Stark&Cie
EN

Writing · Capital

Efficiency is the new moat

DeepSeek wiped $589bn off Nvidia in a day by doing more with less. The lesson for founders is older than AI: whoever needs the least money is hardest to catch.

A single microchip on a light grey surface, lit from the side

On 27 January 2025 Nvidia lost about $589bn of market value in one trading day, the largest one-day loss for any company in American stock-market history. The trigger was a Chinese lab, DeepSeek. It said it had trained its V3 model on Nvidia’s weaker H800 chips for less than $6m.

The figure was disputed, and it covered only the final training run. It did not matter. Investors had priced the AI boom on one assumption: the winners would be whoever spent the most. DeepSeek suggested that a small team with a good idea could get close to the frontier for the price of a villa on the Zürichberg. The panic lasted days, and by October Nvidia was the first company worth $5tn: cheaper models had made more people want chips. The ones left exposed were the labs that had planned to win by spending the most.

What money buys

For a decade the venture industry treated capital as a strategy. Raise more than the rival, hire faster, buy the market and sort out the economics later. That works when money is cheap and the goal is distribution. It works less well when the product is a machine or a model, where the hard part is knowing what to build.

WhatsApp is the old proof. When Facebook agreed to buy it in February 2014 for $19bn, the company had 55 employees, 32 of them engineers, serving 450m monthly users. It had no marketing department. Facebook paid for the users, but the users were the result of discipline: one product, done well, by people who refused to grow the headcount faster than the problem.

What efficiency means at the start

In a young company there is nothing to cut yet. Efficiency there is the ratio between what you spend and what you learn. A prototype that shows the customer does not care is a bargain. An office and a brand agency paid for before that prototype are a waste.

We look for two habits. The team ships something real within weeks, and it can say what it will not do. DeepSeek did not try to beat OpenAI everywhere. It attacked the cost of training and running a model, the place where the incumbents had been least careful.

The moat moves

Large companies are optimised for a world that is ending. When the cost of a capability falls by an order of magnitude, the advantage moves from whoever owns the most of it to whoever needs the least. That holds for compute, energy and people alike. It is why small models and small reactors keep turning up in our portfolio, built by small teams.

Capital still matters. It buys time. But a company that needs ten times its rival’s money to reach the same point owns a bill and calls it a moat.

Sources

  1. Bloomberg: Nvidia’s $589 billion DeepSeek rout is largest in market history, 27 January 2025
  2. Reuters, via The Kathmandu Post: DeepSeek sparks AI stock selloff, 28 January 2025
  3. NPR: Facebook will buy WhatsApp for $19 billion, 19 February 2014
  4. InformationWeek: Facebook’s WhatsApp buy, 10 staggering stats
  5. DeepSeek-AI: DeepSeek-V3 technical report, December 2024
  6. AP, via OPB: Nvidia is the first $5 trillion company, 29 October 2025

More writing