In 2011, Marc Andreessen declared that 'software is eating the world,' defining an era of startup growth. Now, the capital-intensive reality of artificial intelligence is testing every part of that thesis.
The Gospel of Zero Marginal Cost
For over a decade, Marc Andreessen's essay was
the closest thing Silicon Valley had to a sacred text. His argument was simple and powerful: software companies could scale globally with minimal friction. Unlike building a car factory, creating a new software product allowed a small team to serve millions of users at nearly zero additional cost per user. This dynamic produced businesses with incredible gross margins—often 80% or higher—meaning most of each revenue dollar could be reinvested into growth, hiring, and sales. Venture capital flowed to startups that could build a product once and sell it infinitely, disrupting slower, capital-heavy incumbents in industries from music to finance. This was the playbook that built the modern tech landscape.
The AI Cost Anomaly
Artificial intelligence breaks this model. Building and running a sophisticated AI startup is brutally expensive. Instead of near-zero marginal costs, AI companies face significant variable costs that scale with usage. Every time a user generates an image, summarizes a document, or asks a question, it triggers a cascade of computations on power-hungry GPU clusters, often owned by cloud giants like Amazon, Google, or Microsoft. These 'inference' costs can consume a massive portion of revenue. Furthermore, the talent required to build and maintain these systems is scarce and commands premium salaries, while the vast datasets needed for training can cost tens of thousands to hundreds of thousands of dollars to acquire and clean. Recent reports show a single training run for a large model can cost millions.
The Great Margin Squeeze
The result is a direct assault on the beautiful economics of traditional software. While a classic software-as-a-service (SaaS) business might boast 80% gross margins, many AI startups are struggling to hit 50-60%. Some of the fastestest growing 'supernova' AI companies have reported margins as low as 25%, and sometimes even negative, as they burn cash to acquire users. This fundamental shift in unit economics changes the entire venture capital calculation. Lower margins mean less money to fund growth and a much longer, more capital intensive path to profitability. It has led Andreessen himself to comment on a new reality where, in some inefficient deployments, 'humans are cheaper than software,' a direct inversion of his original thesis.
A New Playbook for a New Era
This new economic reality is forcing a strategic pivot across Silicon Valley. The astronomical costs seem to give an inherent advantage to incumbents like Microsoft and Google, who own the underlying cloud infrastructure and can afford to spend billions. Yet, venture firms, including Andreessen's own a16z, are adapting rather than retreating. Investment is pouring into AI startups, but the focus is shifting. VCs are now backing companies building the picks and shovels of the AI gold rush: new types of chips, networking hardware, and compute platforms that promise to drive down costs. For application layer startups, the key to survival is no longer just having a clever model, but possessing a unique advantage that restores margins. This could be a proprietary dataset, a go to market strategy that locks in high value enterprise customers, or an architecture that intelligently routes tasks to cheaper, smaller models. The game has changed from simply building smart software to building a defensible—and profitable—business model around it.
