In this issue:
- The Rocketship—SpaceX is finally going public. The meta-bet is that if you get a kick out of backing Elon, the low cost of capital you provide him gives him a shot at yet another monopolistic business.
- Unit Economics—The cheapest good AI you can use will be funded by ads, but the cheapest AI you can pay for is a different matter.
- Rearrangements—When general-purpose technologies get deployed, they revalue everything else.
- Vibeslop—We're producing more code than we're reading. Good luck!
- Mindshare—Audio's uniquely monopolistic traits.
- Consumption Equality—Waymo as an egalitarian technology.
Talk to this post with Read.Haus.
The Rocketship
At most times in the last few years, if you're reading some document that's full of Elon Musk quotes, pitching some investment, there was a very good chance you were about to lose a bunch of crypto. Who other than Elon Musk could break the rules and produce the SpaceX S-1, which is full of Musk quotes but which also details a real business, or, really, a few tangentially-related businesses? From the beginning, you know you're in for a weird offering document; it's traditional for companies to have a mission statement broad enough that there's no way they can achieve it in any realistic timeframe, which means they'll never have to update it and can keep adding mission-adjacent projects. Lots of things could contribute to getting a computer on every desktop (though at the time Microsoft came up with that mission statement, "popularize skinny jeans so people take their phones out of their pockets when they sit down" probably wasn't on the list). SpaceX n-ups this for some too-large-to-estimate value of N, with a mission "to build the systems and technologies necessary to make life multiplanetary, to understand the true nature of the universe, and to extend the light of consciousness to the stars." But one of the steps between now and that goal is to produce positive free cash flow in large enough amounts to convince future investors to either part with additional money or accept that the company will be reinvesting it for a while yet.
If they can, companies like to go public with two stories. The first is: the businesses we start mature into predictable, profitable ones that will be able to support a buyback. The second is: we're raising capital, rather than returning it, because we have another such business.
The first hard problem for understanding SpaceX is figuring out which is which. They have a rocket business, which they break into Connectivity (launching Starlink satellites into space and selling subscription access to them) and Space (putting other people's stuff into orbit). And they have an AI business, which includes a social network. It's a messy situation, though probably less messy than what would ensue if xAI and SpaceX were independent companies, SpaceX could easily raise cash, and xAI needed it. Two different cases of iffy corporate governance—Musk running multiple independent companies at once, and opaquely sharing resources between them—have roughly canceled out into one oddly-diversified business that can plausibly use a cash-cow side to fund the rest.
If you look backwards, the mature part of the business is the one involving rockets: Space and Connectivity together produced $15.5bn in revenue in 2025, and their adjusted EBITDA was $7.8bn. It's a serious business. And it makes sense to lump these two together: the Space business is, at least so far, a smaller, lower-margin, and less stable business than Starlink.[1] On the other hand, you don't get Starlink without getting to space first. SpaceX takes pains to parse things carefully when they talk about the history of their businesses:
We have a stellar track record of capital allocation and value creation in Space and Connectivity. Since SpaceX’s founding in 2002, we have raised over $9 billion of equity capital to fund the development and growth of these two business segments.
(Emphasis added.)
xAI's trailing financials do not demonstrate such a record; they had revenue of $3.1bn last year, negative adjusted EBITDA of $1.2bn, and a $6.4bn operating loss (most of that gap is the depreciation).
But if you look ahead, you'll actually see a completely different financial picture: SpaceX has signed a $15bn/year deal to sell Anthropic compute, which is ramping up this month and runs through 2029.[2] So, just because of the structure of that deal, they'll be able to print some automatic, high-margin growth for their first few quarters as a public company, before hitting a steadier state.[3] It will be a close race between rapid compounding and step-function growth, but it's possible that SpaceX will have a quarter or two where it's mostly a neocloud by revenue, with a nice sideline in rockets and communications satellites.
