invrs.io
invrs.io is a set of open-source tools for inverse design, a method in which an engineer states the performance a part must have and an algorithm computes the shape that gives it, instead of drawing the shape first. The field here is photonics, the engineering of components that guide and shape light: lenses, color filters, waveguides for AR glasses, micro-LEDs. The main package, invrs-gym, is a collection of standard design problems with one software interface, so that a new AI method can be compared with earlier ones on the same problems. A public leaderboard holds the best known solutions, and a simulator called fmmax computes how light moves through a candidate structure.
Martin Schubert founded invrs.io LLC in 2023 after about seven years at Alphabet's X lab, where he started an AI-guided photonics design project in 2017, and a period at Meta Reality Labs. The first invrs-gym release appeared on PyPI on 26 October 2023, and 40 versions followed by October 2024. He published the paper that describes the gym on arXiv on 31 October 2024 as the only author. The company never sold anything: every package is free under the MIT license, and the Solo Founders newsletter wrote in February 2026 that it was not venture-backed.
In October 2025 Apple told the European Commission that it would acquire certain assets of invrs.io LLC and hire its sole equityholder and employee. The Commission published the notice in February 2026, and MacRumors, 9to5Mac and heise reported it. The buyer disclosed no price, and no source reports any revenue, so the money figure on this page is zero, as the facts table says. The evidence for the sale itself is grade B: a regulator's published list of the buyer's own filing, read by three news outlets.
The numbers, graded
Each row shows the evidence grade and the source numbers from the list at the end of this page. The grades are explained in How Aiolum verifies.
| Fact | Value | Grade | How we know | Sources |
|---|---|---|---|---|
| Outcome, February 2026 | Apple acquired certain assets of invrs.io LLC and hired its only shareholder and employee | B | Apple's notification to the European Commission, as quoted by MacRumors, 9to5Mac and heise online. The Commission's list is the acquirer's own disclosure; this page reads it through the three articles. | [1, 2, 3, 4] |
| Sale price | not disclosed | B | No article and no filing gives a price. The money figure on this page is 0 because no public money figure of any kind exists: no price, no revenue, no funding. | [1, 2, 3] |
| Revenue | none reported | C | Every package is free MIT-licensed code on GitHub and PyPI. No source reports a paid product, a customer or a revenue figure. | [7, 8, 10] |
| Date Apple notified the Commission | October 2025 | B | MacRumors: Apple notified the Commission in October 2025; the Commission published the information after a four-month waiting period. The GitHub repositories were last updated between 20 and 23 October 2025. | [1, 5, 7] |
| Date the notice became public | week of 24 February 2026 | B | 9to5Mac on 24 February 2026, MacRumors and heise online on 25 February 2026. | [1, 2, 3] |
| Team at the sale | 1 person | B | Apple's filing describes the founder as the sole equityholder and employee of invrs.io LLC. | [1, 2, 3] |
| Company founded | 2023 | C | The founder's LinkedIn profile, as quoted by heise online and MacTech. The LLC's registration date was not found. | [3, 5] |
| First public release | invrs-gym 0.1.0 on 26 October 2023 | B | PyPI release history, read in the Wayback copy of 18 October 2024. A 0.0.0 placeholder precedes it. | [10] |
| Releases by 14 October 2024 | 40 versions of invrs-gym, up to 1.4.3 | B | Counted from the PyPI release history. Version 1.0.0 came on 5 July 2024. | [10] |
| Paper | arXiv 2410.24132, submitted 31 October 2024, one author | B | Title: invrs-gym: a toolkit for nanophotonic inverse design research. Author: Martin F. Schubert. | [11] |
| Code, October 2026 | 12 repositories; gym 78 stars, fmmax 21 stars | B | GitHub organization page and repository pages read on 23 October 2026. The code is still public after the sale. | [7, 8, 9] |
| Origin of the simulator fmmax | written and open-sourced at Meta, then forked | C | The fmmax README says it was originally developed and open-sourced at Meta and forked after the primary author left Meta, with significant improvements since. | [9] |
| Founder's background | Micron; 7 years at Alphabet's X, AI photonics design project from 2017; Meta Reality Labs | C | LinkedIn profile as quoted by heise online. MacTech quotes the same profile: about 15 years in the semiconductor industry, more than 6 years leading inverse design programs at Google and Meta, nearly 100 issued patents and 40 peer-reviewed papers. | [3, 5] |
| Outside funding | none reported | C | The Solo Founders newsletter, 27 February 2026: he did not build a venture-backed startup. No funding record was found. | [12] |
| Time from first release to the public notice | 28 months | calc | Calculated: October 2023 to February 2026. From the first release to Apple's notification in October 2025 it is 24 months. | [1, 10] |
How to read this table: a grade A value is the buyer's disclosure to a regulator; grade B is that disclosure as read and quoted by journalists, or a platform record (PyPI, GitHub, arXiv); grade C is a founder or company statement, including a LinkedIn profile as quoted by press. A "calc" value was computed on this page from the rows above it.
