In August 2011, Marc Andreessen published an essay in The Wall Street Journal arguing that software was “eating the world.” It was brilliant, prescient, and — like most Silicon Valley manifestos — contained just enough truth to be dangerous. Fourteen years later, the software that ate the world is itself being eaten. Not by better software, mind you. By something that writes software in its sleep, hallucinates occasionally about database schemas, and charges you per token for the privilege.
What follows is a forensic examination of Andreessen’s original claims, each lovingly dismantled by the arrival of the very AI technology his firm now writes the biggest checks for. We do not do this out of spite. We do it because irony is the only currency that appreciates faster than Nvidia stock.
Point I: ”Software Companies Are Poised to Take Over Large Swathes of the Economy”
“We are in the middle of a dramatic and broad technological and economic shift in which software companies are poised to take over large swathes of the economy.”
Adorable. In 2011, the idea was that if you wrote enough code, you could dominate any industry. Borders lost to Amazon because Amazon was a “software company.” Beautiful narrative. Clean. Investable.
Here in 2026, you don’t need a software company to write software. You need a prompt. The moat Andreessen described — proprietary code built by expensive engineers over years — now has the defensive strength of a sandcastle at high tide. A nineteen-year-old with Claude can replicate your SaaS product over a weekend, ship it on Monday, and be competing with you by Tuesday. The software isn’t eating the world anymore. It’s being extruded — at industrial scale, like sausage, by machines that don’t need health insurance or ping-pong tables.
Point II: The Cost Advantage That Made Software King
“In 2000, the cost of running a basic Internet application was approximately $150,000 a month. Running that same application today in Amazon’s cloud costs about $1,500 a month.”
Andreessen was correct that plummeting infrastructure costs would democratize software creation. What he didn’t anticipate was the logical conclusion of his own argument: if the cost of building software approaches zero, then software itself ceases to be a differentiator.
In 2026, the cost of generating a working application from a natural language description is approximately one latte and forty-five minutes of vibe coding. The $1,500/month cloud bill remains, of course — Amazon always gets paid — but the engineering team that once justified a Series A? Optional. The “software company” premium that inflated valuations for a decade was, it turns out, just the labor cost of a skill that machines were about to commoditize. Andreessen celebrated the 100x reduction in hosting costs. AI delivered a 1000x reduction in creation costs. The infrastructure got cheap. Then the builders got cheaper.
Point III: Amazon, Netflix, and the Software-Eats-Retail Thesis
“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.”
In 2011, Amazon’s advantage was that it had better software than Borders. Today, Amazon’s advantage is that it has better AI than its own software. The recommendation engine, the supply chain optimization, the warehouse robotics, the dynamic pricing — all of it is now driven by machine learning models, not the hand-crafted algorithms Andreessen was celebrating.
Netflix doesn’t win because it wrote great streaming software. It wins because an AI decides what you watch before you’ve finished your previous show. The “software company” label is now a quaint anachronism, like calling Tesla a “battery company.” The software is table stakes. The intelligence layer on top is the product. Borders didn’t lose to software. Borders lost to data. And data, as it turns out, is just AI’s breakfast.
The software engineer was the 21st century’s most overpaid medieval scribe — essential right up until the printing press shipped.— Every recruiter’s LinkedIn post, probably
Point IV: ”Every Company Needs to Become a Software Company”
“Companies in every industry need to assume that a software revolution is coming.”
They did assume it. They spent fifteen years and trillions of dollars becoming “software companies.” They hired armies of engineers. They built “digital transformation” departments. They had Agile coaches. They had Scrum Masters. They had a person whose entire job was deciding which shade of blue a button should be.
And now AI is telling them: you didn’t need to become a software company. You needed to become a company that uses AI. The software was just the awkward intermediary phase. Every enterprise that spent a decade migrating to microservices is now discovering that an AI agent can accomplish what their 200-person platform team built, except faster, and without the passive-aggressive Slack threads about code review standards.
Andreessen’s advice was correct for exactly one economic cycle. “Become a software company” was the 2011 equivalent of “become an electricity company” in 1910. Yes, you needed electricity. No, generating your own wasn’t the point. The software revolution AI revolution means the software layer is becoming invisible — a utility, not a competitive advantage.
Point V: The Talent Wars That No Longer Make Sense
“Qualified software engineers, managers, marketers and salespeople in Silicon Valley can rack up dozens of high-paying, high-upside job offers any time they want, while national unemployment and underemployment is sky high.”
This was arguably Andreessen’s most prescient observation — and the one AI has most brutally invalidated. In 2011, the software talent shortage was real, structural, and seemingly permanent. Companies fought over engineers like medieval lords fighting over arable land.
