Divide the commitment by the site. Thirty-five billion dollars over five years is about $7 billion a year [15]. Set that against roughly 350 megawatts and you get about $20 million per megawatt-year [16]. The $35bn and the 350MW both come from the source Reuters spoke to [1][2]; the five-year term comes from the dev.to account of the deal [9].
Now price the building. Hut 8 said in July that its 15-year lease with an investment-grade customer carried a base-term contract value of $19.6bn [7], and the Financial Times reported Nvidia was the tenant of the 1GW Beacon Point campus [8]. That works out to roughly $1.3 million per megawatt-year [17], about fifteen times cheaper than the capacity price [18]. The two numbers may not describe the same asset, and the sources do not say the Nueces County project is part of Beacon Point. Even so, the direction is unambiguous: the shell and the power contract are a rounding error against the accelerators, the networking, the operators and the margin stacked on top of them.
For that $20 million per megawatt-year to mean anything for your workload, several things have to hold. You have to be buying five-year committed capacity rather than on-demand hours. Your fleet has to be dense H100, H200 and Blackwell of the sort the source described [10]. Power has to be inside the number rather than passed through. And you have to not need the managed storage, identity and orchestration that came free-ish with a hyperscaler contract, because at this layer you build them. A lab with its own training stack clears all four of those conditions; most buyers, by contrast, clear none of them.
What changed here is not independence. Anthropic already holds billions in funding and cloud credits from Google and Amazon [11], and dev.to reads the Lambda deal as a pivot away from depending on firms that ship competing models [12]. But according to the Wall Street Journal the lessee on this data center is Nvidia itself [3], the reseller is Nvidia-backed [1], and the hyperscaler path leads to the same upstream anyway, with AWS reported to be ordering another 2 million Nvidia chips for 2027 and 2028 [20]. The landlord changed; the underlying dependency on Nvidia did not.
What I would underwrite carefully is the counterparty. Lambda has raised $1bn of short-term debt to buy Nvidia chips to rent to Microsoft, and is preparing a further round [13]. A $7bn-a-year book on that balance sheet is a financing question as much as an engineering one. Meanwhile Anthropic added $45bn of Nscale capacity the week before [5], which is $80bn of rent announced in about a week by a company heading for an IPO [19][6]. Nobody in that procurement team is short of paperwork.