| Plan | Per month, billed yearly (US$) | Per month, billed monthly (US$) | Limit |
|---|---|---|---|
| invrs-gym, design challenges | 0 | 0 | MIT license, pip install invrs-gym |
| invrs-opt, optimization algorithms | 0 | 0 | MIT license |
| fmmax, Fourier modal method simulator on JAX | 0 | 0 | MIT license, pip install fmmax |
| leaderboard, dataset of solutions | 0 | 0 | public dataset and website |
The company had no pricing page and no paid plan in any source. The table lists the main packages and their license. The GitHub pages show 12 repositories, each marked MIT where a license is shown. A reader who wants to use them pays only for their own computer or GPU time.
The timeline
From the first package on PyPI to a published acquisition notice in 28 months. Each entry carries its grade.
- 2017CSchubert starts and leads an AI-guided photonics design project at Alphabet's X research lab, where he stays about seven years in total. [3]
- 2022 to 2023CAt Meta Reality Labs he works on inverse design of micro-LEDs for AR glasses. The simulator fmmax is written and open-sourced at Meta. [3, 9]
- 2023CHe founds invrs.io LLC after leaving Meta. The company's stated aim: AI-guided design, first for optics, with an ecosystem open to AI scientists, optimization researchers and optics designers. [3, 5, 7]
- 26 October 2023Binvrs-gym 0.1.0 appears on PyPI: optics design challenges with one common interface. [10]
- 5 July 2024Binvrs-gym 1.0.0. By 14 October 2024 there are 40 released versions. [10]
- 31 October 2024BThe paper describing the gym is submitted to arXiv with one author. [11]
- October 2025BApple notifies the European Commission that it will acquire certain assets of invrs.io LLC and hire its sole equityholder and employee. The GitHub repositories receive their last updates between 20 and 23 October. [1, 7]
- 24 and 25 February 2026BThe Commission's list becomes public after a four-month waiting period. MacRumors spots it; 9to5Mac, heise online, MacTech and Open Source For U report it. No price is given. [1, 2, 3, 5, 6]
- 27 February 2026CThe Solo Founders newsletter lists the sale as an exit by a founder who never raised venture money. [12]
Why it worked
Four causes, in the order they mattered. Each one is a reusable lesson.
A narrow field with few experts
Inverse design of photonic structures is a small specialty inside optics, and the founder had led it at two of the largest companies that need it. The press accounts describe that background as the reason for the deal: Apple bought the assets and hired him, which is what buyers do when the value is a person's knowledge rather than a product's revenue.
The benchmark as the product
invrs-gym is not a design tool for end users; it is a shared set of problems and a scoring rule, with a leaderboard of the best solutions. Whoever defines the benchmark in a field becomes the reference point for everyone who publishes in it. The founder's own paper, with one author, describes the gym and its baselines.
