Axelera AI

The problem

Running AI models next to cameras, robots or factory lines requires chips that are powerful, frugal and cheap, while data center GPUs draw a lot of power and depend on a dominant supplier. IDC puts worldwide edge computing spending at USD 265 billion in 2025, a sign that a growing share of inference is moving out of the cloud.

The solution

Axelera AI designs AI inference processors (AIPUs). Metis, built on 12 nm, is rated at 214 TOPS and 15 TOPS/W and ships as M.2 and PCIe cards and inside partner servers. Europa, launched in September 2026, targets enterprise servers and Europe's AI factories, with up to 629 INT8 TOPS.

How it works

1. The Metis chip combines four AI cores of 53.5 TOPS each, orchestrated by RISC-V controllers.
2. Each core runs matrix multiplications directly inside SRAM (digital in-memory computing), avoiding energy-hungry data movement; the company says these operations make up 70 to 90% of neural network compute.
3. Weights and activations run in INT8 with INT32 accumulation, and more than 52 MiB of on-chip memory limits trips to external LPDDR4x.
4. The Voyager SDK quantizes and compiles PyTorch or TensorFlow models and ships with a ready-made model zoo.
5. Europa scales the architecture up for servers as the Edge 232p and Server 250p PCIe cards, validated in Dell's XE5 and Supermicro's 111AD.

Value proposition

More compute per watt and per euro for local inference: the company claims up to 6x more tokens per second per watt than GPU-based systems for Europa (internal benchmark). The Metis M.2 card was announced at around EUR 149 as a sample price at launch.

Technology edge

Digital in-memory computing avoids the noise and drift of analog approaches while cutting data movement, the main source of power draw. Against Hailo-8 (26 TOPS on M.2), Metis claims 214 TOPS, though CNX Software estimated that within the roughly 8 W M.2 thermal envelope, real-world output would be closer to 120 TOPS. Axelera also offers a European supplier for sovereign AI projects.

Business model

B2B: sales of chips, M.2 and PCIe cards and systems through OEM partners (Dell, Supermicro, Advantech, Lenovo, HPE), plus supply contracts for AI factories. No public pricing beyond evaluation cards.

Market

IDC (February 2026) estimates worldwide edge computing spending at USD 265 billion in 2025, approaching USD 450 billion by 2029 (about 15% a year). Axelera targets embedded computer vision first, then generative AI inference in enterprises and in EU-funded AI factories.

Competitors and alternatives

Nvidia remains the reference for inference, in the cloud and at the edge. Israel's Hailo, a direct edge accelerator rival, signed an agreement to be acquired by Microchip in July 2026 after laying off half its staff. In Europe, France's VSORA, the UK's Fractile and the Netherlands' Euclyd, which announced a EUR 200 million raise in September 2026, also target inference. Axelera stands out with a product line already in production and a catalog of partner servers.

Use cases

• A factory analyzes its security camera footage on site without sending it to the cloud, a use case Reuters cites for the Metis chip.
• A retailer accelerates in-store vision applications on Dell, Lenovo or Advantech systems fitted with Metis, one of the uses the company highlights.
• The EU AI factories IT4LIA (Italy) and MeluXina (Luxembourg) integrate Axelera chips through Dell and integrator E4.

Traction and proof points

More than 600 customers claimed in September 2026 across security, defense, AI factories and enterprise, including Telefónica Tech, OVHcloud, E4 Computer Engineering, Safran, duagon (Knorr-Bremse) and Kudelski Labs. Supply deals 'worth tens of millions' according to the CEO, and USD 1.5 billion in potential sales under discussion (Reuters). System partners: Dell, Supermicro, HPE, Lenovo, Advantech. CES Innovation Awards 2024 Honoree (Metis platform); at CES 2024, 2025 and 2026 in a suite.

What remains to be proven

No revenue is disclosed, and the USD 1.5 billion pipeline is not firm orders. Europa's gains over GPUs rest on internal benchmarks that need independent confirmation. Hailo's slide from a USD 1.2 billion valuation to a distressed sale is a reminder of how capital-hungry AI chips are.