# Ricursive Intelligence

> Source: https://aiwiki.ai/wiki/ricursive_intelligence
> Updated: 2026-11-06
> Fact-checked: 2026-10-27
> Categories: AI Companies
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Ricursive Intelligence." aiwiki.ai, 6 Nov 2026. https://aiwiki.ai/wiki/ricursive_intelligence
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Ricursive Intelligence** is an American [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) company based in Palo Alto, California, that develops AI systems for designing semiconductors [1][2]. It was founded in 2025 by Anna Goldie and Azalia Mirhoseini, the research scientists who co-created [AlphaChip](https://aiwiki.ai/wiki/alphachip), a reinforcement-learning method for chip floorplanning developed at [Google](https://aiwiki.ai/wiki/google) [1][3]. CNBC reported that the company was founded in September 2025 [13]. It emerged from stealth on December 2, 2025, and by January 26, 2026, had raised a total of $335 million across two rounds, reaching a $4 billion post-money valuation with its Series A [1][2][7]. Its stated mission is to use AI to compress the multi-year chip design process and to build a recursive loop in which AI designs better chips and those chips in turn train more capable AI [1][5]. As of mid-2026 the company described itself as being in the first of three planned phases, taking on parts of the chip design flow for chip companies, with an end-to-end chip design model and, eventually, a general frontier AI model co-designed with its own chips as later goals [14].

<figure class="w-56">
<img src="/images/site-2026-10-28-ricursive_intelligence-logo-5b5290915d44-tile.png" alt="Ricursive Intelligence logo" width="480" height="154" />
<figcaption>Ricursive Intelligence logo, from the official website <a href="https://ricursive.com/">ricursive.com</a>. Trademark of its owner.</figcaption>
</figure>

## Overview

Ricursive Intelligence describes itself as a frontier AI lab focused on building self-improving systems, beginning with chip design [5]. Rather than manufacturing or selling chips, the company builds AI software that automates and accelerates stages of the semiconductor design flow, from component placement through design verification [4][6]. The central thesis, which the founders and their lead investor summarize as "AI for chip design and chip design for AI," is that AI can shorten chip development from years to weeks, and that the resulting custom silicon can then accelerate the training and deployment of more advanced [AI models](https://aiwiki.ai/wiki/model), closing a [feedback loop](https://aiwiki.ai/wiki/feedback_loop) the company frames as a path toward [artificial general intelligence](https://aiwiki.ai/wiki/artificial_general_intelligence) and [superintelligence](https://aiwiki.ai/wiki/superintelligence) [3][5][8]. TIME noted that the company's name is a nod to [recursive self-improvement](https://aiwiki.ai/wiki/recursive_self-improvement), the idea that AI could reach a point where it accelerates its own development [16].

SiliconANGLE identified the long-standing electronic design automation (EDA) vendors Synopsys and Cadence Design Systems, both of which sell AI features for chip design, as competition, and noted that AI labs such as [OpenAI](https://aiwiki.ai/wiki/openai) and [Anthropic](https://aiwiki.ai/wiki/anthropic) have also applied their models to chip design and electrical engineering tasks [6]. The founders themselves told EE Times in June 2026 that Ricursive is not an EDA company, will not compete with EDA companies, and will not use standard EDA toolchains, describing its work instead as end-to-end chip design [14]. TechCrunch reported that chipmakers such as [Nvidia](https://aiwiki.ai/wiki/nvidia), [AMD](https://aiwiki.ai/wiki/amd), and [Intel](https://aiwiki.ai/wiki/intel) are among the startup's target customers rather than its competitors [4]. At its December 2025 launch, Electronics Weekly reported that the company aimed to add a "designless" category of company alongside the fabless model coined in the 1980s for firms that design chips but do not manufacture them [9].

Ricursive is a separate company from the similarly named Recursive ([Recursive Superintelligence](https://aiwiki.ai/wiki/recursive_superintelligence)), which TechCrunch described in January 2026 as reportedly founded by [Richard Socher](https://aiwiki.ai/wiki/richard_socher) and also working on self-improving AI systems [2].

