Intro
a16z co-founder Ben Horowitz and partners Martin Casado and Raghu Raghuram introduce the firm's Machine Age Fund and argue the binding constraint on AI has moved below the model, into chips, memory, networking, power, cooling and the buildings themselves. The conversation runs from GPU auctions and hyperscaler capex to per-model ASICs, DC power, the shortage of certified electricians, and why data center construction is heading offshore.
Guest: Ben Horowitz, co-founder and general partner at Andreessen Horowitz (a16z)
Also on: Martin Casado and Raghu Raghuram, a16z general partners launching the Machine Age Fund alongside him
Published: 28 August 2026 on a16z Podcast
Episode page | 55 min
Key Takeaways
The next bottleneck in AI sits below the model, not in it
Raghuram: "Now the bottleneck is all what I call south of the model"
Every GPU coming off the line is already spoken for
Casado: "Basically, every GPU that's being created is already pre-sold"
Supply across the board is booked to 2028, with multi-day auctions running for a few thousand GPUs
Money now converts directly into capability, which never used to work
Horowitz: "It's not hiring 100,000 engineers. It's taking $3 billion and lighting up a magnificent cluster"
Hyperscaler capex is running near $700 billion this year and is expected to reach a trillion next year
A chip built for a single model can now pay for itself
Casado's math: a frontier model costs three to five billion dollars to train, inference has to pay that back at least twice over, and a 20% efficiency saving on $10 billion is roughly what an ASIC costs to build
Above a certain rack density, AC power stops working entirely
Horowitz: "when you get to that level of power per rack, AC power doesn't work anymore"
The shortage has reached the trades
Raghuram: "Only 2% of electricians in the U.S. have been certified on DC power"
Reinforced concrete is among the fastest-rising input prices, alongside memory
Agents work best when the firm treats them as employees rather than as tools
Casado on what the newest agents got right: "how about it is just actually an employee"
Horowitz: they can burn tokens and get nothing done, forget things, make things up and create security problems
Incumbents leave room because the openings are too small to interest them
Horowitz relays Dan Rose's line about silver bricks and gold bricks, and says Nvidia is in the same position
Hard problems have gone from a sliver of top founder pitches to a fifth or more
Casado's own weekend estimate: from maybe 5% to north of 20% or 30%, with Horowitz calling 5% generous
Construction is moving offshore because it is easier to build almost anywhere else
Casado: "we're creating huge jobs, both job and long-term economic opportunity in other countries by banning data centers here"
Demand for tokens is compounding far faster than supply can be added
Horowitz: "the demand for tokens is probably going to grow close to a thousand percent a year"
A New Fund for Everything Underneath the Model
The host opened by reading a line from Marc Andreessen on the scale of the shift: "This is the biggest technological revolution of my lifetime. This is clearly bigger than the internet. The comps on this are the microprocessor, steam engine and electricity, or maybe the wheel."
