Axelera AI
Last reviewed
Jun 8, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 ยท 1,910 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 8, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 ยท 1,910 words
Add missing citations, update stale details, or suggest a clearer explanation.
Axelera AI is a Dutch semiconductor company that designs AI chips for inference, the stage at which a trained model is run to produce predictions. The company builds accelerators around a proprietary digital in-memory computing architecture paired with RISC-V control cores, an approach aimed at delivering high performance-per-watt and performance-per-dollar for computer-vision and, increasingly, generative AI workloads. Axelera AI began with a focus on edge AI inference through its Metis AI Processing Unit (AIPU) and is expanding toward data-center and high-performance-computing inference with a chiplet code-named Titania. Headquartered at the High Tech Campus in Eindhoven, Netherlands, the company has positioned itself as a European challenger to incumbent AI-silicon suppliers such as NVIDIA.[1][2][3]
In February 2026 Axelera AI raised more than $250 million in a financing round led by the European venture-capital firm Innovation Industries, with new investors including BlackRock. The company described the round as the largest ever for a European Union AI semiconductor firm and said it had shipped to its 500th customer.[1][4][5]
| Field | Detail |
|---|---|
| Type | Private company (fabless semiconductor) |
| Founded | July 2021 (incubated as Bitfury AI from 2019) |
| Headquarters | Eindhoven, Netherlands (High Tech Campus) |
| R&D sites | Leuven, Belgium; Zurich, Switzerland |
| Co-founders | Fabrizio Del Maffeo (CEO); Evangelos Eleftheriou (CTO) |
| Core technology | Digital in-memory computing (D-IMC) with RISC-V |
| Key products | Metis AIPU, Europa AIPU, Titania chiplet, Voyager SDK |
| Total funding | More than $450 million in equity, grants, and venture debt (as of February 2026) |
| Customers | More than 500 (defense, industrial, robotics, retail, agritech, security) |
Axelera AI develops hardware and software for running neural networks efficiently, concentrating on the inference stage rather than on training. Its stated goal is to pack the AI compute of a server into a single chip at a fraction of the power consumption and cost of conventional accelerators, making it viable to deploy AI in power- and thermally constrained environments. The firm's products span add-in accelerator cards for the edge, processors for edge servers and enterprise systems, and a forthcoming data-center-class chiplet, all built on the same digital in-memory computing foundation and unified by the Voyager software stack.[2][3][6]
The company markets itself as a European AI-hardware champion. Its early customer base spans physical-AI and edge-AI markets including defense and public safety, industrial manufacturing, retail, agritech, robotics, security, and aerospace. By February 2026 Axelera AI reported shipping to more than 500 customers.[1][4]
Axelera AI traces its roots to 2019, when it was incubated inside the Bitfury Group, a blockchain and bitcoin-mining technology company, as an internal effort then known as Bitfury AI. In early 2020 the team began collaborating with imec, the Belgian nanoelectronics and digital-technology research institute in Leuven, to develop a computing architecture for edge AI. The venture was spun out as an independent company, incorporated in July 2021, with imec and its investment arm imec.xpand among the founding backers.[2][3][7]
The company was co-founded by Fabrizio Del Maffeo, who serves as chief executive officer. Before Axelera AI he was head of artificial intelligence at Bitfury and earlier held managing-director and vice-president roles at AAEON Europe, a subsidiary of the Taiwanese hardware maker ASUS. The other principal co-founder is Evangelos Eleftheriou, the chief technology officer, an IEEE Fellow and former IBM Fellow who spent more than three decades at IBM Research in Zurich working on areas including in-memory and neuromorphic computing. His research background underpins the company's bet on in-memory compute. At launch the startup had recruited more than 20 senior engineers from firms such as Intel, Qualcomm, IBM, and imec.[3][7][8]
Axelera AI is headquartered in the AI Innovation Center at the High Tech Campus in Eindhoven, Netherlands, and operates research-and-development offices in Leuven, Belgium, and Zurich, Switzerland, the last reflecting its IBM Research-Zurich lineage. Some press reports describe the company as Amsterdam-based, but the company's own materials and the imec announcement place its headquarters in Eindhoven.[2][3][7]
The technical foundation of Axelera AI's chips is digital in-memory computing (D-IMC), a variant of in-memory computing. In a conventional processor, weights and activations are shuttled between memory and separate compute units, and that data movement dominates energy use, an effect often called the memory wall. D-IMC instead performs the dominant operation in deep learning, matrix-vector multiplication, directly inside memory cells arranged in a crossbar array, so computation and storage happen in the same place. Axelera AI implements this using SRAM rather than analog memory, which it argues preserves the accuracy and noise immunity of digital logic while capturing the efficiency of computing in memory.[2][3][6]
Around the in-memory compute cores, the architecture uses RISC-V vector cores for pre- and post-processing and for orchestrating dataflow. Building on the open RISC-V instruction set rather than a proprietary one lets the company tailor its compute cores to inference and reduces dependence on third-party licensing. The combination is marketed as a dataflow architecture that delivers high throughput per watt and per dollar for inference at the edge and beyond.[2][3][9]