## Founding and founders

Ricursive Intelligence was co-founded by Anna Goldie, who serves as chief executive officer, and Azalia Mirhoseini, who serves as chief technology officer [1][2]. Both are research scientists who previously worked at [Google Brain](https://aiwiki.ai/wiki/google_brain), [Google Research](https://aiwiki.ai/wiki/google_research), and [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind), and both were early employees at Anthropic [3][4][10][13]. Their paths first crossed at [Stanford University](https://aiwiki.ai/wiki/stanford_university), where Goldie earned her PhD while Mirhoseini taught computer science classes [4]. Goldie told TechCrunch that the two started at Google Brain on the same day, left on the same day, joined and left Anthropic on the same day, rejoined Google together, and left again together before founding the company [4]. According to EE Times, they had worked together for about 10 years by 2026, starting the machine learning persistence team at Google Brain, one of whose projects was AlphaChip [14]. Mirhoseini was also a co-author, with [Noam Shazeer](https://aiwiki.ai/wiki/noam_shazeer), [Jeff Dean](https://aiwiki.ai/wiki/jeff_dean), and others, of the 2017 paper that introduced the sparsely-gated [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) layer [14][22].

The pair are best known for co-creating AlphaChip, a deep [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) system that treats the physical placement of components on a chip as an optimization problem [3][8]. AlphaChip "learns by doing": an agent iteratively places circuit blocks while a reward signal scores the quality of each layout, and the resulting feedback updates the parameters of a [deep neural network](https://aiwiki.ai/wiki/deep_neural_network) so the agent improves over time [3][4]. The method was published in the journal Nature in 2021; the paper reported that it generated chip floorplans in under six hours that were superior or comparable to human-made layouts on power, performance, and area, a task it said otherwise required months of effort by physical design engineers [19]. AlphaChip was subsequently used across four generations of Google's [Tensor Processing Unit](https://aiwiki.ai/wiki/tpu) accelerators and was adopted by external chipmakers, including [MediaTek](https://aiwiki.ai/wiki/mediatek) [3][14].

According to the company, its team's work also includes RL-CCD (Best Paper, DAC 2023), Insta (Best Paper, DAC 2025), and C3PO (Best Paper, ASP-DAC 2026) [5]. Ricursive says its staff come from Google DeepMind, Anthropic, Nvidia, Cadence, Apple, Stanford, MIT, Harvard, and other institutions, with hands-on experience developing [Gemini](https://aiwiki.ai/wiki/gemini), [Claude](https://aiwiki.ai/wiki/claude), TPUs, and Apple silicon [5]. Goldie told CNBC in April 2026 that the founders "got the core AlphaChip team back together" by hiring old collaborators, and that other team members had worked at Google, Anthropic, Nvidia, [Apple](https://aiwiki.ai/wiki/apple), and xAI [13].

Goldie told EE Times that Google had offered to set up an Alphabet spinout for the pair to continue their AlphaChip work, but that they chose to form an independent company so that they could build chips beyond Alphabet and so that other chipmakers would trust them with their data [14]. She also said Ricursive would not license or use any Google intellectual property [14].

## Funding

Ricursive Intelligence raised $335 million in total across a seed round and a Series A, both announced within about two months of its public launch [2][7]. The seed round of $35 million was led by [Sequoia Capital](https://aiwiki.ai/wiki/sequoia_capital) and valued the company at $750 million; it was announced on December 2, 2025, the same day the company emerged from stealth [1][11]. The $300 million Series A was led by [Lightspeed Venture Partners](https://aiwiki.ai/wiki/lightspeed_venture_partners) and was announced on January 26, 2026, at a $4 billion post-money valuation, less than two months after launch [2][7][12]. As of October 2026, no later funding round had been announced; the company's website still cited the $335 million total [5].

