Eigen AI: Revenue, sale and how it was built
Eigen AI (Faster, cheaper serving and post-training of open-weight AI models, sold as a per-token API, United States): Sold for about US$643M, May 2026. Evidence grade A: disclosed by a company or a filing, or verified by a payment processor. Team: 3 co-founders, first public post in May 2025; '30+ team members' on its leadership page in March 2026; 20 employees at the sale in May 2026 (Bloomberg, as Implicator reported). Launched 28 May 2025; sold 11 months after launch.Added 8 November 2026
- Outcome
- Sold for about US$643M, May 2026
- Evidence
- A disclosed by a company or a filing, or verified by a payment processor
- Launch
- 28 May 2025
- Status
- Sold, 11 months after launch
- Team
- 3 co-founders, first public post in May 2025; '30+ team members' on its leadership page in March 2026; 20 employees at the sale in May 2026 (Bloomberg, as Implicator reported)
- Outside funding
- Venture-backed by Tectonic Ventures, E14 Fund, Uncorrelated Ventures, AGI House Ventures and angel investors, as the founders wrote; amount not disclosed
- Main channel
- Not stated
Eigen AI made open-weight AI models run faster and cost less to serve. An open-weight model is a model whose trained numbers (the weights) are published, so anyone can run it on their own computers. Running a model to answer requests is called inference, and its cost is counted in tokens, the word pieces that a model reads and writes. Eigen used quantization, storing each weight in fewer bits (for example 4 instead of 16) so the model needs less memory and computes faster, together with custom GPU code, and sold the result as an API priced per million tokens. It also sold post-training, extra training of a finished model on a customer's own data and goals.
Three co-founders started it: Ryan Hanrui Wang (chief executive) and Wei-Chen Wang, both from the HAN Lab at the Massachusetts Institute of Technology (MIT), and Di Jin, an MIT PhD who had worked on post-training of Meta's Llama 3 and Llama 4. Its first dated post appeared on 28 May 2025. By November 2025 it listed public prices, for example US$0.10 per million input tokens for GPT-OSS 120B, and its own post reported that the benchmark site Artificial Analysis ranked its API first in output speed on three models. In March 2026 it partnered with Nebius, a Nasdaq-listed AI cloud company, and on 1 May 2026 Nebius agreed to buy it.
Nebius stated in its press release that the price was about US$643 million at signing, paid in cash and Nebius shares, and that the deal closed on 10 June 2026. That is the buyer's own disclosure, grade A. The company was no longer a one-person or tiny team at the sale: its own page said it had more than 30 team members in March 2026, and Bloomberg reported 20 employees in May 2026. Revenue, the founding date and the amount raised were never published.
How much did Eigen AI make, and what was it sold for?
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 |
|---|---|---|---|---|
| Sale price | about US$643 million | A | Nebius press release, 1 May 2026: cash and Nebius Class A shares with an aggregate value at signing of about US$643 million, based on the 30-day weighted average share price, subject to adjustments. The split between cash and shares was not given. | [1] |
| Agreement and closing | agreed 1 May 2026; closed 10 June 2026 | A | Nebius announced the closing on 16 June 2026, after the required regulatory approvals. | [1, 2] |
| Days from agreement to closing | 40 | calc | Calculated: 1 May 2026 to 10 June 2026. | [1, 2] |
| Founders | 3 co-founders: Ryan Hanrui Wang (CEO), Wei-Chen Wang, Di Jin | A | Named as co-founders in the buyer's release. The company's leadership page lists the same three under 'Founders'. | [1, 10] |
| Team, March 2026 | more than 30 team members, more than 15 with a PhD | C | Company leadership page, Internet Archive copy of 31 March 2026: '30+ team members, including 15+ PhDs'. | [10] |
| Employees at the sale, May 2026 | 20 | D | Implicator, 1 May 2026: 'Eigen has 20 employees, Bloomberg reported'. Bloomberg's article was not opened (its robots.txt disallows Claude agents). Lower than the company's own '30+' two months earlier; the two may count different people. | [3, 5] |