Open code, in public, under a permissive license
Every package was MIT-licensed on GitHub and installable from PyPI, with 40 releases in the first year. That made the work visible to the small number of people who could judge it, with no sales effort. The newsletter that listed the exit summarized it as: keep putting things out there until someone on the same frequency finds you.
A regulator made the sale public
Under the EU's Digital Markets Act, Apple must report every acquisition to the European Commission, which publishes a list after a waiting period. That is why a one-person sale with no press release became known at all, and why the dates on this page come from a regulator rather than from the founder.
How it was built
These details come from the GitHub pages, the PyPI release history and the arXiv paper. The platform records are grade B; the README statements are grade C. [7, 8, 9, 10, 11]
- Python packages built on JAX, Google's library for array computing with automatic differentiation and GPU execution, so that gradients of an optical simulation can drive an optimizer.
- invrs-gym: a collection of design challenges, each a Challenge object with a Component, a loss function and evaluation metrics, so one code can run many problems. Challenges come from published papers, including Google's Ceviche challenges and problems from the founder's own earlier work.
- fmmax: a Fourier modal method solver on JAX, originally written and open-sourced at Meta, forked and improved after the author left.
- invrs-opt: optimization algorithms with a common interface, including L-BFGS-B. A leaderboard repository holds a dataset of user-submitted solutions with performance and manufacturability scores.
- Distribution: 12 public repositories under the MIT license, pip-installable packages on PyPI with 40 invrs-gym releases by October 2024, and a single-author paper on arXiv.
- No website content was found: invrs.io does not resolve in October 2026 and the Wayback Machine holds no copy, so the GitHub organization page is the company's only surviving self-description.
The dead that came before
Five companies in the Loot Drop database built developer or engineering tools that engineers liked and could not turn into a business. The values below are from that database and are not independently verified.
Mage
The promiseAn open-source data pipeline framework, friendlier than Airflow.
Why it diedThousands of GitHub stars, but few teams paid for hosting, and larger funded competitors took the category.
WunderGraph
The promiseAn open-source framework that joins APIs and databases into one typed interface.
Why it diedEnterprise sales were too slow for the runway, and cloud platforms added the same features.
Pathmind
The promiseAI-driven simulation to optimize engineering and business processes.
Why it diedCustomers needed long education before they could use it, and large AI providers overshadowed the niche.
numericcal, inc.
The promiseAdvanced analytics for engineering product design.
Why it diedCould not stand out against Autodesk and Ansys, and did not move to AI-driven design in time.
Fig DevTool
The promiseAutocomplete and visual command builders for the terminal.
Why it diedDevelopers used it daily and would not pay; AWS bought the team and technology instead.
What changed
All five raised venture money and then had to turn developer goodwill into recurring revenue on a schedule set by their runway. Open code brought users; it did not bring budgets. Fig ended as invrs.io did, with a large company hiring the team, but after US$2.4 million of investors' money and a failed attempt to sell subscriptions.
invrs.io never tried to sell. It had one person, no reported funding and no price list, so there was no runway to run out of. The output was a benchmark, a simulator and a paper in a field where the possible buyers are a handful of companies that build optical hardware. One of them bought the assets and hired the author. That is not a repeatable business model; it is what a published body of work in a scarce specialty can be worth to a company that needs it.
Can one person copy it now?
The structure, for specialists; not as a plan for revenue.
The available part is the method: pick a narrow technical field you already know from the inside, put a shared benchmark and the tools around it into the open under a permissive license, release often, and write the paper. The non-available part is the field itself and the founder's record at Alphabet and Meta, which is what a buyer pays for in a deal with no revenue.
Where it is still open: Any engineering specialty where simulation is slow and expert designers are few: antenna and radio-frequency design, battery materials, acoustic and thermal components, chip packaging.
- 1Choose a specialty where the number of people who can judge your work is small and the companies that need it are large.
- 2Publish the benchmark, not only a tool: a set of standard problems with one interface and a leaderboard makes your work the reference point.
- 3Release on the public package index often, under the MIT license, so that use costs nothing and needs no sales call.
- 4Write the paper as the single author, so the work is tied to one name that a buyer can hire.
