Etched raises $300m Series C at $10.3 billion valuation as Sohu chip boasts 20x H100 inference speed

Etched, a developer of application-specific AI inference chips, has completed a $300 million Series C funding round led by Sequoia Capital, with participation from a16z, Jane Street, Diffusion and SK Hynix. The investment values the company at $10.3 billion, roughly double the $5 billion valuation it reached in its previous round.

Supporting the valuation are a successfully taped-out A0 chip, a frontier inference cluster now under construction and more than $1 billion in customer contracts already in hand. Etched expects to ship its first racks in the summer of 2026 and has begun ramping volume production to fulfil those orders.

Etched was founded by Harvard alumni Gavin Uberti, Chris Zhu and Robert Wachen. Uberti previously worked on AI compilers at OctoML and Xnor. ai and contributed to the Cortex-M backend of TVM; Zhu brings a background in mathematics and high-performance computing; Wachen leads commercialisation, fundraising and organisation building. The team has since expanded across chip architecture, software, rack systems and manufacturing.

The leadership group includes CTO Mark Ross, former chief technology officer of Cypress Semiconductor; Saptadeep Pal, who worked on the architecture of Nvidia’s V100, A100 and H100 GPUs and now leads ASIC and low-level architecture at Etched; Brian Loiler, who built HGX and DGX systems at Nvidia and oversees platform engineering; David Munday, who established Google’s TPU software teams from v1 to v5 and leads the software stack; and Wayne Cao, who brings hardware volume-production experience from Apple and Google and manages production and supply chain.

Etched’s earlier funding rounds trace its rapid ascent. In 2024 it raised a $120 million Series A led by Primary Venture Partners, with backing from Peter Thiel and GitHub CEO Thomas Dohmke. By late 2025 it had closed a $500 million Series B led by Stripe, joined by TSMC, Jane Street, Ribbit Capital, Andrej Karpathy, Fei-Fei Li and Peter Thiel, taking its valuation to $5 billion.

The Sohu chip is an ASIC designed exclusively for transformer-model inference. In 2024, internal benchmarks showed that a server equipped with eight Sohu chips could process 500,000 tokens per second running Llama 70B, 20 times the throughput of an eight-H100 server. Earlier in 2026, the company taped out the A0 revision on TSMC’s process and brought up its first chip cluster within 40 days.

Etched developed a low-voltage inference architecture that runs compute units at less than half the voltage of most AI chips, yielding several times the FLOPs density. When inferencing trillion-parameter sparse mixture-of-experts models, the chip sustains more than 80% of peak FLOPs without triggering clock throttling. For the decode phase, Etched built cluster-level memory with proprietary ultra-low-latency, high-bandwidth interconnects that enable fast cross-chip memory access; a hybrid HBM and SRAM design addresses both memory capacity and latency.

In early customer testing, Etched says its system has delivered state-of-the-art throughput, latency and energy efficiency on inference workloads. According to the company, feedback gathered by lead investors during due diligence indicated that Etched is the first system capable of fundamentally reshaping the unit economics of inference at scale.

The demand backdrop is striking. Google disclosed that the number of tokens processed by its products and APIs each month grew from 9.7 trillion in May 2024 to 480 trillion in May 2025 and surpassed 3.2 quadrillion by May 2026, a more than 300-fold increase over two years. The shift reflects a product evolution from chatbots to reasoning models such as o1 and on to AI agents, which consume far more tokens per task. By 2026, ChatGPT’s weekly active users exceeded 900 million, while Codex and ChatGPT Work together reached 10 million users. Nvidia’s data-centre revenue illustrates the same trend: quarterly revenue rose from $3.62 billion in early 2023 to $22.6 billion in spring 2024 and $62.3 billion by early 2026.

In China, the gap between domestic models and global leaders has narrowed significantly. DeepSeek has challenged top global players on reasoning capability, open-source ecosystem and cost efficiency, while Kimi’s K3 ranks among the top three globally on agentic and long-horizon knowledge tasks. Both companies have open-sourced software-level inference optimisations: DeepSeek released FlashMLA, DeepGEMM, DeepEP and TileLang, and Kimi released Mooncake. Kimi’s K3 also demonstrated the ability to autonomously complete the design and simulation verification of a dedicated chip for a nano-model within 48 hours, using open-source EDA tools and a 45nm process library.