Fifty is the number doing the work, or ten, depending on which ends of the two ranges you pair: a 250 kilowatt AI rack against a 5 kilowatt legacy rack is fifty times the draw, and the cautious pairing of 100 against 10 is still ten times [4][1]. Both readings look backwards. The mandate is set by the forward figure, because a megawatt per rack inside three years is four times today's ceiling and a hundred times the old floor, and a hall sized for the old floor does not get there by adding fans [5][2].
The other number a16z put on the table, 28 times more compute density from Nvidia's H100 to Rubin, invites a comparison the numbers don't actually support [6]. Dividing 250 by 10 gives 25, close to 28, but the 5-to-10 kilowatt rack was never an H100 rack, and the source gives no draw for the H100 generation at all [5]. The two figures describe different populations of steel.
Then the size of the cheque book. The $1.1 billion is about 7.3 percent of the more than $15 billion a16z says it has raised across its vehicles, and about 65 percent of the $1.7 billion Infrastructure Fund 2 that already covered a broader set of investments [1][7][3][4]. Hardware was more than 20 percent of recent deal flow before this close, with prior positions in Unconventional AI, Mind Robotics and SpaceX [8]. A dedicated vehicle at two-thirds the size of one the firm already runs, formalising a fifth of what it was doing anyway, is a scouting budget for component sockets: memory architecture, where the source says workloads are now bound by the speed of feeding the processor rather than the processor itself [3], and the interconnect that has to carry thousands of GPUs talking to each other during training [10].
The view, and it may well be wrong: the power curve is real, and the fund is sized correctly as an option rather than a capacity play. The risk is capture: the layer that reprices is the layer the incumbent keeps inside its own reference design, and a roadmap that moved density 28 times in one generation is a roadmap capable of absorbing a cooling problem [6]. The counter-thesis sits in the same source and arrives hedged: incumbents including Nvidia and CoreWeave are reportedly willing to leave margin on the table for founders attacking these problems from other angles, which, if it holds, is most of the returns case [9]. What would falsify the scouting read is a follow-on hardware vehicle at a multiple of this one. What would falsify the thesis is racks settling near 250 kilowatts because grid interconnection binds before thermals do [4].
