COMPANY PROFILE / MATERIALS INFORMATICS
Radical AI
AI proposes alloys, and its own automated lab makes and measures them: a start-up built around "testing by experiment"
Radical AI is a privately held company in New York whose core business is to use AI to narrow down candidate materials, then make and measure those candidates in a robotic laboratory. Its website reports what happened when AI predictions met experiments, both the successes and the misses. This profile relies only on the company's own announcements, statements by U.S. federal and state authorities, and public paper, model and code pages, and it keeps the company's claims apart from what third parties or papers can confirm.
- Radical AI in 30 seconds, and where it stands in materials informatics
- The logo and how the name is written
- Where the company sits in the materials informatics field
- Scale
- Resources committed to materials informatics
- What it has done so far
- What comes next
- What the company is aiming for in materials informatics
- Map of materials informatics partnerships
- Technologies, products and services
- The risks this company carries
- Glossary, references and claim-to-source audit
1. Radical AI in 30 seconds, and where it stands in materials informatics
MI software & materials AI (secondary)Our own classification Our calculation
(e.g. high-entropy alloys)The State release says the lab focuses on inorganic materials
MATRIX / LitXBenchCode and data released under MIT and similar licences
Materials informatics (MI) is a way of searching for new materials by gathering data, using AI to predict and design candidates, and testing them by experiment. What sets this company apart from other MI companies is that it owns the last step, "testing by experiment", in its own automated lab, and feeds the experimental results back into the next round of predictions; this loop is the centre of its business Sourced. The company writes that simulation alone is not enough and that proven, characterised samples are what matter. It also admits that the lab can process only so many samples a day, far fewer than the candidates AI can propose. This profile therefore concentrates on how far predictions have actually been tested by experiment.
2. The logo and how the name is written
A note on names: the legal entity is Radical AI, Inc. The logo in the website header is the word "RADICAL" with a small square dot; it does not contain the letters "AI" (the image here is the logo graphic taken as it is from the official page's embedded vector file). To avoid confusion with other companies and institutes with similar names, this article writes "Radical AI" throughout.
Official site: https://www.radical-ai.com/3. Where the company sits in the materials informatics field
First, the figure sorts MI companies into four types and shows where Radical AI sits. A "foundation model" is an AI model trained in advance on a very large amount of data so that it can be reused for many purposes.
Next, the MI workflow (data, then prediction and design, then testing by experiment, then feeding results back) and which parts Radical AI covers, based on the company's own account.
4. Scale
The company is privately held, so there are no statutory financial filings. Only what can be confirmed is shown Sourced.
| Item | Detail | Source and type |
|---|---|---|
| Founded | 2024 | State of New York release (27 January 2026). A founding date could not be found on the company's own website |
| Headquarters | New York City. A new headquarters and research site is planned in Building 20 at the Brooklyn Navy Yard | Same release. It is being fitted out; no completion or opening date could be confirmed |
| Employees | About 34 (scientists, engineers and technologists). The company expects to create 115 jobs at the new site | Same release. The 115 is a plan Not yet confirmed |
| Revenue | Not disclosed | Not stated in any official material. This article does not estimate it |
| Funding raised | Seed round of US$55 million (announced 18 July 2025) | Company announcement. No valuation is given in official material and this article does not discuss one |
| Listing | Privately held | No share information on the official website |
The headcount is about 34 as of January 2026; later changes have not been published. Investor names (RTX Ventures, NVentures [NVIDIA's venture arm] and others) appear only in the "About Radical AI" paragraph at the end of the State of New York release. The company's own seed announcement of 18 July 2025 does not name investors. The boilerplate paragraph looks like text supplied by the company, but that could not be confirmed Not yet confirmed.