| Sale price per employee | about US$32 million | calc | Calculated: US$643 million divided by the 20 employees Bloomberg reported. The price also paid for the technology and the existing customers. | [1, 3] |
| Time from first public post to the agreement | 11 months | calc | Calculated: 28 May 2025 to 1 May 2026 (338 days). | [14, 1] |
| Outside funding | venture firms and angels; amount not disclosed | C | The founders' post of 1 May 2026 thanks Tectonic Ventures, E14 Fund, Uncorrelated Ventures, AGI House Ventures and angel investors who backed Eigen 'when it was an idea on a whiteboard'. startupsacquired.com: neither company disclosed how much it raised. | [6, 5] |
| Revenue | not published | C | No revenue, customer count or funding amount appears in the Nebius releases, the founders' post or the articles read. Nebius said 'Existing Eigen AI customers' would get access to its infrastructure. | [1, 6, 3] |
| Serverless prices, November 2025 | from US$0.10 per million input tokens (GPT-OSS 120B) | C | Model catalog with a 'Launch console' button for each model; see the pricing table. | [7, 8] |
| Benchmark ranking, November 2025 | first in output speed on 3 models, about 800 tokens per second | C | Company post of 11 November 2025: first among GPU-based providers on Artificial Analysis for DeepSeek-V3.1-Terminus, Qwen3-VL-235B-A22B and GPT-OSS-120B. | [9] |
| Benchmark ranking, March 2026 | first in output speed on 25 open models; up to 911 tokens per second on GPT-OSS-120B | C | Company post of 17 March 2026 on the Nebius partnership. Implicator cites the same 911 figure. | [11, 3] |
| Product lines, November 2025 | EigenData, EigenTrain, EigenInference, EigenDeploy | C | Data generation, training and alignment, inference acceleration, and deployment (serverless, on-demand or dedicated, on Eigen's cloud or the customer's). | [8, 9] |
| Customer quotes on the home page | Wyze, Butlr, HP Inc. | C | Wyze and Butlr quotes carry a named person; the HP Inc. quote carries only a job title. Not confirmed by the customers. | [8, 12] |
| Company | Eigen AI, Inc., Palo Alto, California | C | Terms of service last updated 28 August 2025, under California law. The terms have a section '4. Fees & payment': fees are invoiced under an order form and payable within 30 days. | [13, 11] |
| What the buyer does with it | built into Nebius Token Factory; a Nebius office in the San Francisco Bay Area | A | Nebius release, 1 May 2026. Token Factory is Nebius's managed service for running open models. | [1] |
How to read this table: grade A values come from the buyer's press releases. Grade C values come from the company's own pages, read in Internet Archive copies, or from press articles. Grade D is a press report that could not be checked at its source. A "calc" value (grade X) was computed on this page from the rows above it.
| Plan | Per month, billed yearly (US$) | Per month, billed monthly (US$) | Limit |
|---|---|---|---|
| GPT-OSS 120B, text | US$0.10 input and US$0.50 output per million tokens | ||
| DeepSeek-V3.1, text | US$0.40 input and US$2.00 output per million tokens | ||
| Qwen3 235B VL Instruct, text and images | US$0.50 input and US$2.50 output per million tokens | ||
| Qwen Image Edit and Flux 1.Kontext, images | US$0.025 per image | ||
| WAN2.2 video models (TI2V 5B Turbo, A14B Animate) | US$0.06 per video second | ||
| Whisper V3 Turbo, speech to text | US$0.0009 per audio minute | ||
| Text-to-speech models (Higgs Audio V2, Chatterbox, CosyVoice) | US$0.10 per audio minute | ||
| Dedicated or custom deployment | custom | quoted per customer; no public price |
Use was paid per unit with no monthly plan, so the per-month columns are empty. The same prices appear on the home page copy of 1 November 2025. Eigen-Banana image editing showed US$0.025 per image on 1 November and US$0.0 on 16 November; on 16 November the TI2V video model was hidden in the page code and only the Animate model was shown. Dedicated and custom deployments had no public price. After the sale, Eigen's optimized models are sold inside Nebius Token Factory at Nebius's prices, which this table does not show.
When did Eigen AI launch, and how did it grow?
From a domain that was for sale in April 2025 to a US$643 million sale in May 2026. Each entry carries its grade.
