Martin Casado Says AI’s Real Bottleneck Isn’t Software — It’s Running Out of Power, Memory, and Copper — BigGo Finance

One of the fastest-growing markets in history is hitting a wall — and it is not a wall of code, but of concrete, copper, and voltage. In a recent conversation on the a16z podcast, Martin Casado, a general partner at the venture firm, put the situation in bluntly physical terms: “We’re also out of power and cooling. And then on top of that, it’s really hard to build because there’s these incredible political headwinds going into it.” His assessment, shared with co-hosts Ben Horowitz and Raghu Raghuram, marks a sharp departure from the software-centric worldview that has defined Silicon Valley for decades.

The three partners were announcing the firm’s new Machine Age Fund, a dedicated vehicle for AI infrastructure investing. The central argument is simple to state but radical in its implications: artificial intelligence has stopped being an engineering problem and become a resource problem. The models work. The demand is real. What is missing is everything those models run on.

The $1 Trillion Question: Where Does the Money Go?

The demand picture is not subtle. Raghu Ram, one of the architects of the fund, cited hyperscaler capital expenditure figures that would have been unthinkable five years ago. According to his data, the collective spend from the big cloud providers is on track to reach around $700 billion in 2026 and is projected to hit $1 trillion in 2027. That is not a forecast of potential demand; it is money already being wired to suppliers.

This surge is backed by the fastest-growing companies in the history of the industry, as Casado noted. The revenue ramps at AI-native companies are unlike anything seen in prior technology cycles, and they are driving an almost insatiable appetite for compute. The consequences are visible in the supply chain. Casado stated that GPU supply is effectively booked out to 2028, with some allocations so scarce that buyers are participating in multi-day auctions for a few thousand chips.

The memory market is under even more extreme pressure. According to the partners, a leading memory vendor currently holds three years’ worth of existing demand on its books before it even begins to address future orders. That kind of backlog turns a commodity into a strategic asset, and it explains why memory prices, not just chip architectures, are reshaping the entire cost structure of AI.

From Engineering Bottlenecks to Physical Ones

The shift Casado described is not just a matter of scale; it is a change in kind. “It used to be when you built something, it was an engineering problem,” he said. “And here it feels like it really is a resource limitation.” His point is that in the software era, the limiting factor was always human talent — could you hire enough engineers, and could they write enough good code? In the AI era, the limiting factor is the physical world: can you build the chip, deliver the power, and lay the copper fast enough?

This inversion changes the risk calculus for investors and founders alike. Casado framed it starkly: “Normally you worry about growth like can I just get people to buy this stuff. You don’t have to worry about that here. The question is can you do this in a way that’s profitable?” The old startup anxieties about customer acquisition and product-market fit are, for this layer of the stack, largely resolved. What remains is an operational and financial puzzle of staggering complexity.

Ben Horowitz grounded the opportunity in a simple observation: “We have a whole new technology. That’s the most important technology ever. And you need a whole new infrastructure.” The Machine Age Fund exists to finance that rebuild, from data center architecture to the electricians certified to handle the new voltage levels.

The Power and Labor Problem Nobody Trained For

The physical bottlenecks are not evenly distributed. Power is the most acute. By the partners’ estimates, new data centers will need roughly 44 gigawatts of additional power by 2028, while the grid is expected to add only about 25 gigawatts. Casado put that gap in human terms: a single gigawatt is roughly the entire power consumption of a city like Flagstaff, Arizona, with a population of 40,000 to 60,000 people. The industry is essentially asking the grid to add the equivalent of dozens of new cities worth of power in a few years.

The problem extends beyond generation. Transformers and turbines are in shortage, and their lead times are not compressible. The shift from alternating current to direct current power distribution inside data centers — moving to 800-volt DC — introduces a new safety regime that most electrical workers are not trained to handle. Only about 2% of US electricians are certified on DC power at the required levels. Meta has launched a training program to close that gap, but the structural shortage is a multi-year problem.

Then there is the construction supply chain. Reinforced concrete prices are rising faster than any other material in the data center buildout, according to the partners. The physical shell of a modern AI facility is itself becoming a scarce resource.

How Founders Are Changing to Chase Hardware

The shift from software to physical infrastructure is changing who builds companies. Casado noted that the share of top-founder deals in hardware has surged from roughly 5% of incoming opportunities to between 20% and 30% in the past few years. The founders tackling these problems are, on average, older and more experienced than the classic twenty-something software archetype, because the problems demand it.

Raghu Ram described the new breed as “systems founders” — people who can architect a chip, design a manufacturing strategy, and manage a supply chain simultaneously. He pointed to Jensen Huang of NVIDIA as the archetype of this profile, someone who thinks about the entire ecosystem before designing the silicon. Horowitz was slightly more cautious about the age profile, noting that while hardware is unforgiving for first-time founders, the ideal combination is often young founders who can attract experienced operating talent.

The capital intensity of these ventures is also unprecedented. The partners described companies raising hundreds of millions of dollars before reaching anything resembling product-market fit — a funding pattern that would have been a red flag in the software era but is now a necessity of building physical infrastructure. “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,” Casado said.

The Geopolitical Calculus of an Infrastructure Race

Horowitz closed the conversation with a broader frame: this is not just an investment thesis, but a national competitiveness question. “America wins in the AI infrastructure game,” he predicted, framing the outcome as a choice between a country that continues to attract people “with nothing who do something profound” and a world where another power sets the terms of technological progress.

The current regulatory environment, however, is actively working against that outcome in the partners’ view. New GPU capacity is being located in Mexico and Australia because US permitting is so difficult. Horowitz argued for a new standard for data center construction — one where facilities contribute power back to the grid, avoid water waste, and act as good neighbors. He pointed to existing data centers that already operate this way, with electricity rates declining year over year in communities that host them.

The unresolved tension is whether US policy can catch up to the speed of the market. The partners are simultaneously bullish on American innovation and frustrated with American bureaucracy. The buildout now underway will test whether the country can coordinate industrial policy, workforce training, and grid investment at the pace the technology demands. The answer will determine not just the returns of the Machine Age Fund, but which country hosts the next generation of AI infrastructure — and the economic leverage that comes with it. For investors, the signal is clear: the bottleneck has moved from the cloud to the concrete, and the companies that can solve for power, memory, and physical construction will own the next decade of value creation.