The Chip That Could Break NVIDIA's AI Empire
Two Harvard dropouts just raised $120M to build a chip that does one thing: run the AI that powers ChatGPT.
And, should NVIDIA be sweating? Let’s unpack.
In 2024, Etched introduced Sohu — a chip designed for one mission and one mission only: running transformer models faster, cooler, and cheaper than anything on the market. [ANNOUNCEMENT]
While NVIDIA makes chips that can run any type of artificial intelligence, Sohu can only run transformer models—the technology that powers ChatGPT, Claude, and every AI tool you've probably used this week.
It has been a year since that bold announcement, and the AI chip battlefield has never been more chaotic.
👉 Get your brand in front of 35,000+ decision-makers.
The $4 Trillion Fortress
NVIDIA just made history. On Wednesday, the company became the first public company ever to reach a $4T market valuation, surpassing Apple and Microsoft.
NVIDIA controls 90% of the AI chip market. Every tech giant—Meta, Google, Microsoft, Amazon—has to write billion-dollar checks to Jensen Huang's company if they want to stay competitive in AI.
The numbers are staggering. NVIDIA reported $44.1B in revenue last quarter, up 69% from a year ago. The company is worth more than the entire stock markets of Canada and Mexico combined, and exceeds the total value of all publicly listed companies in the UK.
But even empires face challengers.
In March 2025, a Chinese startup called DeepSeek released a chatbot that allegedly used far less computing power than American rivals. NVIDIA's stock price collapsed overnight. Suddenly, everyone started asking the same question: Do we really need these $50,000 chips to run AI?
The answer depends on what you're trying to do.
Training AI models—teaching them from massive datasets—still belongs to NVIDIA. But running those trained models to answer questions? That's a different game entirely. And it's a much bigger market.
"The market for running AI models is going to be many times bigger than training them," says Jay Goldberg, an industry analyst. "NVIDIA might own a smaller slice of a much bigger pie."
This shift has unleashed chaos. At least 60 startups are now hunting NVIDIA's business. Even tech giants are building their own chips rather than paying NVIDIA's premium prices.
The opportunity in dominance: If you want to build anything with AI, you need their chips. There's no real alternative.
Meta wants to compete with ChatGPT? Write a check to NVIDIA. Microsoft wants to keep up with Google? Write a bigger check. It's the classic monopoly playbook, and it's made NVIDIA the most valuable company in the world.
But market dominance, this completely always creates massive opportunities for companies willing to think differently.
The One-Trick Pony Strategy
While everyone else tries to build chips that can run any AI model, Etched's founders—Gavin Uberti and Chris Zhu—went completely the other direction.
Sohu is as an ASIC tuned exclusively for transformers. This chip runs exactly one type of AI model: transformers. It can't run older AI models. It can't run different types of AI models. It just runs transformers.
Back in 2017, Google researchers accidentally created what became the most important AI technology ever. They called it "transformers" and it quietly took over the world.
This sounds crazy until you realise that transformers power everything you actually use: ChatGPT, Claude, Gemini, video generation, image creation, and pretty much every AI breakthrough that's made headlines since 2022.
"In 2022, we made a bet that transformers would take over the world," Uberti said. That bet is looking pretty smart right now.
By the data: By building a chip that only runs transformers, Etched claims performance that sounds impossible:
500,000 tokens per second (incredibly fast for AI)
20 times faster than NVIDIA's current best chips
10 times faster than NVIDIA's next-generation chips
One Sohu server replaces 160 NVIDIA chips
The secret? They removed everything that doesn't help transformers run faster. No support for other AI models. No general computing. Just laser focus on the one thing that matters most right now.
Three things happened that made Etched's approach suddenly viable:
Chips stopped getting better. For four years, chip performance has barely improved. NVIDIA's solution? Make chips physically bigger.
AI became a massive business. Training AI models costs $1B. Running them costs $10B. At this scale, even small improvements are worth millions.
Everyone's betting on Transformers. AI companies have spent hundreds of millions optimising for transformers. They're not switching architectures anytime soon.
By only running transformers, they can pack way more math power into the same space. They get over 90% efficiency compared to 30% on NVIDIA chips.
The Risk
Etched's strategy has some massive holes:
