Rain AI
Last reviewed
May 21, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 4,510 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 21, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 4,510 words
Add missing citations, update stale details, or suggest a clearer explanation.
Rain AI, originally incorporated as Rain Neuromorphics, is a San Francisco AI accelerator startup founded by Gordon Hirsch Wilson, Jack Kendall, and Juan Claudio Nino, with technical roots in the University of Florida materials science and computer engineering departments where the three founders met.[^1][^2] The company designs energy-efficient processors for neural-network training and inference, marketing them as brain-inspired alternatives to mainstream GPUs. Across its eight-year history Rain has pivoted at least twice on substrate technology, first showcasing memristive nanowire networks and end-to-end analog circuits, then moving to a digital compute-in-memory (D-IMC) architecture coupled with a RISC-V control core.[^3][^4][^5] The firm became widely known outside the chip community in late 2023, when reporting by WIRED revealed that OpenAI had signed a non-binding letter of intent in 2019 to spend roughly fifty-one million dollars on Rain chips while OpenAI chief executive Sam Altman was personally invested in the startup, an arrangement that figured into broader scrutiny of Altman's portfolio after the November 2023 OpenAI board crisis.[^6][^7] By May 2025 the company was reported to be running on bridge financing and exploring a sale after a planned one-hundred-fifty-million-dollar Series B failed to close.[^8][^9]
| Field | Value |
|---|---|
| Legal name | Rain Neuromorphics, Inc. (operates as Rain AI) |
| Founded | 2017 (University of Florida); Y Combinator Summer 2018 batch[^2][^10] |
| Headquarters | San Francisco, California[^7][^11] |
| Founders | Gordon Wilson, Jack Kendall, Juan Claudio Nino[^1][^2] |
| Architecture (current) | Digital compute-in-memory (D-IMC) cores with RISC-V control[^4][^5] |
| Lead Series A investor | Prosperity7 Ventures (Aramco Ventures), Feb 2022[^3][^12] |
| Notable angel investors | Sam Altman, Y Combinator, Daniel Gross, Jaan Tallinn, Airbus Ventures[^3][^10] |
| Reported OpenAI LOI | ~$51M, non-binding, signed 2019[^6][^7] |
| Status (as of mid-2025) | Exploring sale; bridge financing reported[^8][^9] |
The Rain story begins in materials science labs at the University of Florida, where Juan Nino headed a research group studying electroceramic materials and where graduate student Jack Kendall worked on what would become the company's earliest published architecture: a memristive nanowire network in which thousands of randomly-arranged resistive switches form synaptic interconnects.[^1][^13] Gordon Wilson, then an undergraduate with a background in entrepreneurship, joined the team to commercialize the idea, and the trio formed Rain Neuromorphics in 2017.[^1] The defining technical claim of this period, captured in a 2020 paper "Deep Learning in Memristive Nanowire Networks," was that random three-dimensional crossbar arrays could implement matrix-vector products at extremely low energy if matched with appropriately analog-friendly training methods.[^13]
Rain Neuromorphics was admitted to Y Combinator's Summer 2018 batch and used the demo day cycle to close roughly five million dollars in seed financing from a roster of AI-leaning angels.[^10][^11] Documented seed-round investors include Y Combinator itself, Sam Altman of OpenAI, Daniel Gross (then a YC partner and angel investor alongside Nat Friedman), Skype co-founder Jaan Tallinn, Airbus Ventures, Liquid 2 Ventures, Deepwater Asset Management, and FoundersX Ventures.[^10][^11] The company relocated from Gainesville, Florida to the San Francisco Bay Area after graduating from the YC batch, occupying offices less than a mile from OpenAI's headquarters in the Mission District, a geographical proximity that would later attract notice when the OpenAI letter of intent became public.[^7][^11]
