Somewhere a platform team is filling in a vendor review this week, and one line on it asks where the inference runs and who owns the building. For this slice of Anthropic's future capacity, the answer on the public record runs like this: a site in Nueces County, Texas, developed by Hut 8, into which Lambda deploys Nvidia chips [5]. Lambda is in talks to raise up to $3bn at a valuation it has discussed at $12bn or more, Bloomberg reported [19]. The commitment is roughly 2.9 times that valuation [26].
The two write-ups of the same Journal scoop do not agree on the landlord. The Next Web has Lambda holding the lease, with Nvidia separately contracting with Hut 8 to secure the capacity [5]. Bloomberg's account, carried by the Straits Times, says the Journal reported that Nvidia would hold the lease itself [6]. In one version Lambda rents from a developer; in the other it operates inside a building its chip supplier controls. Anthropic declined to comment, and Nvidia, Lambda and Hut 8 did not immediately respond [22].
The scale is easier to establish than the structure. More than $100bn to AWS over a decade for five gigawatts and Trainium3 access [10], $50bn to Fluidstack and $45bn to SpaceX [12], $45bn over six years to Nscale [9], $35bn to Lambda [1], $10bn to a week-old cloud startup and $9.1bn to Riot Platforms on a 20-year term [11]. That is at least $294.1bn of disclosed commitments [24], against a run rate that passed $30bn earlier in the year [14], or close to ten times current revenue in forward obligations [25]. None of it is money already spent [15].
Former bitcoin miners keep appearing here for an unglamorous reason. They hold the grid connections, land and power contracts that took years to assemble and are now harder to acquire than chips, per The Next Web, which counts Hut 8 as the second miner in Anthropic's supply chain after Riot [16][17]. Texas keeps winning the siting decision on cheap land, a separate grid and permitting that moves faster than most [18].
The procurement slide for all this reads well: many suppliers rather than one, which is a rational preference for a company whose product is a single model family [28]. The failure modes do not divide as neatly as the contracts do, because most of those providers deploy the same vendor's silicon in buildings that vendor helped secure, which is the concentration question The Next Web puts underneath the diversification [27]. Nvidia has spent the year assembling exactly that apparatus, helping cloud providers get the financing and property they need to deploy its systems [30].
The test worth applying to any dependency shaped like this uses two columns instead of a headcount. Column one: how many suppliers you have signed. Column two: how many independent ways the capacity stops arriving or reprices. Anthropic scores high on the first and, on what has been reported, lower on the second, since Nvidia's balance sheet sits behind the tenant, the site and the hardware in a transaction it is not formally party to [7]. Run the same two columns on your own critical vendor and the useful conclusion is that supplier count is not insurance. What you can price instead is contract language you are allowed to read, notice periods on price changes and term length, plus keeping the eval and prompt layer portable enough to run against another model. That portability costs you the model-specific tuning that probably made Claude the right pick, and you pay it whether or not the Texas grid ever misses a beat.
