A rocket company spent $15.8 billion on data centers

Hey {{first_name|Investor}} -

SpaceX ($SPCX) spent $18.4 billion on capital projects last quarter. Of that, $15.8 billion went to AI compute.

Rockets, launch pads, satellite production and ground stations split the remaining $2.6 billion.

So a company whose stated mission is Mars put roughly 6 of every 7 capital dollars into data centers. That number is why I spent this week reading the SpaceX transcript against everything else that reported since late June: Microsoft, Alphabet, Amazon, Meta, Oracle, Micron, Samsung, SK Hynix, SanDisk, Western Digital.

I went in looking for the quarter where the AI buildout finally cools off. Four things came out instead, and I want to walk you through all of them.

SpaceX reports 3 segments: Space, Connectivity, and AI.

Q2 revenue came in at $7.8 billion, up 92% from $4.1 billion. That headline gets quoted everywhere. The segment math underneath it tells a different story.

Space did $962 million, up 29%. Connectivity did $4.3 billion, up 66%. AI did $2.6 billion, up 247%.

Pull AI out and the rest of the business grew somewhere around 58%. Of the roughly $3.7 billion of revenue added year over year, AI contributed about half, Connectivity about 46%, and the actual space business about 6%.

The profit picture tilts even harder. Adjusted EBITDA was $3.5 billion: Connectivity $2.6 billion, AI $1.1 billion, and Space at a $205 million loss. The rockets are the only segment losing money, and they're about 12% of revenue.

The customer list is what turns this from a curiosity about Elon into a data point about the whole sector. CFO Bret Johnsen said new cloud services agreements at their Colossus sites contributed $1.6 billion of incremental revenue in the quarter. President Gwynne Shotwell named the counterparties: "other leading companies such as Google and Anthropic."

Then this, from Johnsen, about the first few weeks of July:

That's $6.7 billion booked in about 3 weeks, by a company that was private 2 months ago and has no history as a cloud vendor. Musk also said the plan is to rent most of it out, with "maybe 10% of our compute to be used for Grok training."

Management's line is that the newest compute pays back in about a year. Treat that as a claim rather than a fact, because none of this has been through a full cycle yet. What's harder to wave off is the revealed preference. A launch company looked at where it could put $18.4 billion and sent 6 of every 7 dollars to compute.

Every management team that reported this cycle said the same thing, unprompted, in plain language, on the record. This is the part that changed how I'm thinking about the next 18 months.

Amazon's Andy Jassy, after raising full-year CapEx to roughly $220 billion:

Microsoft CFO Amy Hood: "Customer demand continues to exceed available capacity."

Alphabet's Sundar Pichai: "we continue to be supply constrained."

Mark Zuckerberg: "The high-level observation is that there's just nowhere near enough compute for all the demand."

And Johnsen, on the SpaceX call, describing the same imbalance from the seller's side: "we expect the supply-demand imbalance in the compute market to continue."

Talk is cheap, so look at what's contracted behind it. Microsoft's remaining performance obligation reached $678 billion, up 84%. Oracle's backlog hit $638 billion. Alphabet's cloud backlog is around $514 billion. Amazon's is roughly $496 billion.

Those are not the same instrument, and I don't want you reading them as one number. They differ on duration, on cancellation terms, and on how concentrated the customer list is. Microsoft's is the one worth pulling apart, because it's the biggest and it got quoted the most. Hood put the weighted average duration at 2.3 years, said about 30% converts to revenue inside 12 months, and said backlog grew 25% year over year once you strip out OpenAI, against an 84% headline.

So the visibility is real and the growth is far more concentrated than the headline suggests, and both of those can be true at once.

Jassy went further and said the lion's share of 2027 capacity is already reserved, with some of 2028 spoken for too.

The CapEx guides moved the same direction. Alphabet raised its 2026 range to $195 to $205 billion from $180 to $190 billion. Amazon raised to about $220 billion from about $200 billion. Oracle guided fiscal 2027 to roughly $70 billion net, up from $55.7 billion. Meta narrowed its range upward at the bottom end. Microsoft's headline calendar-2026 number came down, and Hood attributed the entire change to extending the useful life of data centers from 15 to 25 years, which shifts future leases out of the CapEx line. Her words: "Outside of this useful life impact, our calendar year 2026 CapEx investment expectations remain unchanged."

Nobody in the group guided underlying spending lower.

By my own arithmetic, the 4 largest spenders put roughly $170 billion into capital projects in a single quarter. On August 3, Morgan Stanley put 2027 cloud CapEx near $1.4 trillion, about 17% above a consensus sitting around $1.2 trillion. One house against the street, and a house estimate is not a fact.

I had two other broker numbers in an earlier draft of this and cut them, because when I went back to source them properly they didn't hold up the way they'd been quoted. Worth saying out loud: a lot of the CapEx forecasts circulating right now are cumulative multi-year figures being passed around as annual ones.

What the spread does tell you is that the sell-side argument is about magnitude, not direction. Though I'd add that being bullish on aggregate spend and being bullish on returns from that spend are two different positions, and the second one is where the real disagreement is.

That's the framework I'd carry into the rest of this. Demand for compute is contracted years out, the buyers say so on the record, and the constraint sits on the supply side. Which raises the question I care about as an investor: if supply is the binding constraint, where exactly is it binding, and who collects the rent?

That's where the last 6 weeks of earnings get specific.

The rest of this issue is for Insiders.

Below the line I show where that rent is actually landing, with the margin figures from all five memory makers, the contract structures underneath them, and the company now collecting cash deposits from customers for delivery years out. I also cover what complicates it, from pricing growth that has slowed two quarters running to roughly $820 billion of commitments sitting off the balance sheet. It closes with the six things that would tell me this thesis is breaking, and the full stack of names I'm reading through the theme.