OpenAI's Leaked Financials, SpaceX's Valuation, and the AI Bubble

Ed Zitron

So last year OpenAI spent about $34 billion to make about $13.07 billion, and had about $22 billion in cash at the end of the year. They lost about $21 billion from operations. That $38–39 billion net-loss figure is what I'd describe as GAAP voodoo — there's some very strange stuff going on in the balance sheet, and I'll be going into it in future episodes. They converted from a non-profit to a for-profit, and remained profitless last year regardless. Their net loss is quite strange — it comes to roughly $38.5 billion — but the number I keep coming back to is that they lost $21 billion from operations, and they spent astronomical amounts on sales and marketing. They spent $7.5 billion on cost of revenue, $19.18 billion on R&D, and then that sales-and-marketing cost — $5.73 billion — is the one I really like to hammer on. There's a possibility some of that gets passed through to Microsoft under the revenue-share arrangement; I genuinely don't know, I'm just guessing. But it's an extreme cost regardless.

Another thing to note: $867 million, about 6.6% of OpenAI's revenue last year, came from SoftBank. I don't know exactly what that's for — they have a product called Crystal Intelligence, spelled the same way as the gemstone, I'm not kidding. It was announced in February 2025 and took until the end of the year to hear anything more about it, yet they still paid nearly a billion dollars. That's a huge amount of revenue from a single partner that doesn't appear to have driven their cost of goods sold up at all, and it suggests OpenAI's growth isn't quite as fast as we've been led to believe.

Ed Zitron

I didn't speak to the person familiar with the matter, and I don't know who they are — I think that's nonsense, frankly. You spend $34 billion to make $13.07 billion; you didn't just lose $8 billion. When I read that comment I had a good chuckle, and I knew all the boosters would treat it as the proof they'd been waiting for. The numbers to look at are the actual costs and the actual revenues. These companies — OpenAI especially — have lived high on the hog, spreading confusion around their numbers. Focus on how much they're burning: $1.57 billion just on general and administrative costs, on people alone, and I don't even think that captures everyone, because I don't have the underlying definitions from the sheets. This is a company with spiralling costs, dramatically beyond what we thought. They put a lot of it under research and development because — this is my own belief — they want people to think it's a temporary situation that will go away once there's been enough R&D. It won't. That number is going up; every cost is increasing. It's a genuinely frightening picture of what the inside of these companies looks like.

Consider that OpenAI has already said, as part of the Musk trial, that they'll spend $50 billion on compute in 2026 alone. Their losses next year are going to be astronomical. I don't have privileged knowledge of the 2026 numbers, but based on these, they could be burning $80–90 billion. It's genuinely horrifying. I'm glad I got this story out before the S-1 is filed, because people need to see what the inside of this company looks like before it gets a chance to be massaged.

Ed Zitron

If OpenAI had SpaceX's losses, they'd be thrilled. Then again, SpaceX doesn't have OpenAI's revenue, so who knows. I think the reason SpaceX has flown the way it has comes down to the Musk reality-distortion field, and his ability to play Goldman and JPMorgan off each other. Everyone involved in pumping this — "AI revenue will increase 300x in four years" — should be ashamed of themselves. Putting that aside, these are fundamentally different companies. Even with its AI cash-burn problem, SpaceX still has Starlink, it still has the rockets, and it has the social network too — though I wouldn't call that a plus. It still has a functioning business. OpenAI has ChatGPT and the API — that's it. They dance around it: "oh, we've got Jony Ive's device," "oh, we're going to do an everything app" — but they're dancing around the fact that they have one product, with a few offshoots, and no real way to invest more money to make more money.

You mentioned profitability. I don't think it's possible — their costs are so severe that I don't think profitability is achievable for any AI lab. I think Anthropic is going to look similarly bad. The market seems to believe, for some reason, that things won't look this bad. I've been talking to people all day, and everyone's saying, "wow, I didn't know it would be that bad." Of course it was that bad — do people think I was kidding?

I don't think OpenAI and Anthropic have the nuance, or the aggressive bully mentality, that Musk brought to the SpaceX IPO — and I deeply dislike Musk, to be clear. In the S-1 they'll also have to lay out a growth trajectory, what they could grow into, and I think that's going to be difficult to sell. They could pull it off, theoretically — given how SpaceX has done, I'd say they have a higher chance than none — but SpaceX's numbers were bad, and these numbers are terrifying. I've been watching the cope Olympics on Twitter: "well, actually, it's fine, you saw the quote that said $8 billion, they didn't really spend $34 billion, nothing bad happened." It reminds me of Comical Ali insisting everything was fine. It's all very silly, because this is a company that spent $34 billion to make $13.07 billion, and $867 million of that revenue looks like it came from SoftBank. That is not a stable or thriving business — it's a business consuming capital at an alarming rate with no sign of stopping. They just raised $122 billion at the start of this year, and I don't think that's because they're about to become profitable.

