Unconventional AI

Unconventional AI (legally Unconventional, Inc.) is an American AI hardware startup, headquartered in Mountain View, California, that is trying to build computers for artificial intelligence that compute with the physical dynamics of analog and mixed-signal circuits rather than with conventional digital arithmetic.^[1][2][3] It was founded in 2025 by Naveen Rao, the former head of AI at Databricks who had earlier co-founded the chip startup Nervana Systems and the model-training company MosaicML, together with MIT professor Michael Carbin, Stanford professor Sara Achour and analog chip designer MeeLan Lee.^[4][5][6] The company says its goal is a 1,000-fold improvement in the energy efficiency of generative AI inference compared with state-of-the-art models running on state-of-the-art conventional hardware; press coverage has stressed that this is a target, not a demonstrated result.^[7][8]

In December 2025, about two months after it was formed, the company announced a $475 million seed round at a $4.5 billion valuation, co-led by Andreessen Horowitz and Lightspeed Venture Partners.^[1][4] Its first public model, Un-0 (June 2026), is an open-source image generator built on a software simulation of coupled oscillators.^[9][10] In September 2026 Rao said the company had taped out and tested its first physical prototype chip, and in October 2026 Unconventional agreed to acquire the Israeli AI inference chip company NeuReality.^[11][12][2]

Overview

History

Founding (2025)

Rao joined Databricks in 2023 when it bought MosaicML for $1.3 billion, and he ran AI at the company until September 2025.^[13][4] On September 12, 2025, Bloomberg reported that he planned to leave Databricks to launch a new venture "making a novel type of computer". A Databricks spokesperson confirmed that Rao was moving into an advisory role and said Databricks planned to invest in the new startup, without disclosing the amount.^[17] Later that month Rao acknowledged the company on X, describing it as "rethinking the foundations of a computer to build a new substrate for intelligence that is as efficient as biology. Brain Scale Efficiency without the biological baggage!"^[13]

On October 3, 2025, TechCrunch, citing four people familiar with the discussions, reported that Rao was in talks to raise $1 billion at a $5 billion valuation for "Unconventional, Inc." Two of the sources said Andreessen Horowitz had agreed to lead and that Lightspeed and Lux Capital were participating. According to the sources, Rao had already raised hundreds of millions of dollars and planned to raise the rest in installments (a tranched round), and the company would compete with Nvidia by designing a new AI machine with custom silicon chips and server infrastructure.^[13]

Seed round (December 2025)

The company announced itself on December 8, 2025, with a blog post credited to its four co-founders. It said it had raised $475 million in seed funding at a $4.5 billion valuation, in a round led by Lightspeed and Andreessen Horowitz with participation from Sequoia Capital, Lux Capital, DCVC, Future Ventures and Jeff Bezos, among others, and that Rao would personally invest $10 million.^[1] Bloomberg, which described Unconventional as "a two-month-old startup", added that Databricks also took part and that Rao invested his $10 million on the same terms as the other investors. Rao told Bloomberg the $475 million was the first portion of a financing of up to $1 billion, and said: "The next several years are going to be about trying out a number of ideas and prototypes and coming up with the exact paradigm of what we believe will scale most efficiently and cost effectively."^[4] TechCrunch noted that the valuation was slightly below the $5 billion its sources had described in October.^[16]

Lightspeed partners Guru Chahal, Ravi Mhatre and Bucky Moore published an investment note the same day, arguing that the gap between the power drawn by today's AI hardware, with the latest Nvidia GPUs using more than a kilowatt each, and the roughly 20 watts used by the human brain was "one of the largest opportunities in technology".^[5] Job postings described by SiliconANGLE at launch said the company wanted to build an AI accelerator with 1,000 times better efficiency than current silicon, based on a system-on-chip with a CPU, third-party IP blocks and mixed-signal circuits; another posting sought an engineer familiar with emerging nonvolatile memories such as RRAM.^[18] According to BigDATAwire, which also drew on a video interview Rao gave Andreessen Horowitz, Rao said the company was giving itself five years to build a new analog chip, that it would partner with TSMC, and that the first prototype would be "probably one of the larger, maybe the largest analog chip people have ever built."^[19]

Building the company (2026)

Ali Esfahani, a former technology banker and investor who holds the title "[un]CFO", wrote that the team had 14 people when he joined in January 2026, that the cap table included more than 70 investors, and that the round announced in December had not yet closed at that point. He said the company then held more than 100 investor meetings and raised an additional $60 million, bringing the round to $535 million.^[15] Rao later said the company had "started in earnest in January".^[11] By June 2026 TechCrunch put the headcount at fewer than 50.^[10]

