The S-1 Scorecard

The Setup

A publication that only revisits the calls it got right is not running a scorecard. It is running a highlight reel. So this piece grades the misses first and at full strength, because that is the part that determines whether anything else here is worth reading.

In The S-1: Who Is This IPO Actually For? we made a set of specific, checkable claims about SpaceX (SPCX) based on the prospectus filed May 20. The AI segment was shrinking relative to its competitors. Losses were accelerating. Cash was running out. Capital expenditure on AI infrastructure was the structural risk nobody was pricing. Starlink average revenue per user was falling. And the governance structure left public shareholders without a vote.

On August 4, after the close, the company answered several of those questions with reported figures for the second quarter of 2026. Some of what we wrote held up. Some of it did not.

What We Got Wrong

Three claims from the May piece were contradicted by the print, and one of them was contradicted decisively.

The AI growth comparison. We wrote that the AI segment grew revenue 22 percent year over year while OpenAI grew more than 230 percent over the same period, and we used that gap to argue the segment was losing the race. In the second quarter, AI segment revenue was $2.561 billion, up 247 percent year over year and up 213 percent sequentially. Against a consensus estimate of roughly $2.18 billion, it beat. On the headline number, the framing in our May piece was wrong. The composition of that growth is a separate question, and we take it up below, but the number itself is not in dispute.

The loss trajectory. We described a company where losses were accelerating, citing a $4.28 billion net loss in the first quarter. In the second quarter the net loss was $541 million, an improvement of $467 million against the prior year period. Adjusted EBITDA reached $3.5 billion, up 191 percent. The AI segment narrowed its operating loss to $1.3 billion and turned adjusted EBITDA positive at $1.1 billion. Loss per share was $0.09 against an expected $0.26. That is not deceleration at the margin. That is a material improvement, and our piece did not anticipate it.

The raise. We noted a target of up to $75 billion and framed it as a company reaching for capital it needed rather than capital it wanted. The IPO raised approximately $85.7 billion, alongside roughly $25 billion in investment grade debt. Whatever else the offering was, it was not undersubscribed. Cash and equivalents stood at $93.5 billion at quarter end.

It is also worth recording what the sell side thought. Analyst Gene Munster called the quarter a blowout and argued the market was missing the point. Twenty seven analysts rate the stock a buy against two sells. The bear case in our May piece is not the consensus, and it was not vindicated by the headline figures.

The Number That Moved the Stock

Against all of that, the stock fell. It closed Tuesday at $125.33, dropped more than 8 percent in extended trading, and traded near $111 before Wednesday's open. A company that beat on revenue in all three segments, narrowed its losses, turned a segment EBITDA positive, and raised full year guidance for the first time in its twenty four year history lost roughly a ninth of its market value overnight.

The line that did it was capital expenditure. Second quarter capex was $18.37 billion, more than six times the prior year period, against an average analyst estimate of $13.22 billion. Of that total, $15.83 billion went to AI infrastructure. Debt and finance leases rose from $22 billion to $36.8 billion in a single quarter.

This was the claim our May piece made most directly. We calculated an annualized AI capital expenditure run rate of roughly $30.8 billion from first quarter figures and flagged it as the structural risk in the filing, noting for scale that Meta's entire 2025 capital budget across every business line was around $35 billion. The second quarter run rate is roughly $63 billion annualized. We did not overstate the risk. We understated it by about half.

Nothing about that observation requires the company to be doing something wrong. Capital intensity is not fraud. But it is the variable the market repriced, and it is the one the prospectus discussed least clearly.

Where the AI Revenue Came From

The 247 percent AI growth figure contradicts our May framing on its face. The company's own disclosure of where that revenue originated is more complicated.

On the earnings call, the chief financial officer stated that the company entered into cloud services agreements in the second quarter providing access to compute capacity at its Colossus and Colossus 2 sites, and that the initial ramp from those agreements contributed $1.6 billion of incremental AI infrastructure revenue in the quarter.

