I thought someone was propping up SPCX, but no one was
I had been watching a particular stock’s order book every day for a while.
It was SpaceX (SPCX), which went public on June 12. IPO price $135, with proceeds from the offering of roughly $8.57 billion. One of the largest raises in history.
I had a reason to watch it. On the day it listed, I wrote an article about this company’s multiple, asking how a valuation of 245 times Starlink’s EBITDA could possibly hold together. Having made that much of a fuss, I figured I owed it to myself to see how things turned out.
After listing, the stock fell in nearly a straight line, from a June high around $225 down to the low $100s by early August. Roughly half its value.
The decline itself was entirely expected. I had publicly stated my unease, so let me say this plainly: I wasn’t surprised.
What caught my attention was something else.
The unease of the same pattern, every single day
For several days, the exact same shape repeated itself.
At the open, the price would drop about 5% from the previous close. Once it fell below $110, buying would suddenly come in, pushing it back to $115. From there it would barely move for the rest of the day, drifting to a close.
The next day, the same thing.
This didn’t look like natural price action.
In a stock that kept falling, buying kept showing up at a specific price, and it never went higher than that. It looked like someone was “defending a level.”
My hypothesis was simple: someone with an interest in propping up the stock, and the capital to do it, was buying to defend the price.
There was a basis for this. This stock’s float was only about 5% of shares outstanding at listing. Relative to daily trading volume, that gives a large player more room than usual to move the price. A structure in which a small amount of money can set the price does exist.
Let me state the conclusion up front: this hypothesis was wrong.
More precisely, the phenomenon was real, but there was no one behind it. What follows is the record of me working that out.
It was doing things a price-defender wouldn’t do
The first thing that fell apart was the official stabilization angle.
Underwriter stabilization takes the form of buying back unexercised over-allotment shares. But in this deal, the green shoe had been fully exercised within days of listing, leaving no ammunition to buy back. On top of that, the stabilization period is typically around 30 days, and that had already expired by mid-July.
A motive for the company’s own buyback is also thin. Right after raising $8.57 billion, with quarterly capital expenditure running around $18 billion, there’s no logic for a company to spend cash defending its stock price. It was also within the blackout period before earnings. If an individual insider were buying, a Form 4 would appear within two business days, and that could be checked too.
So far, this is a process of elimination. The real problem came next.
If someone were defending the price, the price action wouldn’t look like this.
The level supposedly being defended was, in fact, stepping down day by day: $110.85, $109.53, $108.66, then $104.83. A step down every single day.
A deliberate actor defending a price would hold the line. A defense line that drops every day is not a defense.
And one more thing: this buyer was selling at $115.
Capping the bounce was part of the same pattern. Buy at the bottom, sell at the top. A defender wouldn’t sell at the top.
The phenomenon was real, but it wasn’t “a force setting a floor” — it was “a force compressing the range.” That’s the heart of where the hypothesis went wrong.
The real cause was direction-blind inventory adjustment
It clicked when I looked at the options market side. I’ll describe it without jargon.
There are three players: the person who buys the option, the dealer who sells it, and the stock market itself.
A dealer’s business is to sell options and earn something like a commission, not to bet on whether the stock goes up or down. That’s the starting point.
A dealer who sold “the right to buy at $115,” for example, is in trouble if the stock rises, because the obligation to deliver shares becomes real. So the dealer buys the stock in advance, but not all of it. If the stock is still at $110, the chance that right gets exercised is low.
So the dealer holds only a probability-weighted amount: 30% chance, hold 30%; 50% chance, hold 50%; 70% chance, hold 70%.
As a result, the dealer buys more every time the stock rises, and sells every time it falls.
The relationship between a dealer’s holdings and the stock price traces an S-curve: the higher the price, the more shares the dealer needs to hold.
Shares that need to be held (out of 100)
30 shares 50 shares 70 shares 95 shares The higher the price goes, the more the dealer buys
$110 $115 $118 $125 Stock price
Figure 1: The relationship between shares a dealer needs to hold and the stock price
There’s a mirror image too. A dealer on the other side of “the right to sell at $110” benefits when the stock falls. To stay neutral, the dealer buys stock to cancel out that gain. The lower it goes, the more gets bought. When it recovers, the dealer sells it off.