From April 2026 the company published a series of research and position posts on its blog, covering hardware and software co-design, analog computing, memory technology and the efficiency target, and it ran an academic grant program with a May 2026 deadline.^{[14][20][7][21]} It released its first model, Un-0, on June 25, 2026, followed by a sparsity study in August.^[9][22] In June it also published a post by Ang Li, a researcher who joined the company after the work described, on DS-ISA, a proposed instruction set architecture for "dynamical system units" accepted at the ISCA 2026 computer architecture conference; the post describes a nine-instruction ISA built around five phases (connect, load, lock, evolve and store).^[23]

At the All-In Summit in September 2026, Rao said the company had taped out its first physical prototype on June 1, 2026, and that the returned chip had produced its first results.^[11][24] On October 26, 2026, Unconventional announced a definitive agreement to acquire NeuReality.^[12]

Founders

Lightspeed's note described Achour and Carbin as having deep expertise in programming novel computing substrates such as analog devices, quantum systems and neuromorphic architectures.^[5] The company describes itself as "founded by experts in AI systems, analog circuits, computing theory, and neuroscience".^[30]

Funding

Sources describe the $4.5 billion figure as the company's valuation without stating whether it is pre-money or post-money.^[4][16] Sequoia also lists Unconventional AI among its portfolio companies.^[31]

Technical approach

Energy as the limit on AI

The company's founding post argues that demand for AI computation, if projections hold, will be constrained by global energy supply within three to four years, and that the human brain, which runs on about 20 watts, shows that computers for AI can be far more efficient.^[1] Its central claim is that neural networks are stochastic machines that today run on deterministic digital abstractions layered over analog circuitry tuned to behave digitally. The founders proposed instead to give software an interface to "the inherent physics of the silicon", so that "we will run the neural network on the physics directly rather than simulating some physical system."^[1] SiliconANGLE likened the idea to the neuromorphic processors Intel has developed, which also mimic biological neurons to improve efficiency.^[18]

Reuters described the method as what Rao calls a "dynamical systems" approach, in which simple computing elements interact by simple rules and evolve over time to reach a result, rather than following programmed step-by-step logic.^[2] In his September 2026 talk Rao called this "4D computing": the time dimension of the dynamics plus the three physical dimensions available through planar layout and die stacking, in a design where each computing element also serves as memory, so there is no separate memory interface.^[11] The NeuReality announcement describes the same goal as a "4-dimensional compute substrate".^[12]

Neural co-evolution

An April 2026 post credited to the four co-founders and four other researchers argues that neural networks came to be designed around the matrix multiplication and vector operations that GPUs handle well, and that the separation between hardware and software design has created what it calls an innovation logjam in AI hardware. The company calls its alternative "neural co-evolution": designing the neural network and the physical system together, with time treated as "a first-class citizen" in computation.^[14] The post cites earlier co-designed hardware, including Nervana's chip, Google's TPU, Cerebras wafer-scale engines, Groq chips and AWS Trainium and Inferentia, as earlier attempts at co-evolution, each aimed at specific bottlenecks of the GPU paradigm.^[14]

The 1,000x target and the memory problem

A May 2026 post by Peter McMahon sets out what the 1,000x target means. The company measures energy as joules per token (for text) or joules per image, at equal output quality to state-of-the-art models on state-of-the-art GPUs or TPUs, with a focus on data center inference; it explicitly rejects TOPS and TOPS per watt as the main metric, and says its baseline is what conventional hardware is expected to reach around 2030.^[7] The post argues that the energy of inference is dominated by storing and moving data rather than by arithmetic: it cites Google figures for 7 nm CMOS showing a more than 500x gap between an 8-bit integer multiply and an 8-bit read from off-chip high bandwidth memory, and says that moving model weights from off-chip HBM to on-chip SRAM gives at best a 10x advantage, even under unrealistically favorable conditions.^[7] Because of Amdahl's law, it says, nearly every part of the system has to be optimized, including attention layers and reads of the KV cache as well as the linear layers, and the 1,000x might be assembled from "10 factors of 2x, or 3 factors of 10x".^[7] Approaches it lists include specializing hardware to a model architecture, hard-wiring model weights (an approach it associates with Taalas), reusing parameters through recurrent dynamics, sparsity, replacing attention with dynamics in the manner of state space models such as Mamba, and letting circuits compute whatever input-output functions they compute efficiently rather than forcing exact matrix-vector multiplication.^[7]

