RT Etched: We're coming out of stealth. We've built our first racks after a successful A0 tapeout, $1B+ in customer contracts, and $800m raised. Early...
Etched has exited stealth mode, announcing that it has built its first racks after a successful chip tapeout, secured over $1 billion in customer contracts, and raised $800 million. Early customer tests show state-of-the-art throughput, latency, and power efficiency on inference workloads. Their first racks are scheduled to ship this summer.
Background
- **Etched** is a startup building a specialized chip (ASIC) designed exclusively for running transformer-based AI models — the architecture behind ChatGPT, Claude, Gemini, etc. Unlike GPUs (e.g., Nvidia's H100), which are flexible general-purpose processors, Etched's chip is hardwired for transformers only, sacrificing versatility for extreme speed and efficiency on inference (running trained models, not training them). The "A0 tapeout" means the design has been finalized and sent to fabrication.
- This post signals strong commercial validation: $1B+ in customer contracts and $800M raised privately already, before first shipments. The chips ship this summer.
- Etched's claim of best-in-class throughput, latency, and power efficiency on inference directly challenges Nvidia's dominance. If true, these chips could dramatically lower the cost of serving AI models, a key bottleneck for widespread deployment.
- This is part of a broader trend: multiple startups (Groq, Cerebras, d-Matrix) are building custom inference chips to compete with Nvidia, betting that transformer models will dominate AI for the foreseeable future, making extreme specialization a winning bet.