- 2018 to April 2025CThe domain eigenai.com belongs to others: a different business, an AI crypto-trading site, in 2018, a parked page in 2021, and from November 2024 to April 2025 a redirect to a domain-sale page. [15]
- 28 May 2025CFirst dated post on the company's site: WorkForceAgent-R1, a research paper on training web agents with reinforcement learning. [14]
- 17 July 2025CA joint post with the SGLang team on speeding up the SGLang inference engine with multiple-token prediction. [8]
- 9 August 2025CServing of OpenAI's open GPT-OSS models on the day of their release, with a free online playground. [8]
- 28 August 2025CTerms of service of Eigen AI, Inc. dated. The first archived copy of the company home page follows on 23 September 2025. [13, 15]
- 1 November 2025CThe home page shows four product lines, per-token and per-image prices, and customer quotes from Wyze, Butlr and HP Inc. [8]
- 11 November 2025CIts serving API is listed on Artificial Analysis, first in output speed on three models. [9]
- 17 March 2026CPartnership with Nebius: Eigen-optimized open models go into Nebius Token Factory, sold per token. Eigen reports first place in output speed on 25 open models. [11]
- 31 March 2026CLeadership page: three founders and '30+ team members, including 15+ PhDs'. [10]
- 1 May 2026ANebius agrees to buy Eigen AI for about US$643 million in cash and shares. [1]
- 1 May 2026DBloomberg reports that Eigen has 20 employees, as Implicator relays. [3]
- 10 June 2026AThe acquisition closes; Nebius announces it on 16 June. [2]
- 24 October 2026CAn Internet Archive capture of eigenai.com records HTTP status 503; the product continues inside Nebius Token Factory. [15, 2]
Why did Eigen AI work?
Four causes, in the order they mattered. Each one is a reusable lesson.
Research that the industry already used
Nebius described Wei-Chen Wang's AWQ as the standard method for serving models with 4-bit weights and Ryan Wang's SpAtten paper on sparse attention as the most cited HPCA paper since 2020. Nebius called AWQ 'now the standard for 4-bit model serving in production deployments', so the company sold methods that buyers already knew by name.
Optimization at three layers at once
The founders wrote that they tuned the model (quantization, speculative decoding, KV-cache layout), the serving system (batching, scheduling, parallelism) and the GPU code (custom CUDA and Triton kernels) together. A <dfn>kernel</dfn> is a small program that runs on the GPU; the <dfn>KV cache</dfn> is memory that stores earlier tokens so a model does not compute them again. The founders wrote that pulling on all three at once produced results that no single layer could.
A public benchmark as the sales page
Artificial Analysis measures the speed and price of many API providers on the same models. Eigen's posts reported first place on three models in November 2025 and on 25 models in March 2026. A buyer of inference can check such a ranking without trusting the seller.
A partnership before the deal
In March 2026 Eigen's optimized models went into Nebius Token Factory. Nebius said the jointly optimized endpoints ranked among the fastest on Artificial Analysis, and about six weeks later it agreed to buy the company. The founders wrote that every joint engineering sprint led to the same conclusion.
How it was built
These details come from the company's own pages and blog, read in Internet Archive copies, and from the buyer's releases. Company statements are grade C; the buyer's statements are grade A. [9, 8, 6, 1]
- Quantization: FP4 and NVIDIA's NV-FP4 formats, post-training quantization (PTQ) with calibration data, and quantization-aware training (QAT) when quality drops. The company reported up to 4 times faster image editing with FP4 on 24 GB GPUs.
- Throughput: continuous batching, speculative decoding (a small model guesses the next tokens and the large model checks them), and KV-cache paging, eviction and quantization.
- Kernels: fused attention and MLP operations and custom CUDA and Triton kernels. Eigen also published joint work with the team of the open-source SGLang inference engine.
- EigenDeploy served models in three modes: serverless, on-demand and dedicated, with autoscaling, hot restarts and canary releases, on Eigen's cloud, a customer's private cloud or on-premises GPUs.
- EigenTrain ran supervised fine-tuning, preference optimization and online reinforcement learning; EigenData generated and checked training data with several cooperating agents.
- The models were open-weight models from other labs, including GPT-OSS, DeepSeek, Qwen, Llama, Gemma, Nemotron, GLM, Kimi and MiniMax, as Nebius listed them.
Tools named in the sources
Each one opens its page in the Tools tab, with every story that uses it.
Who tried this before Eigen AI, and why did they fail?
Four companies in our database of failed startups sold computing power or a managed platform for machine learning and lost to larger clouds or chip makers. The values below are from that database and are not independently verified.
