"We've actually seen multi-day auctions for a few thousand GPUs."
That's Martin Casado, describing the current market. Not a metaphor. Founders and hyperscalers fighting for days over a slice of silicon that would fit in a single rack, because supply across the board is "basically all booked out to 2028."
Hold that image, because it's the catch buried inside a16z's most heretical claim to date.
Last week we noted, in the abstract, that capital had started converting directly into capability — that money was buying outcomes software labor used to gate. This week a16z put a name and a number on it: a $1.1 billion Machine Age Fund, and an explicit economic law to justify it.
Here's the law, from Ben Horowitz. "The one thing we all knew in startup world is that if I have a two-year lead on you and you try to catch me by hiring 1,000 engineers, you're gonna wreck your company. That never works. It's a Mythical Man-Month. Nine women can't have a baby in a month."
Then the inversion: "Okay, now that works."
But read the next sentence closely, because it's where the whole thesis quietly narrows. "But it's not hiring 1,000 engineers. It's taking $3 billion and lighting up a magnificent cluster, and then all of a sudden, whatever, Grok can come out of nowhere and, oh, all of a sudden it's real."
The Mythical Man-Month didn't break. It broke for one specific thing: model training, where a magnificent cluster substitutes for the two years you'd otherwise spend closing a lead. That's real.
Because the same pitch that says "you can throw money at almost any problem" also says the money doesn't work unless the physical thing exists to spend it on. And it mostly doesn't.
Raghu Raghuram's framing is the one that matters for founders: "The model is no longer the bottleneck. Now the bottleneck is all what I call south of the model." Memory. Interconnect. Power. Cooling. The copper mines, in his words.
So the $3B cluster only lights up if you can get the GPUs, the rack power, the behind-the-meter electricity, the interconnect that copper can no longer carry. And those are the things being auctioned over multiple days for a few thousand units at a time.
Which means the actual scarce asset isn't capital. Capital is abundant — hyperscaler CapEx hits roughly $1 trillion next year, up from about $700 billion this year. What's scarce is a defensible position south of the model that isn't already sold out to 2028.
Capital is abundant — hyperscaler CapEx hits roughly $1 trillion next year. What's scarce is a defensible position south of the model that isn't already sold out to 2028.
This is the reframe every product person should internalize. "Capital equals capability" holds only where physical supply exists to absorb the capital. Everywhere else, capability is gated by supply-chain access, and access is the moat.
So where does startup value actually get captured? a16z's own logic answers it, almost by accident, through a quote about gold bricks.
Their partner Alex Rampell apparently pitched the silicon incumbents on a market opportunity. The response, as relayed: "Alex, that's great. It sounds like you can collect a lot of silver bricks. But I have so many gold bricks, I can't even pick them all up. So the last thing I'm doing is looking at a silver brick."
Casado says Nvidia is in exactly that position today. Multi-trillion-dollar market cap, growth pouring in from the 90 percent of the stack it already owns. Why would it stoop to pick up the marginal problem?
That's the opening. "Even 5% of that is a massive private company," Casado notes. The natural law of markets: at sufficient scale, the incumbent rationally ignores opportunities that would still be enormous for a startup.
He's said this before, in a different decade: "We asked these questions during the cloud days. 'Why wouldn't Amazon do this?'" Amazon could have. It didn't, because it was busy with bigger gold bricks.
So the product opportunity is not "build the next model." That's the gold brick — the crowded, capital-intensive fight where the $3B cluster substitutes for engineers. It's also where the incumbents are pointed.
The opportunity is to own a specific physical constraint the incumbents are too big to bother fixing. A silver brick to Nvidia is a gold brick to a founder. Cheaper, higher-bandwidth memory. Faster interconnect where copper has hit its wall. Rack power that has to climb from 100–250 kW today toward 1 MW in three years. Cooling. The behind-the-meter power that data centers increasingly supply themselves because the grid can't build fast enough.
Each of those is a constraint currently rate-limiting the entire trillion-dollar buildout — and none of them is the model. Casado's own numbers reportedly show founders already redirecting toward exactly this: hardware went from what Casado calls "maybe 5%" of top-founder deal flow to "north of 20% or 30% right now."
That's, per Casado's approximate figures, roughly a four-to-six-fold shift in what the best teams choose to build. It's the leading indicator of where value capture is moving, and it points south of the model.
The demand under all of this is not speculative. Casado uses Grok Bot only as demand-side proof — the shift from AI as a search bar, to AI that chats, to AI that behaves like an actual employee with its own computer and browser doing your email triage. Each rung, per Raghuram, "unlocks an order of magnitude more demand." A billion knowledge workers. Then robots. The token intensity of the work keeps climbing.
That's why the supply chain is "under duress" all the way down to the materials used to make memory. The demand signals, in Casado's words, "have never been clearer." The bottleneck is, per a16z, largely on the supply side — which is precisely why supply-side positions are the ones worth owning.
Now the honest caveat, and it's a big one.
This is not a play most software PMs can run, and no raise fixes that. a16z is explicit: hardware founders are older. Per Horowitz, "You don't go intern and build a chip."
Horowitz's own logic on why: "When you're learning how to build a company, it's hard enough if you completely understand the product. If you don't completely understand the product and have to learn it while you build the company, that's just such a steep learning curve for a brand new entrepreneur."
Even Musk and Kalanick, he reportedly notes, started with software companies and needed experience before graduating to "the more elaborate domains." Hardware has been "defocused by the entire industry and academia for the last 20 years."
The profile that wins here is hardware-native — and increasingly, SpaceX-alumni. Horowitz argues the greatest legacy of that company may be the entrepreneurs it produced who are now "changing the entire industrial complex," and he expects the same wave for computer science and hardware.
Which is the part the fund's own framing won't say plainly: capability is now gated by two things, and neither is code. One is supply-chain access booked out to 2028. The other is a founder profile you can't manufacture with a bigger check.
So the takeaway is sharper than "AI infrastructure is hot." It's this: the Mythical Man-Month broke for exactly one activity — buying a training run with a cluster — and that activity is the crowded gold brick everyone can see.
The defensible positions are the silver bricks. The specific physical constraint — memory, interconnect, power, cooling — that's rate-limiting a trillion dollars of demand and that Nvidia won't stoop to pick up.
Owning one of those is a company. But it's a company built by someone who understands the product before they start — which, for once, is a moat the money can't buy.