After hours on August 4, Elon Musk delivered SpaceX's first quarterly report card since going public. The numbers are explosive, but the market reaction was nuanced.

First the good news: Q2 total revenue hit $7.814 billion, up 92% year-over-year, significantly beating consensus estimates of ~$6.9 billion; adjusted EBITDA turned positive for the first time (Q1 was -$600 million); AI business revenue reached $2.561 billion, surging 247% YoY and accounting for roughly one-third of total revenue. Starlink subscribers passed 7 million; launch business remains global number one.

Now the bad news: quarterly AI capital expenditure was approximately $15.8 billion, averaging $170 million per day spent on AI. Musk explicitly stated SpaceX uses "only NVIDIA GPUs," and the Colossus data center is still expanding furiously. After the earnings release, SpaceX stock dropped 7–8% after hours — revenue beat, but the market was terrified by the pace of spending in the tens of billions.

Selling GPU Compute Is More Profitable Than Launching Rockets?

The most interesting thing about this earnings report isn't rocket launches; it's that SpaceX's identity is fundamentally shifting: it's going from a "rocket company" to an "AI compute company."

Where does $2.56 billion in quarterly AI revenue come from? The answer is compute resale from the Colossus data center. SpaceX's Colossus supercomputing cluster in Texas is already one of the world's largest GPU clusters (over 100,000 H100/B200 GPUs), with primary customers being Anthropic, xAI, and other AI companies needing massive scale. Simply put: SpaceX buys NVIDIA GPUs, builds data centers, and resells GPU hours to AI companies — essentially a "compute wholesaler."

How profitable is this business? Gross margins tell the story. Rocket launch gross margins run around 20–30% (rocket manufacturing, fuel, launch pad operations are expensive); Starlink gross margins are around 40–50% (satellite depreciation is the main cost); compute resale gross margins are reportedly 60–70% — once the data center is built, ongoing costs are mainly electricity and operations with very low marginal cost. No wonder Musk is pouring this much money into AI; it really is a better business than launching rockets.

Musk himself admitted: "We didn't expect selling GPU compute would be this profitable." Colossus Phase 1 was fully booked by Anthropic and other customers immediately upon going live; Phases 2 and 3 are still under construction with reservations stretching to 2028. AI compute is currently a seller's market — if you have power, GPUs, and can deliver fast, customers are lined up throwing money at you.

$170 Million Per Day: What Is the Market Afraid Of?

If AI compute is this profitable, why did the stock drop after earnings? Because the market saw a spine-chilling number: $15.8 billion in quarterly AI capex.

What does $15.8 billion mean? It's more than the full-year revenue of many Fortune 500 companies. That's quarterly capex, not annual. At this burn rate, SpaceX is spending over $60 billion a year on AI — multiple times OpenAI's full-year revenue.

The market has three main concerns.

First, uncertainty around ROI. AI compute is indeed in short supply right now, but how long will this supply-demand imbalance last? If massive data center capacity comes online over the next 2–3 years, GPU production ramps up, or algorithm efficiency improves dramatically (e.g., small models catching up to large ones), compute prices could decline steadily the way cloud services prices have. Could the data centers being built for tens of billions today become expensive scrap metal three years from now?

Second, customer concentration risk. SpaceX's compute revenue depends heavily on a small number of large customers (Anthropic is the largest, with xAI reportedly also a major user). If Anthropic switches to other suppliers in the future, builds its own data centers, or burns through its funding, SpaceX's compute revenue could cliff-dive. Over-reliance on a single customer is a cardinal sin in any B2B infrastructure business.

Third, Musk's stretched management bandwidth. Musk is currently running Tesla, SpaceX, xAI, Neuralink, The Boring Company, and Twitter/X simultaneously — six companies, with SpaceX and xAI both burning cash building data centers at a ferocious pace. A person's bandwidth is finite; running this many multi-billion-dollar mega-projects simultaneously — how can it not go wrong? History shows few entrepreneurs can successfully run two $100B+ companies concurrently, let alone six.

But Musk clearly disagrees. His tone on the earnings call was aggressive: "AI compute is the oil of the 21st century. We're not spending too much — we're spending too little. We will continue accelerating Colossus construction; AI capex over the next four quarters may actually increase."

This is classic Musk-style all-in: bet everything for winner-take-all, or go down together. Wall Street loves this narrative but also fears it. During the dot-com bubble in 2000, telecom companies were similarly pouring money into fiber optic cables and data centers, and when the bubble popped, 90% of capacity went idle and stocks lost 90%+ of their value. Will AI compute repeat history? Nobody knows right now.

But one thing is certain: SpaceX is no longer the company you knew that only built rockets. It's now one of the most aggressive players in the AI compute arms race.

See you tomorrow.