When Marc Andreessen declared in 2011 that “software is eating the world,” he gave the technology industry more than a memorable line. He provided the defining investment thesis for the next decade and a half.

Software moved into practically every industry. Retailers became software-powered logistics operations. Banks became software companies with banking licenses. Entertainment, communications, transportation and advertising were reorganized around applications and digital platforms. Cloud computing allowed businesses to rent infrastructure instead of building it, while software-as-a-service companies demonstrated the financial allure of recurring revenue and asset-light growth.

Andreessen was right. Software did eat the world.

Which is why it is hard to ignore the symbolism of Andreessen Horowitz launching a $1.1 billion fund dedicated to hardware.

The firm’s new Machine Age Fund will invest in AI processors, memory, networking, storage, data centers, robotics, home AI appliances and other pieces of the physical infrastructure required to build and operate artificial intelligence. According to The Wall Street Journal, it is A16z’s first dedicated hardware-infrastructure fund.

The headline almost writes itself: When the man who said software is eating the world starts a hardware fund, something fundamental must have changed.

It has. But this is not a repudiation of Andreessen’s original thesis. It is the next chapter.

Software ate the world. Now AI is eating software.

When Software Becomes Raw Material

AI is not eating software in the same way software consumed newspapers, travel agencies or video-rental stores. Software is not going away. If anything, we will probably produce more code than ever.

But software’s position in the value chain is changing.

AI models train on vast bodies of software. AI systems now write code, test applications, review changes, find vulnerabilities and operate production environments. Agents increasingly allow users to request an outcome without opening, navigating or perhaps even knowing which traditional applications were used to produce it.

The software is still there. It is simply moving behind the intelligence layer.

That can turn software from the finished product a customer purchases into an ingredient used by an AI system. Code becomes training material, generated output, agent infrastructure or one component among many inside a larger system designed to deliver an outcome.

We should not get ahead of ourselves. Traditional applications are not about to disappear; every model is not interchangeable and frontier-model performance still matters. But the direction is becoming visible. AI is absorbing more of the work previously performed through conventional software and more of the user relationship that software companies once controlled.

For years, the technology industry moved value from hardware into software. AI may be pushing some of that value somewhere else again.

AI Has a Physical Appetite

Traditional software benefited from making the machines underneath it easy to ignore. Developers could call an API, spin up a cloud instance or add capacity without worrying about where the server sat, which power plant supplied it or how the underlying processor was manufactured.

AI is tearing through that abstraction.

AI needs compute in much the same way appliances need electricity. The model or agent is what the user experiences, just as the refrigerator or television is what the consumer sees. But the appliance is useless without a power system behind the electrical outlet.

In this analogy, models and agents are the appliances. Compute is the electricity. Chips and servers are part of the generating equipment. Networks and interconnects perform a role similar to transmission infrastructure, while data centers increasingly resemble industrial-scale power plants for intelligence.

The analogy is not perfect, but it helps explain why A16z believes the entire physical infrastructure underneath AI must be rebuilt “all the way down to the electricity.”

The firm says compute density per rack has increased approximately 28 times between an Nvidia H100 rack and a Rubin rack. Rack power has climbed from roughly 5 to 10 kilowatts to between 100 and 250 kilowatts for current systems, with one-megawatt racks potentially arriving within three years. Data center development is moving from facilities measured in tens of megawatts toward campuses measured in hundreds of megawatts or even gigawatts.

Memory bandwidth, networking capacity, cooling, materials, electricity and real estate are all becoming limiting factors. A16z says hardware startups have grown from a small part of its deal flow to more than 20%.

The firm is not entirely new to physical technology. It previously invested in SpaceX, Anduril, Skydio and Waymo, among others, and more recently backed companies including Unconventional AI, Nexthop, Volta, Atoms, Heron Power and Mind Robotics. The Machine Age Fund formalizes that activity and gives it a much larger pool of dedicated capital.

A16z is following scarcity. AI can generate a million lines of code, but it cannot prompt another gigawatt of electricity, a semiconductor fabrication plant or a million accelerators into existence.

That makes the constrained physical layers of the AI economy extremely valuable. At least for now.

The Warning Hidden in the Electricity Analogy

The electricity comparison explains the opportunity A16z sees. It also reveals the risk.

