# Naveen Rao

> Source: https://aiwiki.ai/wiki/naveen_rao
> Updated: 2026-06-08
> Categories: AI Companies, People
> From AI Wiki (https://aiwiki.ai), a free encyclopedia of artificial intelligence. Quote with attribution.

Naveen Rao is an American entrepreneur, computer architect, and neuroscientist who has founded three [artificial intelligence](/wiki/artificial_intelligence) companies and twice sold a startup to a large technology firm. He is best known as the founder and chief executive of [Nervana Systems](/wiki/nervana_systems), a deep-learning chip company acquired by [Intel](/wiki/intel) in 2016, and of [MosaicML](/wiki/mosaicml), an efficient large-model training startup that [Databricks](/wiki/databricks) acquired in 2023 for about 1.3 billion dollars. After the MosaicML deal he led generative AI at Databricks as a vice president. In 2025 he left to found Unconventional AI, a hardware startup building custom silicon and systems intended to compete with [NVIDIA](/wiki/nvidia).[1][9][11]

Rao is unusual among AI hardware founders in combining a long career as a chip designer with formal training in neuroscience, a background that has shaped his recurring ambition to build computers that approach the energy efficiency of the brain.[1][11]

The table below summarizes his principal ventures and roles.

| Company | Role | Period | Outcome |
| --- | --- | --- | --- |
| Nervana Systems | Co-founder and CEO | 2014 to 2016 | Acquired by Intel, 2016, about 408 million dollars |
| Intel | Corporate VP and GM, AI Products Group | 2016 to 2020 | Departed March 2020 |
| MosaicML | Co-founder and CEO | 2021 to 2023 | Acquired by Databricks, 2023, about 1.3 billion dollars |
| Databricks | Vice president of AI / generative AI | 2023 to 2025 | Departed September 2025 |
| Unconventional AI | Founder and CEO | 2025 to present | 475 million dollar seed at 4.5 billion dollar valuation |

## Education and early career

Rao earned a bachelor's degree in electrical engineering and computer science from [Duke University](/wiki/duke_university).[1] He then spent roughly a decade working as a computer architect and chip designer in Silicon Valley, holding engineering positions at companies including Teragen, CALY Networks, Kealia, [Sun Microsystems](/wiki/sun_microsystems), and W&W Communications, where his work spanned algorithms, microarchitecture, logic design, and verification.[1][7]

After his years in industry he returned to academia and completed a PhD in computational neuroscience at [Brown University](/wiki/brown_university) around 2011, studying how biological neural circuits compute.[1] He carried that interest into a research role at [Qualcomm](/wiki/qualcomm), where he worked on neuromorphic machines, brain-inspired computing architectures that depart from conventional [von Neumann](/wiki/von_neumann_architecture) designs. Two colleagues from Qualcomm Research, Amir Khosrowshahi and Arjun Bansal, would later join him as co-founders of his first company. All three held doctorates in neuroscience alongside backgrounds in physics, computer science, and engineering.[7]

## Nervana Systems

In February 2014 Rao co-founded Nervana Systems in San Diego with Amir Khosrowshahi, who became chief technology officer, and Arjun Bansal, who led algorithms. Rao served as chief executive.[2][6] The company set out to build hardware and software dedicated specifically to [deep learning](/wiki/deep_learning), rather than repurposing general-purpose [GPUs](/wiki/gpu). Its planned [AI chip](/wiki/ai_chip), the Nervana Engine, was a custom [ASIC](/wiki/asic) that stripped out the components needed for graphics work and added high-bandwidth stacked memory, an approach Nervana argued could deliver far more throughput per chip than contemporary GPUs.[3] Alongside the silicon, Nervana offered a full-stack cloud service called Nervana Cloud and released an open-source deep-learning framework named neon.[2]

On August 9, 2016, Intel acquired Nervana Systems. Reported prices ranged from about 350 million dollars to more than 408 million dollars, and the roughly 48-person team joined Intel's Data Center Group.[2][3][4] At the time, Intel framed the deal as a way to catch up in deep learning, a field where it had fallen behind NVIDIA, and positioned Nervana's technology as the foundation for a new line of purpose-built AI processors.[4]

## Intel

Rao joined Intel through the acquisition and initially served as vice president and general manager of AI solutions. By 2017 he had been promoted to corporate vice president and general manager of the company's Artificial Intelligence Products Group, making him Intel's most senior executive devoted to AI hardware and software.[1][5]

Under Rao, Intel developed the Nervana Neural Network Processor family, which paired a training chip code-named Spring Crest, sold as the NNP-T, with an inference chip code-named Spring Hill, sold as the NNP-I.[5] The processors traced their lineage directly to Nervana's original Engine design. In December 2019, however, Intel paid roughly 2 billion dollars to acquire Habana Labs, an Israeli AI-chip company, and soon decided that Habana's architecture, not Nervana's, would anchor its data-center AI roadmap. Intel cancelled further development of the Nervana NNP line.[5]

Following that strategic reversal, Rao left Intel; his last day was March 10, 2020. The decision to favor Habana over the technology he had brought into the company was widely reported as the proximate reason for his departure.[5]

## MosaicML

In 2021 Rao co-founded his second startup, MosaicML, again teaming with a fellow Intel veteran. Hanlin Tang, formerly a senior director in Intel's AI Lab, became chief technology officer. The founding team also drew on academic research into efficient deep learning: [Jonathan Frankle](/wiki/jonathan_frankle) of the [Massachusetts Institute of Technology](/wiki/mit), known for the "lottery ticket hypothesis," served as chief scientist, and MIT professor Michael Carbin was among the founders.[6][8]