Analog and digital

In an April 2026 post, Boris Murmann argued that analog dot products can be more efficient than digital ones in the low-precision regime of about 1 to 4 bits, while their energy cost rises steeply at high precision; that the measured gap between the best analog designs and digital in-memory computing is smaller than hoped; and that external DRAM accesses can account for half of system energy during large language model decoding. He suggested that a clean-slate, mixed-signal fabric in which analog and digital blocks compute together might succeed where analog alone has not.^[20] Later posts covered training physical systems such as springs and coupled nonlinear differential equations for sequence classification, and the trade-offs among memory technologies from phase-change memory to embedded DRAM.^[30]

Un-0

Architecture

Un-0, released on June 25, 2026, is a class-conditional image generator whose core is a large population of simulated Kuramoto oscillators: each oscillator has a phase that rotates at its own natural frequency and is pulled toward or away from alignment by every other oscillator. The learnable parameters are the coupling strengths between every pair of oscillators and each oscillator's natural frequency.^[9] To generate an image, the model sets every phase to a random angle, drives the main population with a smaller group of class-conditioning oscillators, lets the system evolve, reads out all phases at a fixed time, and passes that latent grid through a conventional upsampling decoder that accounts for under 13 percent of the parameters.^[9] The company chose oscillators because rhythmic synchronization is common in the brain and because a coupled-oscillator system can, in principle, be built directly in CMOS so that the circuit's physics computes the dynamics.^[9]

The models were trained end to end with the "drifting" loss of Deng and colleagues (2026), using DINOv2 features and the AdamW optimizer, with the dynamics integrated by an explicit Euler scheme. Training ran on NVIDIA B200 GPUs: the largest ImageNet model took 640 B200 hours.^[9] Ablations compared the full model with a decoder alone and with a "reservoir" version whose couplings were frozen at random values; the company reported that trained dynamics beat both, and concluded that the oscillators were doing useful computation rather than leaving the work to the decoder.^[9] The weights, together with training, evaluation and ablation code, were published in an MIT-licensed GitHub repository written in plain PyTorch, with checkpoints loadable from the Hugging Face Hub.^[32]

Results

Figures are from the release post.^[9] The repository's results table, which uses clean-FID for CIFAR-10 and the OpenAI ADM evaluator for ImageNet, gives slightly different numbers for some released checkpoints (for example 8.86 for the largest CIFAR-10 model), and the same 6.74 for the largest ImageNet model.^[32]

The company called Un-0 "to our knowledge, the most capable image generator to date to use a simulation of a physical dynamical system". It said Un-0's Frechet Inception Distance placed it alongside or above early conventional generators such as NCSN, WGAN-GP and BigGAN, and that at small parameter counts it extended the quality-versus-size frontier among the models it compared, but that it still trailed later diffusion models such as EDM and improved more slowly with scale than conventional methods.^[9] TechCrunch's launch coverage described the model as performing as well as state-of-the-art diffusion models and quoted Rao calling it "the 'hello world' of a new kind of computer"; it reported that Un-0 ran on a software simulation of the company's oscillator chips, that the company planned to release schematics for a real chip, and that it eventually intended to sell compute capacity from its own systems.^[10]

Sparsity study

Because every oscillator in Un-0 is coupled to every other, the number of couplings grows with the square of the oscillator count, which is hard to build in hardware. An August 2026 post tested random sparsity and a modular layout (all-to-all connections only within groups), removing between 50 and 98.4 percent of the couplings. The company reported that both forms of sparsity improved trainability up to a critical level, beyond which more sparsity hurt, and that a 16-module version of its 6,656-oscillator ImageNet 64x64 model reached an FID of about 7.1 (7.15 in the post's summary, 7.11 in its results table), which it described as about 1.9 better than the dense baseline. In that results table the dense baseline scored 9.07, rather than the 8.41 the Un-0 release reported for the same-size model.^[22][9]

Hardware prototypes

When it raised its seed round, the company expected about three years of research and development and five years to a product, Rao told Reuters in October 2026.^[2] At the All-In Summit in September 2026, he said the company had sent its first prototype design to the fab on June 1, 2026, that the chip was back in the lab, and that it had produced images at roughly 500 nanojoules per image, compared with what he described as millijoules on a GPU. He called it "the first physical dynamical computer ever built", and said that faster-than-expected progress had led him to revise the 1,000x efficiency goal from five years to three and a half.^[11] In October 2026 Rao told Reuters that, helped by its use of AI tools, the company had developed its first test chip and was nearing completion of its second.^[2] These are the company's own statements; as of October 27, 2026, its blog had not published a technical description of the chip.^[30]

