A sharp supply-demand mismatch is now defining the global AI infrastructure landscape. Demand is multiplying at an unprecedented rate, yet the supply of chips, memory, power, and data centers lags by three to five years. According to leading venture capital firm a16z, the traditional computing stack has hit its physical limits, triggering a comprehensive rebuild of infrastructure designed for the AI era.
Recently, Silicon Valley's top venture capital firm a16z announced the launch of its "Machine Age Fund," a dedicated vehicle focused on AI infrastructure investments. In a public video conversation, a16z co-founder Ben Horowitz, general partner Martin Casado, and former VMware CEO Raghu Raghuram laid out the fund's strategic rationale. Their core thesis is centered on a widening chasm between AI compute demand and supply capacity.
Why demand appears infinite while supply struggles to keep pace
The trio's assessment is unified: the AI infrastructure shortage is not isolated to one component. It cascades from chips and memory to networking, data centers, and further to power, cooling, transformers, and even copper mining. "The impact has reached copper mines," Raghu noted, highlighting the scale.
Ben Horowitz offered a simple but powerful framing: "Any problem you have can be solved with enough infrastructure—basically GPU plus money. As long as problems remain unsolved, demand won't stop."
a16z believes this infrastructure boom differs fundamentally from the speculative internet-era buildout. "Remember dark fiber? A lot of what was laid underground was speculative—it was dark," explained Martin Casado. "Today, virtually every GPU is pre-sold before it's manufactured."
Data points reinforce this view: Hyperscaler capital expenditures are around $700 billion this year, projected to collectively surpass $1 trillion next year. Critical component supply is booked solid through 2028. Martin revealed they have witnessed multi-day auctions for thousands of GPUs, with scalpers fetching four times the original price. At the industry flagship conference Hot Chips, a leading memory maker stated that fulfilling today's demand orders alone would require three years of production capacity—not even accounting for future growth.
GPU prices have made a rare rebound. "GPU prices have always trended downward... then they stopped and came back up," Raghu observed, a historical anomaly. Power shortages are equally acute. New data centers are projected to need around 44 gigawatts of additional power by 2028, while grid additions are expected to be just 25 gigawatts. "We're short on GPUs, electricity, cooling, memory—you name it, we're short," Ben added.
Resource constraints have replaced engineering challenges
Why does demand expand without hitting a ceiling? Martin Casado offered a central insight: "In the past, building something was an engineering problem—you threw engineers at it, but that doesn't scale linearly—there's the 'Mythical Man-Month' constraint. Now, this is genuinely a resource problem. We pour massive capital into systems, and they produce results. The bottleneck is these systems' ability to truly absorb the resources we invest."
Each AI generation—from chatbots to reasoning models to agents and multi-agent systems—exponentially increases token consumption. Raghu described the direct trajectory: "If a chat needs 100 tokens, an agent might need thousands. Demand expands on two fronts: individual tasks consume more tokens, and the user base widens—from developers to knowledge workers to the broader public."
Ben Horowitz predicted: "Demand for tokens could grow nearly 1000% per year. That pace is something infrastructure supply simply cannot keep up with." Raghu added, "The only way AI gets better is by using more AI. Reasoning is the fundamental building block it repeatedly calls upon—that's why token counts keep multiplying."
Rebuilding the entire computing stack from first principles
a16z's key argument is that existing AI infrastructure faces a core challenge: "It's not just a matter of insufficient quantity—the entire architecture wasn't designed for AI."
Raghu systematically outlined the rebuild logic: "You have to go category by category: What does compute need? What does memory need? How do they interconnect? How much power does each require? How do you cool it? How do you assemble it? That's what the industry is doing now—breaking problems down to the most basic components and reconstructing from first principles."
Martin Casado used a business case to illustrate the urgency: "Training a frontier model now costs roughly $3-5 billion. You need at least 2x that in inference revenue to recoup—around $10 billion. If you could save 20% efficiency in inference, that's $2 billion. And $2 billion fully justifies developing an ASIC. This means building a custom chip for a single model is now economically viable. That has never happened before—we've never poured $5 billion into creating a single digital artifact."
Physical rebuild pressures are mounting: rack power is surging from 5-10 kilowatts to 100-150 kilowatts, a roughly 70x increase in compute density. Cooling is shifting from air to liquid as a mandatory requirement. Rack voltage rising to 800 volts places it in high-voltage territory—only about 2% of U.S. electricians are DC-certified. Ultra-high-density racks impose new floor-loading requirements on data centers, and even building materials like concrete are seeing rapid price inflation.