FloydHub
The promiseA managed cloud for training machine learning models, with notebooks and GPUs ready to use.
Why it diedAWS SageMaker and Google AI Platform offered infrastructure plus tools at scale, and FloydHub could not stand apart from them.
Vectordash
The promiseCheap GPU time for AI developers and gamers, rented from idle GPUs in many places.
Why it diedLarge cloud providers offered integrated GPU services with more reliable performance than a peer-to-peer network could.
Vathys
The promiseNew processors that run AI work more efficiently than CPUs and GPUs.
Why it diedEstablished chip makers had the market and the money; Vathys could not raise enough to produce and sell its chips.
Sixa
The promiseA powerful computer in the cloud, so professionals could run heavy software on any device.
Why it diedAWS and Google bundled similar services and won on cost and reliability.
The failed startups come from our database of 1,749 dead startups, retold here in our own words. Its entries are AI-assisted summaries of public sources, so these cards carry no evidence grade.
What changed
All four sold access to computing power or a platform on top of it, and the large clouds offered the same thing with more capacity and lower prices. Eigen did not try to win on capacity. It owned a skill that, in Nebius's words, most teams do not have in-house: making a given model run faster on a given GPU, from the weights down to the kernel.
That changed the ending. Instead of competing with a cloud, Eigen became the speed layer of one. The partnership with Nebius in March 2026 put Eigen's models into Token Factory before any deal; the founders called it 'a Partnership before It was a Deal'. The founders wrote that software-only optimization runs into a ceiling without hardware underneath, and the sale gave them the hardware.
Can one person copy Eigen AI now?
Not by one person.
Eigen's value rested on years of published research by its founders and on a team that grew to 20 or more people, many with PhDs, plus GPUs to serve the models. A solo founder cannot copy that. The smaller pattern can be copied: take one optimization method that you understand deeply, apply it to the newest open models on release day, and publish independent benchmark results.
Where it is still open: Tuning for one narrow case that large providers serve poorly: one model family, one accelerator, speech or video models, or models that run on laptops and phones.
- 1Pick one measurable bottleneck in running open models, such as memory per token or time to first token, and become the best at it.
- 2Serve new open models on the day they are released, and publish the speed on a third-party benchmark.
- 3Sell per token with public prices, so a customer can test without a sales call.
- 4Partner with a company that sells the capacity and needs the speed.
What could go wrong for Eigen AI?
The risks are the mirror image of the causes above.
What is not verified
The price is the buyer's own disclosure and is grade A. Three things are not verified. First, the founding date: no source read gives it; the first dated post on the site is from 28 May 2025, and the domain pointed to a sale page until at least April 2025. Second, the amount raised and the revenue, which neither company published. Third, the headcount: Bloomberg's figure of 20 employees was read only as Implicator quoted it, and it is lower than the company's own '30+ team members' in March 2026. A reader could look for Nebius's annual report on Form 20-F for 2026, which may describe the acquisition.
Team size
Three co-founders started the company, as the Nebius release and the leadership page name them; no early source gives a headcount at the start. By March 2026 the company described more than 30 team members, and Bloomberg reported 20 employees at the sale in May 2026. So the business was a small funded team, not a solo business, when it sold.
Trust signals on the home page
The November 2025 home page shows a 'Trusted by teams at' row of logos as images with no names in the text, and lists 'SOC 2, HIPAA, GDPR' under enterprise features with no report or certificate linked on the pages read. One customer quote is attributed only to 'Senior Research Scientist, Personal Systems, HP Inc.'; the Wyze and Butlr quotes give a person's name. In April 2026 the page said the feedback 'All Comes From Real Customers Of Eigen AI'. None of the customers was asked.
What the catalog hosted
Besides language models, the catalog hosted open text-to-speech models (Higgs Audio V2, Chatterbox, CosyVoice) and a WAN2.2 'Animate' video model, and the April 2026 home page listed a 'Real-Time Avatar' use case for reception, customer support and tutoring. The pages read do not advertise voice cloning or animating real people; they sell the models as general services priced per minute or per second.
Product and custom work
The self-serve API had public prices, but the model pages also said 'Need something custom? Work with Eigen AI to fine-tune models, deploy on dedicated infrastructure'. No source says how revenue split between the per-token API and custom enterprise work.
