AI That Designs Its Own Chips: Ricursive's Anna Goldie and Azalia Mirhoseini
Co-founders of Ricursive Intelligence, Anna Goldie and Azalia Mirhoseini, outline their thesis that AI should design the chips that train AI. They detail their three-phase plan to first accelerate chip design with AI tools 100,000x faster than current software, then become a 'design-less' platform for custom silicon, and finally achieve vertical integration by building their own chips and models.
At Ricursive Intelligence, the core thesis is that a recursive, self-improving loop can be created between AI and its physical substrate: AI should be used to design, optimize, and automate the creation of the chips that power AI itself. This vision stems from a decade of foundational work by co-founders Anna Goldie and Azalia Mirhoseini at Google Brain, DeepMind, and Stanford, which culminated in the creation of AlphaChip. AlphaChip is a deep reinforcement learning agent capable of generating superhuman chip layouts, and it has been productionized and used in the tape-out of the last four generations of Google’s TPUs, Axion data center CPUs, Pixel phone chips, and autonomous vehicle hardware.
The Challenge in Modern Chip Design
The current chip design process is a significant bottleneck for the advancement of AI. Two key stages, physical design (placing and routing billions of components on a chip) and design verification (ensuring the logical correctness of the chip), are the longest poles. Each can take up to a year and requires the expertise of hundreds or thousands of engineers. The financial stakes are enormous, with an estimated one-day delay of a major chip like NVIDIA’s Blackwell potentially costing a company over $200 million in lost opportunity.
A Three-Phase Plan for the Future of Silicon
Ricursive Intelligence is tackling this challenge with a phased approach to transform the industry.
-
Phase 1: Accelerate Chip Design The immediate goal is to build AI-powered tools that dramatically accelerate the chip design workflow for existing chipmakers. By creating software that runs up to 100,000 times faster than today’s commercial tools, Ricursive aims to help companies get to market faster with chips that are more performant, cheaper, and more energy-efficient.
-
Phase 2: Democratize Chip Design with a “Design-less” Platform The second phase is to become a platform for hardware design, ushering in a “design-less” era analogous to how TSMC enabled the “fabless” era. The platform will take a specific workload, such as a next-generation AI model, as input and automate the entire design process, from architecture to a “GDS2 clean” file ready for manufacturing. This would unlock custom silicon for any company with a sufficiently large-scale workload, even those without dedicated teams of hardware experts.
-
Phase 3: Vertical Integration With the ability to rapidly design highly performant, custom chips, the final phase is vertical integration. Ricursive plans to build its own chips, train its own models, and co-evolve hardware and software together. This would allow them to provide intelligence at a price and capability point that would be impossible for competitors to match.
Technical Approach: Speed and Co-Optimization
The core of the technical strategy is to redesign foundational chip design tools for speed, making them suitable for AI-driven optimization. For example, they have developed a Static Timing Analysis (STA) engine that is 1,000x faster than commercial equivalents while maintaining high fidelity.
This speed is critical because AI agents, particularly in reinforcement learning, thrive on fast iteration loops. With these accelerated tools providing near-instant feedback, an “outer loop” AI can explore a vast design space, co-optimizing across the entire stack to unlock massive improvements in performance and time to market.
Enabling a “Cambrian Explosion” of Chips
The ultimate vision is to enable a “Cambrian explosion” of custom silicon. Today’s AI landscape is dominated by a few mainstream chips, but future progress will require a much wider variety of hardware customized for specific needs—from massive frontier models to low-power, high-throughput inference chips. By democratizing design, Ricursive aims to be the platform that makes this diverse ecosystem of custom hardware possible.
A New Design Paradigm: Organic, AI-Generated Layouts
A fascinating outcome of this AI-driven approach is the nature of the chip layouts themselves. When human engineers design a chip, they tend to create rigid, grid-like, regular-looking placements. In contrast, the AI-generated layouts produced by systems like AlphaChip are “organic” and “curved.” This non-intuitive design approach was initially shocking to physical design engineers, but it is demonstrably superior, as it naturally minimizes wire length and improves overall performance. This suggests how AI is poised to redesign not just chips, but the entire field of physical engineering.