Buy when it falls, sell when it recovers.
This was exactly the pattern I’d been watching for days.
The dealer hadn’t turned bullish or bearish. It was mechanically adjusting an inventory meant to honor a commitment, in response to price. From the market’s perspective, it still looks like a buy order came in, but there’s no intent behind it.
Defending $110 and capping $115 on the same day isn’t a contradiction either. The $110 right and the $115 right are different products, and which side the dealer sits on differs between them, so the behavior changes depending on the price range.
The sign flips depending on the price range: below $110, buying increases as the price falls; above $115, buying increases as the price rises.
Buying increases as price rises Hedging the call side. The rise accelerates $115
Buying and selling balance out The intraday range compresses $110
Buying increases as price falls Hedging the put side. The decline is held back
Figure 2: Even on the same day, the sign flips depending on the price range
This mechanism isn’t loyal to any price level. It follows wherever the open option positions happen to be concentrated. That’s why the defense line drops every day.
I should note that I haven’t confirmed the actual open interest figures for this stock. All I can say is that this is the most straightforward explanation consistent with the order book’s behavior.
My instinct that “someone is defending the price” was right as a phenomenon. Where I was wrong was in assuming the actor was human.
The test case was earnings
Explanations of this kind can be fitted to almost anything after the fact, so they need to be put in a form that can be tested.
As it happened, the first earnings report since listing fell on August 4, with a massive lockup expiration scheduled two business days later: 911.5 million shares, or 1.4 times the roughly 639 million shares sold in the IPO.
My prediction was this:
- Even if the numbers are good, the stock won’t recover, because what’s setting the price isn’t earnings but the supply schedule
- If the company made an announcement touching the supply side (such as voluntarily extending the lockup), that would change things — but it won’t, because doing so would mean forcing early investors and employees who already have the right to sell to forgo their gains for the sake of outside shareholders
- So, with maybe a bump along the way, the stock ultimately falls
The earnings numbers were a substantial beat.
Revenue of $7.81 billion (versus expectations of about $6.8–6.9 billion). Net loss of $541 million, a loss of 9 cents per share, well ahead of the expected 26-cent loss. Operating loss narrowed from $970 million in the same period a year earlier to $143 million. All three business segments beat consensus. Starlink subscribers doubled year over year to 12 million. Full-year guidance was raised for the first time in the company’s 24-year history.
The stock fell 8% after hours.
It landed right back at the level from the day before earnings.
Inside the “AI company” was an equipment-leasing business
Why does a beat send the stock down? Here another structure comes into view.
This IPO was sold as a “rocket company,” but by revenue it breaks down to 55% communications, 33% AI, and 12% space. Rockets are the smallest piece.
The AI segment drawing attention had revenue up 247% year over year to $2.56 billion. The headline is flashy.
Reading the breakdown, of the $1.82 billion increase, $1.6 billion came from leasing out data centers. It wasn’t Grok usage growing, and it wasn’t AI products selling. It was revenue from renting out facilities full of GPUs to other companies. Advertising revenue from the former Twitter, which sits in the same segment, was down 14% year over year to $367 million.
The contract structure matters even more. One major customer is leasing a data center in Memphis under a fixed monthly fee, three-year, exclusive arrangement. Being a fixed amount means it doesn’t grow from there. Revenue ramps up in steps tied to when contracts start, then flattens out.
And because it’s exclusive, the same facility can’t be sold to another customer. Taking on new customers requires building new facilities.
Lining up the numbers makes the nature of the business clear. The total value of cloud contracts signed this quarter was $14.1 billion. AI-related capital expenditure in the same quarter was $15.8 billion. The capital deployed exceeds the total value of the contracts it brought in. And the contract value is recognized over three years.
Comparing capital expenditure to contract value: against $63.2 billion in annualized AI capital expenditure, the annual revenue expected from contracts is $4.7 billion.
