SpaceX's S-1 Just Revealed That xAI Is Losing $2.5 Billion Per Quarter. Grok Is the Most Expensive AI Infrastructure Bet in History.
SpaceX filed its S-1 with the SEC on May 20, and the AI segment numbers tell a story the press coverage mostly ignored. In Q1 2026, SpaceX's AI segment (which includes Grok, the X platform, data licensing, and the Colossus compute infrastructure absorbed in the February xAI merger) generated $818 million in revenue and posted a $2.5 billion operating loss. Annualized, that's $3.3 billion in revenue against a $10 billion operating loss rate.
For context: SpaceX spent $12.7 billion in 2025-2026 CapEx on AI infrastructure alone. They built what they're calling the world's first coherent gigawatt-scale AI training cluster. And then they rented most of it to Anthropic.
If you're building on Grok's API, this financial picture is the infrastructure risk you need to understand.
The merger math that produced these numbers
xAI (Elon Musk's AI company, founded in 2023) merged into SpaceX in February 2026. The deal valued xAI at $250 billion and SpaceX at $1 trillion, for a combined $1.25 trillion. Post-merger, SpaceX's AI segment houses Grok, the X platform (Twitter/X advertising and subscriptions), and the Colossus cluster in Memphis and Mississippi.
The S-1 breaks out the AI segment separately from SpaceX's core space business (rockets, Starship) and connectivity business (Starlink). That's important because Starlink is profitable: $11.4 billion in revenue in 2025 with $7.2 billion in EBITDA. The space segment is burning money on Starship R&D ($3 billion in 2025). The AI segment is the biggest money pit in the company.
Full-year 2025 AI segment: $3.2 billion revenue (up 22%), $6.4 billion operating loss, R&D up 331% to $5 billion. Q1 2026: $818 million revenue, $2.5 billion operating loss. The losses are accelerating relative to revenue growth.
Why they're renting Colossus to Anthropic
SpaceX built Colossus 1 (220,000+ NVIDIA GPUs, 300+ MW capacity) and announced in late 2025 that Anthropic would rent it at $1.25 billion per month. The deal runs through May 2029 with a 90-day exit clause on either side. I wrote about this deal on May 23 from Anthropic's perspective: your Claude API call is now downstream of a SpaceX data center.
From SpaceX's perspective, the deal looks different. They built a gigawatt-scale cluster for Grok training, ran into the same compute availability crisis everyone else did, and then found out that Colossus 1 generates better return on investment rented to Anthropic at $1.25B/month than sitting idle between Grok training runs. Grok 5 training is happening on Colossus 2 (the Mississippi cluster). Colossus 1 is essentially Anthropic's data center now.
That's not a sign of weakness on its own. Hyperscalers rent unused capacity all the time. But the combination of renting out your primary cluster while posting $2.5B quarterly losses on the business that cluster is supposed to support raises a question: is xAI generating enough from Grok to eventually justify the infrastructure cost, or is Anthropic's rent check the most reliable revenue in the AI segment?
What Grok's trajectory actually looks like
The S-1 doesn't break out Grok API revenue separately from X platform advertising. It's consolidated in "AI segment." But Grok's commercial API has been available since late 2024, and based on developer community signals, adoption has been slower than xAI's marketing suggested.
Grok 3 benchmarks were strong on coding tasks when released in early 2025. The API pricing was competitive. But the developer ecosystem around Grok (tooling, documentation, community support) hasn't caught up with Claude or OpenAI. And the model releases since Grok 3 have been gradual, while Anthropic has shipped Opus 4.7 and Sonnet 4.6 with demonstrably better performance on complex reasoning.
Grok 5 is training on Colossus 2 now. The company is clearly not giving up. But at the current loss rate, xAI needs Grok 5 to be a significant leap, not a marginal improvement, to change the commercial picture.
The risk for developers building on Grok
Public company shareholders seeing a $10 billion annualized operating loss on a $3.3 billion revenue segment will push for one of three outcomes: aggressive monetization, scaling back, or divestiture.
Aggressive monetization means API pricing goes up. Scaling back means the developer API gets deprioritized relative to the enterprise deals that move the revenue needle. Divestiture is unlikely but not impossible: the xAI assets could theoretically be separated again.
For a solo developer who's built on Grok's API, the most likely near-term outcome is not "API disappears." It's "API pricing changes in ways that make the math less favorable, faster than Anthropic's or Google's pricing would." SpaceX needs the AI segment to justify its IPO valuation. Starlink can't carry that math alone.
The honest counter-take
The Starlink segment generates $7.2 billion in annual EBITDA. Starship, when it eventually reaches full commercial launch cadence, is a multi-billion dollar business. SpaceX is not going to go under because xAI is running losses. This is a company with genuine cash-generating businesses that can subsidize infrastructure investment.
And Grok 5 could change the narrative completely. If it ships on Colossus 2 and benchmarks above Claude Opus and GPT-5 on frontier tasks, the commercial picture for xAI's API business improves rapidly. AI model quality is the forcing function for API adoption, and SpaceX has the infrastructure to train a frontier model.
The S-1 losses are a risk signal, not a verdict. But "our primary cluster is rented to our main competitor's biggest infrastructure partner" is not what a confident AI segment story looks like. File that in the risk register and diversify your API dependencies accordingly.
Author
Lukas
@lukcombinator