In 2019, OpenAI signed a non-binding letter of intent to purchase roughly fifty-one million dollars in chips from Rain once the startup reached production silicon, according to deal documents and Rain disclosures to investors reviewed by WIRED in late 2023.[^7] Sam Altman, who as a private investor had put more than one million dollars into Rain at the seed and subsequent rounds, was both the chief executive of the customer (OpenAI) and a financial stakeholder in the supplier, an entanglement that critics later flagged as a structural conflict of interest.[^6][^7] At the time the LOI did not attract public attention. Letters of intent are typically non-binding, and one venture investor quoted in the WIRED reporting cycle observed that such documents are "non-binding 99%+ of the time" and can be unwound at any moment.[^6] Rain had not yet shipped silicon when the letter was signed, so the document functioned as a forward commitment contingent on the company actually fielding production parts.[^7]
In February 2022 Rain Neuromorphics announced a twenty-five-million-dollar Series A led by Prosperity7 Ventures, a venture vehicle funded by Saudi Aramco.[^3][^12] Co-investors included Buckley Ventures, Gaingels, Loup Ventures, Metaplanet, Pioneer Fund, and a long list of operator-angels such as Jeff Rothschild (Facebook), Oliver Cameron (Cruise), Amar Shah (Wayve AI), and Scott Gray (OpenAI).[^3][^12] In the press release Rain unveiled what it described as the world's first "end-to-end analog, trainable AI circuit," a Neuromorphic Processing Unit projected to be roughly one thousand times more energy-efficient than contemporary digital processors on the workloads the company targeted.[^3] Around this period the company began trading publicly under the shorter "Rain AI" name and a new visual identity, with the Rain Neuromorphics legal entity remaining the corporate parent.[^4]
The Series A coincided with two technical milestones. First, EE Times reported in October 2021 that Rain had taped out a demonstration chip implementing a three-dimensional memristor array in collaboration with Mila, the Quebec AI institute led by Yoshua Bengio.[^14] Second, in 2022 Rain co-authored with Bengio's group and Sandia National Laboratories a Nature Electronics paper titled "Activity-difference training of deep neural networks using memristor crossbars," demonstrating on-chip learning on a memristor-based crossbar array using a local learning rule compatible with analog hardware.[^15] These two threads, the analog hardware tape-out and the local-learning algorithmic work, together constituted what Rain marketed as its "end-to-end analog" thesis: a chip on which both inference and training execute in continuous-valued physical quantities (Ohm's-law currents through memristive conductances) rather than in discrete digital arithmetic.[^14][^16]
In November 2023, Bloomberg reported that the Committee on Foreign Investment in the United States (CFIUS) had compelled Prosperity7 Ventures to divest its Rain stake on national-security grounds, citing concerns over Saudi-Chinese AI collaboration and the potential for sensitive AI hardware to reach restricted end users.[^17][^18] CFIUS, an inter-agency committee within the U.S. Treasury that reviews foreign acquisitions of American businesses, did not publicly disclose the specific transaction or rationale, but the unwinding played out over the course of 2023.[^18][^19] The action followed broader August 2023 export restrictions on advanced AI chips from Nvidia and AMD to several Middle Eastern countries.[^18] Public reporting did not identify which subsequent investors purchased Prosperity7's divested shares.[^18]
OpenAI's board removed Sam Altman as chief executive on 17 November 2023, citing that he had not been "consistently candid" with the board.[^6] After a multi-day standoff Altman was reinstated. Within two weeks, on or around 4 December 2023, WIRED published its report on the Rain letter of intent, drawing on deal documents and Rain investor disclosures.[^7] The publication of the LOI did not surface during the firing itself, but it crystallized a broader narrative about overlapping personal investments and OpenAI commercial decisions, alongside reporting on Altman's stakes in Helion (fusion), Worldcoin, and Reddit.[^6][^7] Reporting outlets including Computerworld, Data Center Dynamics, and The Decoder picked up the WIRED disclosure and emphasized the non-binding nature of the document, the absence of any concluded purchase, and the conflict-of-interest framing.[^6][^7][^20]
