"You Don't Actually Need an Abilene to Do That": Crusoe's $3.9B Bet on Factory-Built AI Data Centers
Crusoe just pulled off one of the largest private financings in AI history — and it plans to spend the money making its own biggest achievement optional. The company, which spent the last two years building one of the largest AI data-center campuses on earth in Abilene, Texas, has raised $3.9 billion at a post-money valuation of about $30.9 billion, the Wall Street Journal reports.
The round was co-led by Atreides Management, Valor Equity Partners and Mubadala Investment, with Founders Fund, TPG, the Qatar Investment Authority and several other sovereign-wealth funds participating, according to the Journal. The headline number roughly triples Crusoe's valuation in about eleven months: the company was worth a little over $10 billion when it closed a $1.38 billion Series E last October.
But the real story is what the money is for. After two years of mega-construction in Abilene, Crusoe is now betting heavily on the opposite of Abilene: smaller, factory-built modular data centers it calls "Spark" — manufactured in its own plants, shipped on flatbed trucks, and deployed wherever power is available. Faster and cheaper than large construction projects, the Journal reports.
The pivot: data centers you can put on a truck
The logic, as chief executive Chase Lochmiller explained it to the Journal, is about the difference between training AI models and running them. Giant clusters suit training, but inference "can often be done with considerably smaller amounts of chips," he said. Running inference from a facility like Abilene, he added, "can be a bit of overkill."
His punchline: "You don't actually need an Abilene to do that."
Spark is the physical expression of that argument. Crusoe's first manufacturing facility sits outside Denver in Brighton, Colorado — 352,000 square feet, eventually capable of producing up to 1 gigawatt of Spark capacity a year, the Journal reports. The company has put more than $200 million into the plant and its initial Spark fleet, according to AI Weekly. A second, 400,000-square-foot facility in Tulsa, Oklahoma opened on September 14, bringing more than 300 local jobs and producing switchgear, switchboards, electrical enclosures, industrial controls and copper busbar. Spark units are already operating in Reno, Nevada, powered by old EV batteries and solar.
Lochmiller's ambition for the program is characteristically blunt: "We're going to put them everywhere."
The Abilene behemoth isn't going anywhere
For all the talk of going small, Crusoe's giant campus remains the foundation of the business. The company developed eight buildings in Abilene for Oracle, which rents the compute to OpenAI; after Oracle and OpenAI signed deals with other developers, Crusoe leased the additional buildings to Microsoft instead, the Journal reports.
The campus is planned for roughly 2.1 gigawatts of power capacity, including a 900-megawatt on-site gas plant for Microsoft's portion, according to Lochmiller via the Journal. And the pipeline keeps growing: a project in Armstrong County, Texas with Google (the Journal cites people familiar with the project), two more Texas developments and one in Missouri. In total, Crusoe says it has more than 6 gigawatts of data-center capacity under contract.
OpenAI's own financing ambitions put that demand in perspective — the lab's reported $1.2 trillion pre-IPO round, which we covered in our coverage of OpenAI's $1.2 trillion pre-IPO round, is the kind of capital formation that makes even a $3.9 billion raise look like a down payment.
"We sell data centers, GPUs and tokens"
Crusoe's business model is unusual for a so-called neocloud. Most of its peers do one thing — buy chips, rent them out. Crusoe does three, in Lochmiller's formulation: "We sell data centers, GPUs and tokens."
The third item on that list is the fastest-growing. Crusoe runs customer models and charges per token — "managed inference" — a business that went from roughly zero revenue at the start of 2026 to more than $100 million in annual run rate by summer, according to Lochmiller via the Journal. Its cloud customers include Perplexity; its data-center customers include Meta, Microsoft, OpenAI and Oracle, per Bloomberg and TechCrunch background reporting.