| Round | Date | Amount | Valuation | Lead investor |
| --- | --- | --- | --- | --- |
| Seed | December 2, 2025 | $35 million | $750 million | [Sequoia Capital](https://aiwiki.ai/wiki/sequoia_capital) [1][11] |
| Series A | January 26, 2026 | $300 million | $4 billion (post-money) | [Lightspeed Venture Partners](https://aiwiki.ai/wiki/lightspeed_venture_partners) [2][7] |
| Total | | $335 million | | [2][4] |

The company said it would use the Series A to scale its research and engineering team and to significantly expand its compute infrastructure [7]. EE Times reported in June 2026 that the money would largely be spent on GPU compute hours for training and testing the company's model [14].

## Investors

The Series A was led by Lightspeed Venture Partners, with participation from a group of venture investors and Nvidia's venture capital arm [2][7].

| Investor | Role |
| --- | --- |
| [Lightspeed Venture Partners](https://aiwiki.ai/wiki/lightspeed_venture_partners) | Lead, Series A [2][7] |
| [Sequoia Capital](https://aiwiki.ai/wiki/sequoia_capital) | Lead, seed; Series A participant [1][7] |
| DST Global | Series A participant [2][7] |
| [NVentures (Nvidia)](https://aiwiki.ai/wiki/nvidia_nventures) | Series A participant [2][7] |
| Felicis Ventures | Series A participant [2][7] |
| 49 Palms Ventures | Series A participant [2][7] |
| [Radical Ventures](https://aiwiki.ai/wiki/radical_ventures) | Series A participant (listed as "Radical AI" in the company's press release) [2][7][12] |

TechCrunch reported that the founders said they had heard from every major chipmaker and would not name early customers [4]. Goldie told CNBC that potential customers were more likely to see a new company as a neutral partner than as a competitor [13].

## Focus and technology

Ricursive Intelligence builds a full-stack platform intended to automate and optimize the semiconductor design process [3]. Its systems evaluate large numbers of candidate chip layouts while balancing competing physical constraints at once, such as heat thresholds, power consumption, surface area, and wirelength, the total length of the wiring that connects components [6]. Like AlphaChip, the platform uses a reward signal that rates design quality and allows an AI agent to improve through repeated trials, and the company says its models learn across many different chip designs so that performance compounds over time [4]. The platform also uses [large language models](https://aiwiki.ai/wiki/large_language_model) [4]. The founders have described the eventual goal as a system capable of creating its own silicon substrate and accelerating each subsequent generation of [AI chips](https://aiwiki.ai/wiki/ai_chip) [2][7].

The company frames this as recursive self-improvement applied to hardware: AI optimized silicon yields performance gains for specific AI workloads, and those workloads in turn speed the design of still more efficient compute [3]. Ricursive argues this would make custom chip design accessible to organizations that lack the very large engineering teams currently required, broadening who can build specialized silicon [3].

### Three-phase roadmap

In a June 2026 interview with EE Times, Goldie and Mirhoseini set out a three-phase plan [14]:

| Phase | Goal (as described by the founders) |
| --- | --- |
| One (under way in June 2026) | Take on portions of chip design for chip companies, chiefly physical design and design verification, to speed the path from architecture to GDSII |
| Two | Combine the design stages into an end-to-end model that ingests a workload and outputs GDSII files ready for manufacturing, including for customers without in-house chip teams |
| Three | Co-design models, workloads, and hardware: build the company's own chips and train its own general-purpose frontier models together |

Mirhoseini said the company wants to make chips for customers that do not have in-house chip experts but run an algorithm at scale, and gave scientific and healthcare workloads such as DNA sequencing as possible applications [14]. Of the third phase, she said: "In the third phase, we want to make a frontier AI, period" [14]. EE Times reported that Ricursive trains its model on generally available web data and open-source chip data, can also create and use synthetic data, and was working with chip companies to demonstrate early results on their real designs [14]. In February 2026 Goldie told TechCrunch that customizing a computer architecture to a specific model could yield "almost a 10x improvement in performance per total cost of ownership" [4].

