Axelera AI CEO and co-founder Fabrizio del Maffeo never planned on being an entrepreneur.

But, after launching a successful Kickstarter campaign centered around IoT and Edge AI computing devices in 2015, he found himself playing (and rather enjoying) the role.

“I convinced Intel to give me a chip,” Del Maffeo recalls to DCD. “I said: ‘Give me that chip (Intel’s Movidius Myriad 2 vision processing unit), I will package it in a different form factor and with a specific software stack, and we’ll try to give to customers who want to put AI inside small devices.’”

Dubbed UP Bridge the Gap, the project’s initial offerings were x86-based IoT development platforms, but in 2018, it pivoted to launch the UP AI Edge solution – self-described as “the first artificial intelligence on the Edge computing platform.”

Although UP achieved commercial success amongst developers, del Maffeo says the technology fell short of what was needed to fully support AI workloads at the Edge – “We had thousands of customers buying the product, but the product was terrible, it was awful, because it was not meant for that kind of workload.”

Despite the venture failing to meet expectations, del Maffeo says he saw that there was still a market opportunity for hardware that would allow customers to run large neural networks inside devices, and wondered if he could be the person to develop that solution. “I was 42 at the time, and I had a very good job in the Asus group, but I thought, ‘Why don’t I try?’”

With his background in engineering, del Maffeo started reading papers from places like the European Institute Research Center alongside works published by IBM and Imec, and found himself turning to emerging technologies to see if they held the answer to why UP had failed.

That’s how he came to settle on Digital In-Memory Computing (D-IMC) architecture, an approach that uses digital logic to process data directly within memory arrays, thus eliminating the “von Neumann bottleneck” of moving data between memory and processors.

A year later, del Maffeo had quit the Kickstarter project and begun to focus all his attention on what would eventually become Axelera AI, a Netherlands-based chip startup that designs hardware for high-performance Edge computing use cases.

Analog versus digital

In-memory computing can be digital or analog. However, according to del Maffeo, the digital approach offers higher accuracy, robustness, and easier integration compared to analog, making it ideal for AI inference.

As del Maffeo explains, analog in-memory compute uses the relationship between tension and current within a transistor to make calculations. Essentially, you program the resistor inside the network, program the current, then measure the tension, and use the analog properties of the material to make calculations.

“But it means that you have data which is a bit, so you have to convert it to analog, do everything in analog, and then reconvert it to digital, meaning it’s more efficient but less precise,” he explains. “Because you have to convert the data, you lose precision, then you cannot have more than six or seven-bit integer precision.”

With digital in-memory compute, by comparison, Axelera doesn’t modify the memory; rather, it just places a small digital computing element within it.

As a result, while the efficiency is around 40 to 50 percent lower than with the analog approach, you don’t lose anything, meaning you’re left with full precision.

Officially launched in 2021, Axelera AI develops AI processing units (AIPUs) that allow AI software to be run at the network Edge, as opposed to inside data centers. Built on RISC-V and D-IMC architecture to provide faster data processing and retrieval, the company says that by deploying its technology at the Edge, it speeds up both AI inference and interactions.

Since its founding, the company has raised $450 million in capital from investors including BlackRock, Bitfury, the European Investment Council Fund, and the Samsung Catalyst Fund.

“We believe in inference,” del Maffeo says. “What we do is make AI accelerators that can run inference.”

Even when speaking to DCD in late 2025, del Maffeo could already see what was creeping over the horizon and has become one of the biggest industry talking points in recent months – that inference is going to represent the largest part of the market.

“Once you train a network, you just run it, and especially new neural networks, they don’t need retraining. They can adapt very quickly to different contexts and perform very well,” he explains.

The company’s first product, Metis, was unveiled in 2024 and has been in full production since February 2025. Built with computer vision in mind, while the chip helped Axelera secure some early success, like del Maffeo’s earlier Kickstarter project, the hardware came with limitations, primarily around post-processing memory bandwidth.

Eight months later, in October 2025, Axelera followed up with its Europa chip. Designed to support high-performance Edge computing use cases, including generative AI and computer vision applications, the AIPU delivers up to 629 TOPS at INT8 precision, setting a new industry benchmark for performance, the company claimed at the time.

