Etched Sohu Chip Hits 500K Llama Tokens/Sec in $10.3B Round

Etched closed a $300 million Series C funding round to scale production of its transformer-specific Sohu chip, pushing the hardware startup’s valuation to $10.3 billion. The round, led by Sequoia Capital, marks the highest valuation recorded for a Sequoia-led Series C. The investment follows a private technical diligence phase involving AI researchers Andrej Karpathy, Noam Brown, and Geoffrey Hinton, who validated the architecture before joining as backers.

The company has amassed $1 billion in pre-booked orders for its frontier inference clusters. Etched focuses entirely on AI inference rather than model training, moving away from general-purpose GPUs toward Application-Specific Integrated Circuits (ASICs).

Fixed-Function Transformer Silicon

The Sohu chip is manufactured on TSMC’s N4P 4nm process node. Unlike highly programmable GPUs, Sohu is a fixed-function ASIC that hardcodes the transformer attention mechanism directly into its silicon circuitry. This specialization eliminates the software overhead and generalized compute units required by traditional graphics processors.

To manage the thermal constraints of continuous operation, Etched developed a Low Voltage Inference (LVI) architecture. The LVI design runs core mathematical operations at less than half the voltage of standard AI accelerators. This prevents thermal throttling and allows the hardware to maintain over 80% Peak FLOPs utilization during sustained inference workloads.

Data movement bottlenecks are managed by a proprietary Cluster-Scale Memory (CSM) system. CSM creates a deterministic, shared memory pool across entire hardware clusters. By removing intermediate inter-node software layers, the system achieves SRAM-level decode speeds across distributed chips.

Hardware Performance Comparison

Hardcoding the transformer architecture provides a structural throughput advantage for text generation tasks. An 8-chip Sohu server processes the 70-billion-parameter Llama model at a significantly higher token rate than equivalent NVIDIA configurations.

Hardware (8-Chip Server)	Architecture Type	Llama 70B Throughput (Tokens/sec)
Etched Sohu	ASIC	500,000
NVIDIA Blackwell B200	GPU	43,000
NVIDIA H100	GPU	25,000

Despite the physical hardcoding of the attention mechanism, the Sohu silicon retains compatibility with the broader ecosystem of transformer variants. The chip supports inference for major open-weight families including DeepSeek, Qwen, and Llama. Etched also confirmed the architecture successfully executes non-transformer state-space models like Mamba.

Operations and Infrastructure

To support continuous deployment cycles, Etched brought a new 10-Megawatt hardware testing lab online near its headquarters. The facility provides the necessary power density to run rapid prototyping for future chip generations and multi-rack cluster configurations.

The engineering organization has expanded to 400 personnel, incorporating teams heavily recruited from Google’s TPU division, NVIDIA, Broadcom, and TSMC.

If you deploy high-volume multi-agent systems or consumer AI features, evaluate your ratio of transformer-based inference to general-purpose compute. Transitioning static production workloads to fixed-function ASICs will structurally reduce LLM API costs, provided your core models do not migrate to fundamentally different architectures in the near term.