If there's a coherent way to describe the entire SpaceX company, highlighting genuine synergies between the business of hurling heavy objects into space and the business of maximizing the ad revenue from people asking Grok if the moon landing really happened, it's this: SpaceX is uniquely good at knocking down (or obviating) barriers to big infrastructure projects.[4]They've reduced the cost of putting a kilogram into orbit by 92%, and with Starship they're aiming for 99%+.[5] And, terrestrially:
We brought the first cluster of COLOSSUS online in 122 days, repurposing the shell of an existing factory, and the first cluster of COLOSSUS II online even faster in 91 days. As an illustrative comparison, an industry benchmark to bring online a 100 megawatt greenfield data center is approximately two years.
This is a big advantage! Anyone who has ever raised capital and reported an IRR knows that there's an agonizing gap between when you get money and when it turns into something that produces returns; for institutional investors, this problem has led to a whole ecosystem of financial products basically designed to distort the numbers.[6] In SpaceX's case, there's another interesting possibility: if they really do go public with a trillion dollar-plus valuation, and they can support it for a while, they'll have the lowest cost of capital for anyone investing in frontier technology at that scale aside from certain national governments. At which point the relevant question is over who has the biggest comparative disadvantage in capital allocation: the US government, with all of its weird processes and sensitivity to the needs of various interest groups, or Elon Musk, with his weird processes and tendency to get distracted and found new companies.
It's just unavoidable that this company is a bet on Elon Musk's capital allocation skills. In fact, if they raise $75bn, then if you count up the equity and debt raised by SpaceX and xAI, it's still true that about 60% of the money Musk will be able to deploy through SpaceX will come from this IPO, and the rest from two decades of capital raises by other means. There's a conservation of Elon Musk weirdness at work: my view of Elon Musk is that he's pretty technical, great at fundraising, incredible at recruiting talent, even better at motivating that talent to live on remote south pacific atolls or in literal ghost towns, and has a high risk tolerance. He's also vain, wildly underestimates the difficulty of some problems (running a social network, reducing Federal spending, hitting deadlines), but has sorted himself into domains that maximize the payoff from these traits. Other people disagree, but given that Musk has founded more than one of the most valuable companies in the US, any time you downgrade one skill, you have to mentally upgrade another one. So, if he's actually nontechnical, he must be that much better at recruiting technical talent; if he doesn't have an eye for talent, he must be really good at convincing investors to let him take another swing, etc. Eventually you can reach the point where the most coherent Theory of Elon is that the universe is a simulation, he's the player character, and that he keeps reloading old save files when the random number generator gives him a bad result. But, probably, he's a very smart guy who thinks he's somewhat smarter and bets accordingly, and that he hasn't gone through the last of his massive wealth drawdowns just yet.
Should you trust him to go after a total addressable market of $28.5tr? Should you bet that there really will be synergies between xAI and SpaceX, by way of orbiting datacenters? The good news is that you don't have to: part of the Elon Algorithm is setting insanely ambitious goals and pushing people to their limits to achieve them, even if Musk periodically takes a break for a ketamine vision quest or to pick a president.[7] Musk is good at setting these plans, but he's also really good at getting out of the way when people execute the first few steps for these grand ambitions. It's very hard for SpaceX to support its proposed $1.25tr valuation; even taking some aggressive growth assumptions, that's probably about 25x run-rate revenue for a company with some serious fixed costs and unavoidable margin expenses. (Even if there's no marginal cost for a Starlink satellite, the launch costs get capitalized; you'll be depreciating the satellites, the expenses, even the fuel used, over many years.) What you're actually underwriting is something very meta: will investors keep being happy to back Musk's various ventures? Will they treat SpaceX as his main focus, as he tends to—SpaceX seems to be the senior claimant on Musk's time and money, both of which are pretty valuable. That's really all you're betting on: there will probably be some point in the future where the best guess about the return on some space-related project is, say, 7%, and if SpaceX's cost of capital is 6% it'll happen, while at 8% it won't. Somehow, achieving outlier success in reducing the cost of space travel and reducing the timeline for big construction projects creates an investment opportunity that's mostly a bet on future investor relations.
