AI for Chip Design & Chip Design for AI | Ricursive Intelligence | RAISE Summit 2026

Founders and Origins

• Recursive Intelligence is co-founded by Anna Goldie and Azalea Amir-Hosseini, who have collaborated since 2016 on the "Machine Learning for Systems" team at Google Brain.
• The company launched in September 2023 to scale the "AlphaChip" project, transitioning from a single-stage layout optimization to full end-to-end chip design automation.
• Founders state their vision is to close a "recursive self-improving loop" where AI designs chips that fuel the next generation of AI models.

Core Technology and Performance

• AlphaChip is a deep reinforcement learning agent capable of generating superhuman chip layouts in hours, replacing weeks or months of human effort.
• The company utilizes a dual-loop optimization system where proprietary inner-loop algorithms run 100 to 1,000x faster than traditional commercial tools.
• Recursive's static timing analysis engine achieves 0.999+ correlation with leading tools while operating 1,000x faster, enabling millisecond-level assessments.
• An outer-loop language model leverages these fast inner tools to explore the full chip design stack, from architecture to GDS2 fabrication formats.
• The team includes experts from frontier AI projects (Claude, Gemini, Grok) and hardware development (iPhone chips, TPUs), enabling cross-disciplinary problem solving.

Proven Track Record and Deployment

• Google's previous iterations of AlphaChip were deployed in the last four to five generations of TPUs, Axion CPUs, Pixel phones, and Waymo chips.
• MediaTek has adopted this technology at scale for their own chips.
• External applications include data center accelerators, embedded devices, and general-purpose CPUs, covering a diverse hardware spectrum.

Company Strategy and Phases

• Phase One (Current): Accelerate physical design and verification for existing chip manufacturers (e.g., NVIDIA, AMD, Intel) to reduce design cycles from years to shorter durations.
• Phase Two: Enable "design-less" era for application companies, allowing them to specify workloads (e.g., a new LLM) and receive a co-optimized computer architecture and quick tape-out.
• Phase Three: The company intends to build its own chips and train its own models to co-evolve, achieving the full recursive loop.
• Recursive aims to unlock "multiplier performance gains" (e.g., 20x improvements) via microarchitectural customization rather than the 20% gains typical of physical layout optimization.

Addressing Market Questions

• Target Users: Phase 1 targets traditional chip designers, while Phase 2 aims to democratize chip design for companies without in-house hardware teams.
• Scope of Customization: The technology supports a spectrum from general-purpose CPUs/GPUs to extreme customizations like Talos or weight-baked chips, focusing heavily on microarchitectural decisions (systolic array dimensions, memory hierarchy).
• Software Stack: While AI can replace compiler stacks (PyTorch to PTX), the strategy prioritizes co-design where the software stack is optimized for one or few models, simplifying programming.
• Competitive Landscape: The company aims to enable a "hardware lottery" breakthrough for frontier labs whose AI architectures are currently underserved by existing hardware.