Three Harvard Dropouts Just Raised $300 Million to Build the AI Chip That Rivals Keep Calling Impossible

Key points

• Etched raised $300 million in a Series C round led by Sequoia, valuing the company at $10.3 billion as of July 2025.
• The startup's valuation doubled from $5 billion in roughly seven months after a $500 million raise in December 2024.
• Etched has booked $1 billion in orders and shipped its first full systems to early customers for testing.
• The chips, made by contract manufacturer TSMC, are designed to speed up AI inference, the computing work that happens every time you type a prompt and wait for an answer.
• Co-founder Robert Wachen says the systems can run any modern AI model, not just one type.

Etched, an AI chip startup co-founded by three Harvard dropouts in 2022, just closed a $300 million Series C funding round at a $10.3 billion valuation, first reported by TechCrunch. Sequoia led the deal. Andreessen Horowitz, South Korean memory giant SK Hynix, trading firm Jane Street, and Diffusion Capital also participated.

The numbers move fast. In December 2024, Etched was valued at $5 billion after raising $500 million. Seven months later, that figure has more than doubled. The company calls this the highest valuation ever for a Sequoia-led Series C, which would be a notable record in a market that has seen plenty of large rounds lately.

What does Etched actually make?

Etched builds chips and sells them as complete rack systems, the kind of computing hardware that big AI companies slot into data centres. The chips are designed for one specific job: inference.

Inference is the computing work that happens after you type something into ChatGPT or a similar tool and hit send. It splits into two stages. First, the system reads and understands your prompt, a step called prefill, which is heavy on raw computing power. Then it writes the answer back to you, a step called decode, which needs less computing power but enormous amounts of fast memory.

Etched built a separate chip for each stage. The prefill chip runs at unusually low voltage, which generates less heat and lets engineers pack in more transistors (the tiny switches that do the actual computing). The decode chip uses a new kind of shared memory that lets many chips work together across what Etched calls "cluster scale memory," cutting the time each chip spends waiting for data.

The promise is faster answers at lower cost for AI companies, and eventually for their customers.

Should anyone be sceptical?

Yes, genuinely. The company has kept its hardware behind closed doors, showing demos only to investors and early clients. Public benchmarks do not yet exist.

Co-founder and COO Robert Wachen acknowledges the scepticism. The systems can run any AI model, he says, including newer Mixture of Experts models such as DeepSeek and Qwen (an architecture that splits tasks across several smaller specialist models rather than one giant one) and non-transformer designs like Mamba. Critics had assumed the chips were locked to a single type of AI, which would be a serious limitation.

Famous names have tried the hardware and come away impressed, Wachen says, citing AI researcher Andrej Karpathy, OpenAI's Noam Brown, and Nobel laureate Geoffrey Hinton. That list is credible. It is also, by definition, a curated one.

Survivorship bias is real here. You are hearing about the startup that raised $300 million, not the ones that ran the same race and lost.

What happens next?

Etched has 400 employees, runs a 2-megawatt data centre, and is now working with some of the largest AI companies in the world, according to Wachen. The next step is mass production and delivery of rack systems, which is where hardware startups most often stumble.

"We still have to be humbled by what it will take to actually get to scale," Wachen said.

If Etched delivers, faster and cheaper inference could quietly lower the cost of every AI tool you already use. That is the bet $300 million just bought.