"Most existing data centers will become completely obsolete once we enter the next generation of compute," Ben Horowitz stated bluntly.
Why top-tier entrepreneurs are migrating to hardware in droves
This structural shift is producing clear signals in the startup ecosystem. Martin Casado noted an internal a16z informal metric: historically, hardware-related deals from top founders represented only 3-5%; now that figure has exceeded 20-30%.
"The founder community is typically smarter than the VC community. They've identified this as a highly active innovation area and are acting on it," Martin said.
The profile of these "hardware founders" differs from traditional software entrepreneurs. Raghu described them as "system-level founders"—"You can't just be a researcher or excellent computer scientist. You need to architect and design chips or systems, and also figure out manufacturing and supply chains—things software founders typically don't consider. Jensen Huang embodies this mindset perfectly, thinking about the entire ecosystem from day one."
Frontier AI labs, desperate for capacity, are already signing agreements with early-stage hardware startups "before products even exist," Martin noted—a significant shift that reduces early commercialization risk. The funding environment is also evolving: seed rounds now reach hundreds of millions of dollars, unprecedented for hardware investments.
The future: decades of compute demand ahead
On the new fund's name, Martin Casado candidly remarked: "The term 'Artificial Intelligence' was a misnomer from the start—it should have been called Machine Intelligence." He elaborated: "It's a 70-year-old computer science term loaded with science fiction baggage. Moreover, it doesn't truly mimic human thought—it leverages everything humanity has already learned, rather than reconstructing language from scratch."
The "Machine Age" name also serves as a tribute to hardware. "There's a deep irony here: the people who said 'software is eating the world' ended up discovering that the real constraint lies in the physical machines below the software. The name acknowledges that in this wave, hardware's importance is unprecedented."
Martin offered a long-term perspective: "We're just getting started—we have language, code, and the nascent beginnings of computer use. But over the next 30-40 years, science, materials, biology, creativity—all these domains will be areas where we apply compute power to problems. We're in the very early stages of a very long journey. Be prepared: this compute demand will persist for decades."
The full transcript of the a16z conversation follows below.
The Founding Logic of the Machine Age Fund
Ben Horowitz (a16z co-founder): We are in an era of a brand-new technology, the most important there has ever been. And to support it, you need an entirely new infrastructure.
Raghu Raghuram (former VMware CEO / tech executive and advisor): Usually when we talk about infrastructure, it's servers, storage, networking. But this time, it extends all the way down to mines—copper mines. You can sense the breadth of the impact.
Martin Casado (a16z general partner): Previously, building something was an engineering problem. Now, it feels like resources are the real choke point. We're pouring massive amounts of money into these systems, and they produce results. But the systems' ability to absorb resources is far behind the speed at which we're pouring in.
Raghu Raghuram: A leading memory manufacturer said they have demand orders on hand that would take three years of production capacity to fulfill.
Moderator: Ben, Martin, Raghu, welcome. I'd like to start with a quote from Marc Andreessen to introduce this new fund: "This is the biggest technological revolution I've seen in my life. It's clearly bigger than the internet. The analogies are the microprocessor, the steam engine, electricity, maybe even the wheel." This is the Machine Age Fund. Tell us about it, Ben.
Ben: Simply put, we now have a new technology, the most significant ever. Whenever a disruptive new technology arrives, everything we rely on—infrastructure—must be completely overhauled. And the impact this time is deeper than anything before. So we need new chips, new system software, new ways of delivering power, replacing copper lines... almost everything has to change. It's a very exciting era. For all these hardware needs of the new age, we need a fresh investment approach.
Raghu: I completely agree. Normally, computing infrastructure means servers, storage, and networking. But this time, pushing all the way down to mines and copper—the breadth of impact is clear. That's point one. Point two: over the past three years, we've seen model capabilities steadily improve; models themselves are no longer the bottleneck. In fact, with AI's help, these models are getting stronger at an accelerating pace. The constraint now is what I'd call "below the model"—the underlying infrastructure. That's where we need to focus.
Martin: Let me quickly add: we always follow founders. Over the past few years, we've seen a clear trend—more top teams tackling complex, harder problems. I don't have exact numbers, but estimating over the weekend, hardware represented maybe 5% of top founders' deals previously. Now it's over 20-30%. The founder community's instincts are often sharper than VCs. They've identified this as a highly active innovation zone and are moving in.
Ben: I think 5% might even be an overestimate.
Martin: Yes, it's quite low—maybe 3%.