5. Resources committed to materials informatics
The four resource questions are people, money, outside investments, acquisitions and joint ventures, and ties to universities and public bodies. Radical AI is a company specialised in materials AI, so company-wide figures are also the MI figures.
| Item | What could be confirmed | Status |
|---|---|---|
| People | About 34 (January 2026). The company's blog says software engineers, machine-learning specialists and mechanical designers work in the same team as materials researchers. The company expects 115 jobs at the new site (plan) | Sourced (the plan is Not yet confirmed) |
| Money (raised) | US$55 million seed round (July 2025). A U.S. Air Force AFWERX contract of US$1,197,902 for developing high-entropy alloys (Direct-to-Phase II, August 2025) | Sourced |
| Money (facilities) | The new lab is a US$4 million project, expected to receive up to US$2 million in New York State Excelsior Jobs Program tax credits (performance-based). Investment in computing is not disclosed | Sourced (the tax credits are Not yet confirmed) |
| Outside investments, acquisitions, joint ventures, divestments | None stated in official material | Could not be confirmed |
| Universities, research institutes and public funding | A memorandum of understanding with the U.S. Department of Energy (DOE, December 2025); joint research with Sandia National Laboratories on hydrogen-separation materials (September 2026); tests by the Purdue Applied Research Institute (August 2026); the U.S. Air Force AFWERX contract | Sourced |
Status of the main models, software and facilities
The table lists name, type, specification, form of release, stage and date, as far as the company's official material and public pages allow. Numbers measured under different conditions are not ranked against each other.
| Name | Type | Main specifications (including company claims) | Release and licence | Stage and date |
|---|---|---|---|---|
| TorchSim | Atomistic simulation engine (built on PyTorch) | Runs molecular dynamics (NVE, NVT, NPT) and structure optimisation on many systems at once. The company claims "100x over ASE and 100,000,000x over DFT" (no conditions given in the blog post) | Open source, MIT. The repository now sits under a separate GitHub organisation called TorchSim | Released. 4 April 2025 (paper August 2025) |
| EGIP | Machine-learning interatomic potential (first generation) | A pruned GemNet-style architecture that predicts forces directly. The company claims the best result on the thermal-conductivity task of Matbench Discovery, and 3.5x the speed and 12x the number of atoms of eSEN | Weights and code: could not be confirmed as released | Announcement only. 24 April 2025 |
| MATRIX / MATRIX-PT | Benchmark of materials-science reasoning / adapter for further training of a vision-language model | Data: 11,331 training, 1,688 validation and 470 test items. The adapter is a LoRA for Qwen2-VL-7B. Paper's claim: post-training with images improves interpretation of experiments by 10-25% and text-only reasoning by 5-16% | Data: MIT. Adapter: Apache-2.0 (Hugging Face) | Released. 30 January 2026 (paper) |
| LitXBench / LitXAlloy | Benchmark for extracting experiments from papers | 1,426 measured values from 19 alloy papers, organised by hand. Paper's claim: frontier language models beat existing multi-step extraction pipelines by up to 0.37 (F1) | Code: MIT (GitHub), also on PyPI | Released. April 2026 |
| Antimatter | In-house materials AI engine | Company description: reads hundreds of thousands of papers and screens billions of compositions. Specifications and evaluations: could not be confirmed | Not public (no external offering stated) | In use (company description). 20 October 2026 |
| Automated lab (New York) | Lab that links synthesis, measurement and analysis | Company description: over about nine months it collected more than 10,000 SEM images and measured hardness on more than 500 samples. The new lab is planned to run about 100 experiments a day | Operated in house | In operation (as described in June 2026). The new lab is being fitted out |
A figure such as TorchSim's "100,000,000x over DFT" means different things depending on what it is compared with and under what conditions (number of atoms, which machine-learning potential, what accuracy is required). DFT (density functional theory) is a precise method for calculating a material's properties from the state of its electrons; it is trusted but very expensive to run. A machine-learning potential is a light stand-in trained on such results. Being faster is not the same as predicting real materials more accurately, so this article treats the multiple as the company's claim and does not use it to rank anyone Not yet confirmed.