By 2024 Rain had publicly retired the pure-analog framing and was talking instead about digital compute-in-memory cores coupled with a RISC-V control processor for general-purpose neural-network operators.[^4][^5] In June 2024 the company announced two events that signaled this shift. First, Rain licensed the Andes Technology AX45MPV RISC-V vector processor as the general-purpose computational complement to its D-IMC matrix-multiplication unit, with Rain telling press it planned to "unveil its accelerator solution in early 2025."[^5][^21] Second, Rain hired Jean-Didier Allegrucci, a seventeen-year Apple silicon veteran who had overseen the integration of more than thirty system-on-chips across iPhone, Mac, iPad, and Apple Watch, to lead hardware engineering.[^22][^23] CEO William Passo, a former attorney who had succeeded co-founder Gordon Wilson in the chief executive role in late 2023, described Allegrucci's hire as a signal of Rain's transition from research to high-volume product execution.[^22][^24] Around the same time Rain announced a tape-out collaboration with Synopsys, using Synopsys Cloud and IP to bring a first energy-efficient AI accelerator from architecture to tape-out "in under a year," with the company reporting a thirty percent improvement in engineering productivity via the cloud EDA toolchain.[^25]
To extend runway, Rain raised at least one Series A extension in 2024, with Epic Venture Partners contributing roughly 8.1 million dollars on the back of what the company described as accelerated customer demand.[^26] In late 2024 the company opened a one-hundred-fifty-million-dollar Series B at a reported six-hundred-million-dollar valuation, with Sam Altman publicly pitching investors on the round.[^9][^27] The round failed to close.[^8][^9] In Spring 2025 co-founder Jack Kendall, who had taken over as chief executive after William Passo stepped down citing personal reasons, told stakeholders that Rain was "rapidly depleting its cash reserves" and was raising a three-million-dollar bridge to fund operations while exploring an acquisition.[^8][^9] Press reporting in May 2025 named OpenAI, Nvidia, and Microsoft as parties evaluating the company, with OpenAI described as the leading suitor and reported to be interviewing Rain employees for talent assessment.[^8][^9] As of mid-2025 no acquisition had been announced publicly, and the company continued to operate.[^8]
Rain's technical story has two distinct phases. The earlier "end-to-end analog" thesis, dominant from founding through approximately 2023, treated multiply-and-accumulate operations as physical processes inside arrays of memristive devices. The current "digital compute-in-memory" thesis, dominant from late 2023 onward and described on the company's product pages by 2024, retains the architectural idea of co-locating storage and arithmetic in memory tiles but executes the arithmetic with conventional digital logic rather than analog conductances.[^4][^5]
In the analog phase, Rain's signature device was the memristor, a two-terminal element whose conductance changes as a function of charge history and which behaves, under certain bias conditions, as a programmable resistor.[^13][^14] Rain proposed two distinct memristive substrates over the years: an early "memristive nanowire" architecture in which randomly deposited sub-micron resistive filaments form a three-dimensional sparse network, and a later resistive-RAM (ReRAM) crossbar architecture borrowing manufacturing techniques from NAND flash production.[^13][^14] In either substrate, a matrix-vector product is computed by applying a voltage vector across the rows of a crossbar; Ohm's law and Kirchhoff's current law combine output currents that sum, in a single physical operation, the products of input voltages and stored conductances.[^16][^28] Because no clocked digital operation is performed for each multiplication, the energy per multiply-accumulate is dominated by leakage and DAC/ADC conversion rather than by the multiplication itself, which is the source of the order-of-magnitude efficiency claims Rain made in the 2020 to 2022 period.[^16][^28]