Macro forces driving the hardware influx
Moderator: Can you explain which macro conditions drove this shift? What opportunities are these founders seeing?
Martin: The most obvious point is demand for AI is essentially infinite. Because of that, the entire supply chain faces immense pressure—including raw materials for manufacturing things like memory. What's unique about AI: because demand is infinite and growth is infinite, you typically don't worry about "can we sell it." What you worry about is the company's profit margin—efficiency. Many efficiency problems boil down to physical hardware limits. Existing systems simply weren't designed for AI or these workloads. So AI's business model is putting enormous stress on current systems. I think the world has realized: we must change the foundational components to improve efficiency, sustain business growth, and realize commercial value.
Moderator: How do we know demand truly exceeds supply, rather than another hype cycle?
Raghu: Several indicators suggest it's real. First, the most discerning buyers are placing massive orders. Hyperscalers' capital expenditures are exploding—reportedly reaching $1 trillion combined next year, up from roughly $700 billion this year. These players see demand from all directions: frontier AI labs, AI-native companies, enterprises, domestic and international markets. If anyone has the best visibility, it's them. And their capex growth is unlike anything before—a clear signal. Second, from our portfolio and contacts, application-layer companies are growing at incredible speed. Frontier lab growth is equally well-documented. From the demand side, the signal has never been clearer—this is not hype. Most importantly, all of this is just...
Ben: And chip prices are rising—we've never seen prices do that before...
Martin: GPU prices have been declining—they always do.
Raghu: Right, right—they've always declined. The curve goes down, then comes back up. And if just 5% or 10% of demand stops, the whole market halts.
Martin: On the supply side, available supply is essentially booked through 2028. It's gotten so tight—we've seen auctions for thousands of GPUs lasting several days. On the flip side, as Raghu said, we're seeing the fastest-growing companies in industry history.
Raghu: Also, the value of work units AI can perform keeps rising. But beneath the surface, token consumption is growing by orders of magnitude. A typical conversation might consume 100 tokens; an AI agent might use thousands. Demand expands on two dimensions: more tokens per task, and a broadening user base—from developers to all knowledge workers and beyond. That's the trend we see.
Entire supply chains sold out years in advance
Moderator: You mentioned critical components are sold through 2027, maybe 2028. What does it mean for the industry that everything is sold out that far ahead?
Martin: I don't know if this has ever happened. Think of the internet era—we laid massive infrastructure, but much of what was buried underground was speculative—nobody was using it—remember "dark fiber"? Today, virtually every GPU in production is already sold.
Ben: Yes, that era was entirely different. Around 1998-99, there was indeed bandwidth scarcity, but actual demand was low because few people were online. It was a two-sided problem—companies rushed in for theoretical bandwidth need, but consumer demand wasn't there. And truly consuming serious bandwidth requires high-bandwidth activities like video, which wasn't viable then for many reasons unrelated to data center capacity. So it looks similar on the surface, but it's fundamentally different now. Everything is in shortage—people are reselling GPUs at four times the price. We're also short on power and cooling. Worse, construction itself is extremely difficult due to massive political resistance. In my entire career, I've never seen anything like this—truly unprecedented.
Martin: Let me share a story. I recently spoke with the CFO of a large public company, historically very resistant to cloud adoption, so they run their own servers. During an inventory check, they discovered—the value of memory in their servers had appreciated so much that selling just the memory would fund their entire cloud migration. We're truly in an unusual situation.
Ben: Absolutely. We're short on everything—power, cooling, memory, GPUs. Name it, we're short.
Raghu: Yes. At Stanford's flagship Hot Chips conference, the top memory manufacturer said meeting today's existing demand alone requires three years of production capacity—before counting future demand.
Moderator: So everything is tight simultaneously. Did people underestimate how powerful and useful models would become?
Martin: I think it's more than that. This emerged suddenly—it's only been four years. Even with perfect foresight, once it started working, I don't think...
Ben: We could have built capacity in time.
Martin: Not in time. Absolutely impossible. We're talking about chip cycles of three to four years. Data center construction from groundbreaking takes four to five years.
Ben: Plus grid interconnection, building from scratch, and securing power supply. You either build your own power plant or secure stable power—often you must do both. Not easy.
Raghu: If an industry grows at 20-30%, that's considered excellent. But it's interfacing with the AI software industry, growing at triple-digit rates. You can see the enormous gap, and it's widening.
Why this fund didn't exist years ago
Moderator: Why didn't such a fund exist before? Why wasn't this investment space as attractive five or seven years ago?
