Etched AI chip startup reaches $10.3B valuation on transformer-optimized inference gains

Etched, the AI chip startup founded by three Harvard dropouts in 2022, has raised $300 million in Series C funding at a $10.3 billion post-money valuation, according to TechCrunch. The round was led by Sequoia Capital and included participation from Andreessen Horowitz, SK Hynix, Jane Street, and Diffusion Capital, alongside earlier backers including venture investor Peter Thiel, Tesla AI director Andrej Karpathy, and other prominent founders.

The valuation represents a 106% increase from Etched’s December 2025 Series B, when the company was valued at $5 billion after raising $500 million. This seven-month climb reflects accelerating investor appetite for specialized inference silicon as transformer-based models become computationally dominant in production deployments.

Custom silicon for transformer inference bottlenecks

Etched’s investment thesis rests on optimizing the two distinct computational stages of transformer inference. According to COO Robert Wachen, the prefill phase processes the user’s prompt and context—a mathematically intensive task requiring sustained computation. The decode phase generates output tokens sequentially, demanding less total compute but requiring extremely low latency and high memory bandwidth.

Etched designed custom silicon for each stage. The prefill chip operates at significantly lower voltage than competing AI accelerators, reducing heat dissipation and allowing denser transistor packing. For decode, the company engineered cluster-scale memory architecture that allows multiple chips to share a memory pool at low latency, reducing per-inference cost while maintaining throughput.

Manufacturing and customer traction

In June 2026, Etched announced it had successfully manufactured its custom chips and begun customer testing of complete systems. The company reported $1 billion in booked orders at the time of that announcement, providing revenue visibility that likely accelerated Series C fundraising.

The startup initially faced skepticism that its hardware was locked to specific large language models or transformer architectures. Wachen clarified to TechCrunch that Etched’s systems support any model architecture—including Mixture-of-Experts designs like DeepSeek and Qwen, as well as non-transformer models like Mamba, which use state-space architecture rather than attention mechanisms.

Why This Matters

Etched’s $10.3 billion valuation signals investor conviction that specialized inference silicon, once dismissed as an architectural dead-end, now addresses a real scaling bottleneck. Google’s reported Frozen v2 chip for Gemini—also optimized for specific model families—suggests the industry-wide shift toward inference-focused custom silicon is genuine rather than isolated.

For enterprises operating large-scale LLM inference, Etched’s prefill-decode separation offers a concrete path to reducing latency and cost simultaneously. The $1 billion in backlog indicates deployment is moving beyond R&D labs into production workloads. If Etched’s per-inference cost claims hold up under independent benchmark testing, it could reshape procurement decisions at cloud providers and heavily-inference-dependent applications, particularly for real-time use cases where transformer latency is the limiting factor.