Horowitz's case for the fund is that a new computing era forces a new physical layer under it. New chips, new system software, new ways of doing power, and a replacement for copper
"We have a whole new technology that's the most important technology ever, and you need a whole new infrastructure." — Ben Horowitz
Raghuram said the word infrastructure now stretches much further than it used to. In computing it meant servers, storage and networking; here it runs all the way down to the copper mine
The models are no longer the constraint, in Raghuram's telling — they are improving faster because AI is being used to improve them. "Now the bottleneck is all what I call south of the model" — Raghu Raghuram
Casado said a16z follows founders, and the founder mix has shifted hard toward difficult problems
His weekend estimate: hard problems used to be maybe 5% of the deals coming in from top founders, and are now north of 20% or 30%. Horowitz said 5% was generous and put it nearer 3%
"the founder community, which tends to be much smarter than the VC community, has identified this as a very active area for innovation, and they're responding" — Martin Casado
Horowitz's explanation for the shift is that "the demand for AI is basically infinite", which puts every part of the supply chain under duress, down to the materials that go into memory
Casado added that AI inverts the usual startup worry. Growth is not the question; margin is, and much of the inefficiency is a physical limitation of hardware built for other workloads
How They Know It Is Demand and Not a Hype Cycle
Raghuram's first piece of evidence is the hyperscalers, who see demand from frontier labs, AI-native companies, enterprises and every geography at once. Their capex is about $700 billion this year and is expected to reach a trillion dollars next year
"if anybody has visibility, it is down. And they're jacking up their capex, like it's never been seen before" — Raghu Raghuram
His second is the portfolio itself: application companies and frontier labs are all growing at rates he called insane
Casado's evidence is the price curve, which has broken its historical direction. GPU prices always went down, and now they are going up
"If you look at the supply across the board, it's basically all booked out to 2028." — Martin Casado
He said they have seen multi-day auctions run for a few thousand GPUs
Raghuram added that the value of a unit of AI work keeps rising while the tokens consumed per unit rise by orders of magnitude — a hundred tokens for a chat becomes thousands for an agent
The user base expands underneath that too: developers first, then all knowledge workers, then the back office
Sold Out to 2028, and Why It Is Not the Dark Fiber Era
Casado said he cannot recall an industry being sold out this far ahead. In the internet build-out most of what went in the ground was speculative and dark
"Basically, every GPU that's being created is already pre-sold." — Martin Casado
Horowitz drew the distinction directly: the late-1990s bandwidth glut was two-sided, and this is not. Companies rushed to build capacity that theoretically had demand, but there were not enough people on the internet to consume it, and the high-bandwidth uses like video were not viable yet
"we're flat out, and people are reselling GPUs for four times what they bought them for" — Ben Horowitz
On top of that, he said, the industry is out of power and cooling, and building is hard because of political headwinds. He called it unprecedented in his career
Raghuram offered an anecdote about a CFO at a large public company that had resisted the cloud and kept its own servers. An inventory check found "the memory in their servers had increased so much it could fund the entire migration to the cloud"
"We're out of many things. Power, cooling, memory, GPUs, you name it. We're out of it." — Ben Horowitz
Raghuram, relaying from the Hot Chips conference at Stanford, said a supplier there reported that "the demand they have today will take them three years of capacity to supply. It's just today. It's not even future demand."
Why the Fund Would Not Have Worked Five Years Ago
Casado said nobody could have built ahead of this even with perfect foresight. The industry is four years into it, chip cycles run three to four years, and breaking ground on a data center runs four to five — before the power problem
"So either have to build your own power or usually both. You've got to build your own power and have power source, which is not easy." — Martin Casado
Raghuram framed the mismatch as an industry growing at 20 or 30% being connected to an AI software industry growing far faster, so the gap keeps widening
Casado's historical point is that every epoch produced an independent infrastructure winner, but the change was small enough that the opening was narrow. Client-server, then the internet with Cisco and Juniper, then Arista in the era of mega data centers driven by cloud providers verticalizing
The difference now is that everything is changing at once, not one chip company and one switch company
Horowitz said the firm could probably have started a couple of years earlier
He put the growth rate that supply has to chase at close to a thousand percent a year. "the demand for tokens is probably going to grow close to a thousand percent a year" — Ben Horowitz
