Taking on Nvidia and betting on transformers with Gavin Uberti from Etched

20 Aug 2024 (29 days ago)

Etched's Business Model and Market Position

Etched is a company that produces specialized AI chips designed specifically for running Transformer models, aiming to be significantly faster than competitors by focusing on a single domain. (00:02:27)
Etched is developing hardware specifically designed for transformer models, unlike competitors like Nvidia whose hardware supports a broader range of models.
Etched believes they have an 18-month lead on competitors like Nvidia and Google in developing transformer-specific hardware and liken the potential market shift to the adoption of ASICs in Bitcoin mining.
Etched is going to focus on building specialized chips for one specific AI model.
There are few areas outside of cybersecurity where a single-feature startup can be successful.
There is a lack of case studies to determine if a company focused on one specific AI chip can succeed.
Etched, a company focused on building specialized AI chips, may expand to build other types of chips in the future.

The company was founded in 2022 when Transformers were a risky bet, with no major use case established yet. (00:03:26)
The release of ChatGPT, while not fully validating Etched's approach, demonstrated the demand for high processing power in AI, leading to successful fundraising and hiring for the company.
The cost of training large language models is increasing, with models like GPT-5 projected to cost over a billion dollars to train, making specialized hardware like Etched's a more economically viable option.
Etched has secured tens of millions of dollars in contracts for their new AI chips before the first production run has even been completed.
Some potential customers who were unable to secure a contract for the first production run became investors in Etched.
Fundraising was not a significant challenge for Etched, as the company was able to secure funding from existing investors and new investors, including Hummingbird, Fundo, Oceans, Two Sigma, and Skybox.
Etched was able to raise a significant amount of money very quickly after the release of ChatGPT-4.

Gavin Uberti traveled the world working for a company called octoml where he developed AI models. (00:15:47)
Uberti believes that only 3% of the transistors in an Nvidia h100 are used for the math blocks that perform multiplication.
Uberti believes that gpt3 was a turning point in AI development because it demonstrated that scaling up models could lead to significant improvements.
Etched is building a system, not just a chip, and requires a platform, high-quality software, a customer success organization, and talented individuals to run these operations.
Gavin Uberti, CEO of Etched, prioritizes hiring exceptional talent, emphasizing the importance of a fully in-person company culture with dedicated individuals passionate about building future technology.
The speaker believes their job as CEO has three components: taking care of the employees, setting the long-term vision, and ensuring the company doesn't run out of money.
The speaker believes that building computers is a social problem and that building unique products requires a team effort. (00:28:03)
Etched's founder, Gavin, believes their valuation is meaningless unless they create a product that people want to buy and use.

Meta published a paper on their large language model, Llama 3, revealing that their training clusters crashed every 3 hours on average.
Etched has a larger stock option pool than other companies at a similar stage because they want to recruit the best talent and compensate them fairly.
Chris had to wait two months to see a doctor.
Gavin and Chris's health insurance start date was moved from September 1st to August 1st. (00:30:25)
Etched is a young company whose founder did not initially think to get health insurance for himself or his employees.