Acquisition of NeuReality

On October 26, 2026, Unconventional announced a definitive agreement to acquire NeuReality, an AI inference infrastructure and silicon company based in Caesarea, Israel. The deal had been approved by both boards and was expected to close in the fourth quarter of 2026, subject to customary conditions. NeuReality co-founder and CEO Moshe Tanach and the founding team are to join Unconventional, which said it would keep an engineering presence in Israel.^[12] Rao said the deal would bring in a team that "can build real silicon and the systems around it", and described the 1,000x target as "a systems problem as much as a physics problem".^[12]

Financial terms were not disclosed. Reuters described it as a cash-and-stock deal; Globes reported market estimates of $150 million to $200 million in cash and shares, which neither company confirmed; and Calcalist reported that the deal was expected to return more than the roughly $105 million invested in NeuReality.^[2][33][8] According to Globes, about 40 NeuReality employees in Caesarea and Tel Aviv are to join a similar number of Unconventional staff in California, while some sales, marketing and finance staff were to be laid off.^[33] Rao told Reuters that the faster-than-expected arrival of its test chips had prompted the company to look for an engineering team with experience delivering finished products: "We built up the research bones and now we've got to build up the production bones."^[2]

NeuReality was founded in 2019 by Moshe Tanach, Zvika Shmueli and Yossi Kasus and raised about $105 million from investors including Arm, Samsung Ventures, Varana Capital, Cleveland Avenue and OurCrowd.^[8] Rao had a prior connection to the company: he joined its board when it came out of stealth with an $8 million seed round in February 2021, and Globes listed him among its early backers.^[34][33][2]

Unconventional Grant

In 2026 the company set up a $500,000 Unconventional Academic Research Fund offering up to five unrestricted grants of $100,000 each to professors at US universities, for one-year projects testable in theory or simulation. Focus areas included unconventional circuits, systems architecture for heterogeneous unconventional cores, neural network architectures that exploit analog dynamics or minimize data movement, theory of AI and dynamical systems, and 3D integration of computing and memory. Pre-proposals were due May 15, 2026.^[21] The company said it received nearly 100 submissions and announced five awards on September 10, 2026, asking only that the work be published openly.^[35]

Source for the table: the company's award announcement.^[35]

Reception

The size of the seed round drew attention. Bloomberg grouped Rao with other prominent founders whose young startups had multibillion-dollar valuations, naming Mira Murati's Thinking Machines Lab, Ilya Sutskever's Safe Superintelligence and Bret Taylor's Sierra, and Globes called the raise one of the largest seed rounds ever.^[4][33] The Next Web's coverage of Un-0 stressed that the model ran on a simulation of hardware that did not yet exist and called the 1,000x claim aspirational, and Calcalist's report on the NeuReality deal said the improvement "remains a target, rather than a demonstrated result."^[36][8] BigDATAwire noted that Unconventional was not the only company pursuing analog AI chips, citing Mythic's analog processors and a Peking University RRAM-based analog matrix computing system.^[19]

Debarghya Das's "Ultimate Neolab List" includes the company but rates it a borderline case: researcher-led and pre-revenue with a huge early raise, but focused on hardware rather than mainly software.^[37]

Timeline

See also

• List of neolabs
• Neolab
• Naveen Rao
• Neuromorphic computing
• Thermodynamic computing
• Extropic
• Normal Computing
• AI energy consumption
• AI chip