The deal flow around the company has been frenetic. Bloomberg reported in early September that Crusoe had signed a $13 billion, five-year GPU and cloud contract with Jane Street — a figure that remains reported rather than confirmed — and Axios reported in August that the company had met with bankers at Goldman Sachs and Morgan Stanley about a potential near-term IPO. That IPO chatter got fresh oxygen on Thursday: Bain Capital Ventures, which led Crusoe's Series A back in 2019, raised a new $1.6 billion fund — its eleventh, 14% larger than its prior $1.4 billion vehicle — and in TechCrunch's coverage of the fund, Crusoe was described as "reportedly valued at $30 billion and viewed as a near-term IPO candidate."
BCV partner Kevin Zhang gave the venture-side thesis its most quotable line: the firm wants to fund compute infrastructure until intelligence becomes "too cheap to meter," he told TechCrunch, adding that "BCV can support founders not just with equity capital, but with debt facilities, infrastructure partnerships, and real-economy relationships." Zhang's framing — that AGI "has already arrived" and the firm is now backing "post-AGI era" themes including infrastructure, healthcare, physical AI and security — is as close as venture capital gets to a mission statement for the buildout.
From oilfields to AI
Crusoe's origin story explains its unusual shape. Founded in 2018, the company started by putting compute next to oil-and-gas sites to soak up excess energy, then served crypto miners, before pivoting to Nvidia-GPU AI compute, the Journal reports. That history — chasing cheap, stranded power wherever it sits — is the direct ancestor of the Spark strategy: if the electrons are in Armstrong County or Reno, build the data center in a factory and truck it to the electrons.
The company has also worked conspicuously to stay on good terms with the communities hosting its megaprojects. It contributed roughly $20 million to the Abilene economy last year, including meals for construction workers; its annual tax contributions to the local school district are expected to exceed the school's current annual budget; and in Warrenton, Missouri, it covered football-game admission for spectators, the Journal reports. In an industry increasingly defined by local fights over power and water, that is not charity — it is permitting strategy.
Analysis: the training-to-inference turn
The following is analysis, not reporting. Step back, and Crusoe's raise is the physical embodiment of 2026's most important architectural shift in AI: the industry spent the last three years building ever-larger training clusters, and is now discovering that the money is in running models — inference — which needs to be everywhere, cheap, and close to power, not concentrated in a handful of monuments.
That is what makes Crusoe's strategy genuinely differentiated. Pure neoclouds buy chips and rent them out; Crusoe is vertically integrating the whole stack — the buildings, the electrical gear, the GPUs, the tokens. Factory-built modules dodge the two constraints choking the industry: construction timelines and skilled labor. A 352,000-square-foot plant that can stamp out a gigawatt a year of data-center capacity is, in effect, an assembly line for the inference era.
The risks are real. A $30.9 billion valuation — tripled in eleven months — prices in flawless execution on manufacturing, deployment and the managed-inference ramp. The reported $13 billion Jane Street contract and the IPO-banker meetings are still just reports. And the $3.9 billion figure itself is single-sourced to the Journal as of Thursday.
But compare it with the other neocloud raising money this week: also read our coverage of CoreWeave's $3B convertible and $2.9B stock program. One company is financing the buildout through Wall Street's debt and equity markets; the other is trying to industrialize the buildout itself. Both are bets on the same underlying fact: the world is about to need a staggering amount of compute, and whoever can deliver it cheapest — per token, per megawatt, per truckload — wins.
Sources: The Wall Street Journal (Sept. 17, 2026) — $3.9B round / ~$30.9B post-money valuation, co-leads and participants, Spark strategy and manufacturing details, Abilene/Microsoft/Oracle facts, Lochmiller quotes, community contributions; AI Weekly — $200M+ Brighton investment, Tulsa facility details (400,000 sq ft, 300+ jobs, product lines); TechCrunch (Sept. 17, 2026) — Bain Capital Ventures $1.6B fund, Kevin Zhang quotes; Bloomberg via TechCrunch (Sept. 3, 2026) — reported $13B Jane Street contract, data-center customer list; Axios (Aug. 2026) — reported Goldman Sachs/Morgan Stanley IPO-banker meetings; Reuters/Bloomberg (Oct. 2025) — $1.38B Series E at $10B+ valuation; SiliconANGLE — $2.77B total raised to date.
Comments
Post a Comment