The AI segment went from $818 million in the first quarter to $2.561 billion in the second. That is an increase of roughly $1.74 billion. If $1.6 billion of it came from compute leasing, then approximately 92 percent of the quarter's AI growth came from renting hardware rather than from selling AI products. The remainder, covering Grok, X subscriptions, and advertising, grew roughly 17 percent sequentially. That figure is much closer to the growth rate our May piece described than the headline 247 percent is.

Only one of the announced compute contracts was generating revenue in the quarter. Alphabet's agreement, reported at roughly $920 million per month for approximately 110,000 GPUs, does not begin until October. Reflection AI's $150 million monthly contract started July 1 and lands in third quarter results. The contract that was flowing in the second quarter was Anthropic's: $1.25 billion per month for exclusive access to Colossus 1 in Memphis, a facility reported to house between 220,000 and 325,000 Nvidia GPUs across roughly 300 megawatts of capacity, running through May 2029.

Our May piece devoted a section to this arrangement under the heading The Anthropic Paradox, observing that a company positioning itself as an AI leader was renting its compute to a direct competitor because its own fleet was running at roughly 11 percent utilization. That observation now describes the largest single driver of the quarter the market just repriced.

There is a second order point here that cuts in the company's favor. Compute leasing revenue against installed hardware carries high incremental margin, which is part of why segment adjusted EBITDA turned positive. The company also disclosed a further $6.7 billion of cloud services revenue under contract beginning to ramp in October. This is a real business with real contracted demand. It is simply a different business than the one described in the prospectus.

Starlink Works, ARPU Still Falls

The May piece said Starlink was the genuinely transformative asset inside the conglomerate, and the second quarter supports that. Connectivity revenue reached $4.291 billion, up 66 percent year over year and ahead of a $3.83 billion estimate. Segment adjusted EBITDA rose 64 percent to $2.6 billion. Subscribers reached 12 million, roughly double a year earlier, with 1.7 million net additions and 17 percent sequential growth. This remains the profitable engine of the company.

We also claimed average revenue per user was falling 23 percent year over year as the service expanded into lower priced international markets. Dividing reported Connectivity revenue by period end subscribers, the implied figure is roughly $358 per subscriber for the quarter against roughly $431 a year earlier, a decline of about 17 percent. That calculation is approximate. Connectivity revenue includes enterprise, maritime, aviation, and government business alongside consumer subscriptions, and period end subscriber counts are not the same as average subscribers. The direction holds and the magnitude has moderated. Growth is being purchased with price.

What a Four Percent Float Does

The May piece argued that a public float of roughly 4 percent, with insiders holding the rest, would produce price discovery that reflected supply mechanics as much as fundamentals. The two sessions around this earnings report are a clean illustration.

On Monday the stock rose 5.68 percent to $114.53. On Tuesday, the day of the report, it rose another 9.43 percent to $125.33 on roughly 144 million shares, about two and a half times its recent average volume. That is a 15.6 percent advance into a print that nobody had seen. By Wednesday morning the entire move was gone and then some.

Positioning ahead of an unknown result, then a complete round trip on the result itself, is what a thin float looks like. It is worth holding in mind before reading any single session as a verdict on the business. It is also worth noting that 911.5 million shares are scheduled to unlock on August 6, a quantity larger than the existing public float. That event deserves its own analysis once there is trading data to examine rather than speculation to offer.

A Different Business Than the One That Filed

Read the graded claims together and a pattern emerges that neither our May piece nor the prospectus stated plainly.

The revenue beat and the capital expenditure blowout are not two findings. They are one. The $1.6 billion of compute revenue exists because of the $15.83 billion of AI infrastructure spending, and the contracted revenue ramping in October requires capacity that has to be built, powered, and financed first. Debt and finance leases rising $14.8 billion in three months is how data centers get funded. It is not how launch vehicles get funded.