Interestingly enough, while it's lumpy, its economics are more predictable than quarter-to-quarter variance implies, since so much of that revenue is tied to government contracts. ↩︎
It’s worth noting that either Anthropic or SpaceX could terminate this agreement at any point along the way with 90 days notice. So it might be prudent to impute a healthy discount rate to this revenue. ↩︎
The Anthropic deal means that one of the only constants in AI is that, whether it's 2016, 2021, or 2026, Andrej Karpathy will be doing research using GPUs that Elon Musk paid for. ↩︎
One of the big barriers they are hoping to obviate with their orbital data center dream is power lead times (whether that’s gas turbine blades or interconnect queues). ↩︎
Starship V3 can launch 100 metric tons into orbit per launch and SpaceX believes that Starship will be capable of launching up to 200 metric tons at some point, possibly as soon as V4. For reference, this training cluster contains 56 NVL72 GB300s racks. Each one of those racks weights ~1.5 metric tons. So all of the compute required for this cluster could theoretically go up to orbit in a single Starship launch, with 15 tons to spare! ↩︎
One popular product is a NAV loan: an investor borrows against their existing portfolio to make an investment, and calls capital from its limited partners later on in order to pay back the loan. The effect of this is that the fund starts accruing returns on day one, but the clock for investors starts when they actually send their funds. So a PE firm that makes 2.9x after a six-year holding period, but uses a NAV loan to make that investment six months before calling capital, flips from getting an IRR slightly worse than the S&P to a point or so better. Those NAV loans tend to be pretty safe for banks, since they're doubly-collateralized by the assets of the existing portfolio and acquired companies, and by commitments from limited partners. The net effect of all of this is that future LPs invest a little more, and existing LPs risk a little more. ↩︎
Other aspects of the Elon Algorithm are well-defined in the S-1 as “The Algorithm”. It has 5 steps: 1) make requirements less dumb, 2) delete unnecessary processes or parts (embracing the principle that the best part is no part), 3) optimize the necessary processes or parts, 4) accelerate cycle time, and 5) automate only proven processes after the first four steps are completed. Elon companies are incredible bottleneck identifiers and destroyers. And one obvious reason for that is they are simply laser-focused on doing so! ↩︎
You're on the free list for The Diff. Last week, paying subscribers read about the noisy AI productivity picture ($), a new kind of corporate escape velocity ($), and why pod shops are particularly well-positioned to figure out when different AI use cases suddenly work ($). Upgrade today for full access.
Diff Jobs
Companies in the Diff network are actively looking for talent. See a sampling of current open roles below:
- Series A startup building multi-agent simulations to predict the behavior of hard to sample human populations is looking for a founding recruiter who’s able to attract and close the best research and engineering talent in the world. Experience building high-quality teams as a former founder, VC, or operator a plus. No formal experience in a “recruiting” function required. If you have experience communicating and persuading smart, disagreeable counterparties of your vision, this is for you. (NYC)
- Well-funded, frontier AI neolab working on video pretraining and computer action models as the path to general intelligence is looking for researchers who are excited about creating machines that learn from experience, not text. Ideally you have zero-to-one pre-training experience and/or are a high-slope generalist who’s frustrated that the big labs aren't doing this. (SF)
- A Fortune 500 cybersecurity company with decades of proprietary security data is running an internal incubation with a pre-seed startup mentality and a mandate to build something new in AI. They are looking for a founding product engineer who can ship fast, an engineer with a security background who’d be excited to contribute to OpenClaw’s security efforts, and a generalist (ex-banking/consulting/PE background preferred) who wants to wear a bunch of different hats. Comp is FAANG+ and cash heavy. If you want to build something new in AI, but also need runway, this is for you. (SF/Peninsula)
- High-growth startup building dev tools that help highly technical organizations autonomously test and debug complex codebases is looking for senior product managers who enjoy defining developer-facing APIs and abstractions. Experience with fuzzing or property-based testing a plus! (London, D.C.)
