A power-system event, not a server room
Ten gigawatts does not fit in a colo cage. It is the kind of load that makes grid planners clear a calendar. That was the eventual scale Crusoe and Tallgrass, a Blackstone-backed infrastructure company, put on a campus in southeastern Wyoming.
Bloomberg reported in June 2026 that Crusoe had sold the unfinished project to Google. TechCrunch reported in early November 2026 that the Denver developer had confirmed the sale and that Google is taking over development. That account said Google had raised concerns about cost and construction timeline under Crusoe. Google has not disclosed the purchase price. TechCrunch also reported that Google plans to invest $17 billion to develop the campus.
That is a large check for an unfinished site, and an incomplete story. There is no public inventory of what was late, what was over budget, or a Google document that spells those concerns out in Google’s own words. Whether the 10-gigawatt target, Tallgrass’s role, or the original schedule survives the handoff is not established. A sale does not tell you which pieces of the announcement still stand.
The bottleneck moved down the stack
Most of the AI conversation still orbits models and GPUs. The Wyoming handoff is about the layer underneath: who can turn a power number into a building that actually comes online.
A single gigawatt is already a serious industrial load. Ten is a regional planning problem. Specialized electrical gear, long construction lead times, and interconnection queues sit between a rendering and a live cluster. Those constraints explain why this kind of project slips. They do not prove why this one did. The reporting points to cost and schedule. It does not name a local dispute.
Power did not replace the chip as the only scarce input. GPUs are still hard to get. The constraint moved down the stack, onto substations, switchgear, and calendars. Hyperscalers can absorb a redesign and a multi-year wait. An independent developer lives closer to the calendar it promised customers and lenders.
TechCrunch described Google as a prospect Crusoe had tried, and failed, to sign as a tenant for this particular campus. Google was not a customer already in the building. Buying the site is a different relationship than leasing it. One is a contract. The other puts construction risk on a balance sheet built to carry it.
Do not turn a sale into a eulogy
The tidy reading is that independents cannot deliver mega-campuses on time and on budget, so the physical layer of AI will belong only to a handful of giants. That is a fair worry. It remains an argument. It is not a finding.
On September 17, 2026, after Bloomberg’s June report and before TechCrunch’s November account of a confirmation, Crusoe announced the initial closing of an anticipated $3.9 billion Series F. The company, in a release datelined Denver, called the round oversubscribed, at a $30.9 billion post-money valuation. Chase Lochmiller, co-founder and CEO, was telling an execution story. “Initial closing” and “anticipated” matter. This is not proof that every dollar of $3.9 billion had already landed.
NVIDIA was among the investors named. When the company selling the scarce chips also sits on the cap table of a data-center developer, the stack is tightening. That is how a supply-constrained industry behaves. By itself, it is not a master plan.
On operations, Crusoe said it had more than 6 gigawatts of gross contracted capacity, 1 gigawatt delivered and running, and more than $140 billion in total contracted value. Those are company figures. No outside audit backs them. They also cannot be read as a post-confirmation accounting of the Wyoming handoff. The release does not mention the sale, and a November confirmation does not reveal what “gross contracted” still included in September.
What the numbers do rule out is simpler. A 10-gigawatt Wyoming plan still sitting inside Crusoe’s contracted book would not fit under a total of more than 6 gigawatts. By September, that 10-gigawatt figure was either an aspiration rather than contracted capacity, or already off Crusoe’s books. Public materials do not settle which. They also do not say whether Google keeps the target.
The reported Google relationship did not end at the state line. TechCrunch still has Crusoe developing a separate data center for Google near Amarillo, Texas, a different site from the Wyoming campus. The same coverage describes an Abilene campus Crusoe is building in phases for Oracle and OpenAI, and a separate 900-megawatt campus Crusoe is constructing next to that Abilene site for Microsoft, with a first building expected in mid-2027. Crusoe’s September release calls Abilene a campus it designed and built, which can sit alongside phased work. Amarillo’s terms are not public, and none of these write-ups itemize which halls are energized.
Crusoe’s answer to the delivery problem is vertical integration: originate power, shorten field construction, and sell compute in more than one form. It describes an energy-first model across grid power, batteries, nuclear, thermal, and renewables. That is strategy. It is not evidence of what happened to electrons or permits in Wyoming. The company says its modular product, Crusoe Spark, cuts field work from years to weeks, and it has been selling inference (compute used to run models) alongside space. Crusoe said Managed Inference, launched late the prior year, had more than $100 million in contracted annual recurring revenue, and that Crusoe Cloud bookings had grown more than 20 times year over year. Those growth figures are the interested party’s account. The useful point is the split between mega-campuses and smaller compute sales.
“The economics flow to the lowest-cost producer of intelligence.”
That line is from Gavin Baker, an investor quoted in the financing announcement. It is a slogan. It is not an engineering study. It is still the right frame. A software margin does not describe this cost. The cost is power price, construction time, and who can carry a campus that slips.
If you buy independent capacity, you are renting someone else’s substation
Most companies will never pour a slab in Wyoming. They will buy inference so AI automation, internal tools, and AI agents actually answer. That is the practical layer of AI for business. Some will lease a hall and bring their own GPUs. Some will rent the GPUs too. The commercial form matters less than a blunt question: is the capacity real, or is it a rendering with a logo on it?
A campus announcement is a promise. An energized megawatt is a fact. If you are on a hyperscaler that already owns the yard, you are not renting a stranger’s substation. If you are buying from an independent developer or a wholesale provider, you might be. And if that vendor sells the site halfway through, your roadmap just inherited someone else’s construction risk. That can be rational for the developer. It is not automatically fine for the buyer who thought “contracted” meant “online.”
The Wyoming sale, as reported, does not prove independents are finished. It does show that a mega-campus is closer to a utility project than to a software launch, and that, on this site, Google took the construction problem onto its own books rather than waiting on someone else’s schedule.
The questions the headline skips
- Did Google buy Crusoe’s Wyoming project?
Bloomberg reported the sale in June 2026. TechCrunch reported in early November 2026 that Crusoe had confirmed the sale of the unfinished southeastern Wyoming campus and that Google is taking over development. The price has not been disclosed.
- Was 10 gigawatts a committed plan?
Crusoe and Tallgrass had described an eventual scale of 10 gigawatts. Whether Google keeps that target, Tallgrass’s role, or the original schedule is not public. Crusoe’s September figures do not show a 10-gigawatt project still inside its contracted book.
- Did Google formally blame Crusoe for cost and delays?
TechCrunch said Google had reportedly raised concerns about cost and timeline. That is reported attribution. There is no public Google accounting of what slipped.
- Does the sale show independents cannot deliver mega-campuses?
Not as a finding. Cost, equipment, and interconnection make these builds hard, and one unfinished campus changing hands is a real warning. It is not a verdict on every independent developer, and Crusoe’s September financing does not erase it either.
- What should a company buying AI compute check?
Ask who controls the power, who finishes the building if the developer sells, and whether your capacity is contracted or merely announced. A valuation and a gigawatt target are not a live cluster.
The scarce resource in this buildout is not only a clever demo. It is also a substation with a date on it. Leaders who treat AI as a pure software purchase will keep being surprised by concrete, copper, and interconnection queues. The ones who ask who owns the power, and what happens if the developer sells the site mid-pour, will make fewer expensive assumptions.