CUDA was a “bet” that caused Nvidia’s stock market value to plummet by a good 80 percent. CEO Jensen Huang explains how it came about – and why he stuck to the strategy anyway.
Nvidia CEO Jensen Huang recently spoke at length about the history of his company in a podcast. In particular, the CUDA platform launched in 2006 was discussed, which according to Huang almost ruined Nvidia:
That [CUDA] was the first strategic decision that came closest to an existential threat.
The first step over 20 years ago
The technical foundation for CUDA was laid back in 2003, when Nvidia built IEEE-compatible 32-bit floating point calculations – also known as FP32 – into its shader units.
- This meant that scientific code that was actually designed for CPUs could, in principle, also run on an Nvidia GPU.
- Researchers already knew how to take advantage of this at the time, so CUDA (“Compute Unified Domain Architecture”) followed as a logical step in the context of a fully-fledged architecture.
A decision that halved Nvidia’s stock market value
However, the idea of this technology was not the real drama, but the commercial implementation.
CUDA increased our costs by around 50 percent, and at the time we were a company with a gross margin of around 35 percent. Our margin dropped by about one and a half dollars per chip.
The consequence on the stock market was brutal: Nvidia’s market capitalization fell from around eight billion to just under 1.5 billion US dollars after the CUDA launch – downright ridiculous figures compared to today’s stock market value.
The “CUDA bet” paid off
Nvidia’s CEO nevertheless felt it was essential to bring CUDA to customers via Geforce graphics cards: If CUDA was to have a chance as a new computing architecture, it had to end up in the hands of as many people as possible.
In Huang’s eyes, the principle behind this strategy is quickly explained: “The install base defines an architecture. […] Everything else is secondary”.
- It took several years for the tide to turn. In 2012, the neural network “AlexNet” beat all competitors by more than ten percentage points in the ImageNet competition.
- The underlying hardware? Nvidia graphics cards with CUDA – and suddenly everyone was talking about the architecture.
Background:A quarter of a century ago, a student connected 32 Geforce graphics cards to play Quake 3. This is how CUDA was born
Looking back, Huang believes that Nvidia’s success is based on Geforce; after all, it was these graphics cards that brought CUDA “to everyone”.
At least the current market situation proves him right, as CUDA is by far the dominant platform for AI training and inference – and is no longer the face of a self-destructive Nvidia bet, but one of the most valuable tech companies of our time.
