A two-year-old startup just walked out of stealth mode with $800 million in funding, over $1 billion in forward sales contracts, and a chip it claims can replace up to 160 Nvidia H100 GPUs for certain workloads. Etched officially introduced itself to the world on June 30, carrying a $5 billion valuation.

The company’s weapon of choice is the Sohu chip, an application-specific integrated circuit (ASIC) built exclusively for transformer model inference. Rather than building a general-purpose GPU that can do many things well, Etched built a chip that does one thing, running transformer-based AI models, and claims to do it dramatically better than anything else on the market.

What Etched is actually building

Founded in 2022 by Harvard Thiel Fellows Gavin Uberti and Chris Zhu, Etched has assembled a team of over 400 engineers, many poached from Nvidia, Google’s TPU division, and Broadcom. The company isn’t just designing chips in isolation. It’s building full rack-scale inference systems, co-designing the hardware components to squeeze maximum performance out of every watt.

The Sohu chip is fabricated on TSMC’s advanced N4P process node. First-pass silicon has already been achieved, meaning the initial physical chips came back from the fab working as intended.

Etched claims the Sohu can hit throughput of 500,000 tokens per second on Meta’s Llama 70B model. The company also touts advantages in latency, power consumption, and cost per inference. Initial shipments are scheduled for summer 2026.

The money trail tells an interesting story

The $800 million in total funding includes a major tranche of roughly $500 million that closed between December 2025 and January 2026. That round valued Etched at $5 billion post-money.

The investor list includes Jane Street, Peter Thiel, and a venture capital partner linked to TSMC itself.

Then there’s the $1 billion in forward sales contracts. These aren’t vague expressions of interest or handshake deals. Forward sales contracts represent committed revenue, suggesting that major customers have already decided the Sohu’s value proposition is compelling enough to lock in orders before the systems have even shipped at scale.

The ASIC bet versus Nvidia’s GPU empire

Etched’s argument is that the AI inference market is different from training. Training requires flexibility and programmability, which is where general-purpose GPUs excel. Inference, the process of actually running a trained model to generate outputs, is more predictable and repetitive. That predictability is exactly what ASICs are designed to exploit.

By hardwiring transformer operations directly into silicon rather than running them on general-purpose hardware, Etched is betting that specialization will outperform generalization on raw performance metrics. Google built its own TPU chips for similar reasons. Amazon developed Inferentia and Trainium. The difference is that Etched is trying to sell its systems externally rather than keeping them in-house.

There are currently no public independent benchmarks validating any of Etched’s performance claims. The 160-to-1 GPU replacement ratio and the 500,000 tokens-per-second throughput figure come from the company itself. Until third-party testing confirms these numbers under real-world conditions, they remain marketing claims.

Beyond Nvidia, AMD is pushing into AI inference with its Instinct accelerators. Intel has its Gaudi line. Startups like Groq, Cerebras, and SambaNova are all chasing variations of the same thesis: that specialized hardware can outperform Nvidia’s GPUs for specific AI workloads.