Competition
On this site, Wafer also sells an inference API for open models tuned by its own agents, and Chamber sells agents that manage GPU workloads; Artificial Analysis ranks Eigen against other GPU-based API providers, and Implicator named CoreWeave and Nscale as Nebius's competitors.
Dependence on model providers and hardware
Eigen did not train its own base models. Its catalog and the Nebius release list open-weight models from other labs (OpenAI's GPT-OSS, DeepSeek, Qwen, Meta's Llama, Google's Gemma, NVIDIA's Nemotron, GLM, Kimi, MiniMax), so its product depended on those labs publishing weights that customers want. Its speed work targeted NVIDIA GPUs (Hopper and Blackwell, the NV-FP4 format). After the sale it runs on Nebius's capacity.
Site offline after the sale
eigenai.com answered HTTP 503 ('deployment paused') to the pipeline's harvester on 8 November 2026, and an Internet Archive capture of 24 October 2026 records the same status. The links on this page point to an archived copy of April 2026. Nebius said Eigen's optimization layers are built into Token Factory.
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.
- A Nebius press release, 1 May 2026: Nebius agrees to acquire Eigen AI, strengthening Nebius Token Factory as a frontier inference platformRead in full. The buyer, listed on Nasdaq. Source for the price, the payment in cash and shares, the co-founders, the existing customers and the integration plan.
- A Nebius press release, 16 June 2026: Nebius completes acquisition of Eigen AIRead in full. Source for the closing date of 10 June 2026.
- C Implicator.ai, Marcus Schuler, 1 May 2026: Nebius Buys Eigen AI for $643 Million to Strengthen Token FactoryRead in full. Source for the 20 employees, which it attributes to Bloomberg; Bloomberg itself was not opened because its robots.txt disallows ClaudeBot and Claude-User for news pages.
- C Techzine Global, Erik van Klinken, 1 May 2026: Nebius acquires Eigen AI for $643 millionRead in full. Confirms the price, the three founders and the optimization methods.
- C startupsacquired.com, 1 May 2026, updated: Nebius agrees to buy inference optimisation startup Eigen AI for about $643 millionRead in full (downloaded by the pipeline's harvester on 8 November 2026). A summary of the releases and press; says neither company disclosed the founding date or the amount raised.
- C Eigen AI blog, Ryan Hanrui Wang, 1 May 2026: The Next Chapter of Artificial Efficient Intelligence (Internet Archive copy of 9 May 2026)Read in full. The founders' own account of the sale; source for the investors, the three-layer method and the continuity promise to customers.
- C Eigen AI model catalog (Internet Archive copy of 16 November 2025)Read in full. Source for the prices.
- C Eigen AI home page (Internet Archive copy of 1 November 2025)Read in full. Source for the product lines, prices, customer quotes, compliance claims and the list of 2025 posts.
- C Eigen AI blog, 11 November 2025: Eigen AI now on Artificial Analysis (Internet Archive copy of 13 January 2026)Read in full. Source for the November 2025 ranking and the EigenInference and EigenDeploy details.
- C Eigen AI leadership page (Internet Archive copy of 31 March 2026)Read in full. Source for the three founders and the '30+ team members'.
- C Eigen AI blog, 17 March 2026: Eigen AI and Nebius Partner to Accelerate Frontier Open-Source AI Inference (Internet Archive copy of 31 March 2026)Read in full. Source for the partnership, the per-token access in Token Factory and the 25-model ranking.
- C Eigen AI home page (Internet Archive copy of 17 April 2026)Read in full. Source for the use cases, including 'Real-Time Avatar', and the statement on customer feedback.
- C Eigen AI terms of service, last updated 28 August 2025 (Internet Archive copy of 31 October 2025)Read in full. Source for the company name, the state and the fees clause.
- C Eigen AI blog, 28 May 2025: WorkForceAgent-R1 (Internet Archive copy of 31 October 2025)Read in full. The earliest dated post found on the site.
- C Internet Archive capture index for eigenai.com, with the archived pages of 6 August 2018, 30 November 2021 and 14 November 2024Read the capture list and the three early copies. Source for the domain history and the 503 status of October 2026. The pipeline's harvester also received HTTP 503 from eigenai.com on 8 November 2026.
- C Database of failed startups, local exportLocal export. AI-written summaries, not independently verified. Source for the failed predecessors.
Run this company? Send corrections or a verified figure to aiolum@livyden.com