6. What it has done so far
| When | What happened | Meaning |
|---|---|---|
| 2024 | Founded (State of New York release) | No detailed founding date found |
| 10 January 2025 | "Announcing Radical AI" appears in the news list on the company site | The item links to a social-media post, so its text was not checked |
| 4 April 2025 | Released TorchSim as open source | Part of its prediction tooling made public; the company says it is meant to be used together with its automated lab |
| 24 April 2025 | Announced EGIP, a machine-learning potential | Claims speed and accuracy together on tasks such as thermal conductivity |
| 29 May 2025 | Sent comments to the White House Office of Science and Technology Policy (OSTP) on the national AI R&D strategic plan | Proposed AI-driven materials work, a national data infrastructure and pilot manufacturing facilities |
| 18 July 2025 | Announced a US$55 million seed round | Investor names appear only in the State release's company paragraph |
| August 2025 | Published the TorchSim paper (8 August). Won a US$1,197,902 contract from the U.S. Air Force AFWERX (25 August) | The aim is high-entropy alloys for hypersonic flight |
| 20 December 2025 | Memorandum of understanding with the U.S. Department of Energy | A framework for discussing specific projects; not a contract and not funding |
| 27 January 2026 | The State of New York announced the plan for the new lab | US$4 million project, about 100 experiments a day, 115 jobs (all plans) |
| January to April 2026 | Released MATRIX (paper 30 January), LitXBench (paper 8 April) and others | Public evaluation data for materials AI and for literature ingestion |
| 9 June 2026 | Described automatic SEM image analysis (more than 10,000 images in about nine months) | Automates measuring microstructure, to be used to set annealing times |
| 7 August 2026 | Announced comparison results for alloy "RAI-939" | Tested by the Purdue Applied Research Institute. Chapter 10 covers it in detail |
| 26 August 2026 | Announced a sample-cutting fixture co-developed with Makino | Cuts five samples at once |
| 23 September 2026 | Reported six predictions from an electronic-property model and the experimental results | All six met the target; a second model fell short |
| 29 September 2026 | Announced joint research with Sandia National Laboratories on hydrogen-purification materials | Results are still to come |
| 20 October 2026 | The company announced that TIME named its lab one of the Best Inventions of 2026 | A company announcement; the basis for the selection was not checked in the magazine |
7. What comes next
The State release also says the lab's high-throughput approach will speed up traditional materials discovery by "370x", but it does not say what that figure is measured against. This article treats it as a statement by the company and the State Not yet confirmed.
None of the following is stated in official material Not yet confirmed.
The new lab's start date; the current lab's daily throughput; customer contracts and revenue; composition and processing details of RAI-939; when results of the AFWERX contract will be published; specific projects under the DOE memorandum; when any AI-proposed material might reach a product. The company says it will "sell materials to the industries that matter" (TorchSim announcement), but no sales could be confirmed.
8. What the company is aiming for in materials informatics
This chapter includes things the company has not said directly. They are inferences from primary sources, set out with the evidence.
| What can be read from it | Primary evidence | Type |
|---|---|---|
| It wants to develop and supply materials themselves as its revenue base, rather than sell software | The TorchSim announcement says it will sell materials to industry, and the State release says it will develop, test and commercialise new materials. No software pricing or licensing is mentioned | Inference |
| For now it concentrates on inorganic materials that are quick to make and measure, such as refractory and high-entropy alloys | The State release says the lab focuses on inorganic materials; the AFWERX contract is for high-entropy alloys; the RAI-939 and electronic-property examples are also inorganic. The process described is making button-sized samples and cutting them | Inference |
| Its entry points are government agencies and demand from defence, aerospace and semiconductors | The AFWERX contract, the DOE memorandum, the Sandia collaboration and the OSTP submission. A defence-related firm is among the seed investors (State release) | Inference |
| It treats its own experimental data as its source of advantage | It releases computing tools (TorchSim) and evaluation data (MATRIX), but describes its training data as proprietary and does not release Antimatter | Inference |
| It is willing to publish cases where AI predictions missed | The September 2026 post on the electronic-property model reports misses (samples too brittle to cut, a measured value below the prediction) | Sourced |
The reading that "for now it concentrates on high-melting-point alloys" rests on how the experiments are set up. According to the June 2026 post, the lab heat-treats alloy samples at up to 1,500 °C to mix the elements and then images their microstructure. The August 2026 post describes cutting button-sized samples with an electrical discharge machine to measure strength and resistance to oxidation Sourced. "Melt, solidify, cut and measure" lets equipment be shared more easily than, say, drug synthesis or thin-film fabrication, and the same flow can be run again and again. From this, this article reads the company as polishing the speed of its loop first on metals and alloys Not yet confirmed.