Ben: I think we're entering at a good time—or so I hope. Honestly, entering a few years earlier might have been fine too.
Martin: You can find independent companies emerging in every technology era. Moving from mainframes to client-server, companies emerged; with the internet, more followed—that's why we have Cisco and Juniper. Even in the hyperscale data center era—which, by the way, was mostly vertically integrated by a few top cloud providers—you saw Arista emerge. So investing in hardware and chips has always had a lane, but opportunities were more limited because changes were narrower—maybe just a chip company or a switch company. Now, everything is changing at once. I agree with Ben—we could have started slightly earlier, but the scale of change now is so massive that it's the natural thing to do.
Ben: Also, the demand for intelligence is expanding almost infinitely in a vertical direction—there's no visible ceiling. Companies already using AI are growing usage dramatically; most companies haven't deeply adopted AI yet; consumer usage is just beginning. So demand for tokens could grow nearly 1000% annually, and supply can't keep pace. The infrastructure workload ahead is enormous just to match that growth. I think investment opportunities abound along this path. Additionally, existing hardware architectures were designed for a completely different computing era. It's not just "we need more compute"—we need to rebuild infrastructure from scratch, creating numerous opportunities.
Raghu: These technologies have hit the physical limits of their original design. Going through each category, you find—"OK, this technology has reached its end." So breakthroughs are now required to reach the next level.
Moderator: Let's dig into the demand side. We've moved from chatbots to reasoning models to agents to multi-agent systems, with token consumption exploding at each step.
Ben: Nothing likes AI more than AI.
Moderator: Why hasn't this trend plateaued? Do you think it will continue?
Martin: There are a few angles. First, the way we scale today is fundamentally through massive inference—massive token generation. Reinforcement learning is massive inference. Chain-of-thought is massive inference. Long-running agents, of course, are massive inference. It's just the current scaling path. More broadly, previously, building things was an engineering problem—you pile on engineers, but it doesn't scale linearly; engineering has natural laws, as in "The Mythical Man-Month." But now it genuinely feels like a resource-constrained problem. Whether tokens or something else, you pour money into systems, and they produce output. The bottleneck is the systems' ability to consume resources at the rate we're pouring them in. So tokens are just one phase of the current scaling curve, but we no longer have the "natural governor" we had with engineering. This trend will persist; we must build sufficient supply.
Ben: Simply put: any problem you have, sufficient infrastructure—basically GPU and money—solves it. As long as problems remain, demand persists. That's what makes it challenging.
Raghu: The way AI gets better is by using more AI. Reasoning is its most fundamental, repeatedly used building block. That's why tokens keep multiplying.
Martin: And there's a self-catalytic effect—using AI to create more AI, like using AI to write GPU kernel code. Here's a framework: money comes in; before, you faced an engineering problem taking two years, often ending in failure, with a natural governor, before seeing a product. Now, money goes in, hardware directly produces intelligence, with nothing in between. Our only limitation is whether we can build enough supply. It's a completely different dynamic.
Raghu: Even more GPUs may not be the most stable thing...
Martin: As long as you have capital, GPUs, and data, you can scale indefinitely for the foreseeable future.
Capital intensity and the new economics of intelligence
Moderator: It's fascinating. For a decade, I heard the industry-wide refrain: too much money in startups, overinvestment, too much capital in VC. Ben, what does it mean when the opposite is true—market size depends on how much we collectively invest?
Ben: In the startup world, we all know this rule: if I'm two years ahead and you try to catch up by hiring 1,000 engineers, you just ruin your company. That never works—it's the Mythical Man-Month. Nine women can't have a baby in a month. That's fixed. But now, the logic is inverted. Instead of hiring 100,000 engineers, take $3 billion, light up a massive cluster, and suddenly—Grok appears from nothing, or similar, materializes into existence. These leads can be closed with money; almost any problem can be solved with money, and it actually works. This is completely unlike anything we've experienced. And by the way, we're all still psychologically adjusting to it.
Moderator: ChatGPT has nearly a billion weekly active users. Around 30 million developers use it, consuming a significant share of compute. How should we think about current and future compute demand given actual usage?
Raghu: It's an unfolding progression. ChatGPT started as a daily conversation tool, became a professional programming assistant, then unlocked powerful tools for knowledge workers. Knowledge workers are the next frontier—over a billion globally are waiting in line. This demand has a long runway. Consider their workflows, automation, and then back-office operations—all agent-driven. Each unlock grows demand by an order of magnitude. We're just at the beginning.
