Funding and Strategic Backing
Published 7/2/2026, 1:36:49 AM
Etched’s total funding of $800 million—highlighted by a $500 million Series A1 in late 2025 at a $5 billion valuation—signals a shift toward hardware specialization in AI infrastructure. By developing Sohu, a Transformer-only ASIC (Application-Specific Integrated Circuit), Etched is betting that the AI industry will remain consolidated around the Transformer architecture, allowing them to outperform Nvidia’s general-purpose GPUs in inference efficiency.
Funding and Strategic Backing
The capital raise includes significant institutional support from Jane Street (investing over $100M), Stripes, and VentureTech Alliance, a firm closely linked to TSMC [Source: https://www.bloomberg.com/news/articles/2026-06-30/ai-chip-startup-etched-says-jane-street-tsmc-linked-vc-invested]. Etched claims to have secured over $1 billion in signed customer contracts and plans to ship its first hardware racks in summer 2026.
Competitive Landscape: ASIC vs. GPU
Etched enters a market currently dominated by Nvidia, which recently consolidated its position by acquiring the AI chip startup Groq for approximately $20 billion in December 2025 [Source: https://www.cnbc.com/2025/12/24/nvidia-buying-ai-chip-startup-groq-for-about-20-billion-biggest-deal.html].
| Feature | Etched (Sohu) | Nvidia (H100/B100) | Groq (LPU) |
|---|---|---|---|
| Architecture | Transformer-only ASIC | General-purpose GPU | Language Processing Unit |
| Performance | 20x faster (claimed) | Industry Baseline | High-speed sequential |
| Primary Focus | LLM Inference at scale | Training + Inference | Real-time inference |
| Business Model | Full-scale hardware racks | Ecosystem/CUDA software | Neocloud/API services |
While Nvidia is integrating Groq’s LPU technology into its own "Groq 3 LPX" chips, Etched remains an independent hardware provider. Its Sohu chip uses a low-voltage design intended to prevent thermal throttling, claiming a 20x throughput advantage over Nvidia H100s for models like Llama 70B. Meanwhile, Groq (post-acquisition) raised $650 million in June 2026 to scale its "inference cloud" business [Source: https://groq.com/newsroom/groq-raises-usd650m-to-scale-its-ai-inference-cloud-business].
Implications for Crypto Infrastructure
The emergence of specialized AI ASICs like Sohu has three primary impacts on crypto-compute and decentralized infrastructure:
- DePIN Obsolescence Risk: Decentralized Physical Infrastructure Networks (DePIN) like Akash or Render, which largely utilize consumer or general-purpose GPUs, may face a competitive disadvantage. If specialized ASICs capture the projected 45% of the inference market by 2030, decentralized networks will need to pivot toward supporting these ASICs to remain viable for high-end AI workloads.
- Infrastructure Synergy: The power requirements for Etched clusters (often requiring 2MW+ datacenter standards) are highly compatible with existing large-scale Bitcoin mining and high-performance computing (HPC) sites. This may lead to a "merger" of crypto mining facilities and AI inference centers.
- Supply Chain Gatekeeping: Etched’s reliance on TSMC for fabrication underscores that access to leading-edge silicon remains the ultimate bottleneck for both AI and crypto hardware [Source: https://www.bloomberg.com/news/articles/2026-06-30/ai-chip-startup-etched-says-jane-street-tsmc-linked-vc-invested].
Strategic Risks
The primary risk for Etched is architectural rigidity. Because the Sohu chip is "hardwired" for Transformers, any industry shift toward new architectures (such as State Space Models or Mamba) could render their hardware obsolete. Additionally, Nvidia’s aggressive acquisition strategy—demonstrated by the Groq deal—suggests that incumbent players may continue to buy out or suppress architectural threats before they achieve mass-market scale.
In summary, Etched's $800M raise validates the "ASIC-first" approach to AI, forcing crypto-compute providers to choose between remaining general-purpose or investing in specialized, high-efficiency hardware that risks obsolescence if AI model architectures evolve.