Ben Horowitz: We have a whole new technology. That's the most important technology ever. And you need a whole new infrastructure.

Martin Casado: Normally when we talk about the infrastructure world, we're talking about the servers on the storage and the network. Here it goes all the way down to the mines of copper min. That's how widespread this thing is going to be.

Host: It used to be when you built something, it was an engineering problem. And here it feels like it really is a resource limitation. So whether it's tokens or not, we're pouring a ton of money into systems and then those systems are producing a result. And right now we're bottlenecked on the systems ability to actually match the resources we're pouring into them.

Raghu Raghuram: The leading memory vendor said the demand they have today will take them 3 years of capacity to supply.

Host: If this fund does what we think it will do, how do we see the world in 5 to 10 years?

Ben Horowitz: America wins in the infrastructure and that would be awesome.

Host: Ben Martin Ragu, welcome.

Ben Horowitz: Thank you.

Martin Casado: All right. Thank you.

Host: I want to start with a Mark quote to introduce this new fund. This is the biggest technological revolution of my lifetime. This is clearly bigger than the internet. The comps on this are the microprocessor, the steam engine, and electricity, or maybe the wheel. Guys, the machine age fund. Please introduce it. Ben, start us off.

Ben Horowitz: Well, um, basically what's uh happened is we have a whole new technology that's the most important technology ever. And what happens every time um there's a dramatic new way of using all of the things that we love infrastructure um you need a whole new infrastructure and never has it been more high impact as it is on this one. So not only do we need new chips, new system software, we need new ways of doing power, we need to replace copper. I mean like it's absolutely everything. So it's a very exciting time. So you know particularly for the kind of hardware aspects of of this new era um we needed a new approach.

Martin Casado: Yeah I would agree. I mean normally when we at least in the computing when we talk about the infrastructure world, we're talking about the servers and the storage in the network. Here it goes all the way down to the mines, copper mines. That's how widespread this thing is going to be and as Ben said, it's just so massive from both a national security, geopolitical perspective, but also just the type of software we're building and the impact of the world. And by the way, if you take a step back and look at global GDP, computing is a very small part of it. It's like 3%.

Raghu Raghuram: Yeah, it's very low. Is very low. Yeah. 3%. Yeah.

Host: And explain some of the macro conditions that have led to this this trans this change in terms of the surplus of founders pursuing these idea. Like what are they seeing that's that's enabled?

Martin Casado: Well, I mean the obvious is like you know the the demand for AI is basically infinite and as a result, massive CapEx is being poured into it, which is pulling on the entire stack. But from kind of a history of technology perspective, what's so interesting is you know for the last 20 years we've kind of been living in the von Neumann networked world. You know compute is compute and we kind of build all of these abstractions on top of it. Now all of a sudden, we're rethinking computer architecture. We're rethinking memory, we're rethinking processing, we're rethinking interconnect, we're rethinking physical data centers. And this kind of reset has only happened once or twice in the entire history of computing, and it is driving just really, really smart founders into the space to build companies.

Raghu Raghuram: To add to that, the last time we saw a massive infrastructure wave was maybe 20 years ago with the rise of cloud computing. But even with cloud computing, it was built using existing building blocks: standard x86 servers, standard Ethernet switches, standard storage arrays. This time, the workload is so fundamentally different. The CPU architecture is just not suited for tensor processing at scale. You need a completely new distributed computing paradigm where the "computer" is now the entire cluster of tens of thousands of accelerators. To power that density, existing power grids, cooling, rack designs, and interconnects have to be reinvented from scratch.

Ben Horowitz: If you look back at "software eating the world," the VC playbook for the last two decades was asset-light pure software. Hardware and infrastructure were commoditized; you just spun up AWS instances and wrote software. But today, compute is power, and the physical world is the bottleneck again. If your physical infrastructure cannot supply the power, cooling, and interconnect bandwidth, your frontier training and inference ground to a halt. That is why we raised this dedicated fund—the old SaaS playbook is not sufficient for this deep physical and systems reconstruction.

Host: You mentioned earlier that building systems used to be primarily an engineering problem, whereas now it has become a hard physical resource limitation problem. Can you elaborate on how this resource constraint manifests in practice?

Martin Casado: In the past, if you wanted to scale an internet service like Google Search or Facebook, the bottleneck was distributed systems software engineering—caching, load balancing, database sharding. The servers were commoditized and readily available off the shelf. But today in the AI era, we are hitting hard physical walls: first, the power wall, where rack power density jumps from 10kW to 100kW+, and utilities cannot supply gigawatts on short notice; second, the memory and interconnect wall, where compute growth outpaces communication bandwidth; and third, physical supply chain limits in HBM and advanced packaging. This forces a shift back to deep physical innovation.