The economics on display are those of an infrastructure landlord: high incremental margin on installed capacity, brutal capital intensity to add more, revenue concentrated in a small number of very large counterparties, and a business whose binding constraints are power, chips, and permitting. On the earnings call, analysts asked about all three. Those are not questions anyone asks a rocket company.

The prospectus described the AI segment as the source of most of the company's value and pointed toward models, Grok, and eventually orbital compute. What reported on August 4 monetizes primarily by leasing hardware to other AI companies. That is a legitimate business, and at present a fast growing one. It is also a business with entirely different margin structure, capital requirements, competitive set, and risk profile than the one investors were asked to underwrite in May. The company is now reported to be in discussions with the Department of Defense over data center capacity, and it closed a $60 billion acquisition of Anysphere in June.

On valuation, the picture has genuinely improved. Our May piece cited roughly 94 times revenue at the target valuation. At approximately 13.17 billion shares outstanding and a share price near $111, the implied market capitalization is around $1.47 trillion, or roughly 47 times annualized second quarter revenue. Most of that compression came from revenue doubling rather than from the share price falling. Whether 47 times is defensible depends almost entirely on the company reaching its stated $100 billion annualized revenue run rate by the end of 2026, which depends on the compute contracts, which depends on the capital expenditure. The valuation question and the data center question are the same question.

What Could Go Wrong With This Reading

The strongest argument against everything above is that one quarter of data is being asked to carry more weight than it can bear, and that the parts of our May thesis that failed were the parts that were actually testable.

Consider the bull case honestly. Revenue nearly doubled. Every segment beat. Losses narrowed by nearly half a billion dollars against the prior year. A segment that was hemorrhaging cash turned EBITDA positive. Full year guidance rose for the first time in company history. Management pulled forward its internal $1 trillion revenue forecast from 2031 to 2030 and stated that the $100 billion run rate target would be met even without further business wins. If a company delivers that and the stock falls 11 percent, the more parsimonious explanation may be that the market is mispricing a heavy investment year, exactly as Munster argued, and that the capital expenditure funding future contracted revenue is the least alarming kind of spending there is.

The concentration point cuts both ways as well. Describing the AI revenue as landlord economics is accurate, but leasing compute to Anthropic, Alphabet, and Reflection AI means the counterparties are among the best capitalized buyers in the market, on multi year contracts. That is a better receivables profile than most enterprise software businesses carry.

There is also a real risk in our own framing. Calling this a data center company is a reading of the current revenue mix, not a description of the whole enterprise. Starlink remains the largest segment and the profitable one. The launch business grew 29 percent. Starship development continues. Reducing a company with those assets to its compute leasing arithmetic would be its own form of narrative.

Finally, the durability question is genuinely open. Compute leasing prices could compress as capacity arrives across the sector. Or contracted demand could prove sticky and the current spending could look prescient in eighteen months. Two months of Anthropic revenue does not settle it, and neither does one quarter of capital expenditure.

The Bottom Line

The scorecard is mixed, and the misses are not small. We were wrong about the direction of the losses, wrong about the AI segment's headline growth, and wrong about the reception of the offering. Anyone who read our May piece as a prediction that the business was deteriorating got a quarter that said otherwise.

What held up was narrower and, as it turned out, more consequential. We identified AI capital expenditure as the structural risk the filing understated, and it is the single line that took roughly 11 percent off the stock on a day when everything else beat. We described the compute leasing arrangement as a paradox, and it turned out to account for approximately 92 percent of the AI growth that appeared to refute us. We said a 4 percent float would distort price discovery, and the stock ran 15.6 percent into an unknown print and gave all of it back.

The larger finding is not on either side of the ledger. The business that reported on August 4 is not the business the prospectus described. It is substantially an AI infrastructure landlord, with the margin profile, capital requirements, and customer concentration that implies. That is not a criticism. It is a different set of questions than the ones the S-1 answered, and the honest position is that most of them will not be settled until the Alphabet contract flows in the fourth quarter and the lock up supply is absorbed.

We publish the grades because the alternative is a publication that only remembers its good calls. The data will say what it says next quarter too.