## 2026 developments

- **June 2026:** EE Times published an interview in which the founders described the three-phase roadmap and their decision to stay independent of Google [14].
- **July 2026:** At the 63rd Design Automation Conference (DAC) in Long Beach, California (July 26 to 29), Ricursive had an exhibit booth and Goldie was scheduled to speak in an exhibitor forum session on "Closing the Loop Between AI and the Hardware That Fuels It" [17][18]. In an EE Times preview of the show's AI exhibitors, Frank Schirrmeister, program chair of the DAC Engineering Track, called Ricursive "the most audacious address on the map" and described it as an "engine-replacer" because AlphaChip-style placement replaces the algorithmic core of a place-and-route tool [18].
- **August 2026:** TIME named Goldie and Mirhoseini to its TIME100 AI list for 2026 [16].
- **October 2026:** At TechCrunch Disrupt on October 14, the founders argued that more efficient chips are needed to meet compute demand. Goldie said the industry expects "at least 10,000x more demand for compute" within five years and that "we don't want to build 10,000 times more data centers" [15]. They said Ricursive optimizes entire chips rather than partitioning them into dozens of smaller blocks, and recalled that AlphaChip had produced a circular, "donut-shaped" arrangement of components that initially surprised engineers [15].

### Ripple-65B Preview

In November 2026, Ricursive released [Ripple-65B Preview](https://aiwiki.ai/wiki/ripple_65b_preview), a general-purpose version of a language model developed for its chip-design platform. It is a dense Transformer with about 65 billion parameters, pretrained on 12 trillion tokens. The released checkpoints include an instruction-tuned BF16 model, a BF16 base model and an experimental NVFP4 instruction model, distributed under Apache 2.0.[23][24][25] The NVFP4 version quantizes the MLP layers while retaining BF16 attention.[25]

Ricursive reported $32 million in total research and development costs for Ripple, including model-ladder experiments, data mixing, evaluation, debugging, synthetic data generation, recipe development and final pretraining and post-training. This is the company's development-cost report, rather than the cost of the final training run alone. Ricursive also described the released model as a general-purpose variant of the proprietary model powering its chip-design platform; the two should not be treated as the same released checkpoint.[23]

## Significance

Ricursive Intelligence drew attention both for the speed of its fundraising and for what it represents in the broader semiconductor and AI industries. Its $4 billion valuation less than two months after launch was the focus of headlines from TechCrunch and Crunchbase News, and CNBC cited it as an example of senior researchers leaving large technology companies to found their own well-capitalized AI labs [2][12][13]. Tech Funding News called the Series A one of the largest early-stage rounds in AI-driven semiconductor design [8]. The participation of Nvidia's venture arm signaled interest from the hardware industry in AI-assisted design [2][7].

The company's pitch also reflects a wider argument that chip design, long a bottleneck measured in years per generation, has become a constraint on the pace of AI progress, and that automating it could speed the entire field [1][7]. Analysts have raised caveats. A December 2025 analysis by CTOL Digital Solutions argued that incumbents Synopsys and Cadence already deploy AI tools across major customers, that end-to-end automation would require trustworthy formal verification, and that the company's first customers, foundry partnerships, and benchmarks against incumbent tools on real designs would determine its trajectory [11]. As of February 2026 the company had not named its early customers [4].

## Context: AI-designed chips

The design of integrated circuits has historically combined human engineering judgment with electronic design automation software for tasks like placement, routing, and verification. Floorplanning, deciding where major blocks sit on a chip, is a combinatorial problem with an enormous search space, and it strongly affects a chip's performance, power, and area [3][19]. AlphaChip, the work underpinning Ricursive, was an early high-profile demonstration that reinforcement learning could perform this task at or above expert level and was deployed in production on Google's TPU line, which provided practical evidence that machine learning could contribute to real silicon [3][8].