AI capital expenditure (annualized)
$63.2 billion Annual revenue from contracts
$4.7 billion Rough estimate of recognizing the $14.1 billion contract total over three years
Figure 3: Capital deployed versus the annual revenue it can be expected to generate
There is no way to grow revenue other than piling up more equipment. This isn’t software economics — it’s real-estate leasing economics.
The multiple the market pays for AI companies assumes software-like gross margins and economies of scale. I suspect what the market reacted to wasn’t the sheer size of the $18.4 billion quarterly capital expenditure figure itself, but this ratio.
Nobody can say with certainty whether this is overvalued. Since the company is unprofitable, P/E doesn’t apply, and the answer changes by an order of magnitude depending on whether you apply a telecom multiple, an AI infrastructure multiple, or a data-center real-estate multiple. That’s why the price targets from 34 analysts range from $62 to $800.
Uncertainty itself can be an investment opportunity. If you can buy cheaply something nobody can price, that’s an edge. What’s harsh here is that this sits at maximum uncertainty while also carrying one of the highest multiples around. You could take the same bet on a target with a clearer business structure, and much cheaper.
That said, the communications business is the real thing. Quarterly operating profit of $1.7 billion, up 79% year over year. The barriers to entry are physical too. If this were a standalone company, I think it would be worth considering for a long-term hold. The problem is that the value of the communications business alone can’t explain the current market cap.
At the end of my previous article, I wrote that if there were a stock that let you buy only Starlink, I’d buy it. Two months of earnings later, that wish has only been reinforced. Communications is strong, and it’s the structure wrapped around it that’s setting the price.
Closing
I started from “someone is defending the price” and landed on “a mechanism with no intent was trading in response to price.”
I found this fascinating. It felt like searching for a culprit and finding, instead of a culprit, something closer to a law of physics.
Similar mechanisms apparently show up all over the market: flows tied to index rebalancing, risk-parity rebalancing, forced liquidations triggered by collateral value. None of these involve anyone making a judgment call, yet trading occurs as if by identity, in response to price.
Mechanisms are easier to read than intent. Once you know what something is tracking, you can calculate its next move. Reduce it to “someone’s intention,” and you stop thinking right there.
When you see unnatural regularity, look for the constraint first, not the actor. That was the biggest lesson from this. For someone who used to work on circuits, that should have been a familiar way of thinking all along.
That said, even the mechanism has its limits. Options-driven trading volume has a ceiling, and that pressure itself disappears at expiration, when the open positions vanish. Against a scenario with tens of billions of dollars of supply coming, it becomes a rounding error.
As of this writing, that verdict hasn’t come in yet.
References
- CNBC, “SpaceX stock drops after first earnings report as AI costs soar”
- unite.ai, “SpaceX’s Cloud Business Tripled Its Revenue and It Still Loses Money”
- Stocktwits, “SPCX Stock Declines As Investors Weigh AI Investment Scale Despite Upbeat Q2 Earnings”
- Eastern Herald, “SpaceX Q2 Earnings: 92% Revenue Jump, Starlink Leads”
- Tech Times, “SpaceX Q2 Beat Raises Full-Year Guidance for First Time as Lock-Up Looms”
- ZeroHedge, “SpaceX Slides Despite Big Revenue Beat, Tame CapEx”
- 24/7 Wall St., “Will SpaceX Crush Expectations in Its First-Ever Earnings Report?”
- Forbes, “Most Of SpaceX’s AI Revenue Isn’t Coming From AI”
- FXStreet, “Earnings preview: SpaceX”
- Benzinga, “SpaceX Earnings Prediction Market Preview”
- Previous article: Looking at SpaceX’s IPO numbers, my first thought was “245 times is just plain crazy.”
※This article is not a recommendation to buy or sell any particular stock. Stock prices and order book figures are as of the time of observation and may differ depending on the source.
This piece was planned and directed by Yuichiro Kuzuryu, and written with AI.
Originally published in Japanese at https://clazytech.com/2026/08/1743/. Translated with LLM assistance and reviewed before publication.