Etched doubles valuation to $10.3B to scale custom inference chips

The funding signals an aggressive push for specialized ASIC hardware to replace flexible but power-intensive GPUs in high-volume inference tasks. For the streaming and media ecosystem, this shift could dramatically lower the unit economics of generative AI video processing, real-time translation, and personalized content recommendation at scale. By isolating the inference workload, Etched claims its architecture avoids the thermal bottlenecks that limit traditional GPU clock speeds. Industry stakeholders should watch for the performance benchmarks of the first production racks shipping this summer, which will determine if specialized silicon can effectively erode Nvidia’s dominance in the rapidly growing inference market.

The rise of specialized inference hardware comes as the broader market shifts from training massive models to serving them. Per Bloomberg Intelligence, February 2026, the custom AI accelerator market is projected to grow at a 44.6% compound annual growth rate through 2033, more than double the 16.1% rate expected for general-purpose GPUs. This divergence is driven by hyperscalers like Google, Microsoft, and Amazon, which have collectively committed billions to in-house silicon like the TPU v7 and Maia 200 to optimize inference costs that now consume two-thirds of all AI compute energy.

Competition in the merchant silicon sector is also intensifying. While Nvidia maintains a significant market share lead with its Blackwell and upcoming Rubin architectures, startups like Etched are betting on architecture-specific optimizations. According to reporting from EE Times in July 2026, Etched’s ‘Sohu’ chip is specifically tuned for transformer-based models, claiming to provide superior throughput for long-context applications and autonomous agents compared to flexible alternatives.

Strategic partnerships are proving critical for these new entrants to manage supply chain risks. Inclusion of SK Hynix in Etched’s latest round is notable as the firm is a primary supplier of the HBM3E memory required for high-performance AI chips. Simultaneously, per Reuters, May 2026, ASIC-based AI server shipments are expected to reach nearly 28% of the total market this year, as the industry moves away from using expensive training-grade GPUs for routine deployment tasks.