Every company that has adopted AI is growing its usage fast, most companies have barely adopted it, and consumers are just getting started
Raghuram added that the existing categories are hitting the physics limits of what they were designed for, so the next step needs technical breakthroughs rather than more of the same
Why Every Step Up in AI Multiplies Tokens
The host noted that each move — chatbots to reasoning to agents to multi-agents — has multiplied the tokens a single task consumes
"nobody likes to use AI more than AI." — Ben Horowitz
Casado's explanation is that today's scaling method is inference-heavy by construction. Reinforcement learning is a lot of inference, chain of thought is a lot of inference, and long-running agents are a lot of inference
The deeper shift, in his framing, is that software has stopped being an engineering problem and become a resource problem. Adding engineers had a natural governor, which is where the mythical man-month came from; pouring money into systems does not
"we're bottlenecked on those systems' ability to actually match the resource for pouring into them" — Martin Casado
With no natural regulator, he said, the trend should be expected to continue and supply has to be built to support it
"any problem that you have can be solved with enough infrastructure and power and money. And so until we run out of problems, we're not going to run out of demand." — Ben Horowitz
Raghuram put it as AI's answer to getting better being to use more AI, with inference the building block it reaches for over and over
Casado called the loop auto-catalytic: using AI to write a GPU kernel is using more AI to make more AI. In the old model, money went in, an engineering project took two years and usually failed. Now nothing sits between the money and the hardware creating intelligence
Money Now Buys What Headcount Never Could
The host raised the long-standing complaint that too much money was going into startups and venture capital, and the shift to a market that is as big as the industry collectively funds
Horowitz said the oldest rule in startups has just been repealed. A two-year lead used to be uncatchable, because hiring a thousand engineers to close it wrecked the company
"Nine women can't have a baby in a month. That's it. That never works. Okay, now that works." — Ben Horowitz
"It's not hiring 100,000 engineers. It's taking $3 billion and lighting up a magnificent cluster." — Ben Horowitz
He named Grok and Kimi as examples of models that can come out of nowhere on that basis
"you can throw money at almost any problem, and that works. And so that is just completely different than anything we've ever lived through." — Ben Horowitz, who added that everyone is still psychologically adjusting to it
From Chatbots to Computer Use
The host put the scale of current usage on the table as the show's own figures: the ChatGPT app has a billion weekly actives, and about 30 million developers account for a large share of compute demand
Raghuram described the progression as casual use, then coding for professionals, then tools for knowledge workers, then the back office running on agents — each step unlocking roughly an order of magnitude more demand
Horowitz said computer use is the wave now doing to general work what coding agents did to software. He used an agent over the weekend to update his credit card across a set of services and cancel subscriptions he had been too lazy to deal with
His framing: it is not coding, it is a knowledge work case that happens to sit inside the computer
Casado said Marc Andreessen's steam-engine and electricity analogy is the right one, because a new thing that can be turned to work takes decades to find its applications. Language and code are done, computer use is just starting, and science, materials, biology and creativity are still ahead
"we're on the very, very early part of a very long journey" — Martin Casado
"So let's expect this compute need to persist for decades." — Martin Casado
The host noted they had not even reached embodied AI or robots, which will be another source of demand
Managing an Agent Like an Employee
Casado traced three stages in how the industry has thought about putting AI into products. First a feature bolted onto a product, like a search bar; then a chat interface, because that was the traditional way to do it
The next stage, which arrived earlier in the year with OpenClaw, made the agent a standalone extension of the user — sharing your keys, knowing your passwords, doing what you would do
"how about it is just actually an employee" — Martin Casado, on what the newest agents got right: an entity with its own computer and its own browser rather than access to yours
His own test has become whether an agent can just do the task. Calendar management and booking meetings are the obvious cases; reading and triaging his email is the non-obvious one, and it knows to check with him before acting
Horowitz's read is that this is a new class of employee that a company has to learn to work with, the way it spent years learning to work with human ones. They can burn a lot of tokens and spend a lot of money getting nothing done, forget things, make things up, behave well or badly, and create security problems, and they can also be enormously productive