Ben: And now with Grok Bot, the transformation happening in programming is spreading through Grok Bot to general computer use. We're in another demand wave, with more to come. Demand seems nearly limitless. We haven't even discussed embodied AI or robotics—another massive demand source.
Raghu: Martin is the expert here, but as I understand it, Grok Bot uses "computer use"—like a human sitting at a keyboard.
Martin: I actually used it last weekend—had it update credit card info on services I'd been ignoring and cancel subscriptions. Nothing to do with programming; that's genuine computer use.
Raghu: You're essentially creating a workforce of knowledge workers living inside computers.
Martin: I do think Marc Andreessen is right—the best analogy is the steam engine or electricity. We've introduced something that converts into labor. There are obvious applications now, but for the next 30-40 years, we'll apply compute to any problem with a clear reward signal. We've only just begun—currently language and code, and early computer use. But what about science, materials, biology, creative work—all massive application areas? We're in the very early stages of a long journey, and we've removed a key bottleneck—traditional software engineering. Other bottlenecks will emerge, adding complexity elsewhere. But I believe we're at the very early stage of a long era of throwing compute at problems. So be prepared: this compute demand will last decades.
Grok Bot and the evolution of AI interaction
Moderator: Marty, let's talk Grok Bot. You mentioned it impressed you—what's most interesting?
Martin: Our industry has gone through several iterations on how AI enters our lives. Initially: add AI to a product, it's like a search box. Then you chat with it, and it responds—that was conventional. Later, OpenAI introduced something earlier this year that made me think: perhaps "better than Google search" isn't AI's full form. Maybe it's an autonomous entity, but an extension of you—sharing credentials, doing things you'd do. More an extension, yet more human. Grok Bot got it right by reframing: what if it's an employee? Now you have an entity that doesn't need your credentials—it has its own computer, its own browser. Given these are the world's smartest models, it can do what any employee can do. It's interesting because now my first thought when trying to accomplish something is: can Grok handle it? Often the answer is yes, even for things you wouldn't expect. Obvious ones like calendar management and scheduling, but also less obvious—I'll have it read and sort emails; I don't tell it how, but it knows to confirm with me before executing classification. These systems are smart enough for high-level tasks; they figure it out.
Moderator: Ben, you've thought deeply about organizational dynamics and culture. Your perspective?
Ben: It's like adding a new type of employee—and there will be many. We must adapt, just as we took years to learn how to collaborate with human employees. Now we have these other employees. There's a learning curve. They can burn through enormous tokens and money without accomplishing anything useful. They forget, they hallucinate, they perform well or poorly, they pose security risks. All these exist. But simultaneously, they can be extraordinarily efficient. Figuring out how to integrate them with human colleagues is what we're navigating. I'm not here to claim I've found the answer—that we have a perfect loop and I'm replacing all humans because I can. Not at all. We're thinking: how do we make all human employees hyper-empowered without destabilizing the company? Because if robots go off the rails, it's a big problem.
Raghu: It's fascinating. We've tried different approaches to bringing agents into systems. Ultimately, Martin's insight proved most effective: treat them like people and push forward. That's what we do, and it's proven the most durable approach internally.
Redesigning the AI hardware stack from scratch
Moderator: Let's return to the supply side and bottlenecks. Data centers, chip architectures, system software—none were designed for AI. What would a true AI redesign look like? What mental model should we use?
Raghu: Starting from the premise that existing infrastructure models must change, you go category by category, find value dislocations, and unlock each bottleneck. Ultimately, look at what inference engines do—they're memory-intensive, generating tokens while computing. Ask: how do I optimize this entire process? What form should memory take? What form should compute take? How do they communicate? How much power does each require? How do you dispatch with those power needs? Then how do you assemble it all? That's what many founders are doing now—breaking problems to the most fundamental components, asking: what computation is happening here? Matrix multiplication. How do I optimize compute units for that? Token generation needs increasing memory—what's the optimal memory hierarchy? How is power consumed? Then interconnect—within the chip, across chips, across data centers—each level's power cost? You must decompose everything and rebuild from these building blocks. That's the direction and opportunity we see.
Martin: Let me give a useful mental model for how the landscape shifts. Today, training a frontier model costs $3-5 billion. Post-training, inference needs to recoup at least 2x for commercial viability—roughly $10 billion in inference revenue. If you save 20% efficiency, that's $2 billion—enough to build an ASIC. We've reached a point where custom ASICs for a single model are economically justified because of the capital invested. Unlike traditional software—highly stateful and dynamic—model weights are fixed. We don't know if the world goes fully "an ASIC per model," but it's a useful model for how architecture will evolve—becoming more tailored to massive capital investments. In industry history, we've never poured $5 billion into a single digital artifact. This will demand the highest-ever hardware requirements.
