How to Dropout and Raise $120 Million to Beat Nvidia

Earlier this year, my friend Joshua Rahn called me about a new AI chip company launched by three college dropouts to take on Nvidia.

I have no experience in the semiconductor business. So I ignored Josh.

Luckily, one of my partners also heard about the company, and we decided to dig in. I’m psyched we did.

Today, the three visionary founders, all Thiel Fellows, made a significant announcement. Their company Etched secured $120 million in Series A funding, a testament to their potential. Velvet Sea Ventures and a dozen early Facebook executives and renowned figures like Peter Thiel and Stanley Druckenmiller joined in this exciting journey.

The founders do a much better job than I could ever do explaining the company’s technology here. And Josh provides epic color here, drawing correlations between the early days of Facebook and Etched. You can also see the vast coverage on CNBC, Bloomberg, TechCrunch, Reuters, VentureBeat, and many others.?

Why am I excited about the company??

First, you need to understand just a little about how AI works generally and large-language models like ChatGPT specifically.

Transformers are the backbone of every large AI product. It’s the T in ChatGPT. They are natural language processing (NLP) systems that understand and generate human language.

If chips that power today’s AI applications were baseball players, they would be utilitity players. They can play most positions.?

Nvidia chips that are used to power ChatGPT and other large language models can also be used to render graphics in your favorite video game, run advanced simulations and calculations, mine cryptocurrency, process real-time data from sensors to power autonomous vehicles and more. They are flexible. They can do almost anything you ask them to do.

Etched’s first product, Sohu, is the world’s first transformer-specific AI chip. No company has ever built an algorithm-specific AI chip. New chip designs are expensive to develop, costing up to $100 million and taking years to bring to production.?

The market for transformers is enormous, driven by the massive and unprecedented demand for AI applications like ChatGPT. And the Etched team thought it would be big enough two years ago to create a chip specifically to power them.

The result is impressive. Sohu is the fastest transformer chip of all time, the company reports. And it’s not even close - 20X faster than the fastest Nvidia chip running on Facebook’s popular Llama open-source language model.

How is Sohu so much faster than Nvidia’s most advanced chip? Because Sohu can only run one algorithm, the Etched team could remove hardware that transformers don’t need and streamline the software.

Critics of the technology believe the world will move away from transformers. Etched is taking other side of the bet and is going all in today on transformers.?

I am not qualified to opine. But I have seen the past three decades of tech unfold. Technology with huge moats often dominate for many, many years. And transformers have a huge moat that should endure for many years.?

What is Etched’s moat? The team points to the following:

• Transformers are behind every AI product. And the next-generation of language models rely on transformers.
• As the scale of AI model training increases exponentially, the cost and risk associated with testing new architectures also surge, making it more viable to develop features on existing transformer frameworks.?
• Current software libraries are specifically optimized for transformers, supporting complex functionalities not easily replicated in alternative architectures.?
• Next-generation hardware from the leader, NVIDIA, are tailored to enhance transformer efficiencies, indicating a long-term commitment to this architecture.

Since starting the company after dropping out of Harvard 2 years ago, the founders have made significant progress and writes that it is “on track for one of the fastest chip launches in history.”

Top AI talent has left the largest AI chip projects to join the company. Etched has partnered with TSMC on their 4nm process to secure enough manufacturing capacity and supply from the top vendors to scale in its first year of production. And the market has reacted with early customers having booked tens of millions in reservations for its hardware.

What Etched is trying to do is hard and full of risk like all startups. Many are quick to point out all the reasons why it will fail. Those of us who invested asked a different question: what if they succeed?

And that’s why I’m excited about Etched.

If the team succeeds, AI applications that take 60 seconds to a few minutes today to render a response become real-time, feeling more like Google search than 1996 AOL dial-up. New applications in communications, medicine, finance and more that are technically impossible today will become possible. And all of it will be done at a much lower cost, democratizing AI beyond those who are fortunate enough to be able to spend $20 per month for the latest applications.

Congrats to the Etched team on the Series A round! Now the real work begins.