"I don't want to sit up here and say, I've cracked the code." — Ben Horowitz, rejecting the idea that the firm is quietly automating humans out
"how do we make all our human superhuman without wrecking the place because the bots get out of control" — Ben Horowitz
Raghuram said the firm tried several ways of getting agents into its systems and the one that stuck was Casado's: treat them as people
What a Data Center Designed for AI Looks Like
Raghuram's method is to break the machine into fundamental components and rebuild it around what inference actually does. An inference engine consumes a great deal of memory and generates new tokens alongside the compute
The questions that follow: what the memory should be, what the compute should be, how they talk to each other, how much power each needs, how each is cooled, and how collections of them are assembled
The same questions repeat at every level of interconnect — on the same chip, across chips, and across data centers
Casado offered a mental model for how far the architecture could go: a custom chip per model. A frontier model costs three to five billion dollars to train, inference has to pay that back at least twice over, so on $10 billion a 20% efficiency gain is worth $2 billion — roughly what it costs to build an ASIC
"it actually makes sense to build an ASIC per model just because the amount of capital investment in that model" — Martin Casado
Unlike traditional software, which is dynamic and full of state, model weights are fixed
"I don't think in the history of the industry, we've ever created a digital artifact with something like $5 billion that went directly into that artifact." — Martin Casado
The host laid out the physical changes underway: compute density climbing something like 70X, and liquid cooling becoming a requirement rather than an option
Horowitz's first answer is that the electrical architecture breaks before anything else does. "when you get to that level of power per rack, AC power doesn't work anymore" — Ben Horowitz, which pushes the industry to DC power, which needs its own cooling and is dangerous
He noted the irony that "Edison promoted DC power by claiming how dangerous AC power was, demonstrating it by like electrocuting animals"
He argued the political environment now sets a higher bar than the engineering does. State-of-the-art data centers are already on liquid cooling, but "It's got to be eco-friendly liquid cooling", and power has to contribute to society rather than take from it
He put the share of data centers behaving badly on water at probably 10%, adding that it is not as much as pistachios or almonds despite what gets demonstrated on the internet
He said the industry has gone through a one-way door on this, and the engineering it requires has not been done yet
Denser racks also break the building. Floors have to carry the weight, and walls have to be thick enough to contain the noise or no state will permit them
"once we get to Feynman, a much smaller percentage of the data centers that we have today work" — Ben Horowitz
The Trades Problem Nobody Budgeted For
"Everybody talks about memory prices, but one of the fastest areas that prices is increasing is reinforced concrete" — Raghu Raghuram
Raghuram said sending 800 volts to the rack creates a second problem beyond the danger: there are not enough people qualified to do the work. "Only 2% of electricians in the U.S. have been certified on DC power" — Raghu Raghuram
Meta runs a program to train people up for free, which he compared to a new job corps
"AI is taking all the jobs. AI is going to create a lot of new electricians." — Ben Horowitz
Raghuram added that the large cloud operators are all experimenting with robots to assemble servers and put them into the data center, and expects that to grow
What the Fund Will and Will Not Touch
Casado defined the fund's scope as computer science infrastructure — anything a model runs on. Chips, networking, interconnect, storage, and probably down to the electricity
On robotics, his framing is that AI's breakthrough is letting computers interact with the physical world — seeing, hearing and talking — which creates new platforms
He dismissed "edge device" as a term that does not mean anything. The platform could be a mobile device, a CDN, a laptop, or an embodied device that moves around
As infrastructure investors, he said, they do not do heavily regulated or verticalized industries, but any computer science platform that pushes AI further out is of interest
The Power Bottleneck, and the Word Gigawatt
The host put the grid arithmetic to them as the show's own figure: "by 2028, new data centers are going to need something like 44 gigawatts of additional power against maybe 25 gigawatts of expected grid additions"
Horowitz stopped the conversation to object that nobody using the word gigawatt knows what it means. Casado's answer: multiple football fields, and the power of 50,000 homes
"I grew up in Flagstaff, Arizona, which is a town of 40 to 60,000 people, depending on the universities. We have less than a gigawatt of power consumption." — Martin Casado
"Basically, light up and air condition your entire town for a gigawatt." — Martin Casado
"everybody talks about the gigawatt. There's very few gigawatt data centers that are actually up. We've got a long way to go." — Ben Horowitz