Moderator: Rack power is jumping from 5-10 kilowatts to 100-150. Compute density is up roughly 70x. Cooling shifts from air to liquid as mandatory. What investment opportunities lie within?
Ben: At those rack power levels, AC power doesn't work anymore. It's crazy. So you switch to DC—which, by the way, requires separate cooling and is genuinely dangerous. Ironically, Edison promoted DC by arguing AC was dangerous—demonstrating with animals, remember the horse. He was right, but about DC itself—it's extremely powerful. Starting with power: this area will change dramatically, especially combined with cooling. We're shifting from air to liquid cooling—for any frontier data center, liquid cooling is effectively standard now. That's mostly decided. But there's more nuance. Given the political environment, liquid cooling isn't enough—it must be environmentally responsible. DC isn't enough—the electricity must power society, not just draw from it. Some data centers behave poorly—about 10%—wasting massive water. People compare to pistachios and almonds' water use, but data centers can be more efficient. Others are "power parasites"—consuming without contributing back to the grid. These must end because we've passed the tipping point. This requires serious engineering that many haven't invested in—opportunity awaits. At high rack densities, other issues arise—floor designs must support the weight, real engineering challenges. Also, huge resource needs. These facilities are extremely loud, so walls must be thicker to avoid disturbing communities—no state would allow otherwise. Many current architectural and design approaches are obsolete. Most existing data centers won't survive the next generation—only a small fraction will be usable.
Raghu: Everyone talks about memory prices, but one of the fastest-rising costs is reinforced concrete for data centers. Also, when delivering 800 volts to racks—first, danger rises significantly; second, we lack enough qualified electrical contractors handling 800-volt DC inside data centers—that's high-voltage territory.
Ben: Only 2% of U.S. electricians are DC-certified. That says it all. Meta now has a training program—free training to fill these roles. It's like a modern "job corps." Funny—they say AI takes jobs, and AI ends up creating a new class of electricians.
Computing's physical limits and future regions
Moderator: We seem to be building something big.
Raghu: Yes—operators of large cloud data centers are aggressively experimenting with robotics for server assembly and data center installation. As AI evolves, this trend will accelerate.
Martin: By the way, the fund we're raising targets computer science infrastructure—anything models depend on falls under that umbrella: chips, interconnects, storage, down to power itself.
Moderator: How do you think about robotics, especially Arm architectures, or U.S. manufacturing?
Martin: Any platform AI will run on is our focus. A major AI breakthrough is computer interaction with the physical world—seeing, hearing, speaking. This creates entirely new platforms. People say "edge devices," but that's ambiguous—phones, CDNs, laptops, or physical devices that move. We don't do heavily regulated or overly vertical industries, but any computer science platform extending AI outward is of great interest.
Moderator: Back to data centers—by 2028, new centers need 44 gigawatts, grid supply is only expected to add about 25.
Ben: Wait—we throw around "gigawatts." Martin, what's a gigawatt?
Martin: How big is it? It covers several football fields. Huge. Equivalent to powering 50,000 people.
Ben: What does that mean?
Martin: About 50,000 households. I grew up in Flagstaff, Arizona—45,000-60,000 residents in university season. Our whole city used under a gigawatt. So one gigawatt...
Ben: ...powers an entire city with air conditioning. Exactly.
Martin: I mean it.
Ben: He says it casually.
Martin: No—but everyone talks gigawatts, yet few actual gigawatt-scale data centers are operating. We have a long way to go.
Moderator: Why can't power companies and tech giants build faster?
Ben: There are so many issues. First, data centers are built manually, so construction has standard problems. But more challenging: permits and grid connection. You're juggling multiple things: securing power interconnection rights—an enormous regulatory and bidding struggle. Natural gas grid connections are severely limited. And you must build power plants yourself. Guess what? Transformers, turbines—everything needed is in shortage. You must secure all of it. It's not a software problem—it's not about hiring more engineers and working weekends—that method never worked well anyway, but it doesn't apply here. These are real bottlenecks with uncompressible lead times. The world's smartest people are trying to compress them—it's extraordinarily hard. And demand shows no sign of slowing. We're already behind. Demand grows 10x annually; supply can't match it.
