Last week a16z has launched a $1.1 billion fund for the physical infrastructure underneath AI: chips, networking, memory, cooling, data centers, robotics, energy. Jen Kha, who runs global partnerships at the firm, spent twenty-five minutes on the a16z Podcast explaining the reason, and it isn’t a market observation. It’s a confession that the model doesn’t work anymore, not for this: everything below the software stack has been “largely an uninvestable category for the most part for the last 30 years,” built for a previous era’s internet and now being rebuilt “from a blank sheet of paper.”
Measured against a “classic” deal (software), the “new” deal (hardware) loses the internal capital-allocation argument every time, structurally. You cannot fix that with a bigger check inside the old fund. You have to build a new one.
So, what has changed?
Capital-only engagement is gone, replaced by hyperscaler navigation and market access founders can’t buy elsewhere.
Government moved from compliance line to structural counterparty, the way South Korea’s president made Silicon Valley his only U.S. stop.
Lastly, the capital structure itself concedes the point, $25–35 million at seed for maximum ownership rather than a growth check chasing an inflection point the fund can no longer afford. Read blind, you’d guess this firm was halfway to a different model entirely.
The emperor’s old clothes
It’s still called a fund, and a fund still does one thing: invest capital for equity, at a valuation, on a schedule. The vocabulary hasn’t moved — seed, Series A, “the inflection point” — all of it borrowed intact from the model Kha just spent twenty minutes explaining doesn’t fit the terrain. The mechanics underneath the vocabulary haven’t moved either: a founder has an idea, develops it largely alone, and then *pitches* it. That word is doing unexamined work. Pitching is the language of an investor evaluating a story someone else already wrote, not the language of a venture structured, from inception, around a demand signal the fund itself helped establish.
Three tells, in Kha’s own account:
The favorite example undercuts the thesis. Unconventional, the chip-redesign company she’s proudest of backing at seed, was two people from Databricks “going off in the cave,” found and funded once the concept was already legible. That’s the oldest story in venture capital, wearing a new fund’s name.
The diligence method never got rebuilt. Ask how the fund actually diligences a hardware deal this unfamiliar, and the answer is relationship density: people who already worked at Arista, at VMware, and at Intel.
The doctrine survived the pivot intact. “Follow the nerd energy”, Chris Dixon’s line, repeated approvingly: back the person, trust the network, let conviction about the founder stand in for a named buyer. That’s the software era’s underwriting logic, unmodified, now pointed at chips and data centers.
The wrapper got rebuilt completely. The thing inside it (bet on the founder, let the market discover demand after the fact) never made it onto the table. It’s very hard to interrogate the water you’re swimming in, even while pointing at the sea and saying, correctly, that it has changed.
Two rooms, one blind spot
This shape has shown up once already this year. A few months ago I wrote about Marc Rowan from Apollo and David Haber from a16z, each independently identifying that capital-intensive, physically real ventures had outgrown venture equity and didn’t yet fit private equity either. Blind men, each with a real piece of the elephant: Apollo is a financing-and-acquiring machine with no mechanism for originating a venture from inception. A16z, in that conversation, was a funding-and-advising machine that could accelerate a pipeline but not transform the risk profile inside it.
Rowan and Haber were describing the outside of the problem - how capital should sequence across a venture’s life, who originates, who hands off to whom. Kha’s episode describes the inside of the same failure: not market architecture but underwriting mechanics, not the handoff but the judgment applied before a dollar moves. Neither pair is talking about the other. Put them together anyway and the thing none of them states on its own comes into view.
Rowan’s fix: Proximity and offices near founders, earlier engagement, a tighter pipeline into PE’s acquisition machinery.
Haber’s fix: The same instinct from the other side of the handoff, optimize portfolio companies’ capital structure sooner.
Kha’s fix: A bigger, earlier check, aimed at a harder asset class, sourced the same way software deals always were.
Three real corrections, one unexamined premise underneath all of them: that the founder’s independent bet, discovered and backed rather than designed and structured, is still the right unit of risk. That correspondence is the encouraging part, not the damning one: it means the anomaly is visible from inside the two best-resourced rooms in the industry, in none of the same language, arriving at structurally identical half-measures.
The engine nobody in the room can build
Set the three corrections side by side and they stop reading as isolated fixes made by three firms for three separate reasons. They read as fragments of one alternative none of them has assembled: a venture engineered from inception around a named demand signal, not discovered afterward through a founder’s independent bet. Capital structured across its full instrument range — equity, debt, retainer, engagement fee — instead of defaulting to dilutive equity because that’s the only instrument the fund was built to hold. Government and regulators as structural participants in the venture’s design from day one, not stakeholders managed after the fact. Failure priced in at structuring, so a write-off is a modeled outcome, not a surprise.
Nobody currently holding a fund gets there, however sharp the diagnosis, because getting there means abandoning the thing that makes a fund recognizable as a fund: funding as the primary mechanism, the founder’s pitch as where the process begins. A16z rebuilt everything around that premise. Apollo rebuilt everything around it too. Neither has asked whether the premise itself is what needs rebuilding.
The theory of what replaces it is A New Science of Venture Building. The instrument it produces is the PE-embedded venture builder: demand confirmed before capital moves, the full instrument stack in play, government designed in at inception, failure conditions priced at structuring rather than discovered at write-off. Not a better VC fund. Not a friendlier PE shop. A different animal, built by people willing to leave the fund wrapper behind entirely.
If a16z, or Apollo, or anyone else builds that instrument in the next two years, this argument is wrong, and should be treated as wrong. Until then, the gap isn’t in their diligence. It’s in what their diligence was ever built to see.