ElastixAI, a Seattle-based AI infrastructure startup founded by former Apple and Meta machine learning researchers, has emerged from stealth with $18 million in seed funding to address the systemic inefficiencies and high costs of generative AI inference.
The company is launching a software platform that converts off-the-shelf FPGA-based servers into high-efficiency AI supercomputers. ElastixAI says its software-ML-hardware co-design approach delivers up to 50x lower total cost of ownership and 80% lower power consumption for large language model inference compared to legacy GPU-based systems.
According to the company, the AI inference market is expected to reach $255 billion by 2030, but existing infrastructure remains fundamentally mismatched for generative AI workloads. While LLM inference is memory-bound, standard GPUs are designed for compute-bound tasks such as training, leading to low compute utilization, wasted capital, and excess energy consumption. The company also noted that custom silicon development cycles can exceed three years, often lagging behind rapid advances in machine learning techniques.
ElastixAI positions its platform as a drop-in replacement for traditional GPU workflows, maintaining compatibility while enabling higher-density execution of LLM operations. The company says its approach eliminates “dark silicon” by activating only the circuits required for inference and enables cutting-edge AI implementations on current hardware without waiting for next-generation chip cycles.
The platform is currently available to select enterprise partners, data center operators, and AI model providers.
KEY QUOTE:
“The industry is currently leaving an order-of-magnitude of performance on the table because hardware can’t keep up with the advances in ML. We’re moving away from ‘one-size-fits-all’ hardware. By applying proprietary post-training optimizations to FPGAs, we let hardware adapt to the model rather than forcing the model to struggle on the hardware.”
Mohammad Rastegari, PhD, Co-Founder of ElastixAI