“Europa covered these gaps [that existed in Metis],” del Maffeo explains. “Europa, essentially, is a chip that is a multi-core solution – it has eight cores, each of which can potentially run a network, or can be combined to distribute the network among the eight cores.”

Inside every Europa core are two elements: a RISC-V vector processing unit and a digital memory processing unit. Why these two elements? Del Maffeo says that if you look inside neural networks, 70-90 percent of the operations are simple vector matrix multiplications.

What Axelera AI has done, therefore, is design a digital in-memory computing engine to carry out the vector multiplication. As a result, up to 90 percent of the calculation goes directly into this compute engine, which is “super optimized” for this task, and the rest of the operations go to the RISC-V vector processing unit.

This, the company says, eliminates memory bottlenecks and delivers “three to five times performance efficiency over the leading industry solutions in the same product category.”

In addition to Metis and Europa, Axelera AI also has plans to launch an AI inference chiplet dubbed Titania, built using the company’s D-IMC architecture. Where Europa brings compute to the Edge, Titania will allow customers to scale up deployments in the data center, supporting inference workloads on a much larger scale.

Del Maffeo says that while Axelera currently has its computer vision solution in production, the company is also looking to target vision and large language models with its future hardware offerings. “Today, with computer vision, we are very strong in defense, retail, surveillance, and we’re getting traction in medical and automotive.”

When it comes to the challenges of being a startup trying to take on some pretty hefty incumbents, del Maffeo says it ultimately comes down to cost and performance.

“If you are not competitive in cost at the Edge, you don’t win the Edge market. And we are competitive. We have the highest throughput per dollar on the market, by far,” he says. “We published benchmarks, and you can see we’re up by a factor of two, three times. At the Edge, if you don’t have that, you die, because it’s all about performance per dollar, performance per watt.”

However, for larger-scale deployments, like those being targeted by Titania, he says that the company has to be realistic, acknowledging that there will also be a gap between new startups and the big players already in the field.

“I strongly believe in specialization. In the future, for each different workload, you will have a specialized chip because you cannot generalize, particularly as inference becomes more important,” he says. “Hyperscalers have already been doing this for many years. Google has developed a TPU for this specific [inference] workload. And for video content, they have ideas.

“But I strongly believe that there is a space for us.”

Going global

For del Maffeo, ensuring Axelera is a truly global outfit is almost as important as the technology it develops.

That said, the company has strong European roots: its headquarters are in Eindhoven, the Netherlands, the same town where photolithography equipment maker ASML is based, and, during its short life, Axelera has already become involved with a number of European projects. This includes the Digital Autonomy with RISC-V for Europe (DARE) Project, which provided the company with a €61.6 million ($66.9m) grant to support the development of its Titania AI inference chiplet.

Most recently, in April 2026, the European High Performance Computing Joint Undertaking announced that Italy’s new IT4LIA AI factory in Bologna would, in part, be powered by inference accelerators from Axelera AI.

However, when it comes to fabricating its chips, the company has looked even further afield, with the 12nm Metis manufactured by TSMC in Taiwan and the 5nm Europa produced by Samsung in South Korea.

“Our chips are on a node which cannot be manufactured in Europe,” del Maffeo says. “Not yet, anyway. They’re too advanced.”

He adds: “I would like to manufacture my chip everywhere – to have the chance to do it in Asia for each Asian customer, in Europe for Europeans, and in the United States, for Americans. I think that in the future, supply chains will be more localized.

“Realistically, our first chip, Metis, which is now in production, could potentially be manufactured in Dresden, Germany, from 2028 because TSMC is working on a fab there, in partnership with NXP and Infineon.”

Ironically, for Europa, there isn’t the possibility of having it produced this side of the Pacific in time for its anticipated release later this year.

For del Maffeo, however, that’s not necessarily a bad thing. While he might have visions of watching Metis chips come out of a fab in Dresden by the end of the decade, that doesn’t mean Axelera will shut itself off from other continents should the opportunity present itself.

“I don’t believe in isolationism,” he says. “We want to be a global player. And you have to be if you want to win this market.”