Casado's answer on why utilities and hyperscalers cannot simply build faster is that none of the constraints are software constraints. Construction still needs humans, and before that come permits and access to power
Getting grid or natural gas access is a regulatory bidding struggle with limited capacity to tap
Building your own power runs into shortages of transformers, turbines and everything that goes into them
"This is not a software problem. It's not just like a bunch of engineers, can't you like work weekends and that type of stuff." — Martin Casado
"The demand is growing 10x a year right now. And the supply just can't grow that fast." — Martin Casado
Why Data Centers Are Going to Mexico and Australia
Horowitz said new companies are frequently siting GPUs in Mexico, Australia or another country because building in the United States is so difficult
"we're creating huge jobs, both job and long-term economic opportunity in other countries by banning data centers here" — Martin Casado
Horowitz's proposed fix is a standard rather than a ban. "the right answer would be to set a standard where a data center contributes back to the community" — power gets better, there is no noise, no water issue, and it adds jobs, with everyone held to that bar
He said this is not futuristic: there are data centers doing it now, where energy rates have gone down every year
The mechanism he described: the facility generates its own power and gives power to the state during the day, then borrows at night when the state does not need it. Power plants always generate peak capacity, and a data center's flat day-and-night draw is symbiotic with a city's peaky one
Why "Machine Age" and Not "Artificial Intelligence"
Casado's first argument is that the field was misnamed. "Artificial intelligence was the wrong word. We shouldn't have called it. It's machine intelligence." — Martin Casado
His reason: it is not how humans think. Nobody knows how to take an AI with no knowledge, put it in the world and have it reconstruct language. What has been built learns off everything humans have already learned
He noted the term goes back 70 years in computer science and carries baggage from science fiction and from Nick Bostrom
The second argument is the irony of the position the software-eating-the-world camp now finds itself in — pouring money into something and being limited by the machines below it. The name is a nod to how significant the hardware component has become
"It's also a cool name. It sounds good. It's futuristic." — Ben Horowitz
Why Nvidia Does Not Take All of It
The host asked why incumbents like Nvidia would not simply take the lion's share, given how capex-intensive the businesses are
Raghuram's answer is that the next 10x on the metrics that matter has to come from first-principles innovation. Tokens per second per dollar, tokens per watt, tokens per rack — any of them
Casado called it the law of markets. The silicon incumbents are worth multiple trillions, so even 5% of that is a massive private company, and the incumbent has no reason to chase it while the other 90% is growing just as fast
The same question was asked about Amazon during the cloud years and about Microsoft before that
"Once you get to a certain scale, there's tremendous opportunity for innovation at the margins." — Martin Casado
Horowitz told the story behind the principle. a16z partner Alex Rampell was trying to sell his startup TrialPay's services to what was then Facebook, and Dan Rose, then head of corporate development, turned him down
"It sounds like you can collect a lot of silver bricks, but I'm like, I have so many gold bricks. I can't even pick them all up." — Dan Rose, as relayed by Ben Horowitz, who said Nvidia is in that position
Casado's structural claim is that growth fragments a market and only a slowdown consolidates it. He used Ford's Rouge River plant, where water, coal and rubber went in and cars came out, and the company city Ford built in the Amazon to own the rubber supply
That worked until Ford made people show up to things on time. Today the car industry has multiple tiers of suppliers
"as markets expand, they fragment. And then once that growth slows down, they tend to consolidate" — Martin Casado, with consolidation coming through acquisition or through new challengers
Horowitz added that the biggest company can reach the biggest use cases, but the use cases are multiplying faster than any one firm can serve them
Casado's final point is that good margins may no longer be automatic. Software delivered them once the business worked; AI may not, which makes hardware optimization directly meaningful to the value of the business
The Kinds of Companies, and the Kinds of Founders
Raghuram listed the subsectors: compute chips first, but a chip is no longer enough — the company has to build a full system. Memory innovation, networking innovation and power tooling all follow, and each category can support a public-company-scale business
On top of that sits a layer of software to automate and manage the fleets
The host noted that first rounds for these companies have been massive, in the hundreds of millions
Horowitz said the defining difference is that a lot of money goes in before there is a product, with more risk and more capital than the firm's usual deals
"a lot of the chip founders are here from the past. Like the guys who know how to make memory, they're not young." — Ben Horowitz