MosaicML's premise was that training large neural networks had become needlessly expensive, and that algorithmic and systems improvements could cut the cost and time of training by large factors. The company exited stealth on October 13, 2021, with an initial 37 million dollar round led by Lux Capital and DCVC, with participation from Future Ventures, Playground Global, AME Cloud Ventures, Correlation Ventures, and E14; it raised about 64 million dollars in total.[6] MosaicML released open-source tooling including Composer, a [PyTorch](/wiki/pytorch) training library, and StreamingDataset, a system for streaming training data from object storage.[8]

As the [large language model](/wiki/large_language_model) boom accelerated, MosaicML shifted toward helping enterprises train their own models on proprietary data instead of relying on closed APIs. In May 2023 it released MPT-7B, the first in a family of open-source, commercially usable MPT (MosaicML Pretrained Transformer) models, including a StoryWriter variant with an extended context window. In June 2023 it followed with MPT-30B, trained on NVIDIA H100 GPUs, which the company positioned as surpassing the original GPT-3 in quality. The MPT models were downloaded millions of times and helped establish MosaicML as a leading open-model lab.[8]

## Databricks

On June 26, 2023, Databricks announced a definitive agreement to acquire MosaicML for approximately 1.3 billion dollars, inclusive of retention packages, in one of the largest generative AI acquisitions of the year. The deal closed the following month.[9][10] Databricks, led by chief executive [Ali Ghodsi](/wiki/ali_ghodsi), folded MosaicML's training technology into its platform and rebranded it as Mosaic AI, using it to let enterprise customers build and serve custom models on their own data.[10]

Rao stayed on to run generative AI at Databricks. His role was reported variously as vice president of AI and vice president of generative AI, reflecting a deliberately broad remit.[9] Under his oversight, the former Mosaic team built DBRX, an open [mixture-of-experts](/wiki/mixture_of_experts) language model with 132 billion total parameters and about 36 billion active per input, which Databricks released on March 27, 2024 as a demonstration that enterprises could own high-quality models trained on their own data.[10]

## Unconventional AI

In September 2025 Rao left Databricks to start his third company, Unconventional AI (legally Unconventional, Inc.). The startup aims to rethink the foundations of computing hardware for AI, designing custom silicon together with the surrounding server systems. Rao has described the goal as building "a new substrate for intelligence" that is "as efficient as biology," a direct return to the brain-inspired ambitions of his neuroscience training, and has framed the venture as a challenge to NVIDIA's dominance in AI chips.[11][12]

In October 2025, reporting indicated that Rao was raising toward a target of up to 1 billion dollars at a valuation around 5 billion dollars, with backing from [Andreessen Horowitz](/wiki/andreessen_horowitz), Lightspeed Venture Partners, and Lux Capital, and that Databricks was also among the investors.[11] On December 9, 2025, the company confirmed a 475 million dollar seed round at a 4.5 billion dollar valuation, led by Andreessen Horowitz and Lightspeed, with Lux Capital and DCVC participating. The financing was described as a first installment toward the larger target.[12]

## Other roles

Rao has also held outside positions reflecting his standing in the field. In September 2017, while at Intel, he was named to the board of directors of Square, the payments company later renamed [Block](/wiki/block_inc), to advise on its machine-learning strategy.[1] He has additionally served as a scientific advisor to the Allen Institute and has been a frequent speaker and commentator on AI hardware and the economics of model training.[7]

As of June 2026, Rao is founder and chief executive of Unconventional AI and is regarded as one of the most prolific founders in AI infrastructure, having built and exited two companies in deep-learning hardware and software before launching a third aimed at reinventing the computer for AI.[11][12]

## References

1. Square (Block), "Square Names Naveen Rao, Leader in Artificial Intelligence, to Board of Directors," press release, September 21, 2017.
2. Wikipedia, "Nervana Systems" (accessed 2026), covering the 2014 founding, co-founders Amir Khosrowshahi and Arjun Bansal, the Nervana Engine and Nervana Cloud, and the 2016 Intel acquisition at an estimated 408 million dollars.
3. TechCrunch, "Intel buys deep learning startup Nervana Systems for a reported $350 million," August 9, 2016.
4. MIT Technology Review, "Intel Buys a Startup to Catch Up in Deep Learning," August 9, 2016.
5. Synced (SyncedReview), "Intel AI GM Naveen Rao Leaves After Habana Chosen Over Nervana," 2020, covering his title, the NNP-T (Spring Crest) and NNP-I (Spring Hill) chips, the Habana Labs decision, and his March 10, 2020 departure.
6. SiliconANGLE, "Naveen Rao's startup MosaicML exits stealth with a mission to improve machine learning training," October 13, 2021, on the 2021 founding, CTO Hanlin Tang, and the 37 million dollar round.
7. Business Insider, "These three neuroscientists went from leading a hot startup to helping lead a key Intel initiative," 2019, plus Directors & Boards roster on his pre-PhD computer-architecture career.
8. Databricks (MosaicML) Blog, "Introducing MPT-7B: A New Standard for Open-Source, Commercially Usable LLMs," May 5, 2023, and "MPT-30B: Raising the bar for open-source foundation models," June 22, 2023.
9. Databricks, "Databricks Signs Definitive Agreement to Acquire MosaicML, a Leading Generative AI Platform," newsroom press release, June 26, 2023.
10. TechCrunch, "Databricks picks up MosaicML, an OpenAI competitor, for $1.3B," June 26, 2023; Databricks, "Databricks Launches DBRX," March 27, 2024.
11. TechCrunch, "Naveen Rao's new AI hardware startup targets $5B valuation with backing from a16z, sources say," October 3, 2025.
12. TechCrunch, "Unconventional AI confirms its massive $475M seed round," December 9, 2025.