Aug 28, 2026 · 53m · a16z
Why Top Founders Are Racing Into AI Infrastructure
Ben Horowitz · 17m spoken Martin Casado · 15m spoken Raghu Raghuram · 9m spoken Erik Torenberg · 5m spoken
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of The a16z Show, Andreessen Horowitz partners Ben Horowitz, Martin Casado, and Raghu Raghuram announce the Machine Age Fund and explain why the frontier of artificial intelligence is now governed by physical infrastructure, silicon, power, and datacenter constraints rather than software alone.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The host holds 10.5% of the talking time here. How this is scored →
The host as informed peer 3.7 Guest teaching 4.7 Guest disagreement 0.6 The host pushing back 0.3
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 34:15 Ben challenges gigawatt metric usage
Ben playfully interrupts Erik and Martin to challenge the casual tossing around of gigawatts, demanding a concrete physical breakdown.
Hardest push from the host ▶ 4:23 Erik presses on real demand vs hype cycleErik directly challenges the assumption of infinite demand by asking for concrete proof that this is not simply another ephemeral hype cycle.
Biggest teaching moment ▶ 26:44 Martin lays out economics of per-model ASICsMartin educates the host on the unprecedented digital artifact economics where spending billions on models justifies custom per-model ASICs to recoup inference costs.
The host holds their own ▶ 28:07 Erik cites precise datacenter density metricsErik demonstrates deep subject familiarity by citing specific shifts in rack kilowattage, compute density multipliers, and liquid cooling transitions.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| The a16z Show Animated Title Sequence | 3 | 4 | 1 | 0 | Erik introduces the new Machine Age fund citing Marc Andreessen. Ben, Raghu, and Martin collaboratively break down the fundamental hardware and supply bottlenecks driving the initiative. | |
| Infinite Demand and the Global Supply Crunch | 4 | 5 | 1 | 1 | Erik asks how they know demand is real rather than a hype cycle. Raghu, Ben, and Martin explain hyperscaler capex, sold-out supplies through 2028, and unprecedented chip price increases. | |
| Why AI Infrastructure Scaling is Unique | 3 | 6 | 1 | 1 | Erik probes why the fund was not created earlier. The guests explain how prior computing epochs had minor hardware shifts compared to the current thousand-percent token growth and physical limit bottlenecks. | |
| The Mechanics of Demand and Direct Resource Scaling | 4 | 5 | 0 | 0 | Erik asks about the multi-agent token multiplication effect. Martin and Ben explain the shift from engineering bottlenecks governed by the mythical man-month to direct capital-to-compute resource scaling. | |
| The New Startup Paradigm and Long-Term Compute Horizon | 4 | 4 | 0 | 0 | Erik connects past VC skepticism about excessive capital to today's massive compute clusters. Raghu, Ben, and Martin walk through expanding demand across knowledge workers and embodied AI. | |
| GrokBot and the Rise of AI Coworkers | 3 | 4 | 0 | 0 | Erik asks about GrokBot and organizational dynamics. Martin and Ben discuss treating bots as autonomous coworker entities rather than simple software extensions while managing security and behavior. | |
| Engineering the AI Datacenter: ASICs, DC Power, and Physical Limits | 5 | 6 | 1 | 0 | Erik asks detailed technical questions citing specific power densities and liquid cooling figures. Raghu, Martin, and Ben detail bespoke ASICs, DC power risks, electrical contractor shortages, and structural building obsolescence. | |
| Fund Mandate, Robotics, and Datacenter Power Constraints | 4 | 5 | 2 | 1 | Erik asks why hyperscalers cannot build power capacity faster. Ben interrupts teasingly to define what a gigawatt actually means before outlining severe grid, turbine, and regulatory bottlenecks. | |
| The Philosophy Behind the Name 'Machine Age' | 4 | 5 | 1 | 0 | Erik asks about the name Machine Age and whether incumbent giants like NVIDIA will dominate. Martin and Ben point to the law of market fragmentation and specialized use cases. | |
| The Profile of Infrastructure and Systems Founders | 4 | 5 | 0 | 0 | Erik references Patrick Collison's observation regarding the age of tech founders. Ben, Raghu, and Martin explain why physical manufacturing and complex supply chains reward systems experience. | |
| Partner Expertise and the Future of American Infrastructure | 3 | 3 | 0 | 0 | Erik wraps up by asking about firm expertise and the long-term vision. Ben delivers an inspiring closing message about American technological competitiveness. |
Statements from this episode (24)
Raghuram: AI bottlenecks have shifted from models to physical infrastructure
“What we have seen over the last three years is the steady increase of the capabilities of the models. Where the model is no longer the bottleneck. And in fact, using AI, these models are getting better faster and faster and faster. Now the bottleneck is all wh…”
Casado: Top founder hardware pitches jumped from 5% to over 30%
“I don't know the actual numbers, but I was trying to estimate it over the weekend. So I think we'd get, you know, maybe five percent of the deals from top founders would come in, would be harder before. Now it's the north of 20% or 30% right now.”
Raghuram: Hyperscaler capex hits $700 billion this year, $1 trillion next year
“Next year, supposedly, it's going to reach a trillion dollars collectively across the bay.
