A megawatt in one rack is 100 to 200 times the draw of the 5 to 10 kilowatt rack it replaces [1], and four times the top of the 100 to 250 kilowatt band the current generation already sits in [2]. The cabinet keeps its footprint and takes two orders of magnitude more electricity. That, rather than the fund size, is the part a product plan has to survive.
The density number needs a closer look. a16z pairs a 28-fold rise in compute density between Nvidia's H100 generation and its Rubin racks [3] with a power range whose starting vintage it never states [4]. The two ratios do not share a denominator [4], so the pitch establishes that the cabinet consumes more without telling you whether the token coming out of it costs fewer watts.
The part closest to a product decision is memory and interconnect, which a16z lists among its target areas alongside power-efficient edge devices and the cooling, materials, electrical and real estate work around a facility [11]. Its own description of the failure mode is worth quoting: a rack of accelerators that cannot be fed data fast enough is an expensive way to generate heat, and the firm says the industry spent two years discovering how often that is the actual limit [9]. Teams tend to read a slow agent as a model problem and go shopping for a better model, but a16z's data suggests the same feature is often bandwidth-bound, a fault no model swap will fix.
Where the cheques land says more than the mandate does. One recent one is a Series A into Netris, which automates the networking that slows down GPU clouds [16], and the portfolio the fund builds on reads "systems" widely enough to put Skydio, SpaceX, Anduril and Waymo next to Nexthop, Volta and Mind Robotics [17]. At $1.1bn against the more than $15bn a16z announced across new funds in January [15], this is roughly seven percent of the year's declared firepower [3], which makes it a specialisation rather than a reallocation.
So the forcing function for the next roadmap review is two numbers written next to every AI feature: tokens consumed per user action, and the number of seconds after which your users actually abandon, measured rather than assumed. High tokens with a short fuse is the quadrant that inherits the substation queue and the cooling bill, and it either gets smaller (distilled model, cached retrieval) or comes off the interactive path. High tokens with a long fuse can be batched toward whichever campus got power. Low tokens with a short fuse is the argument for the edge silicon a16z says it wants to fund [11]. Low tokens and a long fuse, do what you like. Most teams have never written the first number down, which is how a feature ships with an unpriced dependency on a rack that does not exist yet.
