DeepSeek Changed the Economics of AI Overnight

In January 2025, DeepSeek released R1. Within weeks, it became the most downloaded app in the world, overtaking ChatGPT. The model performed on par with OpenAI's o1 on most benchmarks. The cost to train it? Reportedly under $6 million. For context, GPT-4 cost over $100 million.

That is not a marginal improvement. That is an order-of-magnitude cost collapse. And it broke something fundamental in the prevailing narrative around AI.

The Narrative That Broke

For two years, the dominant AI thesis was simple: whoever spends the most on compute wins. Scaling laws. Bigger clusters. More GPUs. The arms race was measured in billions of dollars and hundreds of thousands of H100s.

DeepSeek shattered that thesis. Not with more money, but with better ideas. Mixture-of-experts architectures. Smarter training recipes. Ruthless efficiency in how they used limited hardware.

The market reacted violently. Nvidia lost nearly $600 billion in market cap in a single day. Broadcom, AMD, and the entire semiconductor supply chain followed. Not because AI was over, but because the assumption that AI required infinite capital to compete was suddenly in question.

What This Means for Founders

If you are building an AI startup, the DeepSeek moment is the best thing that could have happened to you. Here is why:

The cost floor just dropped. Inference costs that were $30 per million tokens two years ago are now under $1 for comparable quality. If your startup's unit economics depended on model costs declining, the timeline just accelerated by 18 months.

The moat moved. It is no longer about access to the biggest model. It is about what you build on top of the model. Application layer. Vertical expertise. Data flywheels. Distribution. The things that were always supposed to matter now actually do.

China is not behind. If you are building with the assumption that US labs have an insurmountable lead, update that assumption. The talent density in Beijing and Hangzhou is real. The competitive pressure will only intensify.

Where We See Opportunity Now

The companies we are most excited about right now are not building foundation models. They are building specific, vertical products that take advantage of near-free intelligence. Healthcare diagnostics. Legal document analysis. Financial compliance automation. These businesses just became 10x more viable.

We are also watching the infrastructure layer closely. If everyone can train competitive models cheaply, the bottleneck shifts to deployment, evaluation, and monitoring. The picks-and-shovels thesis still holds. It just got repriced.

The DeepSeek moment is a reminder of something we have always believed: the most important breakthroughs do not come from the biggest budgets. They come from the sharpest minds working under constraints. That is exactly the kind of founder we back.