Five drones hovering above a forested valley at dusk, a glowing scan mesh covering the far ridgelines beyond them, their data running back to a rugged edge compute unit on a rock in the foreground.
Article Icon
Allyson Klein
@
TechArena
Nov 18, 2026

Built for the Edge, Axelera AI’s Europa Scales Up

Inference will account for roughly two-thirds of all AI compute in 2026, according to Deloitte’s 2026 Technology, Media & Telecommunications Predictions. Alexis Crowell, global chief marketing officer at Axelera AI, has firm views on where it should run, and she brought two chips to explain when she recently joined me for an episode of “In the Arena.”

“It’s always more fun to show versus just talk about,” Alexis said. On the table sat Metis, the company’s first-generation AI processing unit, and Europa, the larger chip Axelera AI launched that week.

Europa and Metis share one design philosophy: solve the hardest problems at the edge first, and the answers carry into bigger compute.

Solving for the Edge First

Both chips use a digital in-memory compute architecture that keeps memory and compute together, which Alexis said cuts power draw and raises throughput. Axelera AI started at the edge on purpose. She described three walls that get steeper as deployments move toward the data center: power and the thermals that come with it, performance and cost. “If you can solve for that at the edge, you can scale it into data center,” she said.

That approach traces back to founder and CEO Fabrizio Del Maffeo, who saw AI spreading across thousands of edge locations, where it only scales if it is affordable, easy to program and manageable from afar. Solving those constraints on day one, Alexis said, beats taking big compute and trying to shrink it.

Edge Workloads Are Changing

For close to a decade, edge inference mostly meant vision models running security applications. Now Axelera AI has customers placing Metis, which runs models of up to 8 billion parameters, in a battery-powered backpack for medical triage. Alexis noted that a specialized language model job like that doesn’t need a massive frontier model.

She also sees capacity limits in the data center pushing more compute toward the edge, where data stays private and responses come in real time. IDC estimates that global edge computing spending reached nearly $261 billion in 2025 and will grow 13.8% a year to $380 billion by 2028.

Bringing Inference Back On Premises

The same pressures are pulling inference workloads out of the cloud and back on premises. “A lot of it boils down to control,” Alexis said. “We’ve all seen the headlines of what’s happening with these jailbreak models, these agents that have decided they don’t care what their rules said.”

Companies want a firmer hold on their data and on their spending. Cloud inference means paying for the instance and then paying per token. With owned hardware and open-source models that now rival frontier models in efficacy, the cost becomes a capital purchase plus a monthly power bill.

Real-Time Response and Real Use Cases

Many customers never think about the SRAM built into the chip, and Alexis agrees they shouldn’t need to. “They should care about what sort of business problems this is going to solve for me,” she said. Keeping memory and compute together cuts data transfer and shortens response time. That matters for a factory robot that must stop the instant a person is in danger, or for drones that need to flag a threat the moment it crosses a perimeter.

Her favorite applications are medical. Alexis pointed to a World Health Organization figure of roughly 10 million fewer clinicians than the world needs, and the WHO’s current projection puts the shortfall at 11 million health workers by 2030. She sees AI helping with faster diagnosis and drug discovery. For her own productivity, she likes having a language model trained on her company’s data sitting on her desk, with nothing sent elsewhere.

Data sovereignty came up throughout the summit, and Alexis tied it to Europa’s high performance on a low power budget. One chip runs a model with 32 billion parameters, and customers who need more can combine several chips in one system to run larger models. Because that compute sits in their own environment, the data never has to go anywhere else.

Software That Spans Both Chips

Axelera AI also announced AxeleraScript at the summit, which lets customers bring any model, operator or weights to its hardware. Shortly after Alexis joined the company two years ago, she told me Axelera AI had a great software strategy. Now I understand what she meant. Customers can also use Voyager Wingman, an agent partner that opens up the full software stack through ordinary conversation, so even someone who hasn’t coded in years can put it to work.

Europa uses the same Voyager toolchain as Metis, and having worked in silicon, I know how much that continuity matters. Metis draws single-digit watts, while Europa averages around 30 watts. In her example, a fire department scouts California for wildfires with drones.

“Metis would sit on your drone, it could sit on hundreds of drones, and then Europa would sit back in kind of the command center,” she said.

Each drone identifies fires on board and sends back only the data that matters, and one Europa chip handles the whole fleet. Because one software stack runs across both chips, customers can split a pipeline between the two architectures and still run it as a single solution.

Where Inference Goes Next

Alexis expects inference to make up more than 80% of the market within a few years, embedded so deeply in electronic design automation, medical tools and other workflows that people stop calling it inference, much as no one thinks of predictive text on a phone as AI.

The TechArena Take

While a lot of inference will stay data centers on costly, power-intensive chips, Axelera AI is betting a meaningful share will run where data is created, and Europa is its case that edge-first engineering holds up as models grow. The company’s strongest advantage may be software, with one Voyager toolchain and AxeleraScript tackling the programmability problem that has held back many accelerator challengers. To learn more, visit Axelera AI’s website or view the full podcast episode.

‍

Subscribe to Our Newsletter

Read the latest in the world of AI, data center, and edge innovation.