Raghu Raghuram: This has also completely changed customer procurement behavior. Previously, enterprises bought IT gear according to annual budgets and depreciation cycles. Today, frontier AI labs and hyperscalers are racing to lock in multi-gigawatt power agreements, chip allocations, and data center capacity 3 to 5 years in advance. This panic-driven resource land grab is unprecedented in industrial history.

Host: Given these massive resource constraints, how do you see the chip landscape evolving? NVIDIA has a dominant position, but where are the opportunities for new startups and custom hyperscaler ASICs?

Martin Casado: This comes down to the laws of markets. NVIDIA has built an incredible moat with CUDA and full-stack system integration. But in a multi-trillion dollar compute market, capturing even 5% of a specialized segment creates a massive standalone company. Furthermore, AI workloads are diverging: frontier training still requires universal GPUs and massive cluster scale, but inference—especially with test-time compute and reasoning agents—has wildly different latency, cost, and power profiles, creating enormous room for custom silicon and near-memory computing.

Ben Horowitz: We are seeing the rise of the systems company. You cannot just design a neat silicon die anymore; you have to build a full system. You need to co-design from semiconductor physics, advanced packaging, and optical interconnects, all the way up through compilers, runtime software, and data center cooling. The complexity is immense, but so is the defensive moat.

Host: Let's return to data center infrastructure. By 2028, new data centers are projected to require roughly 44 gigawatts of additional power against perhaps 25 gigawatts of expected grid additions. How should we conceptualize a gigawatt and the challenge it presents?

Martin Casado: People throw around the word "gigawatt" casually now. But what is a gigawatt? I grew up in Flagstaff, Arizona, a town of 50,000 to 60,000 people. Our entire town ran on less than a gigawatt of power! A single gigawatt can power and air condition an entire city. Now the industry is designing multi-gigawatt AI campuses. That means you literally need dedicated nuclear SMRs, massive natural gas plants, or geothermal facilities co-located on site.

Raghu Raghuram: When rack power density hits 100kW+, traditional air cooling is physically obsolete. You must shift entirely to liquid cooling, whether direct-to-chip or immersion. Furthermore, facility power delivery is transitioning to 800V DC architectures to minimize resistive losses. Copper interconnects are hitting physical distance-bandwidth limits inside clusters, accelerating the transition to co-packaged optics (CPO). The physical data center is undergoing the most radical reconstruction since its inception.

Ben Horowitz: This is why compute deployments are expanding into unconventional geographies with abundant energy. If you cannot secure power interconnects and cooling permits, having allocations of GPUs is meaningless—they are just cold silicon. Physical infrastructure is the ultimate governor on the rate of AI progress.

Host: Why name it the "Machine Age Fund"? What is the philosophical and technological thinking behind this name?

Martin Casado: I think Ben is completely right: "Artificial Intelligence" is almost a misnomer. It really is Machine Intelligence. It is not necessarily how biological humans think; it is a cache and synthesis of human thought, executed across physical machines made of silicon, electricity, and optics. The next breakthroughs in intelligence will be driven by the raw capability and efficiency of the underlying machines. The machine is the vessel and multiplier of intelligence.

Ben Horowitz: It is also just a great, futuristic name. It honors the heritage of the Industrial Revolution while signaling that we are entering a new epoch where machines and compute systems rebuild the physical world.

Host: Given that infrastructure is extremely CapEx intensive with long R&D cycles, what archetype of founders can succeed here? How does this differ from traditional SaaS or software founders?

Ben Horowitz: The defining difference is that massive amounts of capital must be deployed before there is even a finished product. This demands extraordinary capital strategy, technical vision, and systems execution under uncertainty. Moreover, in hardware and silicon, domain expertise takes decades to accumulate. Great young founders in this space must have the maturity and charisma to recruit and lead seasoned industry veterans who have built chips and physical infrastructure for thirty years.

Raghu Raghuram: They must be true "systems founders." You cannot just be an algorithm researcher or an isolated chip designer. You must understand the holistic system tradeoffs spanning semiconductor physics, networking topologies, distributed runtime compilers, and data center operations.

Martin Casado: Another key environmental factor is customer desperation. Frontier labs and hyperscalers are so bottlenecked that they actively partner with early-stage startups to test alternative silicon, optical switches, and cooling designs. This creates early commercial validation and feedback loops that were impossible in previous hardware cycles.

Host: If the Machine Age Fund accomplishes what you set out to do, how do you see the world transformed in 5 to 10 years?

Ben Horowitz: Hopefully, America wins in the infrastructure race. We want to see an abundance of clean, efficient power, hyper-efficient sustainable data centers, and an abundance of chips and memory. It goes back to why America is special: it is the best place in the world for ambitious people to come with nothing and build something profound. Sustaining that open innovation ecosystem requires maintaining leadership in frontier technology and foundational infrastructure.

Host: Fantastic wrap-up. Martin, Ben, Raghu, thank you so much for joining us.

Ben Horowitz: Thank you.

Raghu Raghuram: Thank you.

Martin Casado: Thanks, yeah.