9. Map of materials informatics partnerships
10. Technologies, products and services
Prices and contract terms for anything Radical AI offers to outside customers are not disclosed Not yet confirmed. What has been published is software and data released openly, and test results for materials made in its own automated lab. For background, see our explainers on machine-learning interatomic potentials, self-driving laboratories, closed-loop optimisation, and Bayesian optimisation and active learning.
(1) A materials test result: alloy "RAI-939"
| Item | What the company says | Type |
|---|---|---|
| Target | A new alloy aimed at replacing "C103", an existing aerospace alloy. The company says C103 has been used since the Apollo era, degrades quickly at high temperature, needs coatings and frequent replacement, and has long lead times | Company's claim |
| How it was made | Over 16 weeks, AI-generated hypotheses and the autonomous lab produced more than 300 novel materials. The company says only 3,493 unique alloys had been studied experimentally in this space over 20 years or more | Company's claim |
| The test | The Purdue Applied Research Institute compared the two in a torch test that simulates flight environments and can heat materials beyond 2,500 °C. After one minute RAI-939 gained 0.25% in mass while C103 lost 32%. The company calls this a "125x" increase in environmental survivability | Company's announcement (with a quote from the test lab's director) |
| Conditions | Specimen size, atmosphere, temperature history, composition and processing are not given in the post. How the "125x" was calculated is not stated either | Could not be confirmed |
| Next step | Make larger test pieces and gather quality-control data closer to production methods | Not yet confirmed |
This is an announcement on the company's blog; no peer-reviewed paper could be found. "Mass gained" and "mass lost" point in different directions: oxidation building up on the surface versus material being lost. But a single one-minute comparison with undisclosed conditions says nothing about service life or reliability in actual use. Until independent repeat tests or a detailed third-party report appear, it is best read as a promising early result.
(2) The electronic-property model: six predictions, six materials on target
The September 2026 post is notable because it reports both success and failure Sourced.
| Item | What the company writes |
|---|---|
| The model that worked | In about seven weeks the team built the model, screened millions of materials, made the samples and had them measured by a third party. Of six materials the model proposed, all six met the electronic target. One contained an unexpected element, and the company itself suspected the model was wrong |
| The model that did not | A second electronic-property model did not reach the same success. Two of the targeted samples were too brittle to be cut to the size needed for measurement. In one case the measured value was below the prediction, and the company says the model needs work |
| Measurement | Samples were polished and cleaved, then measured with an atomic force microscope |
| What is not published | The type and value of the target property and the compositions are not stated |
(3) Software and data released openly
TorchSim is software that lets GPUs run many atomistic simulations, such as molecular dynamics (following how atoms move in a computer), at the same time. MATRIX is a question set that tests how well AI can answer materials-science questions, including questions that require reading microscope images and X-ray diffraction plots. LitXBench is a question set that tests how accurately a machine can pull experimental conditions and results out of papers. The licences for all three were checked on their public pages (the table in chapter 5).
Radical AI's posts offer three points a materials engineer can take away Sourced.
(1) It tries to compute quantities that experiments can measure. The March 2026 post says it computes an alloy's melting point, ductility and yield stress with structure-aware calculations rather than the rule of mixtures (a simple approximation that averages properties by composition). It does not assume that elements sit randomly; it takes into account "short-range ordering", the tendency of certain elements to prefer certain neighbours. Hardness is easy to measure and correlates with yield strength, so the company says it has measured more than 500 samples and uses them to tune its model.