- A leading AI transformation & PE investment firm (think private equity meets Palantir) that’s been focused on investing in and transforming businesses with AI long before ChatGPT (100+ successful portfolio company AI transformations since 2019) is hiring experienced forward deployed AI engineers to design, implement, test, and maintain cutting edge AI products that solve complex problems in a variety of sector areas. If you have 3+ years of experience across the development lifecycle and enjoy working with clients to solve concrete problems please reach out. Experience managing engineering teams is a plus.
Even if you don't see an exact match for your skills and interests right now, we're happy to talk early so we can let you know if a good opportunity comes up.
If you’re at a company that's looking for talent, we should talk! Diff Jobs works with companies across fintech, hard tech, consumer software, enterprise software, and other areas—any company where finding unusually effective people is a top priority.
And: we're now actively deploying capital into early-stage companies through Anomaly. Our focus is on defense, logistics, robotics, and energy. If you'd like to chat, please reach out.
Elsewhere
Unit Economics
One of the gaps in how companies think about temporarily losing money is that some companies are more willing to tolerate high upfront costs in order to scale, and others tolerate bad unit economics—either because they expect unit costs to continuously drop as they grow, or because they expect their competitors. There are some fun examples of the latter—San Francisco in the mid-2010s, or mattress companies a few years later ($, WSJ). But AI companies have clearly mastered this. Their basic attitude seems to be that no new launch should have a positive gross margin, but that every worthwhile model will eventually pay its own way. DeepSeek is making a significant bet on this, by indefinitely extending what was originally an introductory 75%-off discount. As AI gets used more, it will get increasingly hard to identify the bottom of the market: for discrete tasks people do, it's already an order of magnitude cheaper than the alternative. Sometimes, the optimal move is to cut prices fast enough that the price is too cheap to pay attention to, at which point the benefit of all future cost reductions accrues to whoever's selling tokens.
Rearrangements
Last year, The Diff wrote about how new general-purpose technologies revalue existing resources. Natural gas, whey, gasoline, and Reddit comments all went from negative-median-value annoyances to useful feedstock for value-added processes. One place where this can show up is a spike in M&A ($, FT), particularly when there's an industry that gets a positive demand shock and there's a debate about how durable it is. A one-point increase in long-term growth rates means a one-point drop in equilibrium earnings yield, and for a stable industry like utilities, that can lead to a big shift in valuations—and can divide the industry into bidders who believe in the growth story and sellers who don't.
Vibeslop
The makers of Pi, the harness underlying OpenClaw, warn that widespread use of AI coding tools is leading to more technical debt ($, WSJ). One thing this suggests is that the right way to iterate is to build a prototype first, then break it into microservices with carefully-considered incoming and outgoing data structures before customers use it. If there are independent components of a bigger program, it's easier to see which one is buggy, either because it's producing broken outputs or because it's too slow. It's much easier to build something monolithic, but that also makes it opaque. The microservice approach adds overhead, but that's a relatively safe bet to make a time of high cloud capex: if the biggest companies on earth are competing to buy more infrastructure, then either they're right about demand (in which case you should pay up for reliability) or they have that completely wrong (in which case you'll pay a lot less for redundancy).
Mindshare
Spotify is offering audio versions of long magazine articles within their app, and isn't specifically disclosing whether or not they're using AI (though the article phrases it in a way that implies off-the-record confirmation). It's another example of what makes Spotify unique: audio is uniquely easy to consume while doing something else (commuting, chores, exercise) and uniquely hard to consume alongside any other kind of media. So their goal, particularly in the age of liquid content, is to make everything audio so they get uniquely concentrated attention.
Consumption Equality
Waymo turns out to be very nice for people with disabilities, which is a common pattern with new technologies: they raise variance at the right tail of the distribution, but reduce it at the left tail by making technological substitutes for variable human skills more available. This will be an underrated aspect of AI for a while, but it's important: it's very easy not to think about how the things you regularly do would be inconvenient or impossible if you had a disability, and those disabilities also make people less visible. Living in a blander, more mass-produced world has its aesthetic downsides, but raising the baseline for certain kinds of consumption disproportionately benefits people who were previously furthest below that baseline.