AlphaChip's results were also disputed. Goldie, Mirhoseini, and Jeff Dean wrote in November 2024 that a non-peer-reviewed invited paper at ISPD 2023 had questioned its performance claims without running the method as described in Nature, for example by skipping pre-training and using substantially fewer compute resources [20]. Chung-Kuan Cheng, Andrew B. Kahng, and colleagues, whose assessment of the method was first posted in February 2023, released an updated version, last revised in March 2026, that evaluated the latest open-source release against alternative macro placers on new public benchmarks and listed confirmations it considered still missing, such as of the method's scalability and pre-training methodology [21].

Ricursive Intelligence extends that lineage by aiming at the full design flow rather than placement alone, and by tying chip design explicitly to the development of the AI models that the chips will run [5][14]. Its formation in late 2025 coincided with intense demand for [AI accelerators](https://aiwiki.ai/wiki/ai_accelerator) and growing competition among AI labs and chipmakers to design custom silicon, the environment in which strategic investors such as Nvidia's NVentures chose to back the company [2][6]. Other startups pursued related ideas at the same time; TechCrunch noted that Naveen Rao's [Unconventional AI](https://aiwiki.ai/wiki/unconventional_ai) was also working on an intelligent substrate [2].

## See also

- [AlphaChip](https://aiwiki.ai/wiki/alphachip)
- [ChipNeMo](https://aiwiki.ai/wiki/chipnemo)
- [ChipAgents](https://aiwiki.ai/wiki/chipagents)
- [Unconventional AI](https://aiwiki.ai/wiki/unconventional_ai)
- [Recursive Superintelligence](https://aiwiki.ai/wiki/recursive_superintelligence)
- [List of neolabs](https://aiwiki.ai/wiki/list_of_neolabs)
- [Neolab](https://aiwiki.ai/wiki/neolab)