(2) It cuts the effort of measurement by automating it. Measuring the spacing of dendrite arms (tree-like crystals) in a microscope image takes a specialist about 15 minutes per image, and the company says two people doing the same task can differ by ±16%. It says automation takes about a second per image. That spacing is a guide to how long a sample needs to be annealed (heat-treated).
(3) It admits the gap between the number of computed and tested candidates. Computation proposes far more candidates than the lab can make. The company calls this the "next problem" and says experiments that disagree with simulation will also be used for learning. From the materials side, how the certainty (uncertainty) of a prediction is handled is the key to making this approach practical.
11. The risks this company carries
| Risk | Detail | Type |
|---|---|---|
| Little outside verification of results | The RAI-939 result is a blog announcement; no peer-reviewed paper, test conditions or composition are public. For the electronic-property example, the content of the target is not public | Sourced |
| Lab throughput | Far fewer candidates can be tested than AI can propose (the company says so itself). The new lab's roughly 100 runs a day is a planned figure and its start date is unconfirmed | Sourced (the plan is Not yet confirmed) |
| Computed predictions can miss | The company acknowledges that another model fell short in the September 2026 post and that one measured value came in below prediction | Sourced |
| Distance to practical use | The company says it will sell materials, but no customers, revenue or volume production could be confirmed. Between a small test piece and a usable part lie scale-up and quality control | Inference |
| Reliance on public funding and defence demand | The disclosed contracts and collaborations are mostly with government bodies; no private customers could be confirmed | Inference |
| Handling of the open-source project | TorchSim has moved to a separate GitHub organisation, and the company's current relationship to it could not be confirmed in official material | Not yet confirmed |
| Dramatic multiples | Multiples such as "125x", "100,000,000x" and "370x" either lack a stated baseline and conditions or are the company's own claims. This article does not use them to rank anyone | Sourced |
12. Glossary
- Materials informatics (MI)
- Searching for new materials efficiently with data and AI, in the flow of data, then prediction and design, then testing by experiment.
- Self-driving laboratory (SDL)
- A lab in which AI chooses the next experiment, robots and automated tools make and measure the material, and the results go back to the AI.
- High-entropy alloy
- An alloy that mixes several metallic elements in roughly equal proportions. The number of possible compositions is huge, which makes it a popular target for search.
- Machine-learning interatomic potential (MLIP)
- A fast model of the forces between atoms, learned by machine learning from precise calculations (DFT).
- DFT
- Density functional theory: a precise way to calculate a material's properties from the state of its electrons. Computationally heavy.
- Rule of mixtures
- A simple approximation that estimates an alloy's property by averaging its components by proportion. It ignores the effect of the actual microstructure.
- Benchmark
- A common question set and scoring method for comparing methods under the same conditions. Always check what it actually measures.
- SEM
- Scanning electron microscope: an instrument that images the surface or cross-section microstructure of a sample at high magnification.
- C103
- The name of a high-temperature alloy used in aerospace, taken as the comparison material in the company's announcement. It is generally known as a niobium-based alloy (our own note).
- ASE / GemNet
- ASE is a widely used Python toolkit for atomistic simulation; GemNet is a neural-network design for learning forces between atoms.
- Memorandum of understanding (MoU)
- A document setting out a framework for future cooperation. It does not necessarily carry legal obligations or funding.
- SBIR
- The U.S. Small Business Innovation Research programme, through which government agencies fund research and development at small companies.
