Renting out data center capacity, not chatbot usage, is where SpaceX’s AI segment found most of its second-quarter money, according to The Decoder’s Maximilian Schreiner. The segment covers xAI and X, and its quarterly revenue reached $2.56 billion, against $818 million three months earlier. A single buyer, left unnamed in the disclosures, contributed close to $1.52 billion toward that total. The Decoder’s own reading is that the buyer is probably Anthropic, an inference it presents as such rather than as established fact.
That detail outranks the headline growth figure. A company known publicly for a consumer chatbot is now earning most of its AI revenue as a landlord, and the biggest tenant may be a direct competitor to the product it sells. If the inference holds, Anthropic would be paying SpaceX for the same Colossus infrastructure that trains and runs Grok, the model Anthropic’s own Claude is built to beat.
The segment is still unprofitable at scale. SpaceX’s AI business absorbed a $1.26 billion operating loss in the second quarter, narrower than the first quarter’s $2.47 billion shortfall against that same $818 million revenue base. Across the first six months of 2026, the segment’s combined operating losses reached $3.73 billion. Revenue is climbing faster than losses are shrinking, but the gap remains wide.
The compute buildout behind those numbers keeps expanding. SpaceX currently operates 1.4 gigawatts of capacity, a figure disclosed in its most recent quarterly filing. Musk told investors on the company’s Q2 earnings call that total capacity should clear two gigawatts before 2027 is out. By his own account, Musk expects the real number to land nearer ten gigawatts than five gigawatts once that year closes. He offered no firm commitment, and The Decoder treats the target as guidance rather than a locked plan.
The chip count underneath that target is already substantial. Regulatory filings tied to SpaceX’s initial public offering put the total Colossus chip count, current and planned combined, at around 540,000 GPUs. Older Nvidia H100 units make up close to 100,000 of that base. GB200 and GB300 chips already installed or committed add roughly 110,000 apiece, and SpaceX has flagged more than 220,000 additional GB300 units still to arrive. The next expansion phase points to Vera Rubin, Nvidia’s forthcoming chip platform that Musk described as the strongest architecture on the market today. Scaling toward the stated power target would push the total GPU count well past one million. If the entire expansion ran on Vera Rubin chips exclusively, the ceiling would sit above two million GPUs, a total worked out by The Decoder itself from Musk’s capacity goal rather than a number SpaceX has published. SpaceX has not disclosed what fraction of that future buildout will run on Rubin chips versus its existing hardware.
The compute spending sits on top of a corporate structure reshaped earlier this year. SpaceX absorbed xAI, Musk’s separate AI company, through a stock-heavy merger that closed in February 2026. The deal priced SpaceX near $1 trillion and pegged xAI separately at $250 billion, putting their combined worth close to $1.25 trillion. Musk has also raised the idea of stationing data centers in orbit as a longer-term ambition, though nothing in the Q2 disclosures ties dollars to that plan yet.
Anthropic and OpenAI are building differently, and the contrast shows SpaceX’s ambitions are large but not unique in scale. Claude trains and runs across three separate chip suppliers, Amazon’s Trainium chips, Google’s TPUs, and Nvidia GPUs, rather than one. Anthropic’s disclosed pipeline includes as much as five gigawatts of added Amazon capacity, divided across Trainium and Graviton hardware, with another 3.5 gigawatts from Google’s TPUs. OpenAI leans more on Nvidia but still spreads its bets: roughly ten gigawatts of committed Nvidia capacity, six gigawatts from AMD, and two gigawatts of Amazon’s Trainium chips. Measured against those combined totals, one company chasing ten gigawatts on its own is ambitious, but not an order of magnitude beyond what OpenAI and Anthropic are already assembling across multiple suppliers.
Today’s newsletter also covers Anthropic signing a reported $10 billion, six-year cloud deal with a separate provider and building out its own chip design team. Taken together with the SpaceX numbers, the pattern is a frontier lab buying, leasing, and now designing compute from every direction at once rather than betting on one supplier.
For operators tracking AI infrastructure costs, the SpaceX numbers are a signal to watch compute leasing terms as closely as chip purchase announcements: capacity contracts between labs, not just training runs, are now moving the revenue lines that matter.
This account is based on reporting by Maximilian Schreiner for The Decoder, published August 5, 2026.
