Fifteen years ago, tech investor Marc Andreessen published his famous essay, “Why Software Is Eating the World.” He predicted at the time that technology companies were tremendously undervalued, and that low startup costs and almost infinite scalability would lead software-based companies to dominate every industry.
You can see what he means. Today, the “Mag 7” stocks dominate the S&P 500 with market capitalizations in the trillions. Even startups like Anthropic and OpenAI are valued at hundreds of billions of dollars. Meanwhile, massive investment in data centers is reshaping industries from construction to energy.
But not so fast. While recent advances in machine learning have been exciting, it’s still unclear how much real value is being created. The truth is that we still live our lives largely in the realm of atoms and that isn’t changing. That’s why software is unlikely to ever eat the world, and why many of the most exciting technologies of the future will be rooted in physical space.
The economy is not digital
In his essay, Andreessen wrote, “Today, the world’s largest bookseller, Amazon, is a software company—its core capability is its amazing software engine for selling virtually everything online, no retail stores necessary.” Well, not really. While software remains a core part of Amazon’s business, today the company is firmly ensconced in the physical world, with not only retail stores but also hundreds of warehouses and a massive fleet of trucks.
It’s not just Amazon. Most of our economic lives are rooted in atoms, not bits. A quick examination of your monthly bills will likely show that most of your spending goes to things like housing, transportation, energy, food and, depending on your age, health care. That dwarfs what most people spend on phones, computers, and internet services.
In fact, a report by the International Data Center Authority found that the digital economy accounts for a mere 15% of global gross domestic product. That’s a lot of money in nominal terms, but it’s still dwarfed by the other 85%. That’s why Amazon went to the expense and trouble of investing in physical spaces. Even Netflix, another company Andreessen touted, is opening up real-life entertainment centers.
As much as we may seem glued to our phones, the physical world is where we live. It’s where we eat, work, meet each other, and have fun. It’s what nature evolved us for, which is why Zoom calls are never quite as satisfying as real-life encounters. Software has a big appetite, but the real world is simply too big and complex to be eaten.
Still, there is genuine opportunity in using software to shape the physical world in ways that unlock enormous value.
Matter is not digital
Materials are something we interact with constantly, often without thinking about them. We want our clothes to be soft and warm, our tools to have high tensile strength so they can do work without breaking. Some things require specific properties, such as the ability to conduct electricity or resist shattering on impact.
This has long been the realm of a fairly obscure field called materials science, and, traditionally, it has been something akin to a cottage industry. Scientists would begin with a set of desired properties and then, through a painstaking process of trial and error, often involving the testing of thousands of candidates, eventually find something useful.
But in the early 2000s, an MIT professor named Gerd Ceder began developing computational methods to predict new materials. That eventually led to the Materials Project at Lawrence Berkeley National Laboratory. Now, rather than testing thousands of candidates, scientists could eliminate most of them through digital simulations and then test the ones that remain.
As more materials data became available, two Stanford graduate students started applying machine learning to materials databases and found that they could dramatically improve development economics. The company they founded, Citrine Informatics, became a pioneer in the space and has attracted large players such as Dassault, Schrödinger, and Microsoft.
Still, materials are not digital, so there will always be some loss in information when digital systems are used to model physical reality. However, we are beginning to see the emergence of non-digital architectures, such as quantum computers, that can model the physical world with far greater fidelity.
Biology is not digital
In 2024, Demis Hassabis and John Jumper won the Nobel Prize for their development of AlphaFold, an AI model that can predict, with incredible accuracy, the structure of proteins. This was a breakthrough of historic proportions because, much like computational approaches in materials science, it allows scientists to identify potential drug candidates hundreds, if not thousands, of times faster than with conventional methods.
The potential is mind-blowing. In 2023, Insilico, a Hong Kong-based biotech startup, advanced the first AI-generated drug candidate into human clinical trials. And there are currently dozens of potentially life-changing drugs in the pipeline that were discovered in a mere fraction of the time that it would take using conventional methods.
That’s impressive. But like materials, biology is not digital. No matter how ingeniously conceived and constructed, we still need to see how a therapy works on humans in the real world. We need to be sure that proposed cures are safe, nontoxic, and an improvement on existing molecules and methods. That, and not drug discovery, is what makes up the bulk of development costs.
A 2024 paper suggested that AI discovery could double the overall success rate from 5% to 10% to 9% to 18%, which is significant. Still, the claims of the tech optimists that “AI will cure cancer” are more than overblown. Anybody who has spent any time in a hospital will tell you that healthcare remains incredibly labor-intensive, requiring capable, caring professionals. And there is an extreme shortage of them in the U.S., which software will do little to solve.
We need to focus more on atoms, less on bits
Fifty years ago, in 1976, life expectancy in the U.S. was 72 years, versus 78 today. American families typically had one car and one television. Houses were smaller, nutrition was worse, we polluted like hell, and there was no internet. We spent much less time with our screens and more time with each other.
Today, it’s easy to see how many things have gotten better, but it’s just as easy to see how others have gotten worse. While in the aggregate, incomes have improved, most of that has gone to top earners, leaving many households feeling worse off. While we have amazingly cool gadgets, costs for basic needs, like housing, healthcare, and education, have soared.
The truth is that we’re very good at innovating in the digital space because it’s fast, cheap, and low risk. But the real opportunities are in the messy, physical world. So we’re ending up with lots of incremental digital innovation and not enough transformational change in the real world.
In sum, it’s hard to see how we’ve become meaningfully better off over the last 50 years. For all of the Silicon Valley blather, most American families are materially struggling and our mental health is declining. This isn’t because of some exogenous shock, but because of choices we’ve made. We have the technology to improve our lives, but the benefits are not accessible to most.
What we have to reckon with is that the world is not digital. We live, eat, travel, and breathe in physical spaces, and no amount of algorithms and data centers will change that. As philosopher Martin Heidegger pointed out long ago, technology is less a creation than it is an uncovering. It brings us possibilities, but it is our responsibility to enframe and direct them in ways that will benefit us.
We live in a world of atoms, not bits. Technology matters only if it makes our lives better.