13. References
- Radical AI Official website (home page and Vision page): the company's own description, news list and logo. Checked 22 October 2026. https://www.radical-ai.com/
- Radical AI "Radical AI announces $55M in new funding" (18 July 2025): the company's announcement of its US$55 million seed round. https://www.radical-ai.com/news/series-seed
- Radical AI "Introducing TorchSim" (4 April 2025): release of the open-source atomistic simulation engine, with the company's speed claims. https://www.radical-ai.com/news/introducing-torchsim
- Radical AI "EGIP: Another step forward in ML-first materials" (24 April 2025): the first-generation machine-learning interatomic potential. https://www.radical-ai.com/news/EGIP
- Radical AI "Beyond Composition: Computing What Materials Will Actually Do" (9 March 2026): structure-aware property calculation. https://www.radical-ai.com/news/beyond-composition-computing-what-materials-will-actually-do
- Radical AI "Confirming Superior Performance of Materials Discovered in Radical AI Self-Driving Labs" (7 August 2026): the torch test of alloy RAI-939. https://www.radical-ai.com/news/purdue-applied-research-institute-confirms-superior-performance-of-radical-ai-materials
- Radical AI "6 Predictions, 6 Validated Materials" (23 September 2026): six predictions from an electronic-property model and what experiments found. https://www.radical-ai.com/news/6-predictions-6-validated-materials-enabling-frontier-electronic-property-materials
- Radical AI "Automated Microstructure Analysis Moves Different" (9 June 2026): automatic measurement of microstructure from SEM images. https://www.radical-ai.com/news/scale-speed-precision-automated-microstructure-analysis-moves-different
- Radical AI "Radical AI Awarded U.S. Air Force Contract" (25 August 2025): contract value and purpose of the AFWERX award. https://www.radical-ai.com/news/radical-ai-awarded-us-air-force-contract
- Radical AI "Radical AI Signs Memorandum of Understanding with the U.S. Department of Energy" (20 December 2025). https://www.radical-ai.com/news/radical-ai-signs-memorandum-of-understanding-with-the-us-department-of-energy
- Radical AI "Radical AI's Proposal to the White House OSTP" (30 June 2025): comments on the national AI R&D strategic plan. https://www.radical-ai.com/news/RFI
- Radical AI "Sandia National Laboratories Partners With Radical AI" (29 September 2026): announcement of joint research on hydrogen-separation materials. https://www.radical-ai.com/news/sandia-national-laboratories-partners-with-radical-ai-to-discover-new-materials-for-hydrogen
- Radical AI "Breaking the Bottleneck: Radical AI Automated Sample Prep with Makino" (26 August 2026): co-developed sample-cutting fixture. https://www.radical-ai.com/news/breaking-the-bottleneck-radical-ai-automated-sample-preparation-with-makino
- Radical AI "Radical AI's self-driving lab named one of TIME's Best Inventions of 2026" (20 October 2026): the company's own description of how it works (Antimatter). https://www.radical-ai.com/news/radical-ai-self-driving-lab-named-one-of-time-best-inventions-of-2026
- Radical AI "A Benchmark for Extracting Experiments from Literature" (13 April 2026): LitXBench and LitXAlloy, and how much literature the company ingests. https://www.radical-ai.com/news/litxbench-a-benchmark-for-extracting-experiments-from-materials-literature
- Radical AI "Leveraging Experimental Data Beyond Language: A Multimodal Benchmark" (10 February 2026): the MATRIX announcement. https://www.radical-ai.com/news/leveraging-experimental-data-beyond-language-a-multimodal-benchmark
- State of New York "Governor Hochul Celebrates Radical AI Establishing New York's First Fully Autonomous Materials Science Labs at the Brooklyn Navy Yard" (27 January 2026): founding year, headcount, the new-lab plan and the seed investors named in the company boilerplate. https://www.governor.ny.gov/news/governor-hochul-celebrates-radical-ai-establishing-new-yorks-first-fully-autonomous-materials
- Paper (arXiv) "TorchSim: An efficient atomistic simulation engine in PyTorch" (submitted 8 August 2025, arXiv:2508.06628). https://arxiv.org/abs/2508.06628