Fast-forward to 2026, and the equation has inverted. AI coding assistants have made individual developers 5–10x more productive, which means companies need 5–10x fewer of them. The remaining engineers are more valuable than ever — if they know how to architect AI systems. But the vast middle tier of “write CRUD apps and attend standups” developers? Their roles are evaporating faster than a puddle in the Mojave.
The great irony is that Andreessen identified the problem — not enough people could participate in the software revolution — and AI solved it. Not by training more engineers, but by making engineering itself semi-automated. The solution to the talent shortage was, it turns out, to eliminate the need for most of the talent. Be careful what you disrupt; you might be next.
Point VI: The Defensibility Illusion
“We believe that many of the prominent new Internet companies are building real, high-growth, high-margin, highly defensible businesses.”
Let us pour one out for “defensibility.” In 2011, Andreessen listed his portfolio: Facebook, Groupon, Skype, Twitter, Zynga, Foursquare. Let’s check the scoreboard. Groupon: a coupon company cosplaying as tech. Zynga: a memory. Foursquare: a trivia answer. Skype: absorbed into the Microsoft Borg. Twitter: purchased by a man who renamed it after a variable.
The “defensibility” of software was always more fragile than VCs admitted. Code can be replicated. Network effects can be disrupted. And AI has accelerated this to absurd speeds. The average SaaS product now has a competitive lifespan somewhere between a fruit fly and a Snapchat story. When anyone can build a “highly defensible” software business in a weekend, the definition of “defensible” needs some urgent revision. Perhaps “defensible” now means “we have a moat of regulatory compliance so byzantine that not even AI wants to deal with it.”
Software ate the world. AI ate the software. VCs ate the AI companies’ margins. The circle of Silicon Valley life.— The ouroboros accelerates
Point VII: Healthcare and Education: “Next Up for Transformation”
“Health care and education, in my view, are next up for fundamental software-based transformation.”
Reader, it is 2026. Healthcare is still running on fax machines. Education still uses Blackboard. Software did not transform these industries because software was the wrong tool. You can’t CRUD your way out of a system where a hospital’s billing department has more employees than its emergency room.
But AI — AI might actually do it. Not because it’s better software, but because it’s not software at all in the traditional sense. An AI that can interpret medical imaging, explain a diagnosis in plain language, or tutor a student at their individual pace isn’t a “software application.” It’s a cognitive prosthesis. Andreessen was right about the destination but wrong about the vehicle. Software couldn’t eat healthcare because healthcare fights back. AI might succeed because it doesn’t write code — it reads X-rays.
Point VIII: The American Advantage
“Our combination of great research universities, a pro-risk business culture, deep pools of innovation-seeking equity capital and reliable business and contract law is unprecedented and unparalleled in the world.”
This remains true, but with a twist Andreessen didn’t see coming. The American advantage in AI isn’t software engineering — it’s compute access and capital concentration. The U.S. leads in AI not because of Stanford CS graduates (though they help), but because Nvidia is an American company, the hyperscalers are American companies, and the venture capital willing to light billions on fire for foundation model training is American capital.
The software-era advantages Andreessen listed — universities training engineers, a culture of risk-taking startups — are being replaced by a new set: whoever has the most GPUs wins. The pro-risk business culture now means “willing to spend $10 billion training a model that might or might not work.” The great research universities are important not because they produce programmers, but because they produce researchers who understand transformer architectures. The game has shifted from “who can write the best code” to “who can afford the most electricity.”
Point IX: ”I Know Where I’m Putting My Money”
“Instead of constantly questioning their valuations, let’s seek to understand how the new generation of technology companies are doing what they do.”
The most delicious irony of all: Andreessen Horowitz, the firm built on the thesis that software eats the world, is now one of the largest investors in AI — the technology that eats software. a16z’s portfolio page reads like a confession that the 2011 essay’s thesis has an expiration date. They’re not betting on software companies anymore. They’re betting on AI companies that make software companies obsolete.
And to his credit, Andreessen would probably argue this is exactly what he meant all along — that the best software would eventually write itself. That’s the beauty of being a venture capitalist. Your thesis is always right. It’s just that the timeline and the mechanism and the entire category of investment were slightly different than described. Details.
The Updated Thesis
So here is the corrected version of the essay, updated for 2026:
Software AI is eating the world software, which already ate the world, so technically AI is eating the world at one degree of separation.
The companies that will dominate the next decade are not “software companies.” They are intelligence companies — organizations that harness AI to do what previously required thousands of engineers, or millions of lines of code, or a McKinsey engagement. The software layer hasn’t disappeared. It’s just no longer the point. It’s the plumbing. And nobody writes breathless Wall Street Journal op-eds about plumbing.
Andreessen was right that technology would transform every industry. He was right that the transformation would be led by companies in Silicon Valley. He was even right that “the best of the new companies will grow even faster” once conditions stabilized. He just didn’t realize that the new companies would be the ones making his old companies’ core competency — writing software — about as special as knowing how to operate a photocopier.