Electricity became indispensable to the modern economy. Demand grew beyond anything its original producers could have imagined. Yet indispensability did not give electric utilities unlimited freedom to set prices or determine who received service.

It produced the opposite response.

The more essential electricity became, the less society was willing to tolerate unfettered private control over it. Governments imposed rate regulation, reliability requirements, service obligations and rules governing access. Electricity remained enormously valuable to the economy, but the returns available to its providers were constrained.

That is the central argument of my forthcoming book, The Indispensability Trap: Whatever becomes essential eventually has its returns capped.

The trap does not require demand to disappear. Demand can explode while unit prices and producer margins decline. The important question is not merely which technology becomes indispensable. It is where durable margin eventually comes to rest.

AI infrastructure would appear to be a prime candidate for the same process. If compute becomes as essential to economic activity as electricity, governments and customers will not happily accept permanent scarcity, unpredictable access or dependence on one supplier. They will encourage new capacity, alternative architectures, interoperability, domestic production and multiple sources of supply.

National-security concerns add another layer. Advanced processors, data centers and power infrastructure are already subject to export controls, industrial policy and government intervention. The more strategically important AI becomes, the more government will shape who can build it, buy it and use it.

Indispensability creates demand. It also attracts competition, regulation and strategic capture.

The Appliances Fell Into the Trap, Too

The history of appliances supplies the second half of the analogy.

Once a standardized electrical infrastructure existed, appliance manufacturers no longer needed to include a private generator with every refrigerator, washing machine or television. They could assume their products would plug into an outlet and draw electricity from a common system.

That standardization unlocked enormous innovation. It also made it easier for more manufacturers to enter appliance markets. Over time, many product categories became crowded and increasingly interchangeable. Value migrated toward brands, design, retail distribution, customer relationships and whatever new layer could provide differentiation.

There was no rung on the ladder where a company could climb, declare itself indispensable and rest forever.

AI may travel a similar path. Compute becomes more available and utility-like. Models plug into compute. Applications plug into models. Agents connect those models to business processes. Orchestration platforms decide which model gets called, which data it can access and which actions it is permitted to take.

As each layer becomes standardized, value looks for somewhere else to go—toward proprietary context, permissions, workflow integration, evaluation, governance, control and distribution.

Intelligence may become indispensable without making any individual model, chipmaker, cloud provider or infrastructure supplier permanently indispensable.

Is A16z Funding the Platform or the Bottleneck?

This is the strategic question hanging over the Machine Age Fund.

The shortages A16z identifies are real. AI requires more processors, memory, networking, storage, cooling and electricity. The physical buildout could become one of the largest capital-investment cycles in history.

But identifying where demand is growing is not the same as determining where durable returns will settle.

Scarcity attracts capital. Capital attracts competitors. Engineering finds alternatives. Standards reduce switching costs. Customers develop second sources. Governments subsidize new capacity and intervene when dependence becomes politically unacceptable.

Some Machine Age investments may create platforms with durable architectural control. Others may supply components that are desperately needed today but increasingly interchangeable tomorrow. Hardware’s capital requirements, longer development cycles and manufacturing dependencies make that distinction especially unforgiving for venture investors.

A16z is therefore making two bets. The first is that AI will require an enormous physical infrastructure buildout. That seems increasingly difficult to dispute.

The second is that the firm can identify which parts of that infrastructure will retain pricing power after today’s bottlenecks ease and the Indispensability Trap begins to close. That is a much harder wager.

What Digital Leaders Should Take From This

Enterprise leaders should not read the announcement and conclude that every company needs to build a data center or design its own processor. The broader lesson is that AI strategy can no longer be separated from infrastructure strategy.

CEOs, CIOs and boards need to understand where their AI capabilities ultimately come from and how dependent the organization is on a particular model, cloud, accelerator or infrastructure provider. They need to know whether workloads and data can move, whether contracts preserve negotiating leverage and whether proprietary context and workflow integration belong to the enterprise or its vendors.

Buying access to the best available AI system is not the same as building a durable AI position.

Andreessen was right the first time. Software ate the world by making the machines underneath it increasingly easy to ignore. AI is now eating software while making those machines impossible to ignore.

Andreessen Horowitz is following value back into the physical stack. The question is whether it is investing in the next indispensable platform—or merely the next layer destined to fall into the trap.