Spun out of Dutch unicorn Bitfury, Eindhoven-based stealth startup Axelera AI stepped into the spotlight with a €10M seed investment. Leveraging expertise from IBM, Imec, Intel and Qualcomm, the ambitious young company wants to lower the AI threshold for European SMEs by supplying a cost-effective range of plug-and-play products based on its own chip.
“It’s like using a Formula 1 engine in a Fiat 500 because you have nothing else,” says Fabrizio Del Maffeo, describing the existing artificial-intelligence solutions available to small and medium enterprises. “SMEs are the heart of European business. Most of them could greatly benefit from AI. Unfortunately, they’re held back in applying it for the simple reason that it’s too expensive and too inefficient. Existing solutions based on Nvidia GPUs, for example, cost many hundreds of euros and most SMEs really don’t need all that computing power.”
With his new Eindhoven-based startup, Axelera AI, Del Maffeo wants to lower the AI threshold for SMEs by supplying a cost-effective range of plug-and-play products based on its own chip. “But we’re not a chip company,” he stresses. “Our target customers aren’t looking for a chip. AI isn’t a hardware-driven market; it’s a market driven by data, by data scientists, by software engineers – and software engineers need more than a chip. We’re delivering complete solutions, consisting of hardware and a software stack. Our vision is that customers buy our chip in a Mini PCI Express module, for example, which they can plug into their system board to take care of the AI part of their application.”
Still, the plan to also develop a chip scared off a lot of investors. “When I was pitching my idea for the new company, I told them that there would be more to it than that, that we were going to build up a European AI technology company and that we would need 100 million euros for that. They called me crazy and far too ambitious. You’re doing semiconductors, they said, which isn’t a European business anymore, and you’re asking way more than is typically invested in Europe. So, I got many rejections, which is why it took two years to get the venture off the ground.”
Eventually, the Amsterdam-based emerging-technologies expert Bitfury decided to back Del Maffeo’s plans, together with Imec, through Imec.xpand, and Innovation Industries. Last week, Axelera AI could finally announce the closing of its seed investment round with a haul of 10.15 million euros (12 million dollars). These funds will be used to support the startup’s ambitions to develop industry-defining products that bring the AI computational power of an entire server into a single chip, powering AI applications at the edge.
Truly European
Axelera AI was incubated by Bitfury in 2019, operating as Bitfury AI until coming out as an independent entity last July. “Bitfury is a Dutch unicorn,” explains Del Maffeo. “They were one of the first to go all in on bitcoin. Realizing that mining on PC computers was inefficient, expensive and environmentally unfriendly, they designed their own chip in 2013 and subsequently their own board and system, which they started using to mine themselves. This made them a lot of money. After the first bitcoin crash in 2017, they decided to diversify and they invested in blockchain software solutions, in immersion cooling for data centers and in what they saw as the next best thing after mining – artificial intelligence.”
Del Maffeo joined Bifury in 2019 to head the AI venture. He moved there from Eindhoven-based Aaeon Europe, an Asus company, where he was VP and managing director. “At Aaeon, I was also responsible for worldwide AI development and worked closely with Intel and its AI subsidiary Movidius to enable an AI ecosystem,” he says, outlining his background in artificial intelligence. “At some point, I came into contact with Bitfury and they offered me the opportunity to bring new AI technology to the market, leveraging their vast resources.”
However, Bitfury had no AI product to offer yet. So, Del Maffeo set out to talk to Europe’s brightest minds in computer science. “I went to Imec, to KU Leuven, to ETH Zurich, I reached out to IBM in Zurich, and from these conversations, I concluded that it would be possible to develop cutting-edge technology with a truly European footprint. With IBM and Imec being big rivals in research, I also saw that I needed to convince competing teams in different countries to join hands. It took some time, but I managed to bring them together and it worked so well that Bitfury decided to accept my proposal to spin off the company and put money in.”
Currently, Axelera AI has a team of 17, distributed across its headquarters in the AI Innovation Center of the High Tech Campus Eindhoven and offices in Leuven, Zurich and Pisa. By the beginning of November, according to Del Maffeo, the headcount will already be 28. The company has recruited senior engineers and developers from world-leading AI companies and research centers, not only from IBM and Imec but also from Intel and Qualcomm. “We’re ramping up very fast, especially here in the Netherlands, where the heart of the company is.”
Too big a shift
To lower the AI threshold, Del Maffeo’s European team is building on proven technology. “99.9 percent of the data scientists are using standard networks,” the CEO points out. “Leveraging the in-memory computing expertise of Imec and IBM and the know-how on dataflow architectures from Intel and Qualcomm, we’re going to give them an efficient technology that allows them to reuse what they have.”
Axelera AI has all the competences to jump on advanced developments when the time is right. The Imec hires, for example, are the very same people who developed the Ania analog inference accelerator the Leuven institute demonstrated last year. Del Maffeo: “We have the team on board and we have the IP, but we’re not productizing it right now, for the simple reason that the Ania chip employs a ternary network and ternary networks aren’t widely adopted yet. To take that to market, we’d have to build up a completely different software stack and customers would have to change the way they design their networks – that’s too big a shift. The chip is still too far away from what people are actually using, but it shows the capability of the team and the potential of the in-memory compute technology and we’re taking advantage of that.”
Likewise for spiking neural networks. “A lot of startups are doing that,” observes Del Maffeo. “From our point of view, however, spiking isn’t a mature field. We do have a lot of experience with it. In fact, our CTO, who’s going to join us in October, is an expert on the matter. But we don’t believe that it’s the time yet for spiking. It’s time to deliver a standard architecture, but with extremely high performance and efficiency.”
“By the end of the year, we’re going to have an internal chip tape-out, to verify the final configuration,” projects Del Maffeo. “I expect the complete chip to be taped out next summer, after which it will be delivered to partners for finetuning. We’re aiming for volume production and broad availability in 2023.”
War chest
Del Maffeo also foresees a steep rise in personnel. “By the end of this year, we should be at 30-35 people. By mid-June next year, 60-70. By the end of 2022, 100-120. We’re going to have offices wherever in Europe we find talented people to work with. The Netherlands will remain our home base, though, with the majority of our researchers stationed here. We’re now in Eindhoven, but we might open up a site in Amsterdam.”
The ambitious plans evidently require a sizable war chest. “10 million is a drop in the ocean,” argues Del Maffeo. “We need at least 100 million, but we have strong investors: Bitfury is a multi-billion company, Imec can invest up to 60M, Innovation Industries is a 250M investment fund that can go up to 25M. We’re going to open our Series A round in six months, with a Series B to follow probably somewhere in 2023.”
“It’s a big bet, but we consider it a safe bet,” concludes Del Maffeo. “The market is coming, the AI trend from cloud to edge is there, it’s inevitable. And it’s a huge opportunity for Europe. Forty percent of the papers in the field are coming from European researchers not working for European companies. The big European corporations are reluctant to take the risk and invest the money to take it to the next level. We see this gap and are jumping in.”