that logic people default to assuming the AI companies standing at the front of the next wave will also monetize around consumer settings. Musk’s statement stretches many Chinese investors’ imagination — Chinese people find it hard to imagine a B2B company with a market that large.

Raymond: China and America have to be discussed separately. First, in China B2B and B2G are often inseparable; serving enterprises and serving the state and SOEs are mixed together. In America, B2B is B2B, with a very small state-owned portion; apart from Palantir having some defense contracts, most of it isn’t.

Second, America’s B2B market is friendlier overall and enterprises are more willing to pay, because labor is expensive — pay a bit more to save trouble, and I’m willing. That friendliness produced the last decade’s SaaS supercycle. Chinese SaaS basically never survived; every year was the first year of SaaS, and then it wasn’t. Chinese people do consumer extremely well, and Douyin ultimately produced TikTok. America clearly invests more in B2B, and Anthropic and OpenAI are the example: Anthropic chose enterprise from the start, so it cares about security, about data privacy, about the things enterprises care about, and doesn’t do video; OpenAI wants consumer, wants a billion-user application, does Sora 2, then shuts it down.

I can’t predict the future, but as of today, B2B’s willingness and ability to pay are far stronger than consumer’s. Which is why you find Anthropic’s revenue rising extremely fast, and rising without limit. B2B users paying for intelligence may be a very important trend ahead.

Think about it: you’re a small boss with 30 people in the company, and the more you engage with AI the more easily a thought surfaces: whose work in this company could be replaced by AI? Every boss will think that. So today, whether you’re an employee or a boss, you can’t avoid it. As an employee you can prepare: either you use AI well, or that day comes eventually. If what you do every day is highly repetitive labor, your boss is already doing the arithmetic. What replaces you is an invoice from Anthropic.

Tiannan: That invoice reconciles with Musk’s $23 trillion. An agent can work 24 hours, never sleeping, and doesn’t need social insurance contributions; in high-cost regions the substitution effect is fairly obvious. I see many startups, American and Chinese, facing this challenge: I’m building some agents now, or building some hammer-type SaaS tools, and I’ll also wonder whether one day an Anthropic Markdown file and a string of skills disrupts my entire company’s business — the probability of which is very high. And the cost calculus is plain: as long as it’s cheaper than a person, compute might be used.

The second difference is the market. In America, coding is the first validated B2B market; in China, the hundred-billion-yuan market that suddenly grew after AI got hot is actually AI short dramas, which broke ten billion in a few months, and daily consumption on Douyin is now over a hundred million. This year it may surpass the film market, next year it may surpass livestream e-commerce, and the year after it may be a hundred billion. Then look at where the money flows: the script may come from Fanqie Novel, the model may be ByteDance’s own video model, the ad platform is Douyin, and the hosting platform is Hongguo short drama — every fee in the whole process is earned by ByteDance alone. In China it may instead be consumer that explodes first. Everyone loves doing AI equally; the soil differs, and the application-side business models that get explored differ too.

Raymond: But look, even Doubao is starting to charge, so maybe China will also produce vendors breaking out on the B2B side. Chinese SaaS doing badly in the last era had its own historical reasons, and these past few years have entered a new stage, so you can’t categorically say no B2B business in China has a chance. I’m personally very willing to pay for this; the thing clearly improves my productivity. As long as the effect is visible and it helps me get work done, people are willing to pay — it is, after all, a productivity tool.

Tiannan: A final piece of gossip. In the xAI period, did any Chinese institutions hold private-market shares in this company? How did the valuation move?

Raymond: SpaceX still felt like a $200 billion company last year, and climbing to a trillion happened early this year — that xAI target feels like it was engineered upward, not because the business suddenly multiplied. Before there was AI there was only Starlink, and Starlink did fairly well, with a valuation then of two or three hundred billion. Over the past 12 months it’s up four or five times; that part is genuinely real. Further back its valuation went up and down; sometimes fast Starship progress would hype the valuation for a stretch, but it is after all an unlisted private company, and most of the time it moved up and down with the market, with macro and with US equities. The real move is the past year.

Back to your question. Many Chinese individuals and institutions, if they’ve invested in offshore entities — for instance if many internet company bosses have invested in Silicon Valley companies — then SpaceX is certainly an unavoidable name. I believe many have looked at it, and quite a few have invested. xAI, far fewer.

If you're working on this too — or you think we've got it wrong — write to us at [email protected]; if you'd rather not write, just leave your email below.