Raghuram said these have to be systems founders, not researchers. They have to architect the chip or system and simultaneously think about how it gets manufactured and who supplies it
"of course, Jensen is the Michael Jordan with this" — Raghu Raghuram, on founders who think about the entire ecosystem before they start designing the chip
Casado named two conditions that did not exist five years ago. "the labs are so desperate that they will engage with startups", inking deals before the hardware exists, which gives investors early signal
The second is capital availability: there is consensus that this is the moment to rebuild the stack, so follow-on rounds are well funded
Where the Young Founders Went
The host raised Patrick Collison's remark from a few years ago that there seem to be fewer young founders now than in the era of Zuckerberg or Gates, allowing for the Michael Truells of the world
Horowitz's answer is that a complicated supply chain and a manufactured product reward experience. Elon Musk and Travis Kalanick both started with software companies and needed that experience before graduating to more elaborate domains
His test: building a company is hard enough when you completely understand the product. Learning the product while building the company is a brutal curve for a first-time founder
He said Truell built a pure software AI company and could probably do the harder thing ten years from now, but not today
Casado added that hardware was defocused by industry and academia for twenty years, so the pool of people who have done it is thin
"You don't go intern and build a chip." — Martin Casado, who expects the new companies to produce a generation of founders who will be hired in far more junior
Raghuram argued Elon Musk's greatest legacy may be the founders who left, not the companies he built. "the amount of entrepreneurs that have come out of SpaceX that are changing the entire industrial complex, maybe even greater legacy than the companies themselves" — Raghu Raghuram
He said one of the firm's own investments was started by two founders in their twenties, with experienced people visible the moment you walk through their offices
Horowitz's qualifier: the founder does not have to have the experience, but had better be able to tap into it in a real way
Asked what the world looks like in five to ten years if the fund works, Horowitz described an America with abundant chips, memory and power, and "lots of like super eco-friendly, efficient data centers out there"
"It's kind of the best place to come with nothing and do something profound." — Ben Horowitz, who said the lead in technology is what keeps that true
Horowitz's bottom line is that the constraint on AI has moved out of software and into the physical world, and that the firms which relieve it — in chips, memory, power and cooling — decide whether the abundance shows up in America or somewhere with a different set of values.
Products, Companies & Tools Mentioned
Nvidia (The incumbent the fund's thesis has to get past; Horowitz says it is in the position of a company with more gold bricks than it can pick up, so the margins are left to startups)
Grok (Horowitz's example of a model that came out of nowhere once money could be turned into a cluster; he also used a Grok agent over the weekend to update his credit card across services and cancel subscriptions)
OpenClaw (Casado's example of the previous stage of agents — a standalone extension of the user that shares your keys and passwords, rather than an entity with its own computer)
ChatGPT (The host's marker for scale: a billion weekly actives)
OpenAI and Anthropic (The sweet spot the host and Casado agree is a tough place to build into; outside their three-to-five focus areas is where the room is)
Cisco, Juniper and Arista (Casado's proof that each infrastructure epoch produced an independent winner — but only one or two, because the change was small)
Meta (Runs a free training program to certify workers on DC power, which Raghuram likened to a job corps; also the company, then Facebook, that turned down Alex Rampell)
SpaceX, Astranis and Waymo (Earlier a16z checks into hardware and systems, which Casado cites as evidence the competency is not new)
Kimi (Named beside Grok as a model that can appear suddenly at the frontier)
TrialPay (Alex Rampell's startup, and the source of the silver bricks and gold bricks exchange with Dan Rose)
Ford (Casado's case study in vertical integration and then fragmentation — the Rouge River plant, and the company town Ford built in the Amazon to control rubber)
Amazon, Microsoft and Google (The incumbents Casado invokes when asking why the giant does not simply do it itself; Google as the buyer a hardware startup could not have approached five years ago)
Hot Chips (The industry conference at Stanford Raghuram was relaying supply commentary from)
Books & Resources Mentioned
Fordlandia (The book behind Casado's account of the city Ford built in the Amazon jungle to own his rubber supply)
The Mythical Man-Month (The engineering law both Horowitz and Casado invoke and then declare repealed — adding people used to be the natural governor on speed, and capital now is not)
Get the latest market chatter and takes as they happen:
X | Threads | Instagram | YouTube | TikTok | Facebook