a16z fences off $1.1B for power, cooling, memory and networking

Take the rack numbers as a spec sheet rather than a slide. Moving from 5 to 10 kW per rack to 100 to 250 kW is a 10x to 50x increase in power delivered into the same floor position ^[8][13]. The 1 MW rack a16z expects within three years is another 4x on top of that range ^[8][15]. These are a16z's own projections, and the announcement carries no independent forecast ^[8], so the useful question is which of the two numbers you are buying against. At 250 kW you are placing purchase orders. At 1 MW you are filing for a utility interconnect. Those are different calendars, and only one of them speeds up when you add capital.

That distinction is why a fund like this behaves unlike a software fund. a16z's own framing is that computing infrastructure is being redesigned down to the electricity, with economics closer to industrial development than to the low-capital software businesses venture returns were recently built on ^[6]. Industrial development means money in front of revenue for years, on schedules set by permitting and construction. The fund says it wants to "open the throttle and accelerate the physical buildout of AI" ^[2]. Throttles assume the fuel line is already connected.

Now the number I would push on at review. a16z says hardware and infrastructure already accounts for over 20% of its recent startup deal flow, and discloses neither a time period nor a deal count ^[7]. Twenty percent of forty deals in one quarter and 20% of four hundred over three years describe different firms with different conviction. Without a denominator it is positioning.

The roster is better evidence than the percentage. Horowitz co-founded Loudcloud in 1999, turned it into Opsware, software for automating data-center operations, and sold it to Hewlett-Packard in 2007 ^[9]. The lesson a16z draws from that period is what happens when a technically sound product arrives before the hardware and connectivity around it are ready [c9b]. Casado's Nicira made network infrastructure programmable and went to VMware for $1.26 billion in 2012 ^[11]. Ulevitch's OpenDNS went to Cisco for $635 million in 2015 ^[12]. All three were software sold into infrastructure, not manufacturing. a16z seems to know it, which is why the announcement names Guido Appenzeller, Shangda Xu and Erin Price-Wright as hardware and infrastructure specialists alongside the five partners who presented the fund ^[4][3].

Fund arithmetic sharpens the problem. The $1.1 billion reserved here is 87% of what Nicira sold for ^[14]. For one Nicira-scale exit to return this fund, a16z would have to own nearly the entire company at exit, which venture ownership does not deliver. So the thesis requires outcomes larger than the ones on its partners' resumes, in a category where every company carries manufacturing and supply-chain risk stacked on top of engineering risk.

The 2009 franchise was organized around the argument that software would eat the world ^[5]. This one is capital aimed at the layer where the answer is amperage and heat rejection, and where a rewrite buys you nothing.