ETCHED

Etched started with a bet most chip experts thought was reckless. In 2022, Gavin Uberti and Chris Zhu dropped out of Harvard to build it, and Robert Wachen joined as the third founder.

Their reasoning went like this. Nvidia GPUs are flexible, they can run all kinds of AI.

But flexibility costs speed. Uberti and Zhu bet that transformers, the design behind ChatGPT, had already won and were not going anywhere.

So they decided to burn that one architecture directly into the silicon. No flexibility, just raw speed for transformers.

Everyone said it was too risky to bet the whole company on a single architecture. They did it anyway.

Etched sells a specialized AI chip. Most AI runs on Nvidia GPUs, which are powerful but general-purpose, meaning they can handle many kinds of computing.

Etched went the opposite way. Its chip, called Sohu, is built to run only transformer models, the kind that power ChatGPT and similar systems.

By giving up flexibility, Etched claims Sohu can run those models much faster and cheaper than a GPU. The plan is to sell these chips, and the servers built around them, to companies that run huge AI workloads and want to slash their computing bills.

The one product that matters is Sohu, Etched's transformer-only AI chip. The company claims a single Sohu server can replace many Nvidia GPU servers for running transformer models, at a fraction of the cost and power.

Around the chip, Etched is building the full server systems that customers would actually plug in. Everything the company does points at one goal, running transformer models faster and cheaper than anything else on the market.

Etched's growth plan is really a timing bet. The founders decided the transformer had already won the AI race, so a chip built only for transformers could beat Nvidia's do-everything GPUs on speed and cost.

Instead of chasing broad appeal, they went narrow and deep. The pitch attracted big-name money fast.

Their $120 million Series A in 2024 pulled in famous investors, which handed a tiny startup instant credibility in a field ruled by giants. The strategy is to be the fastest option for the one thing that matters most right now.

Etched is making a huge bet on a single idea. Its entire chip only runs transformer models.

If the AI world shifts to a different architecture, that specialized chip loses its reason to exist. It is also going head to head with Nvidia, the most valuable company on the planet, whose chips are the industry default and whose software everyone already uses.

Convincing customers to switch is brutally hard. On top of that, Etched still has to actually ship Sohu at scale and prove the eye-popping speed claims in the real world.

Talk is cheap. Silicon is not.

Etched raised $120 million in a Series A in 2024, led by Primary Venture Partners and Positive Sum. The eye-catching part was the list of individual backers.

Peter Thiel and the legendary macro investor Stanley Druckenmiller both put money in, alongside a crowd of other tech and finance names. For a company founded by two Harvard dropouts a couple of years earlier, that kind of investor list is a serious vote of confidence.

It signaled that some very smart money believed the transformer bet was worth chasing, even against Nvidia.