- GitHub TorchSim/torch-sim (moved from Radical-AI/torch-sim; MIT License). Checked 22 October 2026. https://github.com/TorchSim/torch-sim
- Paper (arXiv) "MATRIX: A Multimodal Benchmark and Post-Training Framework for Materials Science" (submitted 30 January 2026, arXiv:2602.00376). https://arxiv.org/abs/2602.00376
- Hugging Face radical-ai/MATRIX (dataset, MIT License; 11,331 training, 1,688 validation and 470 test items). https://huggingface.co/datasets/radical-ai/MATRIX
- Hugging Face radical-ai/MATRIX-PT (LoRA adapter for Qwen2-VL-7B, Apache-2.0). https://huggingface.co/radical-ai/MATRIX-PT
- Paper (arXiv) "LitXBench: A Benchmark for Extracting Experiments from Scientific Literature" (first submitted 8 April 2026, arXiv:2604.07649). https://arxiv.org/abs/2604.07649
- GitHub Radical-AI/litxbench (MIT License). Checked 22 October 2026. https://github.com/Radical-AI/litxbench
14. Claim-to-source audit
| Claim in the text | Type | Source |
|---|---|---|
| Legal name Radical AI, Inc.; privately held; a New York company | Sourced | Reference 1, 17 https://www.radical-ai.com/ https://www.governor.ny.gov/news/governor-hochul-celebrates-radical-ai-establishing-new-yorks-first-fully-autonomous-materials |
| Founded in 2024; about 34 employees (January 2026); the new-lab plan (US$4 million, about 100 experiments a day, 115 jobs, up to US$2 million in tax credits) | Sourced | Reference 17 https://www.governor.ny.gov/news/governor-hochul-celebrates-radical-ai-establishing-new-yorks-first-fully-autonomous-materials |
| The US$55 million seed round announced on 18 July 2025 | Sourced | Reference 2 https://www.radical-ai.com/news/series-seed |
| Seed investor names (RTX Ventures, NVentures, Noa, Eni, Infinite Capital, AlleyCorp) appear in the company paragraph of the State release | Sourced | Reference 17 Statement in the paragraph; whether the company supplied it could not be confirmed https://www.governor.ny.gov/news/governor-hochul-celebrates-radical-ai-establishing-new-yorks-first-fully-autonomous-materials |
| Release of TorchSim (4 April 2025) and the company's claims of 100x over ASE and 100,000,000x over DFT | Sourced | Reference 3 https://www.radical-ai.com/news/introducing-torchsim |
| TorchSim is open source under the MIT licence and now lives in the TorchSim organisation's repository | Sourced | Reference 19 https://github.com/TorchSim/torch-sim |
| The TorchSim paper (arXiv:2508.06628) was submitted on 8 August 2025 | Sourced | Reference 18 https://arxiv.org/abs/2508.06628 |
| What the EGIP announcement says (Matbench Discovery thermal-conductivity claim; 3.5x speed and 12x atoms versus eSEN) | Sourced | Reference 4 Company's claim https://www.radical-ai.com/news/EGIP |
| Release of EGIP weights or code could not be confirmed | Not yet confirmed | Reference 4 Not stated in the announcement. This article does not assert that no release exists https://www.radical-ai.com/news/EGIP |
| Structure-aware calculation of melting point, ductility and yield stress, short-range ordering, SparseBO, and more than 500 hardness measurements | Sourced | Reference 5 https://www.radical-ai.com/news/beyond-composition-computing-what-materials-will-actually-do |
| The RAI-939 versus C103 torch test (+0.25% / -32% mass), more than 300 novel materials, 16 weeks, 3,493 alloys studied previously | Sourced | Reference 6 Company's claim, with a quote from the test lab's director https://www.radical-ai.com/news/purdue-applied-research-institute-confirms-superior-performance-of-radical-ai-materials |
| Six predictions, all six on target; a second model fell short; some samples too brittle to cut | Sourced | Reference 7 https://www.radical-ai.com/news/6-predictions-6-validated-materials-enabling-frontier-electronic-property-materials |
| More than 10,000 SEM images, manual measurement about 15 minutes and ±16%, automated about one second, heat treatment at 1,500 °C | Sourced | Reference 8 https://www.radical-ai.com/news/scale-speed-precision-automated-microstructure-analysis-moves-different |
| The US$1,197,902 AFWERX contract for high-entropy alloys (25 August 2025) | Sourced | Reference 9 https://www.radical-ai.com/news/radical-ai-awarded-us-air-force-contract |
