Transformers & NVIDIA — The Virtuous Cycle

This is written for my friends with no AIML background who are killing it in the stock market, especially with Tech Stocks.

This blog was initially published at https://sanjaysays.com, followed by Medium.

You know how everyone jokes that NVIDIA is the only company actually making money from AI? There’s a really interesting reason for that, and it’s all about these things called transformers (not the robot kind!).

Think of transformers like super-efficient brains that get smarter the more data you feed them. Unlike older AI models that had limits, these ones just keep getting better with more information. Pretty cool, right?

It’s a pretty sweet deal for NVIDIA — they basically hit the tech lottery by having exactly what everyone needed right when they needed it. And now they’re so far ahead, it’s super hard for anyone else to catch up!

Now, here’s where NVIDIA got lucky — their graphics cards (GPUs) turned out to be perfect for running these transformer AIs. It’s like they accidentally built the perfect tool for the job before anyone knew what the job would be! Their GPUs are really good at doing tons of calculations at once, which is exactly what transformers need.

The tech world went all-in on transformers (think ChatGPT and friends), and NVIDIA had time to make their hardware work even better with them. It’s kind of like having a head start in a race and then also being the only one who knows the track really well.

Here’s the kicker — when companies want to build really big AI models, they need thousands of NVIDIA’s GPUs. Nobody else’s hardware has been tested at that massive scale. So even when competitors like AMD or Intel say they’ve built something just as good, big tech companies stick with NVIDIA because they know it works.

It’s a pretty sweet deal for NVIDIA — they basically hit the tech lottery by having exactly what everyone needed right when they needed it. And now they’re so far ahead, it’s super hard for anyone else to catch up!

Note: For a more technical version, please see The Sequence article