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Naveen Rao, MeeLan Lee, Michael Carbin and Sara Achour, Unconventional AI, Introducing Unconventional AI, December 8, 2025.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Reuters (via The Algemeiner), Unconventional AI Buys Israeli Chip Startup NeuReality to Speed Product Development, October 26, 2026.
3. ^1 ^2 ^3Unconventional, Inc., Open roles (Greenhouse job board), accessed October 27, 2026 (all 18 listed roles based in Mountain View, California, some with remote options).
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Rachel Metz, Bloomberg (via The Mercury News), AI computer startup hits $4.5 billion valuation in seed round, December 8, 2025.
5. ^1 ^2 ^3 ^4Guru Chahal, Ravi Mhatre and Bucky Moore, Lightspeed Venture Partners, Investing In Unconventional AI: Biology-Scale Efficiency For The AI Era, December 8, 2025.
6. ^1 ^2 ^3 ^4 ^5 ^6Lightspeed Venture Partners, Unconventional AI (portfolio page), accessed October 27, 2026.
7. ^1 ^2 ^3 ^4 ^5 ^6Peter McMahon, Unconventional AI, How to improve AI energy efficiency by 1000x, May 7, 2026.
8. ^1 ^2 ^3 ^4Meir Orbach, Calcalist (Ctech), One-year-old AI unicorn Unconventional acquires Israeli chipmaker NeuReality, October 26, 2026.
9. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Unconventional AI, Introducing Un-0: Generating Images with Coupled Oscillators, June 25, 2026.
10. ^1 ^2 ^3 ^4 ^5Russell Brandom, TechCrunch, Databricks' former AI chief thinks he can cut AI's power bill by 1,000x, June 25, 2026.
11. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8All-In with Chamath, Jason, Sacks & Friedberg, Naveen Rao: 4D Computing, AI's Energy Wall & Beating Biology (transcript), All-In Summit talk, episode published September 21, 2026 (machine transcript via HappyScribe).
12. ^1 ^2 ^3 ^4 ^5 ^6Unconventional AI, Unconventional AI to Acquire NeuReality, Accelerating Its Path to 1,000x AI Compute Efficiency, October 26, 2026.
13. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Marina Temkin, TechCrunch, Sources: Naveen Rao's new AI hardware startup targets $5B valuation with backing from a16z, October 3, 2025.
14. ^1 ^2 ^3 ^4 ^5Carlos H. Morales, Peter McMahon, Srenik Mehta, Sriram Krishnamoorthy, Sara Achour, Michael Carbin, MeeLan Lee and Naveen Rao, Unconventional AI, Neural co-evolution, the inevitability of hardware and software co-evolution for AI, April 2, 2026.
15. ^1 ^2 ^3 ^4 ^5Ali Esfahani, Unconventional AI, [The Making of the [un]CFO: Building a Company in the Age of AI](unconv.ai/...o-building-a-company-in-the-age-of-ai), June 18, 2026.
16. ^1 ^2 ^3Marina Temkin, TechCrunch, Unconventional AI confirms its massive $475M seed round, December 9, 2025.
17. ^1 ^2 ^3Bloomberg (via Bloomberg Law), Databricks AI Chief to Exit, Launch a New Computer Startup, September 12, 2025.
18. ^1 ^2Maria Deutscher, SiliconANGLE, Jeff Bezos backs $475M seed round for chip startup Unconventional AI, December 8, 2025.
19. ^1 ^2 ^3BigDATAwire (HPCwire), Unconventional AI Wants to Solve AI Scaling Crunch with Analog Chips. Will It Work?, December 11, 2025.
20. ^1 ^2Boris Murmann, Unconventional AI, Analog is dead, long live analog!, April 30, 2026.
21. ^1 ^2 ^3Unconventional AI, Unconventional Grant, program page, accessed October 27, 2026.
22. ^1 ^2 ^3Unconventional AI, Less is More: Scaling Dynamics with Sparsity, August 13, 2026.
23. ^Ang Li, Unconventional AI, Speaking the Language of Physics: Introducing the First Instruction Set Architecture (ISA) for Dynamical System Hardware, June 30, 2026.
24. ^1 ^2Molly O'Shea, Sourcery, ICYMI: 10 Wild Hot Takes From the All-In Summit 2026, September 27, 2026.
25. ^Michael Carbin, Home page, MIT CSAIL, accessed October 27, 2026.
26. ^Jonathan Frankle and Michael Carbin, The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks, arXiv:1803.03635, March 2018.
27. ^MIT News, Helping nonexperts build advanced generative AI models, June 21, 2024.
28. ^Stanford Profiles, Sara Achour, accessed October 27, 2026.
29. ^Sara Achour, Home page, MIT CSAIL, accessed October 27, 2026.
30. ^1 ^2 ^3Unconventional AI, Blog and home page, accessed October 27, 2026.
31. ^Sequoia Capital, Unconventional AI (portfolio page), accessed October 27, 2026.
32. ^1 ^2Unconventional AI, unconv-ai/Un-0, GitHub repository (MIT License), created June 2026, accessed October 27, 2026.
33. ^1 ^2 ^3 ^4Globes, Energy efficient chip co Unconventional AI buys Israel's NeuReality, October 26, 2026.
34. ^Frederic Lardinois, TechCrunch, NeuReality raises $8M for its novel AI inferencing platform, February 10, 2021.
35. ^1 ^2 ^3Unconventional AI, Announcing the 2026 Unconventional Grant Winners, September 10, 2026.
36. ^The Next Web, Unconventional AI releases its first model, built on an oscillator architecture its founder says could cut power use by a factor of a thousand, June 25, 2026.
37. ^Debarghya Das, The Ultimate Neolab List, accessed October 27, 2026.