The VC Playbook for AI Is Broken

VCs are pattern-matching to SaaS playbooks. AI doesn't work that way. The winners look nothing like the bets.

10 min read

The Wrong Map for the Territory

Venture capitalists are applying the old software playbook to AI investments. It's like using a road atlas to navigate the ocean. The traditional VC model assumes you can build differentiated software with a small team and modest capital. AI breaks that assumption in spectacular ways.

The SaaS playbook that minted billionaires for two decades was elegant in its simplicity: low upfront cost, recurring revenue, expanding margins, winner-take-most markets. AI has none of these properties. The upfront cost is enormous. The revenue model is uncertain. The margins often get worse at scale. And the market structure might be winner-take-all, which means 99% of investments go to zero.

The Old Playbook

Classic SaaS investing formula: Seed at $2M to build MVP. Series A at $10M to find product-market fit. Series B at $25M to scale sales. Growth rounds to pour fuel on the fire. Total to meaningful scale: $50-100M over 5-7 years.

AI investing reality: You need $100M just to train a competitive model. And that's before you've proven anything. Before you have a single customer. Before you know if the model is even good enough to sell.

The mismatch between VC fund structures and AI capital requirements is creating a peculiar dynamic. Most VC funds are $200-500M. A single AI bet at the foundation layer could consume 20-50% of the fund. No rational portfolio manager would concentrate that heavily on one company. So either fund sizes need to increase dramatically, or VCs need to accept they can't play at the foundation layer.

Winner-Take-All Economics

Software markets typically support 3-5 meaningful players. AI foundation model markets support maybe 2. Why? Because the feedback loops are vicious:

Data flywheel: Better models attract more users → more usage data → better models. The rich get richer, exponentially.

Talent concentration: The top 100 AI researchers globally can make or break a company. They cluster around the best-funded, most prestigious projects.

Infrastructure advantages: Once you've built the training infrastructure, marginal costs approach zero. First-mover advantages compound.

In AI, second place isn't the first loser. It's the last company anyone remembers. Look at the history: who was the second-best search engine in 2005? The second-best smartphone OS in 2015? Markets with strong network effects and scale economies don't support runner-ups.

The Portfolio Theory Problem

VCs love portfolio theory: back 20 companies, expect 1-2 massive winners to return the fund. This works when startup outcomes are somewhat independent. But in AI, outcomes are highly correlated.

If OpenAI builds AGI first, it doesn't just kill its direct competitors — it potentially makes every other AI investment worthless. Traditional portfolio: independent bets with uncorrelated outcomes. AI portfolio: highly correlated bets where one winner could make all others irrelevant.

This breaks the fundamental math of venture capital. Diversification doesn't protect you when all your bets are essentially the same bet wearing different clothes. And most AI portfolios are exactly that — variations on "this team can build a better model/wrapper/application than that team."

The Burn Rate Reality

AI startups burn cash at rates that would make even crypto projects blush. A "lean" AI startup might burn $5-10M per month just on compute. Add talent costs (top ML engineers cost $500K+ annually), and you're looking at run rates that traditional VC math can't support.

VCs are used to companies that become more capital efficient over time. AI companies often become less efficient as they scale — more data requires bigger models, bigger models require more compute, more compute requires more capital. It's the opposite of the SaaS efficiency curve.

What Actually Works

Smart money is adapting. Here's the emerging playbook:

1. Go extremely big or go home. Don't write $5M cheques to foundation model companies. Write $100M cheques or find a different investment.

2. Focus on application layer with strong moats. Find companies that use AI but have defensibility beyond their model quality. Proprietary data, regulatory approval, embedded workflows.

3. Invest in picks and shovels. Infrastructure, tools, and services that benefit regardless of which foundation models win.

4. Look for data monopolies. Companies with unique datasets that can't be replicated or commoditised.

The reckoning is coming. Funds that applied software economics to AI companies are going to get wiped out. The math is simple: if you need $1B to build a competitive foundation model, and traditional VC funds are $500M, you can't play the game. Either the fund sizes need to change, or the investment strategy needs to change. Most VCs are trying to avoid both. The result? A lot of very expensive lessons about why AI isn't just software.