SpaceX acquired xAI, inherited two GPU data centers, and started renting the capacity to Anthropic for $15 billion a year. That single contract is worth roughly 80% of what SpaceX earned from 23 years of building rockets: the company posted $18.7 billion in total revenue for full-year 2025. The neocloud era has a very clear opening act.

The deal’s structure matters. According to SpaceX’s S-1, Anthropic has taken Colossus 1 (estimated at 300 megawatts) and a portion of Colossus 2, totaling roughly 500 megawatts of live capacity. Analyst Jamin Ball, writing in Clouded Judgement on May 22, flags that the implied rate of $30 million per megawatt is well above the $10 to $12 million per megawatt that other neoclouds typically charge. The premium may be explainable: the contract carries a 90-day cancellation clause, and Anthropic is acutely compute-constrained. A buyer willing to pay spot-like rates for guaranteed capacity with an out is not the same buyer as one locking in a five-year agreement.

The SpaceX deal is a data point inside a much larger forecast. Ball runs the math from a top-down capacity estimate: if OpenAI and Anthropic alone are planning to reach roughly 55 gigawatts of capacity by 2030, and they represent 30 to 40% of all new AI builds, the industry needs to add more than 150 gigawatts in the next 4.5 years. At approximately $50 billion per gigawatt for a Blackwell-dominated data center, total capital expenditure comes to roughly $7.5 trillion. Spread across 4.5 years, that is about $1.7 trillion per year, or roughly 5% of annual US GDP.

Ball’s historical comparison is the 1880s US railroad buildout, which ran at approximately 6% of GDP. That comparison deserves more weight than the article gives it. The 1880s railroad boom was followed by cascading defaults, a banking crisis, and the consolidation of hundreds of competing railroads into a handful of survivors. The investors who financed the buildout did not uniformly profit from the buildout. The analogy Ball offers as context for the opportunity is also the template for how an overbuild destroys capital.

The $13.5 trillion enterprise value figure compounds the uncertainty. It is derived by applying current neocloud public-market multiples (roughly $90 billion of enterprise value per live gigawatt, based on CoreWeave and IREN trading data) to the projected 150-gigawatt buildout. Ball acknowledges this is “funny math.” The number is the product of today’s market multiples applied to a forecast that is itself the output of Claude being asked to estimate industry demand. A 20% neocloud capture rate then yields the $2.5 trillion figure. Each of those steps carries a wide confidence interval, and they multiply together.

The structural point underneath the math is sound regardless: hyperscalers cannot build everything, AI labs are compute-constrained on timelines that procurement cycles cannot fix, and independent infrastructure operators willing to carry construction risk can charge premium rates for fast capacity. CoreWeave’s annualized run rate guidance of $18.5 billion by end of 2026, on a target of 1.7 gigawatts, implies a more modest $11 million per megawatt rate. The business model works at that price. The question is whether it works at the scale the forecast requires, or whether a supply glut in gigawatts 50 through 150 compresses rates toward the cost of capital.

The operator-level implication is specific. If you are negotiating a multi-year compute contract in 2026, you are not only pricing for inference cost and availability. You are taking a position on neocloud counterparty solvency three to five years out. The 90-day out in the SpaceX deal is useful precisely because Anthropic appears to understand that concentrating dependency on a single infrastructure provider whose primary business is launching rockets carries risks that standard SLA language does not cover.

Watch CoreWeave’s revenue per megawatt through the back half of 2026. If rates compress toward $11 million as more capacity comes online, the $90 billion-per-gigawatt enterprise value math that anchors the $13.5 trillion forecast will need to be revised downward, and the neocloud boom thesis will look more like a neocloud cycle.

Posted by Jamin Ball in Clouded Judgement on 2026-05-22.