| The DOE memorandum (20 December 2025) is a framework for discussing specific projects | Sourced | Reference 10 https://www.radical-ai.com/news/radical-ai-signs-memorandum-of-understanding-with-the-us-department-of-energy |
| The OSTP submission (29 May 2025) and its content | Sourced | Reference 11 https://www.radical-ai.com/news/RFI |
| Joint research with Sandia National Laboratories on hydrogen separation (29 September 2026) | Sourced | Reference 12 https://www.radical-ai.com/news/sandia-national-laboratories-partners-with-radical-ai-to-discover-new-materials-for-hydrogen |
| A fixture co-developed with Makino that cuts five samples at once | Sourced | Reference 13 https://www.radical-ai.com/news/breaking-the-bottleneck-radical-ai-automated-sample-preparation-with-makino |
| The Antimatter description (hundreds of thousands of papers, billions of compositions) and the company's TIME announcement | Sourced | Reference 14 Company description and announcement https://www.radical-ai.com/news/radical-ai-self-driving-lab-named-one-of-time-best-inventions-of-2026 |
| More than 5 million pages of papers indexed; 100 agents each reading 100 papers | Sourced | Reference 15 https://www.radical-ai.com/news/litxbench-a-benchmark-for-extracting-experiments-from-materials-literature |
| LitXBench / LitXAlloy (19 papers, 1,426 measurements, MIT) | Sourced | Reference 23, 24 https://arxiv.org/abs/2604.07649 https://github.com/Radical-AI/litxbench |
| MATRIX (11,331 / 1,688 / 470 items, MIT) and MATRIX-PT (Apache-2.0); the 10-25% and 5-16% claims | Sourced | Reference 20, 21, 22 https://arxiv.org/abs/2602.00376 https://huggingface.co/datasets/radical-ai/MATRIX https://huggingface.co/radical-ai/MATRIX-PT |
| The MATRIX announcement and the company's explanation of adding visual data | Sourced | Reference 16 https://www.radical-ai.com/news/leveraging-experimental-data-beyond-language-a-multimodal-benchmark |
| Main type autonomous experiments & robotics, secondary type MI software & materials AI | Our calculation | Our own classification, not the company's self-description (this article) |
| The MI loop illustration (data, prediction and design, experiment, feedback) and the three-questions illustration | Our calculation | Our own framing (illustrations in this article) |
| Revenue, valuation, customers and the lab's current throughput | Not yet confirmed | Reference 1 Not stated in official material. This article does not estimate them https://www.radical-ai.com/ |
| The new lab's start date, RAI-939 composition and processing, time to product | Not yet confirmed | Reference 17, 6 Not stated https://www.governor.ny.gov/news/governor-hochul-celebrates-radical-ai-establishing-new-yorks-first-fully-autonomous-materials https://www.radical-ai.com/news/purdue-applied-research-institute-confirms-superior-performance-of-radical-ai-materials |
| The reading that it wants to develop and supply materials as its revenue base | Inference | Our interpretation of the TorchSim announcement and the State release (this article) |
| The reading that it concentrates on alloys that are quick to make and measure | Inference | Our interpretation of the heat-treatment and cutting steps described (this article) |
| The reading that its entry points are public bodies, defence and semiconductor demand | Inference | Our interpretation of its contracts, memorandum and collaborations (this article) |
| The reading that its own experimental data is its source of advantage | Inference | Our interpretation of what it releases and what it keeps private (this article) |
| The risks of distance to practical use and dependence on public funding | Inference | Our interpretation, as no customers or revenue are disclosed (this article) |
Last updated: 22 October 2026 / Troy Technical
The figures in this article are based on Radical AI's official announcements, the State of New York release, papers, and the public GitHub and Hugging Face pages. No market-research estimates or press-only figures are used. The results quoted (125x, 100,000,000x, six of six and so on) are the company's own claims and have not been confirmed by independent verification. Passages marked "Inference" are this article's reading of primary sources and are not statements by the company.