Skip to main content

ElastixAI

Category: AI Infrastructure

A software-ML-hardware co-designed platform that converts off-the-shelf FPGA-based servers into high-efficiency AI supercomputers optimized for generative AI inference. ElastixAI was founded in 2025. The company is led by Mohammad Rastegari. Based in Seattle, USA. Team size: 10-50. Total funding raised: $18.0M. Latest round: Seed round ($18.0M, Feb 2026). Key investors include FUSE (lead investor), Catapult Ventures, Tyche Partners, Liquid 2 Ventures, DNX Ventures.

Founded
2025
Headquarters
Seattle, USA
Team size
10-50
Total funding
$18.0M

Value proposition

Delivers up to 50x lower total cost of ownership and 80% reduced power consumption compared to GPU-based deployments for LLM inference, while fitting within standard 17-19 kW rack power envelopes using air cooling instead of specialized liquid-cooled infrastructure.

Products and solutions

Elastix Rack - FPGA-based supercomputer system, AI Inference Software Platform, vLLM Plug-in with OpenAI-compatible API, Proprietary Model Conversion Tooling, ML-defined Software Configurable Compute

Unique value

Pioneers ML-defined, software-configurable compute using proprietary post-training optimizations on FPGAs. The platform lets hardware adapt to the model rather than forcing models to struggle on hardware. Optimizes for cost-per-bandwidth and cost-per-capacity using commercial off-the-shelf FPGA servers with cost-effective DDR and HBM memory instead of expensive premium GPU memory tiers.

Target customer

Hyperscalers, data center operators, AI model providers, and enterprises deploying large-scale generative AI inference workloads

Industries served

AI Infrastructure, Cloud Computing & Data Centers, Enterprise AI, Generative AI & LLMs, Semiconductor & Hardware

Technology advantage

Achieves three key advantages through unified software-ML-hardware co-design: (1) Cost Efficiency - up to 50x lower TCO per token without sacrificing performance; (2) Energy Efficiency - 80% reduced power consumption per token at production scale with intelligent resource utilization; (3) Elasticity - dynamically scales and optimizes for evolving model demands while remaining fully compatible with existing AI pipelines via drop-in replacement for GPU infrastructure.

How they differentiate

ElastixAI uniquely uses off-the-shelf FPGA-based servers with proprietary software-ML-hardware co-design, delivering 50x lower total cost of ownership and 80% reduced power consumption. Unlike competitors using custom silicon requiring 3+ years development cycles and specialized liquid-cooled infrastructure (120-200 kW racks), ElastixAI's FPGA approach offers hardware reconfigurability that adapts to rapidly evolving ML architectures while operating within standard 17-19 kW air-cooled rack power envelopes.

Main competitors

Groq, Cerebras Systems, SambaNova Systems

Key partnerships

FPGA manufacturers (validation partnerships), Data center operators (deployment partnerships), FUSE (lead investor with deep semiconductor ties), Catapult Ventures, Tyche Partners, Liquid 2 Ventures, DNX Ventures

Major milestones

Founded in early 2025 by ex-Apple/Meta AI researchers with Xnor.ai pedigree (acquired by Apple for ~$200M), Raised $18M seed funding led by FUSE in February 2026, Emerged from stealth with FPGA-based AI inference platform, Validation partnerships established with FPGA manufacturers and data center operators, Proprietary vLLM plug-in with OpenAI-compatible API developed for seamless GPU infrastructure replacement

Growth metrics

Recently emerged from stealth mode (February 2026); platform available to select enterprise partners; first hardware shipments (Elastix Rack) planned for mid-2026

Market positioning

Early-stage challenger in AI inference infrastructure market, targeting cost-conscious hyperscalers and enterprise data center operators seeking GPU alternatives without sacrificing performance or requiring infrastructure overhauls

Geographic focus

Headquartered in Seattle, USA; initially targeting North American data center operators and hyperscalers with global expansion potential

Patents and IP

No registered patents disclosed as of latest update

About Mohammad Rastegari

Co-founder and CTO of Xnor (acquired by Apple for ~$200M in 2020). Senior Technical Manager in AI/ML at Apple (2020-2024). Distinguished AI Scientist at Meta (April-December 2024). Research Scientist at Allen Institute for AI (AI2) for 5 years. Affiliate Assistant Professor at University of Washington. Creator of XNOR-Networks for efficient deep neural networks. PhD in Computer Science from University of Maryland (2012-2015). Published researcher with 34,000+ citations in top-tier venues including IJCV, IEEE PAMI.

Official website: