Summary: Axelera AI announced on February 24, 2026 that it raised more than $250 million, which its press release called the largest investment in an EU AI semiconductor company. The article links the funding to edge inference hardware, including the 10-watt Metis AIPU at 214 TOPS and the in-development 629-TOPS Europa platform, with capital directed toward manufacturing scale and customer support.
Axelera AI, headquartered in Eindhoven, announced on February 24, 2026 that it has closed a round worth more than $250 million — a figure its own press release bills as the largest investment ever made in an EU AI semiconductor company. Innovation Industries led the financing, BlackRock and SiteGround Capital came aboard as new investors, and a group of returning backers — among them Bitfury, CDP Venture Capital, the European Innovation Council Fund, Belgium’s SFPIM, Invest-NL, and the Samsung Catalyst Fund — put in more. The company, founded in 2021, has now raised more than $450 million across equity, grants, and venture debt, and the fresh capital is headed for manufacturing scale, its customer success organization, and its Partner Accelerator Network.
NVIDIA has data-center AI locked up, and Axelera is not fighting on that ground. The company leaves training to others, concentrates on AI inference running on edge devices, and competes on efficiency rather than raw throughput. In its coverage, Sifted put it bluntly: the raise is money “to take on Nvidia.”
In-Memory Computing: The Efficiency Explained
The technology at the heart of the company goes by D-IMC, digital in-memory computing: SRAM organized into crossbar arrays, so computation happens right where the data already sits. On conventional chips, the costliest step is often not the arithmetic but the hauling of data back and forth between memory and compute; D-IMC deletes that haul entirely, and the power budget stays small as a result.
That thinking carries into how the company frames the use of funds. CEO Fabrizio Del Maffeo contends that data centers are running into power and cooling limits while edge devices operate under even tighter energy and bandwidth constraints — and that “inference costs run 15 times higher than training over a model’s lifetime.” Efficiency, in this telling, is the business model.
From Metis to Europa
The flagship shipping today is the Metis AIPU: 214 trillion operations per second at about 10 watts, built for battery-powered jobs such as AI navigation on warehouse robots. It comes in two form factors — a PCIe accelerator card carrying as many as four chips, and a single-chip M.2 version for compact, low-power systems. The software layer is the Voyager SDK, constructed on Apache TVM and paired with a Model Zoo of prepackaged models to cut deployment friction.
Waiting in the wings is Europa, the next-generation platform now in development: 629 TOPS — better than double Metis — with 8 AI cores, 16 CPU cores, and 128 MB of memory. The company claims up to three times the performance per watt of its competitors and points to an internal test that topped 13,168 frames per second on computer-vision workloads. Volume manufacturing rests on partnerships with TSMC and Samsung.
The Case for the Edge
The commercial track record deserves attention: Axelera says it recently shipped to its 500th global customer, spread across defense and public safety, industrial manufacturing, retail, agritech, robotics, security, telecom, aerospace, and the enterprise market. What these deployments share is workloads whose data cannot or need not make the round trip to the cloud: privacy law, sovereignty requirements, latency budgets, and bandwidth cost all steer inference onto local hardware.
For Europe, the round adds an industrial-policy layer on top. With the US and China holding sway over AI compute, the EU can now point to its first AI chip company financed at this scale, with public capital — the EIC Fund, SFPIM, Invest-NL — at its back. A European edge-inference supply chain has outgrown the talking-point stage.
Developer Takeaways
Three consequences for practitioners. First, the shelf of edge-inference hardware is filling out: a 10-watt, 200-plus-TOPS part available in PCIe and M.2 form factors lets vision, anomaly detection, and similar workloads run locally and indefinitely, free of a cloud dependency and of the recurring per-token bill that comes with one. Second, the software stack governs adoption speed — building Voyager on Apache TVM gives teams already inside the TVM ecosystem essentially zero migration cost, the right wager for a challenger that lives or dies on switching costs. Third, anyone building AI products should memorize the “inference costs 15x training” figure: shipping the model is only the beginning, the long-run cost structure is dictated by inference efficiency, and that is exactly why dedicated edge silicon exists.