[296] Raghu Raghuram: Hyperscamers this year, it's about 700 billion dollars, right?”
Casado: Global AI hardware supply is basically booked out through 2028
“I mean, the supply, if you look at the supply across the board, it's basically all booked out to twenty-twenty-eight.”
Horowitz: Market intermediaries are reselling GPUs for four times purchase price
“This is like, we're flat out, and people are reselling GPUs for
[514] Ben Horowitz: Four times what they bought them for”
Raghuram: Current AI memory demand requires three years of manufacturing capacity
“So the flagship conference for the industries, the one called Hot Chips, is going on in Stanford, and the leading memory ranger said the demand they have today will take them three years of capacity to supply.
[588] Raghu Raghuram: It's just today.
[589] Raghu…”
Casado: Perfect foresight couldn't have prevented AI infrastructure capacity shortages
“So even if we had a perfect Oracle, once it started working, I don't think we could have built the capacity. We could have built the capacity. There's no way. And we're talking about like chip cycles, which tend to be three to four years. We're talking about b…”
Horowitz: AI compute demand will not decline until humanity lacks problems
“Any problem that you have can be solved with enough infrastructure and money. And so until we run out of problems, we're not going to run out of demand.”
Horowitz: Billions in compute can now instantly erase a startup's multi-year lead
“The one thing we all knew in startup world is that if I have a two-year lead on you, and you try and cast me by hiring a thousand engineers, you're gonna wreck your company. Like, that never works. It's a mythical man month. Nine women can't have a baby in a m…”
Raghuram: AI agents for knowledge workers will multiply compute demand tenfold
“There are over a billion knowledge workers in the world, right? And with that, it's a long ways to go for that demand. And by the way, the work that they do, all this workflow and automation, So on, and then you get to the back office, which is all the agents.…”
Casado: AI compute demand will persist for decades like electricity
“The right analog here is like the steam engine or electricity. In the following way, like, we've introduced this new thing that you can turn to work, and there's some very obvious applications now, but there's probably 3040 years of throwing computer problems.…”
Casado: AI agents work best as autonomous digital employees with dedicated computers
“GrokBot got really right is, no, how about if it's actually an employee? So now you have this thing that's an entity, and it doesn't have, like, special access to your keys or whatever. It has its own computer, and it has its own browser, and because these are…”
Casado: Multibillion-dollar AI training costs justify building per-model custom ASICs
“Today to build a frontier model costs, let's say, three to five billion dollars, right? So, and let's, you know, and that's to train it. And so the inference has to pay back at least that. Of course, right? You know, in order for any of this stuff to be viable…”
Horowitz: AI datacenters exceeding 100 kilowatts per rack require DC power
“First of all, when you get to that level of power per rack, AC power doesn't work anymore. So like, that's a pretty wild thing. So now, now you're into DC power, which by the way, also requires its own cooling.”
Horowitz: Only 2 percent of US electricians are certified for DC power
“Only two percent of electrical engineer, er electricians in the U.S. Have been certified on DC power.”
Raghuram: Major cloud providers are experimenting with robots for racking servers
“The big guys that own the big cloud data, big data centers, they all are furiously experimenting with robots, right, to do the work of assembling or putting servers into the data center, etc. And so you will see that increasing as a result of the evolution in …”
Horowitz: AI compute demand is growing 10x annually, outpacing physical supply
“The demand is growing. You know, 10 X a year right now, and, you know, the supply just can't grow that fast.”
Casado: US power constraints force startups to deploy GPUs in Mexico and Australia
“By the way, it is so bad that right now if we have new companies going for GPUs, it's often in Mexico or Australia or in other country just because it is so difficult in the United States.”
Raghuram: Achieving 10x AI efficiency requires first-principles hardware startup innovation
“If you want to have a 10 X on those metrics,
You gotta have new innovation.
And new innovation traditionally comes from building founders thinking about solving the problem from first principles in a different way, right?
And that's what's needed here for t…”
Horowitz: Nvidia's massive core growth causes it to ignore smaller AI opportunities
“So the last thing I'm doing is looking at a silver brick, and I think NVIDIA's in that position.”
Casado: Hardware optimization will directly dictate AI business margins and upside
“We may actually be entering an era where the optimization in the hardware is absolutely meaningful to the upside of the business in a way that we haven't seen in the past.”
Casado: Desperate AI labs ink commercial deals before startup hardware even exists
“The first one is the labs are so desperate that they will engage with startups. And so, like, we actually have quite a bit of signal early on because they're, you know, the labs are, Inking deals with companies before they actually have hardware available. And…”
Casado: AI hardware startups will produce a transformational alumni founder wave
“Like, listen, we're gonna create a whole generation of, you know, founders that come from these new companies that will know how to do this, and, you know, they'll be hired in much more junior. Like, I would say, actually, one of the greatest legacies of Elon …”
Horowitz: Losing US technology leadership means competitors will define global values
“We'd like to keep that going, and I think that doesn't continue to go if we lose our lead in technology. I think we'll be in another era, and there'll be another country, and maybe they have a different set of values around that.”