Martin: It's gotten so bad that many new GPU cluster startups are going to Mexico, Australia, or other countries because it's just too difficult in the U.S.
Ben: We're effectively exporting massive job creation and long-term economic benefits because of how we restrict data centers. The right approach: establish a standard—data centers contribute to communities: improved power reliability, no noise, no water issues, job creation. That should be the standard, and all should follow. Some already do—it's not futuristic. Energy rates fall annually because they generate their own power, selling to the grid during day and drawing at night—since baseload plants always run at peak capacity. Data centers have flat demand 24/7, while cities peak during day and dip at night—a complementary relationship.
Why "Machine Age" is the right name
Moderator: We've discussed why this name fits. Why is "Machine Age" particularly apt for describing what we're experiencing?
Martin: First, Ben's right—"artificial intelligence" was a misnomer. It should have been "machine intelligence."
Moderator: Elaborate.
Martin: Because it's not necessarily how humans think, right? It's an archive of human thought—a collection of everything humans have already learned. But so far, we haven't figured out how to put an AI with zero knowledge into the world and have it reconstruct language from scratch—we haven't done that. What we've built: systems that learn from existing human knowledge and apply it effectively. Also, "AI" is a 70-year-old computer science umbrella term covering many different things. Of course, it carries baggage—from science fiction to Nick Bostrom's writings, etc. So first, it's about acknowledging the reality: this is fundamentally machine intelligence. Then you emphasize "machine." There's a deep irony—those who championed "software eating the world" now find themselves where the real constraint is the physical hardware below the software. The name is somewhat a tribute: hardware's importance is unprecedented in this wave, and we want to say that.
Raghu: Yes—the next wave of breakthroughs comes from the quality of underlying machines. Basically, that's the name's origin.
Moderator: It also sounds cool.
Ben: "Machine"—it's punchy. Futuristic.
Why new entrants can still compete with incumbents
Moderator: Given the capital intensity and scale of AI infrastructure investment, have we passed the point where new companies can meaningfully enter? Why wouldn't Nvidia or CoreWeave just take most of the market?
Raghu: They're doing exceptionally well—no question. But going back to our earlier discussion: you don't need fundamental innovation just to sustain growth or maintain improvement rates—whether tokens per dollar, per watt, per rack, or power consumption. Pick any metric. To achieve 10x improvements, you need new innovation. And new innovation traditionally comes from founders who reason from first principles and think differently—that's what the next innovation leap requires.
Martin: It's market dynamics, right? The existing chip giant is worth trillions. Even capturing just 5% of that market creates a massive private company. Nvidia could do it—they certainly could—but if their focus is on the 90% growth segment, why pursue the 5%? You'd ask the same question in the cloud era: "Wouldn't Amazon do this?" Or in Microsoft's era: "Why wouldn't Microsoft?" It's a natural market dynamic—once scale is reached, the edges create enormous innovation opportunities.
Ben: Our partner Alex Rampel has a great story. His startup TrialPay tried selling to Meta (then Facebook). Dan Rose in corporate development said: "Alex, this sounds great and you'll pick up plenty of silver bricks. But I'm surrounded by so many gold bricks on the ground I can't pick them all up. The last thing I want to do is look at a silver brick." That's Nvidia's position now.
Moderator: 100%. We discussed that being in OpenAI's or Anthropic's core capability sweet spot might be a tough position. But outside those 3-5 core domains...
Martin: When markets expand, they fragment—this happens constantly. Think of Ford's early days, the River Rouge plant in 1913—coal, water, rubber trees in, cars out, like an assembly line.
Ben: He even bought entire rubber plantations in the Amazon, right. A book called "Fordlandia" covers it—full vertical integration, he built a city called Fordlandia in the Amazon with American-style bandstands, ice cream, everything. It operated initially until he demanded workers show up on time—everyone said screw you.
Martin: So look at the auto industry today: multiple supplier tiers, dozens of companies. This always happens—markets expand and fragment; consolidation follows when growth slows, either via M&A or new challengers. That's the eternal private-market cycle.
Ben: Yes—use cases keep multiplying. Even if you're the largest company covering the biggest use case, there are so many use cases, many extremely valuable, that large companies can't possibly execute all of them well.
Raghu: Architecture used to be a simple, unified thing. No longer—it's extraordinarily complex. Therefore, optimizing in different ways is inevitable.
