PALO ALTO, Calif. — Ricursive, founded by the former leads of Google’s AlphaChip project, has raised $335 million to develop an end-to-end AI model for chip design, targeting workload-specific custom hardware for third parties. The startup is not an EDA company and will not use standard EDA toolchains, instead aiming to transform chip design by enabling customers without in-house chip expertise to build custom silicon.
Founding and vision
Co-founders Anna Goldie and Azalia Mirhoseini, who led the machine learning persistence team at Google Brain, developed AlphaChip in 2018—one of the first reinforcement learning approaches to solve a real-world engineering problem. AlphaChip was used for macro placement across four generations of Google TPUs and adopted externally by companies including MediaTek. Ricursive’s technology goes well beyond AlphaChip in breadth, performance improvements, and speed, and the startup will not license or use any Google intellectual property.
Goldie and Mirhoseini chose independence over an Alphabet spinout to build chips beyond Google. “We wanted to have that broader impact,” Goldie said, noting that as a standalone company, chip makers trust them more with data, and the team can move faster with a singular mission.
Three-phase rollout
Phase one, currently underway, focuses on accelerating physical design and design verification for chip companies. “The goal is to take on the long pulls—physical design and design verification—so we can move more quickly from architecture to GDSII,” Goldie said.
Phase two will combine chip design stages into a single end-to-end model that ingests workloads and outputs GDSII files ready for manufacturing. This enables fast custom chip designs for AI accelerators and other workloads. “If we customize the compute architecture to the model architecture, we can achieve massive performance improvements,” Goldie said. Ricursive also aims to democratize chip design for companies with at-scale workloads that have not previously considered custom hardware, including applications like DNA sequencing.
Hardware-workload co-design
Phase three represents the ultimate vision: tight co-design of model, workload, and hardware. “Why not build our own chips, train our own models, and co-evolve them?” Goldie said. Ricursive’s models will be generally intelligent, handling everything from chip design to system and infrastructure design. Mirhoseini compared this to what OpenAI and Anthropic do, with a common intelligence layer supporting different specialties. “Chips are enablers of models,” she said. “If you can co-evolve the chip and model together, you speed up training time by multipliers, enabling the AI to evolve faster.”
Ricursive is training its model on generally available web data and open-source chip data, supplemented by synthetic data to avoid real-world data limits. The company is already working with chip companies to demonstrate early results on real designs.
Looking ahead
Ricursive’s approach could fundamentally alter how custom silicon is built, lowering costs and time-to-market for a new class of workload-specific chips. By closing the loop between AI model development and hardware design, the startup positions itself not just as a chip design tool, but as a catalyst for the next generation of AI evolution. If successful, Ricursive may enable applications that were previously impossible due to latency or power constraints, while challenging established chip design paradigms.
