Anthropic commits about eight times its annual revenue run rate to 14.8 gigawatts

The five headline deals add to $505 billion of the $517 billion total ^[4][1]. Capacity reconciles less neatly. Amazon, Google, Microsoft, Lambda and nScale account for 12.46 of the 14.8 gigawatts ^[5]. SpaceX's roughly $45 billion and Fluidstack's $50 billion of data centres in Texas and New York carry no gigawatt figure in cryptobriefing's account ^[7][9]. The two together are $95 billion, 18 percent of the money ^[6].

Where the capacity is stated, the implied prices differ by nearly six times. Amazon and Google's $300 billion for 11 gigawatts works out at about $27 billion a gigawatt, and Microsoft's $30 billion for 1 gigawatt is not far off it ^[5][1][6]. Lambda and nScale's $80 billion for 460 megawatts implies roughly $174 billion a gigawatt, on a six-year term against the decade the hyperscaler contracts run ^[8][2][3][1].

OpenAI has said publicly it wants $750 billion and 30 gigawatts by 2030 ^[13]. Anthropic's 14.8 gigawatts is 49 percent of that capacity for 69 percent of the money ^[7]. Anthropic's book prices a gigawatt at about $35 billion, OpenAI's target at $25 billion ^[8]. The horizons are not the same, since OpenAI's target ends in 2030 and Anthropic's contracts run ten years ^[13][1].

Spread over ten years, $517 billion averages $51.7 billion a year, or about 80 percent of the roughly $65 billion annualized run rate Anthropic now reports ^[11][9]. The whole book is close to eight times that run rate ^[10]. The earlier figure was $180 billion of server leasing through 2029, so the commitment has grown by $337 billion ^[3][12].

The exposure runs both ways. Amazon and Google have, cryptobriefing wrote, "effectively locked in a decade of revenue from a single customer" ^[15]. Behind them, the same account puts Samsung and SK hynix among the memory and foundry suppliers that benefit directly ^[14]. The combined energy draw compares with more than ten advanced nuclear reactors ^[10].

All of it rests on "up to" ^[1]. The report gives contract values and terms in years. Without minimum purchase obligations, a ceiling with milestones and a firm take-or-pay bill look identical from outside. There are two readings. If the hyperscaler tranche is firm, the cost base under Claude inference for years is set by paper already signed, and the listing is the financing event for it ^[12]. If the smaller contracts are ceilings written to be announced, the $517 billion overstates what gets paid, and the $174 billion a gigawatt at Lambda and nScale is where the slack sits ^[2]. I lean to the second for the smaller tranche, because it prices six years at nearly six times the rate two hyperscalers accepted for ten ^[3]. Minimum purchase obligations in an S-1 would settle it.