The classical objection to analog matrix-multipliers is that they are difficult to train. Gradient-based backpropagation requires forward and backward passes that share a precise transposed-weight pathway, but device variability in analog crossbars makes the transposed-weight assumption unreliable.[^16] Rain's published response, in collaboration with Yoshua Bengio's group at Mila and Benjamin Scellier, was to adopt equilibrium propagation, an analog-friendly local learning rule that converges by relaxing the network to two slightly different equilibrium configurations and comparing them.[^16] A 2021 paper "Training End-to-End Analog Neural Networks with Equilibrium Propagation" with Yoshua Bengio as co-author claimed that crossbars of memristors could be trained without a separate digital training apparatus, at least in simulation.[^16] A subsequent 2022 Nature Electronics paper, co-authored with Sandia National Laboratories and Texas A&M, demonstrated on-chip training on a real memristor crossbar using an "activity-difference" local learning rule.[^15]
In October 2021 Rain announced a tape-out of a demonstration chip implementing this approach in a 180-nanometer CMOS process, the first physical embodiment of the company's end-to-end analog architecture.[^14] Rain described the demo as having ten thousand neurons and used it in technical presentations to argue that a substantially larger chip could deliver one-thousand-fold improvements in joules per inference compared with contemporary GPUs.[^14]
By the public 2024 product pages and the Andes partnership announcement, Rain had stopped foregrounding analog memristive computation and was instead describing a "Digital In-Memory Compute" (D-IMC) architecture.[^4][^5] On Rain's "Approach" page the company writes that D-IMC is "scalable to high-volume production and support[s] training and inference," that arithmetic uses a "block brain floating point" numerical format derived from BFLOAT16 with optimized 4-bit and 8-bit matrix-multiplication paths preserving FP32-equivalent accuracy, and that a proprietary interconnect bridges the D-IMC cores with a RISC-V vector control processor.[^4]
Compute-in-memory in its digital form (sometimes called digital CIM or D-CIM) retains the architectural insight that data movement, not arithmetic, dominates the energy budget of a matrix-multiply on modern silicon. In a digital CIM tile, multiplier-accumulator circuits are integrated into or adjacent to SRAM macros, so a row of weights can be fetched and consumed by an arithmetic operation in the same cycle without traversing a global bus.[^29] Unlike the analog approach, the multipliers themselves are conventional digital gates, which preserves the deterministic accuracy that frustrated analog systems but sacrifices some of the per-operation energy advantage of true analog charge accumulation.[^29] Andes CEO Frankwell Lin characterized Rain's contribution as having "designed one of the most energy efficient matrix multiplication units using digital CIM technology."[^5][^21]
The RISC-V Andes AX45MPV is a multi-core vector processor that handles general-purpose machine-learning operators (activations, normalization, attention softmax, layout reshapes) too irregular for the fixed-function D-IMC arrays. Rain layered Andes' Custom Computing Business Unit (CCBU) services on top of the AX45MPV license to add bespoke vector instructions describing the D-IMC operations, using Andes' ACE/COPILOT instruction-customization tooling.[^5][^21] The combination positions Rain's first generation accelerator as a heterogeneous SoC similar in shape to other modern AI inference processors: a fast matrix engine for the dense linear-algebra hot paths and a programmable vector unit for everything else.