Axelera AI's first commercial product is the Metis AI Processing Unit. The Metis AIPU delivers peak performance of up to roughly 214 trillion operations per second (TOPS) at INT8 precision, with an efficiency of about 15 TOPS per watt. It is sold in several form factors: an M.2 module carrying a single AIPU, PCIe accelerator cards, and a quad-AIPU PCIe card reaching on the order of 856 TOPS. Metis targets edge inference for computer vision, robotics, video analytics, and embedded industrial systems. Independent benchmark coverage has highlighted the platform's throughput and efficiency on machine-vision models, with the company and reviewers attributing much of its advantage to the accompanying software.[2][3][10][11]
In 2025 the company introduced Europa, a second-generation AIPU aimed at edge servers, enterprise systems, and rack-mounted deployments running multi-user generative AI as well as computer vision. Europa is specified at up to about 629 TOPS (INT8), built from eight AI cores using second-generation D-IMC, and pairs roughly 128 MB of on-chip L2 SRAM with external memory delivering around 200 GB/s of bandwidth. Axelera AI has indicated that Europa-based PCIe accelerators are expected to ship in the first half of 2026.[2][3][12]
The company's most ambitious product is Titania, a data-center-class AI inference chiplet still in development. Titania is being designed under the EuroHPC Joint Undertaking's DARE (Digital Autonomy with RISC-V for Europe) consortium and is built on the same D-IMC architecture and on RISC-V, intended to scale near-linearly from the edge to the cloud. Axelera AI has said the first systems powered by Titania are anticipated around 2027, with broader deployment targeted for 2028 for data-intensive AI and future large-scale HPC centers. (The earlier description of Titania as a small edge chip should be read instead as a scalable inference chiplet aimed at servers and HPC.)[9][13][14]
All of the hardware is supported by the Voyager Software Development Kit (SDK), which uses the open-source Apache TVM compiler stack to optimize and deploy models, together with a model library, sometimes called the Model Zoo, of pre-packaged networks to simplify edge AI deployment.[3][6][10]
Axelera AI has raised capital across several rounds, grants, and venture debt since its founding. The figures below are drawn from company announcements and press coverage; some totals are approximate.
| Round / source | Amount | Date | Notes |
|---|---|---|---|
| Seed | About $12 million | September 2021 | Led by Bitfury; imec, imec.xpand, Innovation Industries [7] |
| Series A (and extension) | About $50 million total | 2022 to 2023 | Oversubscribed [3][8] |
| Series B | About $68 million | June 2024 | Reported as Europe's largest fabless-semiconductor Series B [2][3] |
| EuroHPC DARE grant | Up to 61.6 million euros | March 2025 | Part of a 240-million-euro consortium award for Titania [9][13] |
| Growth round | More than $250 million | February 2026 | Led by Innovation Industries [1][4][5] |
The February 24, 2026 round was led by Innovation Industries and drew new investors including BlackRock and SiteGround Capital, alongside existing backers such as Bitfury, CDP Venture Capital, the European Innovation Council (EIC) Fund, the Federal Holding and Investment Company of Belgium (SFPIM), Invest-NL, the Samsung Catalyst Fund, and Verve Investments. The company said the financing brought its total raised since incorporation in July 2021 to more than $450 million in equity, grants, and venture debt, and described it as the largest investment ever in an EU AI semiconductor company. No valuation was disclosed. (The starting characterization of the round naming BlackRock, the Samsung Catalyst Fund, and the EIC as lead-level participants is broadly correct, but the round was led by Innovation Industries.)[1][4][5][15]
Axelera AI is frequently cited as a flagship of Europe's drive for semiconductor sovereignty, the ambition to reduce the continent's reliance on foreign-designed AI chips. Its investor base mixes private venture capital with state-linked and pan-European public funds, including the European Innovation Council Fund, Italy's CDP Venture Capital, Belgium's SFPIM, and the Dutch state investor Invest-NL, the last of which has framed the company as building "a true European champion in AI infrastructure." Its participation in the EuroHPC DARE program ties the Titania chiplet to the European Union's broader effort to develop sovereign RISC-V-based high-performance computing.[1][9][14]
The competitive landscape spans edge-AI and inference specialists as well as established players. In the edge and machine-vision segment, Axelera AI's Metis competes with chips from Hailo, SiMa.ai, Google Coral, Kneron, and DeepX, among others, with reviewers generally placing Metis among the fastest M.2 and PCIe options for multi-camera and heavier vision workloads. As the company moves up-market with Europa and Titania, it increasingly contends with data-center inference incumbents and challengers, most prominently NVIDIA, whose GPUs dominate AI acceleration, along with inference-focused firms such as Groq. Axelera AI's differentiation rests on its digital in-memory compute and RISC-V architecture, which it argues yield better performance per watt and per dollar than GPU-based approaches for inference, and on a European supply story that appeals to customers in defense and the public sector.[1][3][9][10]