12 Startups · $2.1B+ · Voice AI + Chips + More

What they're building

The chip design industry has a profound bottleneck. Building a single advanced semiconductor, takes teams of hundreds of engineers, 12-18 months of iterative design work, and anywhere from $100M to $650M in engineering labour. The finished chip is already outdated by the time it ships, because AI models have moved on. Ricursive's insight is structurally simple and staggeringly ambitious: use AI to design the chips that train better AI, then use those better chips to design even smarter AI. A recursive loop; hence the name. Their co-founders Anna Goldie and Azalia Mirhoseini proved this was possible at Google with AlphaChip, which compressed chip layout design from multiple months to just a few hours using deep reinforcement learning, and has been baked into multiple generations of Google's TPUs, including v5e, v5p, and the 6th generation Trillium. Ricursive is building a full-stack platform that handles every phase of chip creation - from placement, routing, verification, and architecture search, to using AI agents that get smarter with every chip they design.

Why this matters

The semiconductor industry is a $600B+ market, and it is structurally stuck. The dominant or the 'Big Three', EDA (Electronic Design Automation) tooling vendors, Cadence, Synopsys, Siemens EDA, have barely changed their software paradigm in 30 years. Human engineers are still the bottleneck. AI cannot advance faster than the hardware it runs on, and hardware cannot advance faster than humans can design it. Ricursive attacks that constraint at its root. If their platform delivers even 10x improvement in design throughput, it doesn't just make existing chips cheaper, it makes custom, purpose-built chips (for robotics, drugs, climate, medicine) economically viable for companies that could never afford bespoke silicon before. The downstream effect on the entire AI industry is difficult to overstate. Medicine: Drug discovery today is constrained not just by biology but by compute. Training a protein-folding model or running a molecular dynamics simulation requires months of GPU time on hardware that was designed for general AI workloads, not biochemistry. A purpose-built chip for protein simulation - optimised for the specific mathematical operations that dominate that workload - could cut that timeline by an order of magnitude. The bottleneck preventing its existence has always been cost: a bespoke chip costs $100M+ to design and fabricate. Ricursive's platform, if it works, brings that cost into range for a well-funded biotech, not just a hyperscaler. Climate and energy: Power grid optimisation, weather modelling, and carbon capture simulation are all compute-bound problems that run on chips designed for advertising inference and image generation. The mismatch is enormous. A custom chip for climate simulation would process fluid dynamics calculations - the mathematical core of weather and grid modelling - far more efficiently than anything currently available. The reason it doesn't exist is the same: design cost. Ricursive changes that equation. Robotics: The physical intelligence problem, giving robots the ability to sense, plan, and act in unstructured environments, requires real-time inference at extremely low power budgets. A robot arm running on a GPU draws too much power and generates too much heat. The robotics industry has been waiting for purpose-built inference chips for years; Qualcomm, NVIDIA, and a dozen startups are all trying to solve it. Ricursive's platform, in theory, could compress the design cycle for exactly these chips from 18 months to weeks. The compounding effect is the thing to understand. Faster chip design → better chips → faster AI → better chip design tools → faster chip design again. This is the loop Ricursive is trying to own. It's not a product company in the traditional sense - it's a bid to own the recursive feedback mechanism at the centre of the entire AI economy. That's why Lightspeed and Sequoia invested at a $4B valuation into a company with no product and four months of existence. They're not betting on a tool. They're betting on a new law of the industry.