Crusoe has closed the initial part of a $3.9bn Series F round, the Denver company said on 17 September, putting its post-money valuation at $30.9bn. Atreides Management, Mubadala Capital and Valor Equity Partners co-led the round, which Crusoe described as oversubscribed. The size matters because the company is betting that AI compute will be built where power is available rather than where land is cheap.
Who is paying for the round
The investor list runs well beyond the three lead funds. Founders Fund, GIC, Nvidia, the Qatar Investment Authority, Radical Ventures and TPG took part, and the full list of further backers reaches 29 names, among them ARK Invest, Baillie Gifford, Fidelity, Salesforce Ventures, Tiger Global and the research firm SemiAnalysis. In July, TNW reported that Crusoe was in talks to raise $3bn; the round closed higher than that target. Crusoe now reports more than $140bn in total contracted value, more than 6GW of gross contracted capacity of which 1GW is operational, and cloud bookings up more than 20 times year on year. The company employs more than 1,800 people across five countries and has opened offices in Bellevue and New York.
How the modular model works
Crusoe builds modular data centres, called Spark, in its own factories and then trucks them to locations with available power, the Wall Street Journal reported. That approach cuts construction in the field from years to weeks, according to Crusoe. The logic rests on a difference between two workloads: serving a model to users needs far fewer chips than training one. Chase Lochmiller, co-founder and chief executive, told the Journal that a site like Abilene is not required for inference and that running inference from such a campus can be a bit of overkill. A plant outside Denver will eventually produce up to a gigawatt of Spark capacity a year, he said, and units already run in Reno, Nevada, on old electric car batteries and solar panels.
The Abilene data centre campus in Texas, which Crusoe built, is where OpenAI trained Astra, the company said. That project represents the large-scale end of the business, while Spark addresses the opposite case. Crusoe earns money at three layers, as Lochmiller put it to the Journal: it sells data centres, GPUs and tokens. Its managed inference product launched late last year and has passed $100m in contracted annual recurring revenue, according to the company. Lochmiller told the Journal that this business went from almost no revenue at the start of 2026 to that run rate by the summer. The strategy has at times unsettled Crusoe's board, which has suggested narrowing the focus, the Journal reported, citing people with knowledge of the conversations.
What this means for business
For companies that buy AI capacity, the practical consequence is a shorter path from decision to running workload. A modular site delivered by truck changes the unit of planning: instead of waiting for a large campus to be energised, a buyer can attach compute to an existing power source, which matters most for inference-heavy products with steady demand. Small firms gain access to capacity that previously required joining a queue for hyperscale space, while large enterprises can place smaller sites near their own operations and keep latency and data flows under control. The constraint shifts from construction schedules to the availability of power and the terms on which it is supplied.
Several questions remain open before this becomes a standard procurement route. Crusoe does not disclose the cost per unit of Spark capacity, the delivery times for a specific order, or how the economics compare with renting from a hyperscaler over a three-year horizon. The Reno deployment relies on repurposed electric car batteries and solar panels, a configuration whose reliability and maintenance profile differs from a grid-connected facility, so buyers should ask about uptime guarantees, service levels and who handles battery replacement. The $100m in contracted annual recurring revenue is a contracted figure, not recognised revenue, and the $140bn in total contracted value spans multi-year commitments. A signed contract also does not by itself prove that inference demand is growing at the rate the company's bookings suggest.
The marker to watch is the Denver plant: whether it reaches the stated ceiling of up to a gigawatt of Spark capacity a year, and whether Crusoe discloses delivery times and unit economics alongside that figure. If modular capacity ships on the promised weeks-long schedule and customers renew inference contracts beyond the first term, the model becomes a credible alternative to conventional data centre procurement. If the plant ramps slowly or the contracted backlog converts into revenue more slowly than booked, the trucking approach stays a niche for edge cases rather than a shift in how AI capacity is bought.
