Project Braid's schedule is being written by transformers, governors, and consent

Project Braid, the internal name for the cloud venture Google and Blackstone formed in May, has started slipping against its own timetable, as Bloomberg reported this week and Data Center Dynamics carried on September 9. The sites that slipped, from Wyoming to Texas, say more about transformers and governors than about the venture's balance sheet.

Blackstone committed $5 billion of equity from its funds to back a cloud platform built on Google's Tensor Processing Units, with the partners aiming to bring 500 megawatts online in 2027 and expand after that. The venture sits outside a conventional Blackstone data-center lease: in a lease, delay leaves a building empty; in a cloud venture, delay leaves the compute itself unbuilt, which is a different category of risk for a $5 billion equity check.

Cheyenne, Wyoming, is the clearest case. Crusoe had been pursuing the site as a 1.8-gigawatt campus with room to scale toward 10 gigawatts, and paused work in June at the request of its customer; Google took over the project the following month, and the development has since been scaled down. Bloomberg's sources attach an internal explanation to the handover, but the sequence is the fact an underwriter should keep: one month a site sits in a developer's pipeline, the next it sits in a hyperscaler's, and the megawatts shrink.

The rest of the list is less dramatic but no easier. One project location was deemed unsuitable because it would require the wrong kind of transformers; in Texas, Governor Greg Abbott's decision to freeze new projects has created a separate obstacle. Equity cannot accelerate transformer production, and a cloud company's balance sheet carries no weight against a governor's freeze.

Benjamin Treynor Sloss, the venture's CEO, told Bloomberg by email that "it is very common for individual potential sites to progress at different rates," that the possibility had been built into the JV's planning, and that the venture holds a "diversified set of options" and is making "strong progress." Bloomberg's sources put the site list at 29 potential locations.

A 29-site list is optionality. But every option carries the same expiration problem: the 2027 target only works if a meaningful share of those sites can move through interconnection, equipment ordering, and local approval on the same calendar, and none of those calendars bend to a $5 billion check.

In the interconnection queues and load rules, grid access is becoming an underwriting variable; Cheyenne and Texas make the same point with a construction schedule attached. Consent, not capital, is the binding constraint on data-center construction: a filed permit is not a stable property right, and a customer's request to pause can rearrange a project in a matter of weeks.

The same constraint shows up elsewhere in Blackstone's infrastructure book: in late August, the EU approved Blackstone's €2 billion Eurowind acquisition, a deal that will be settled by grid connections and planning consent as much as by the assets themselves. The lesson from the Fervo-Google arrangement was that delivery is the test, and Project Braid is now taking that test at a different scale, with a cloud business riding on the outcome.

None of this means the JV is failing. A diversified pipeline of 29 sites should look like this: some options fail, some stretch, a handful carry the project, and the CEO's "strong progress" language is consistent with a portfolio being actively pruned. The underwriting question is whether the 2027 date can survive a year in which Cheyenne shrank, Texas froze, and transformer specifications became a gating item; the safer model treats 500 megawatts as a 2028 event and any 2027 capacity as upside.

Watch the site list itself. If 29 starts shrinking, the 2027 language in the next update will matter less than the list; if it holds, the JV can plausibly get to 500 megawatts later than advertised and in smaller first steps. The next update's site count, more than the CEO's adjectives, will show whether 2027 has already become 2028.