Martin: Interesting—many don't realize that margins historically "fall out" naturally from software as a standard technology approach. It's not a technical problem—once your business works, margins tend to be excellent, whether packaged software or SaaS. AI may not follow this. We might be entering a new era where hardware-level optimization's impact on margins becomes critical in ways we haven't seen before. That's a huge opportunity.
Investment focus and founder profiles
Moderator: Let's discuss what types of companies we'd invest in—perhaps segmentation or disclosed examples. Raghu?
Raghu: As we've discussed, segments span many categories. Most obvious is compute chips—but a chip alone is no longer enough; you must build an entire system. What does that system need? Memory innovation, networking innovation, power chips, etc. Each category could produce public-company-level ventures—that's what we focus on. Integrated together, they need substantial software—automation, cluster management. Software is also a key area. These components are interconnected; no category can be ignored.
Moderator: What distinguishes these companies? From disclosed investments, seed rounds are massive—hundreds of millions. Is it founder type? What's different in building and investing versus traditional software?
Ben: The biggest difference—you've noted it—is the massive capital required before having a product. That's inherent. Large models have similar characteristics, but their path is clearer; this field is riskier, requires more capital, and is more complex than many previous ventures. Also, many chip founders are veterans—people who understand memory, for instance—not young. So that differs too, but it's exciting.
Raghu: Additionally, these must be system-level founders. You can't be just a researcher or excellent computer scientist. You must architect and design chips or full systems, understand mass production—who supplies components—and a host of downstream issues. Software founders typically don't consider these. The best founders—Jensen is the "Michael Jordan" of this—think through the entire ecosystem before designing a chip because bottlenecks are numerous and all parts must interlock. That's a defining characteristic of these founders, very different from other fields.
Martin: Two external factors matter. First, AI labs are so resource-hungry they proactively partner with startups. We get early signals—labs sign agreements with companies before hardware delivery. Completely different from five years ago when selling an immature hardware product to Google was impossible. Second, capital availability is much looser. There's consensus that this is the time to reshape the industry, with ample funding for later rounds. You want to invest where capital flows. So the entire environment is transformed.
Founder demographics: fewer young founders?
Moderator: Patrick Carlson noted a few years ago that young founders seem rarer—no Zuckerberg-dorm-room or Gates-Microsoft equivalents. There are exceptions like Michael, but it's more "veterans" starting companies. Your thoughts?
Ben: As Raghu says, when supply chains are complex, manufacturing is involved, and technical difficulty is extreme, experience truly matters. Look at Elon or Travis Kalanick—early on, they built software companies. Even top-tier talent accumulates startup and technical experience before "leveling up" to tackle harder, bigger domains with their many moving parts. Learning entrepreneurship while already deeply familiar with your product is hard enough. If the product is unfamiliar, the learning curve for a first-time founder is extremely steep. So: on one hand, you have someone like Michael—young and brilliant—in pure-software AI; on the other, experienced operators like Elon or Travis handle far more complex domains. Michael might do the same in a decade, but right now it's too difficult.
Martin: It's also important to note—this field was neglected by industry and academia for nearly 20 years. It existed but was never a growth track; that was software and networking. So very few people have direct experience emerging from universities or large companies. You don't "intern and design a chip." But this is changing. We'll cultivate an entire generation of founders from these new companies who know how to do this. They'll enter at junior levels and grow. In that vein, Elon's greatest legacy may not be just his companies but the SpaceX alumni transforming the industrial system—possibly exceeding the companies themselves. We'll see the same in computer science and hardware.
Raghu: Yes—one of our portfolio companies was founded by two founders in their twenties, but walking their office, you also see many experienced people. An ideal combination.
Ben: Founders don't necessarily need deep experience themselves, but they must mobilize, attract, and collaborate with experienced people—the quality of those people matters enormously. It's complex.
Moderator: On experience—you're launching a large fund without new general partners. You're consolidating. Is it because you and your team have deep experience in this field that's been dormant?
Martin: Interestingly, I feel we almost needed someone to remind us not to drift—because of our backgrounds. It's hard because we've spent so much of our careers in systems and hardware that we're naturally drawn there. We have made hardware investments over the years—SpaceX, Android with very early checks, Astronomer, Waymo. Even early on, we did quite a bit. It's baked into our DNA. So this doesn't necessarily require expanding the team or building new capabilities—they're added bonuses.
Moderator: Thank you for this closing conversation about the Machine Age Fund.
Ben, Martin, Raghu: Thank you.
Ben: Great.
Moderator: Martin, Ben, Raghu—thanks.
Ben: Thank you.
Martin: Thanks.
Raghu: Thanks.
Comments