## References

1. "Ricursive Intelligence Launches Frontier AI Lab to Transform Semiconductor Design and Accelerate Path Toward Artificial Superintelligence," PR Newswire, December 2, 2025. https://www.prnewswire.com/news-releases/ricursive-intelligence-launches-frontier-ai-lab-to-transform-semiconductor-design-and-accelerate-path-toward-artificial-superintelligence-302630776.html
2. "AI chip startup Ricursive hits $4B valuation two months after launch," TechCrunch, January 26, 2026. https://techcrunch.com/2026/01/26/ai-chip-startup-ricursive-hits-4b-valuation-two-months-after-launch/
3. "Investing in Ricursive Intelligence: AI for Chip Design and Chip Design for AI," Lightspeed Venture Partners. https://lsvp.com/stories/investing-in-ricursive-intelligence-ai-for-chip-design-and-chip-design-for-ai/
4. "How Ricursive Intelligence raised $335M at a $4B valuation in 4 months," TechCrunch, February 16, 2026. https://techcrunch.com/2026/02/16/how-ricursive-intelligence-raised-335m-at-a-4b-valuation-in-4-months/
5. Ricursive Intelligence, company website. https://www.ricursive.com/
6. "Ricursive Intelligence nabs $300M to speed up chip design with AI," SiliconANGLE, January 26, 2026. https://siliconangle.com/2026/01/26/ricursive-intelligence-nabs-300m-speed-chip-design-ai/
7. "Ricursive Intelligence Raises $300 Million Series A at $4 Billion Valuation to Accelerate AI-Driven Semiconductor Design," PR Newswire, January 26, 2026. https://www.prnewswire.com/news-releases/ricursive-intelligence-raises-300-million-series-a-at-4-billion-valuation-to-accelerate-ai-driven-semiconductor-design-302670061.html
8. "From Google's AlphaChip to $4B: Ricursive becomes an AI chip-design unicorn," Tech Funding News, January 2026. https://techfundingnews.com/ricursive-intelligence-300m-series-a-4b-valuation-google-alphachip-founders/
9. "'Designless' company category set to join fabless," Electronics Weekly, December 2025. https://www.electronicsweekly.com/news/business/designless-company-category-set-to-join-fabless-2025-12/
10. "AI Chip Innovator Ricursive Secures $300M Series A at $4B Valuation Just Two Months Post Launch," MLQ.ai. https://mlq.ai/news/ai-chip-innovator-ricursive-secures-300m-series-a-at-4b-valuation-just-two-months-post-launch/
11. "Google AlphaChip Founders Raise $35 Million Seed Round to Automate Chip Design with AI," CTOL Digital Solutions, December 2025. https://www.ctol.digital/news/google-alphachip-founders-raise-35-million-seed-automate-chip-design-ai/
12. "AI Lab Ricursive Intelligence Lands $300M Series A At $4B Valuation Less than Two Months After Launch," Crunchbase News, January 26, 2026. https://news.crunchbase.com/venture/startup-ai-lab-ricursive-seriesa-unicorn/
13. "Meta, Google, OpenAI among Big Tech firms seeing top staff leaving to launch AI startups," CNBC, April 28, 2026. https://www.cnbc.com/2026/04/28/meta-google-big-tech-staff-ai-labs-investors.html
14. Sally Ward-Foxton, "Startup Ricursive to Create an End-to-End AI Model for Chip Design," EE Times, June 10, 2026. https://www.eetimes.com/startup-ricursive-to-create-an-end-to-end-ai-model-for-chip-design/
15. Aisha Malik, "Ricursive Intelligence is betting on AI-designed chips to help solve the compute crisis," TechCrunch, October 14, 2026. https://techcrunch.com/2026/10/14/ricursive-intelligence-is-betting-on-ai-designed-chips-to-help-solve-the-compute-crisis/
16. "Azalia Mirhoseini and Anna Goldie," TIME100 AI 2026, TIME, August 27, 2026. https://time.com/collection/time100-ai/2026/azalia-mirhoseini-anna-goldie/
17. Frank Schirrmeister, "DAC 2026: What Does It Actually Take to Create AI Chips?," EE Times, July 24, 2026. https://www.eetimes.com/dac-2026-what-does-it-actually-take-to-create-ai-chips/
18. Frank Schirrmeister, "AI in EDA is Real, it's Now, and is On Show at DAC 2026 in Long Beach" (appendix: map, table, detailed walk), EE Times, July 2026. https://www.eetimes.com/wp-content/uploads/dac2026-part1-ai-landscape-v8-Appendix.pdf
19. Mirhoseini, A., Goldie, A., Yazgan, M., et al. "A graph placement methodology for fast chip design." Nature 594, 207 to 212, June 9, 2021. https://www.nature.com/articles/s41586-021-03544-w
20. Goldie, A., Mirhoseini, A., Dean, J. "That Chip Has Sailed: A Critique of Unfounded Skepticism Around AI for Chip Design." arXiv:2411.10053, November 15, 2024. https://arxiv.org/abs/2411.10053
21. Cheng, C.-K., Kahng, A. B., Kundu, S., Wang, Y., Wang, Z. "An Updated Assessment of Reinforcement Learning for Macro Placement." arXiv:2302.11014 (v1 February 21, 2023; v3 March 10, 2026). https://arxiv.org/abs/2302.11014
22. Shazeer, N., Mirhoseini, A., Maziarz, K., Davis, A., Le, Q., Hinton, G., Dean, J. "Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer." arXiv:1701.06538, January 23, 2017. https://arxiv.org/abs/1701.06538
23. Ricursive Intelligence. [Introducing Ripple-65B Preview](https://www.ricursive.com/ripple). November 2026. Model release, relationship to the chip-design platform and reported development costs.
24. Ricursive Intelligence. [Ripple-65B Preview model card](https://huggingface.co/ricursiveAI/Ripple-65B-Preview). BF16 release, architecture, training and Apache 2.0 license. Accessed November 7, 2026.
25. Ricursive Intelligence. [Ripple-65B Preview NVFP4 model card](https://huggingface.co/ricursiveAI/Ripple-65B-Preview-NVFP4). Experimental quantized checkpoint and mixed precision. Accessed November 7, 2026.

