Etched Raises $800M, Partners with TSMC on AI Inference Chips

What happened

Etched announced it has raised $800 million across four rounds, including a $500 million round in December led by Stripes. The company disclosed its investor base, which includes VentureTech Alliance (linked to TSMC), Jane Street (over $100 million invested), Peter Thiel, hedge funds like Hudson River Trading and Two Sigma, and AI researchers Geoffrey Hinton and Fei-Fei Li. The startup achieved first-pass silicon success on its Sohu chip, manufactured using TSMC's N4P process node. Etched also revealed it has signed $1 billion in sales contracts and plans to begin shipping full inference systems to customers in summer 2026. The company employs over 400 engineers recruited from Nvidia, Google, Broadcom, TSMC, and SK Hynix, and is building full rack-scale systems including custom chips, circuit boards, cooling, and networking.

Why it matters

This development highlights growing competition in the AI inference chip market, where specialised hardware is emerging to challenge Nvidia's dominance. Etched's partnership with TSMC validates the commercial viability of its approach and provides access to advanced manufacturing capacity. The $1 billion in forward sales contracts suggests major customers are already committing to Etched's technology before full-scale production, indicating strong demand for inference-focused hardware. For TSMC, the strategic investment through VentureTech Alliance expands its exposure to the AI inference market beyond GPU manufacturing, potentially creating a new revenue stream as the industry shifts focus from model training to deployment. The successful N4P tapeout demonstrates TSMC's process technology can support specialised AI chip designs, not just general-purpose GPUs. Etched's Low-Voltage Inference (LVI) architecture, which operates at half the voltage of typical AI chips while maintaining 80% peak performance, could differentiate TSMC's manufacturing capabilities in power-efficient AI hardware.

Bigger picture

The AI chip landscape is fragmenting as companies move beyond training workloads toward inference optimisation. Google built internal TPUs, Amazon developed Inferentia and Trainium, Microsoft designs its own accelerators, and OpenAI recently announced a custom chip with Broadcom. Startups like Groq (recently raised $650 million) and Cerebras (completed a breakout IPO) are pursuing similar specialisation strategies. Etched represents another bet that application-specific integrated circuits (ASICs) hardwired for transformer models can outperform general-purpose GPUs on cost, latency, and power efficiency for inference workloads. The company claims its Sohu chip can replace up to 160 Nvidia H100 GPUs for certain tasks and achieve 500,000 tokens per second on Meta's Llama 70B model, though these figures await independent validation. TSMC's involvement through VentureTech Alliance positions the foundry to capture market share regardless of which inference chip architecture wins, while also gaining early insight into emerging AI hardware requirements.

What to watch

Monitor whether Etched delivers systems on schedule in summer 2026 and whether independent benchmarks validate its performance claims against Nvidia GPUs. Watch for customer announcements, as the $1 billion in contracts remains undisclosed. Track whether other TSMC partners or competitors respond with similar low-voltage or inference-specific designs. Observe how hyperscalers balance internal chip development with external procurement from startups like Etched. Any delays in TSMC's N4P production capacity or yield issues could affect Etched's ability to scale, making TSMC's manufacturing execution a key variable. Finally, watch whether Nvidia responds with specialised inference products or maintains its general-purpose GPU strategy.