Etched Closes $300M Series C at $10.3 Billion on 23 July 2026
Etched, an AI chip startup founded in 2022 by Harvard undergraduates who left to build application-specific inference hardware, closed a $300 million Series C on 23 July 2026 at a valuation of $10.3 billion. The round was led by Sequoia Capital, with participation from Andreessen Horowitz, Jane Street, Diffusion, and SK Hynix. The financing doubles Etched's December 2025 valuation of $5 billion, where the company had raised a $500 million round. Entering this Series C, Etched had raised approximately $800 million in total across all prior rounds. The company reports more than $1 billion in signed customer contracts for its Sohu chip system, with first rack shipments scheduled during summer 2026 and production capacity targeted to scale toward gigawatt levels by 2027.
The Sohu Chip: Built Solely for Transformer Inference
Most AI accelerators are general-purpose processors adapted for machine learning workloads. Etched has taken a different approach: the Sohu chip is an application-specific integrated circuit designed to do a single thing exceptionally well — transformer inference, the computation that runs when a trained language or vision model generates an output. Built on TSMC's 4-nanometre process, Sohu hardwires the full transformer computation graph directly into silicon rather than executing it through a programmable GPU pipeline. This eliminates the overhead that general-purpose chips carry when running workloads they were not specifically designed for. The result is a chip that cannot run arbitrary software but can run transformer models at a depth of optimisation that GPU architectures structurally cannot match.
Performance: 500,000 Tokens Per Second versus 25,000 on H100
Etched's published benchmarks place the Sohu system at 500,000 tokens per second on the Llama 70B model. A comparable H100 configuration achieves approximately 25,000 tokens per second on the same workload — making Sohu roughly 20 times faster on that benchmark. For inference use cases, where AI models are served to many concurrent users, throughput is the metric that maps directly to cost: more tokens per second means more queries served per dollar of hardware. Etched has positioned the Sohu system for inference providers and large enterprises that serve high query volumes, and for whom per-token cost at scale is a primary operating constraint.
The Architectural Limitation That Defines Sohu's Market
The specialisation that makes Sohu fast also defines its boundary. The chip is designed exclusively for dense transformer architectures. Mixture-of-Experts models — including DeepSeek V4 and Qwen3-235B-A22B, two of the most widely deployed open-weight models as of mid-2026 — are not compatible with the Sohu chip. This is a significant constraint because MoE models have become a standard architecture choice for developers seeking high benchmark performance with lower active parameter counts. Operators who have built production inference infrastructure around MoE open-weight models cannot migrate to Sohu without also switching their model choices.
Why the AI Inference Hardware Market Is Worth Competing For
The economics of running AI at scale have made inference hardware one of the most valuable positions in the technology stack. Training a large language model is periodic; inference is continuous and scales with every user interaction. As AI applications move from prototype to production and user bases grow into the hundreds of thousands, inference cost becomes the primary operating expense for AI product companies. Nvidia's H100 and successors dominate this market, but their general-purpose GPU architecture carries overhead that a purpose-built inference accelerator can eliminate. Etched is not the only company attempting to capture this margin — Google builds TPUs, Amazon ships Trainium and Inferentia, Groq builds Language Processing Units, and Cerebras produces wafer-scale systems — but Etched's claim to the deepest transformer-specific hardware specialisation distinguishes it from competing approaches.
What Etched Means for Indian AI and Software Teams
For Indian AI product companies and software teams running inference-heavy workloads on dense transformer models, Etched's progress is directly relevant to infrastructure cost planning. Inference cost is often the largest variable expense for AI applications serving significant user volumes, and a 20-times throughput improvement — if verified under production conditions — could substantially reduce per-query cost versus GPU-based deployments. Three questions matter for Indian teams evaluating Etched: first, whether their production models are dense transformers or MoE architectures, since only dense models run on Sohu; second, whether their query volumes are high enough for dedicated inference ASICs to make economic sense; and third, whether Etched's production ramp can serve Indian cloud or on-premises deployments by the time procurement cycles open in 2027. SK Hynix's participation in the round is notable for another reason: the Korean memory company is already embedded in India's AI supply chain through its HBM partnerships with Nvidia, and its investment in Etched signals that the inference ASIC market is large enough to attract the companies already supplying the existing GPU stack.