Public Rain materials and press coverage variously cite "85% greater energy efficiency than Nvidia's" GPUs, "100 times more computing power" with "up to 10,000 times greater energy efficiency" than GPUs for training, and "1000x more energy efficient than today's best processors" for inference.[^3][^6][^20][^27] These claims have not been independently benchmarked in any standard MLPerf-style submission as of the company's public communications through mid-2025, and the order-of-magnitude estimates predominantly assume scaled-up future silicon rather than measured throughput on shipping product.[^29] EE Times reporting in 2022 described an early product roadmap targeting 125 million INT8 parameters with sub-50-watt power for vision, speech, NLP, and recommendation workloads, though that roadmap was associated with the analog generation that Rain has since deprioritized.[^14][^30]
| Date | Round | Amount | Lead investor | Notes |
|---|---|---|---|---|
| 2018 | Seed | ~$5M | Sam Altman / Y Combinator angels | Y Combinator S18 batch; angels included Daniel Gross, Jaan Tallinn, Airbus Ventures[^10][^11] |
| Feb 2022 | Series A | $25M | Prosperity7 Ventures (Aramco) | Pioneer Fund, Loup Ventures, Metaplanet co-invest[^3][^12] |
| 2023 | CFIUS divestment | n/a | n/a | Prosperity7 forced to exit on national-security grounds[^17][^18] |
| 2024 | Series A extension | ~$8.1M | Epic Venture Partners | Bridge to extend runway while pursuing Series B[^26] |
| late 2024 | Series B (attempted) | $150M target | (failed) | Reported $600M valuation; round did not close[^8][^9] |
| 2025 | Bridge | ~$3M | undisclosed | Working capital while pursuing acquisition[^8][^9] |
Rain's cumulative disclosed funding through mid-2025 is on the order of forty to fifty million dollars across the seed, Series A, A-extension, and bridge tranches, exclusive of the failed Series B target.[^11][^26] Sam Altman personally invested more than one million dollars in Rain across rounds, separately from his role pitching the Series B at a six-hundred-million-dollar valuation.[^6][^7][^9]
| Person | Role | Notes |
|---|---|---|
| Gordon Hirsch Wilson | Co-founder; CEO until late 2023; thereafter executive advisor | Stepped down November 2023[^24] |
| Jack Kendall | Co-founder; CTO; CEO from 2025 | Took over as CEO when Will Passo departed[^9] |
| Juan Claudio Nino | Co-founder; University of Florida professor of materials science | Originator of the memristor-materials thesis[^1][^13] |
| William "Will" Passo | CEO November 2023 to 2025; previously COO | Departed for personal reasons in 2025[^9][^24] |
| Jean-Didier Allegrucci | SVP / hardware engineering lead from June 2024 | Seventeen years at Apple silicon, 30+ SoC integrations[^22][^23] |
Rain has remained small by hardware-startup standards, with public team-size estimates in the low double digits through 2024 and approximately seventy employees by mid-2025 based on reporting on its sale process.[^9][^11]
Beyond the Andes Technology RISC-V license and the Synopsys EDA partnership for first-product tape-out, Rain has accumulated a small number of named ecosystem partners.[^5][^25] On the research side, the Mila collaboration with Yoshua Bengio and Benjamin Scellier produced the equilibrium-propagation analog-training paper, and the joint work with Texas A&M University and Sandia National Laboratories produced the 2022 Nature Electronics on-chip-training demonstration.[^15][^16] The OpenAI letter of intent has never converted to a commercial agreement, since Rain had not yet shipped production silicon at the time of writing; however the LOI has remained a recurring data point in industry discussion of OpenAI's hardware strategy.[^7][^9][^20] No major design-win customer has been announced publicly.
Rain operates in a crowded field of AI-accelerator startups attempting to dislodge Nvidia's GPU dominance in inference workloads, and a smaller subset of that field specifically pursues compute-in-memory architectures.