When this company goes public — and it looks like it will, they've filed plans to IPO in the fall — the valuation will hinge almost entirely on the CEO's ability to tell a compelling story that enough people buy into. Elon is someone you could bet on doing that; he has a proven track record of selling markets on what his companies will achieve. I'm not sure Sam Altman has that same ability to catch the imagination of investors, because the numbers clearly aren't going to do it for him.

Ed Zitron

Thanks for having me.

Nicholas Owens

Thanks for having me. The headline number is the weighted average of three scenarios. We valued SpaceX in three parts: the rocket business, Starlink, and the AI piece, which is itself a set of moving parts. The rocket-and-Starlink piece is the more straightforward, more mature part of the business — in most of our scenarios it's worth around $611 billion, roughly $40 a share in enterprise-value terms.

The AI piece is riding on some unproven outcomes. I'm not talking about the science or engineering behind data centres in space — though there's debate about that too — but the financial benefit of having a data centre in space: is there an operating-cost advantage over a terrestrial data centre? We modelled the business primarily as an infrastructure play — renting out computing capacity, the way they've started doing with Anthropic and Google. If Grok takes off, we'd be roughly indifferent between them renting out that capacity or using it for Grok jobs and monetising it another way; we assume there's some market rate for it either way.

The difference between our valuation and the market price comes down to how probable you think it is that the Starship rocket becomes highly reusable — in hours or days — and that data centres in space turn out to be a compelling bargain on operating cost versus terrestrial ones. Both might be true, but we think there's only about a 7% probability that they're both true at once, and that's what would justify a valuation much closer to the market price. The market seems to be pricing in that both will be true for certain. We don't.

Nicholas Owens

There was quite a good internal discussion about this a couple of weeks ago. I'll say this: the SpaceX business that I, as the aerospace-and-defence analyst, was planning to cover until February has almost all the characteristics of a wide-moat business — a very prodigious cost advantage, both through R&D and through economies of scale. They've simply done it more than anyone else, and the satellite business and the rocket-launch business reinforce each other as they march down the cost curve. They're a decade ahead of anyone else, and should stay there as Starship's bigger cargo bay keeps lowering the cost per kilogram launched — you're dividing by a bigger denominator, and that's remarkable.

The company gets a narrow-moat rating from us because of the AI piece, which is indeterminate at best in terms of what its moat would be — and they're investing very aggressively in it. You're taking proven returns on capital from Starlink and the rockets and writing a $60 billion cheque to acquire Cursor, making similarly sized investments to build data centres in space. There could be moats in AI, but we don't see evidence of them here — Grok isn't one of the leading models. There's a pathway, in the moonshot scenario and even in what we call the minimum-viable-product scenario, where they extend some of that cost advantage into data centres in space. If they can, that would be moaty. But it's too soon to say, frankly.

Nicholas Owens

Yes — and I think framing this as an AI business is a rational way to tap into investor appetite for AI.

Nicholas Owens

I think so. I thought about it in terms of supply and demand — as you say, a small float, plus structural demand from passive funds, ETFs, and others that track a handful of indexes. In our moonshot scenario, if you dial the probabilities to 100%, the share price would be about $169. There are hundreds of billions of dollars of potential revenue in a scenario where they become a major player in gigawatts of compute in space. Giving them the benefit of the doubt on all of that gets you to $169. Above that, investors are pricing in options like a city on Mars, which we don't ascribe positive value to — it's a wash. If you're paying more for these shares, you're paying for the option that those things work out, and we think they're more likely than not to fail.

Longer term — starting around the earnings announcement I expect at the end of July or early August — big chunks of shares will come free from lockups on insiders. There'll be a Nasdaq rebalance before then, but then a decent chunk of insider stock comes online, and I think that will be the next real test of supply and demand.

Nicholas Owens

I wouldn't try to predict the stock chart, but it's tens of billions of dollars — more than the IPO float — coming online in that first big chunk after Q2 earnings. The indexes that are buying will do a float adjustment to their weighting as the float increases, so their demand will scale up gradually, but I don't think it will offset that chunk coming online. And think about who's selling: these are investors who've owned the stock privately for more than a decade in some cases. They don't care whether it's $160, $170 or $200 — their cost basis is nothing. The price is set by the marginal seller.

Nicholas Owens

It's mostly a wash. We actually lowered our probability-weighted fair value slightly today, to $62 a share from $63. The idea behind Cursor, as I understand it, is to bring on people who can make Grok better and more useful for coding applications — giving them the benefit of the doubt that this works. It was originally structured as an option to buy if that proved out; they've essentially jumped to the finish line. So the purchase price, the outlay, and the dilution from $60 billion of new equity is offset in our forecast by an adjustment for greater enterprise revenue from that audience.

Nicholas Owens

Thank you.