The Bottom Line
On 23 July 2026, Etched closed a $300 million Series C at $10.3 billion, led by Sequoia with Andreessen Horowitz, Jane Street, Diffusion, and SK Hynix. The round funds production and deployment of the Sohu chip — a TSMC 4nm ASIC hardwired for transformer inference — which claims 500,000 tokens per second on Llama 70B versus approximately 25,000 from a comparable H100 configuration, a roughly 20-times throughput advantage on dense transformer workloads. With $1 billion in signed contracts and first rack shipments expected this summer, Etched is the most capitalised dedicated-transformer-inference startup in the market. The critical constraint for Indian and global teams: MoE architectures including DeepSeek V4 and Qwen3-235B-A22B are incompatible with Sohu, limiting the chip's applicability to teams whose production models are standard dense transformers.
Frequently Asked Questions
What is the Etched Sohu chip and what does the $300M Series C fund?+
The Etched Sohu is an application-specific integrated circuit designed exclusively for transformer model inference, built on TSMC's 4-nanometre process. Unlike general-purpose GPUs, it hardwires the transformer computation graph directly into silicon, eliminating the programmability overhead that GPU architectures carry on inference workloads. Etched closed a $300 million Series C led by Sequoia Capital on 23 July 2026 at a $10.3 billion valuation — double its December 2025 valuation of $5 billion — with participation from Andreessen Horowitz, Jane Street, Diffusion, and SK Hynix. The funding supports production ramp-up and customer deployments of the Sohu system, with first rack shipments expected during summer 2026 and production capacity targeted toward gigawatt levels by 2027.
How does the Sohu chip's performance compare to Nvidia's H100 GPU?+
Etched's benchmarks show the Sohu system delivering 500,000 tokens per second on the Llama 70B model, compared to approximately 25,000 tokens per second from a comparable H100 configuration — a roughly 20-times throughput advantage on that specific workload. This performance gap reflects the benefit of transformer-specific silicon: Sohu executes the transformer computation path without the programmability overhead that makes GPUs suitable for a wide range of tasks but introduces inefficiency on narrow, well-defined workloads. The benchmark applies to dense transformer models only. Sohu is not compatible with Mixture-of-Experts architectures such as DeepSeek V4 and Qwen3-235B-A22B, two of the most widely deployed open-weight models as of mid-2026.
What AI models cannot run on the Sohu chip and why does this matter?+
The Sohu chip does not support Mixture-of-Experts architectures. MoE models route each token through only a subset of expert layers rather than the full parameter set, requiring different computational patterns than the dense transformer architecture Sohu is optimised for. As of mid-2026, the most widely deployed open-weight models include DeepSeek V4 and Qwen3-235B-A22B, both MoE architectures and both incompatible with Sohu. Organisations that have built inference infrastructure around these models cannot migrate to Sohu without also changing their model choices. Dense transformer models including Llama-family models and standard GPT-style architectures are compatible with the Sohu chip.
Why is SK Hynix investing in Etched and what does it signal for the AI chip market?+
SK Hynix is one of the world's two leading suppliers of High Bandwidth Memory, the specialised DRAM used in AI accelerators including Nvidia's H100 and H200. Its investment in Etched signals that the Korean memory company sees the inference ASIC market as large enough to sustain hardware ecosystems beyond Nvidia's GPU platform. For SK Hynix, investing in a Nvidia alternative hedges its memory supply business: if inference ASICs capture a significant share of the AI inference market, SK Hynix wants to be supplying memory to those chips as well as to GPU-based systems. For the broader AI chip competitive landscape, SK Hynix's participation alongside Sequoia and Andreessen Horowitz reinforces the thesis that dedicated inference silicon is developing into a standalone product category rather than remaining a niche alternative to GPU inference.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.