What to watch
The risks are the mirror image of the causes above.
No money figure at all
The buyer disclosed no price, no source reports revenue, and no funding is recorded. This page shows a money outcome of zero. The sale is real, the amount is unknown, and it may have been small.
An acquisition of assets and a hire, not a product sale
Apple's filing says it acquired certain assets and hired the sole equityholder and employee. The press calls it a micro-acquisition or an acqui-hire. The value to the buyer appears to be the person and the code, not a customer base.
The maintainer is gone from the public code
The GitHub repositories were last updated in October 2025, the month Apple filed its notice. The packages stay public under the MIT license, but no source says who will maintain them.
The founder has said nothing in public
No statement by Schubert about the sale was found: not a post, not an interview. The founding year and the career history come from a LinkedIn profile as quoted by press, and the newsletter item is a secondary summary of the same articles.
Sources
The grade next to each source is the best grade that the source can support. "Read in full" means that the text of the page was downloaded and the numbers were checked in it.
- B MacRumors, 25 February 2026: EU Reveals Apple Made a Tiny Acquisition Last YearRead in full. Joe Rossignol's report on the Commission's list. Source for the October 2025 notification, the four-month waiting period, the quoted notice text and the 2023 founding year from LinkedIn.
- B 9to5Mac, 24 February 2026: Apple acquires startup specializing in AI-powered light and optics design toolsRead in full. Marcus Mendes's report, crediting MacRumors. Source for the filing's words: acquire certain assets from and hire the sole equityholder and employee of invrs.io LLC.
- B heise online, 25 February 2026: Apple acquires photonics startup invrs.ioRead in full. An English translation of the German article, editorially reviewed. Source for the founder's career: Micron, seven years at Alphabet's X with an AI photonics design project from 2017, Meta Reality Labs and micro-LED inverse design.
- A European Commission, Digital Markets Act case search: acquisitions notified by gatekeepersCould not read: the page loads its list through a script and an API key. The notice text is cited from the quotations in MacRumors, 9to5Mac, heise online and MacTech, which link to this page.
- C MacTech, 24 February 2026: Apple acquiring invrs.io that develops open-source frameworks for photonics researchRead in full. A summary of MacRumors and 9to5Mac that quotes the founder's LinkedIn profile: founded 2023, 15 years in the industry, 6 years leading inverse design programs, nearly 100 patents, 40 papers.
- C Open Source For U, 25 February 2026: Open Source Photonics Frameworks Drive Apple Acquisition Of invrs.ioRead in full. A secondary report that calls the deal a micro-acquisition and names 9to5Mac as the first reporter.
- B GitHub organization page invrs-io, read 23 October 2026Read in full. The company's self-description, the list of 12 repositories, their MIT licenses and their last-update dates in October 2025.
- B GitHub repository invrs-io/gym, read 23 October 2026Read the README. Source for the package design, the list of challenges and the 78 stars.
- C GitHub repository invrs-io/fmmax, read 23 October 2026Read the README. Source for the statement that fmmax was developed and open-sourced at Meta and forked after the primary author left, and for the 21 stars.
- B PyPI project page for invrs-gym, Wayback Machine copy of 18 October 2024Read in full. The live PyPI page returns a script challenge. Source for the release history from 0.1.0 on 26 October 2023 to 1.4.3 on 14 October 2024, the author field and the MIT license.
- B arXiv 2410.24132: invrs-gym: a toolkit for nanophotonic inverse design researchRead the abstract page. Submitted 31 October 2024, single author Martin F. Schubert.
- C Solo Founders newsletter, 27 February 2026: Solo founder raises $50M, an Apple acquisition, and a new solo launchRead the invrs.io item. A secondary summary: not venture-backed, open-source tools shared on GitHub, Apple acquired the company to bring him in-house, nearly 100 patents and 40 papers. It links to 9to5Mac.
- C Loot Drop database export, failed startupsLocal export. AI-written summaries, not independently verified. Source for the failed predecessors.