| Company | Substrate | Primary target | Notes |
|---|---|---|---|
| Rain AI | Digital CIM with RISC-V | Edge to data-center inference | Formerly analog memristor; pivoted to D-IMC[^4][^5] |
| Mythic AI | Analog matrix-multiplier in flash | Edge AI inference | Long-standing analog-CIM player; multiple product generations[^31] |
| EnCharge AI | Switched-capacitor analog CIM | PC and edge AI accelerators | Princeton spin-out led by Naveen Verma; charge-domain analog[^32] |
| Analog Devices | Analog CIM research and signal-chain IP | Mixed-signal AI | Established mixed-signal manufacturer; AI work largely research-stage |
| d-Matrix | Digital CIM in SRAM | Generative-AI inference | Corsair platform; digital-only CIM[^33] |
| Cerebras Systems | Wafer-scale digital | Training and large-batch inference | WSE-3 takes the opposite "go very large" position[^34] |
| Graphcore | Massively parallel digital IPU | Training | Acquired by SoftBank in 2024 for approximately $500M[^35] |
| Tenstorrent | Multi-core RISC-V plus Tensix matrix engines | Open AI hardware stack | Backed by Hyundai and Samsung; raised over $1B[^36] |
The competitive dynamics most directly relevant to Rain are the analog-CIM camp (Mythic, EnCharge, Sagence and others), which shares Rain's "memory is the compute" thesis but has not yet captured significant commercial volume; and the digital-CIM camp (d-Matrix, Axelera, and Rain itself in its current form), which dilutes the per-operation energy advantage in exchange for digital determinism and easier compilation.[^31][^32][^33] In parallel, the wafer-scale and dataflow camps (Cerebras, Tenstorrent, SambaNova, Groq) have demonstrated production hardware at scale but rely on conventional digital arithmetic and bank on different sources of leverage (interconnect, on-die memory, fabrication scale).[^34][^36]
Rain's funding ($40 to 50 million disclosed) is one to two orders of magnitude smaller than that of Cerebras ($718 million pre-IPO), Tenstorrent (over $1 billion), or SambaNova.[^11][^34][^36] In the inference niche the recently-funded competitors Mythic ($125 million in late 2025) and EnCharge AI also occupy more capitalized positions than Rain.[^31][^32]
Outside the semiconductor press, Rain AI is principally remembered as a node in the network of Sam Altman's personal AI investments, and specifically as the supplier in the 2019 non-binding letter of intent that surfaced in WIRED's December 2023 reporting following the OpenAI board crisis.[^6][^7] Three threads make the episode of broader interest. First, it illustrates the structural ambiguity around an OpenAI chief executive who simultaneously serves as a private investor in companies whose customers include OpenAI; the LOI was non-binding and never converted, but its disclosure became part of the conversation about Altman's "web of personal investments," including stakes in Helion, Worldcoin, and Reddit.[^6][^7]
Second, the CFIUS-driven divestment of Prosperity7 Ventures from Rain in 2023 was an early high-profile application of foreign-investment-review machinery to a private AI hardware startup, foreshadowing tighter U.S. controls over Gulf-region capital flowing into sensitive U.S. AI assets.[^17][^18][^19] The same regulatory thread extends through August 2023 chip-export restrictions on Nvidia and AMD shipments to certain Middle Eastern countries.[^18]
Third, the failure of Rain's 2024 to 2025 Series B and the company's slide into bridge-funding territory, even with sustained Altman backing and a high-profile hardware-engineering hire from Apple, illustrates how difficult it is for capital-intensive accelerator startups with under fifty million dollars in committed funding to traverse the gap between research-stage silicon and production deployment, particularly against competitors with billions of dollars of runway.[^8][^9]
Rain's published technical claims have been recurring sources of skepticism in the semiconductor trade press.[^14][^29] Energy-efficiency comparisons of "1000x" against Nvidia H100 or AMD Instinct MI300X reference architectures have not been substantiated through standard benchmarks such as MLPerf, and depend on assumptions about target workload, batch size, and silicon node that are not always disclosed.[^8][^29] The shift from end-to-end analog to digital compute-in-memory between 2022 and 2024 was unaccompanied by public explanation of why the analog thesis was deprioritized, although industry analysts have observed that Mythic's commercial difficulties and the maturation of digital CIM made the analog path increasingly difficult to justify on a venture timeline.[^31][^32][^33]
On the business side, post-mortem coverage of the failed 2024 Series B has emphasized that Rain's leadership transition (Wilson to Passo to Kendall in roughly eighteen months), the CFIUS-driven loss of a strategic Saudi LP, the absence of disclosed paying customers, and Rain's relatively small team had combined to make a one-hundred-fifty-million-dollar institutional round difficult to assemble even with vocal Altman support.[^8][^9] Reporting in May 2025 framed Rain not primarily as a technology failure but as a sales-and-go-to-market failure: "talented engineers but lacked the sales acumen to close enterprise deals," in one source's characterization.[^8] As of mid-2025 the outcome of the acquisition process and the long-term fate of Rain's intellectual property remained open.[^8][^9]