Venture Capital Videos

How Ricursive Intelligence’s Founders Are Using AI to Shape The Future of Chip Design

The video introduces Recursive Intelligence, a startup founded by Anna Goldie and Aalia Mirhoseni, who previously built Google’s Alpha Chip that helped design four generations of TPUs. They argue that the current chip‑design cycle is a major bottleneck for AI progress because silicon development lags far behind the rapid evolution of neural‑network models. By applying modern AI techniques—particularly reinforcement‑learning agents and large‑scale synthetic data—to the entire chip‑design workflow, they aim to enable a true co‑design loop where hardware and workloads evolve together. Key insights from the founders include the mismatch between repurposed GPUs and purpose‑built AI silicon, the combinatorial complexity of floor‑planning (millions of nodes, strict power‑performance‑area constraints), and how a learning‑based placement agent can iteratively improve by experience, achieving “superhuman” results. Their RL system learns cost functions that matter to chip engineers—congestion, timing violations, power—rather than academic metrics like half‑perimeter wire length. Synthetic data generation is presented as the solution to the data‑privacy constraints of sharing customer designs, allowing the model to scale far beyond what any single client can provide. Notable examples illustrate the transformative potential: the AI‑generated layouts often take unconventional curved, donut‑shaped forms that reduce wire lanes and power consumption—placements a human designer would deem risky. After successive tape‑outs, the TPU team moved from skepticism to trusting the AI for larger chip blocks, with each new TPU generation showing larger area coverage and greater performance gains. The founders describe this as a recursive self‑improvement loop: better chips accelerate AI, which in turn creates more capable AI for chip design. The implications are profound. If Recursive Intelligence can deliver end‑to‑end, design‑less chip creation, it could democratize custom silicon, lower barriers for fabless companies, and dramatically shorten time‑to‑market for next‑generation AI accelerators. Faster, more efficient chips would bend the scaling laws that drive AI breakthroughs, potentially accelerating the race toward artificial general intelligence and reshaping the semiconductor industry’s business models.