• Home
• News

Artificial intelligence has created an unprecedented demand for computing power, but the industry's reliance on enormous cloud data centers does not fit every application. A factory camera inspecting products in real time, a robot navigating a warehouse or a drone analyzing its surroundings may need to make decisions immediately, often under strict limits on electricity, cooling and connectivity.

Dutch semiconductor company Axelera AI, participating in the Expand North Star 2026 ecosystem through ScaleX, is building AI processors around the idea that a growing share of inference will need to happen much closer to where data is generated.

Founded in Eindhoven in 2021, the company has developed an architecture based on Digital In-Memory Computing, or D-IMC, designed to reduce one of the hidden costs of artificial intelligence: repeatedly moving enormous quantities of data between memory and processing units.

Its first-generation Metis architecture established Axelera in edge AI, while the new Europa processor significantly expands the company's ambitions into enterprise infrastructure and more demanding Physical AI workloads.

With more than 600 customers and more than $450 million attracted through equity, grants and venture debt since its incorporation, Axelera is no longer simply trying to prove that its semiconductor architecture works. It is trying to prove that the AI chip market has room for a European architecture built around efficiency rather than ever-increasing power consumption.

The AI Industry Has a Data Movement Problem

Modern neural networks perform huge numbers of mathematical operations, but computation itself represents only part of the challenge.

Data must constantly move between memory and the units performing those calculations. As AI workloads become larger and more complex, that movement consumes energy, generates heat and can restrict performance.

Axelera attacks this problem by bringing computation closer to memory.

Its proprietary Digital In-Memory Computing architecture allows operations used in neural-network inference to be performed within the memory architecture rather than requiring data to travel repeatedly between separate processing and storage components.

Reducing unnecessary data movement can lower energy consumption while allowing substantial amounts of inference processing to take place inside devices with limited power and cooling capacity.

Why Inference Matters as AI Spreads

Artificial intelligence computing can broadly be divided into training and inference.

Training is the expensive process through which a model learns from large amounts of data. Inference begins after training, when the model is used to analyze new information and produce an answer or decision.

A computer-vision model may be trained to recognize vehicles using a large image dataset. Every time the deployed model examines a new video frame and identifies a vehicle, it is performing inference.

The same principle applies to speech recognition, industrial quality control, medical imaging, robotics and generative AI.

As artificial intelligence spreads into millions of real-world applications, inference becomes a continuous computing requirement rather than a one-time development process. Axelera's strategy is built around making that stage more economical.

Metis Took AI Out of the Data Center

The company's first-generation Metis AI Processing Unit was designed primarily for edge inference.

A Metis-based PCIe accelerator can provide up to 214 TOPS of AI performance while operating within a power envelope dramatically smaller than the large accelerators associated with centralized AI infrastructure.

TOPS refers to trillions of operations per second and is commonly used as one measure of AI accelerator performance.

An accelerator embedded inside an industrial computer, intelligent camera or autonomous machine has to operate within strict limits on power consumption, heat dissipation and physical size.

This makes performance per watt particularly important. The engineering challenge is not simply to deliver more AI processing power, but to provide enough performance within the energy and thermal constraints of the device itself.

A Factory Shows Why Edge AI Matters

Consider an industrial production line using multiple high-resolution cameras to identify manufacturing defects.

One architecture could continuously transmit video to a cloud data center, where AI models analyze the images before returning the result.

That approach can increase network requirements, make connectivity part of the critical production infrastructure and create recurring cloud processing costs. Sensitive manufacturing information may also need to leave the facility.

Edge inference changes the model by allowing cameras or nearby industrial computers to analyze much of the information locally.

Only relevant results need to travel elsewhere, reducing bandwidth requirements while allowing the AI application to continue operating even when external connectivity is limited.

For industrial systems requiring immediate responses, the reduction in latency can be even more important.

The Camera Becomes an Intelligent Machine

Moving inference to the edge also changes the function of everyday equipment.

A conventional security camera captures video for humans or another system to examine later. Add sufficient local AI processing and the same device can identify unauthorized access, detect unusual activity or classify objects as events occur.

In manufacturing, cameras can identify defective products before they continue along a production line. Traffic infrastructure can classify vehicles and analyze movement, while retail systems can measure occupancy and customer flows.

Robots and drones can use the same principle to interpret their surroundings without transmitting every sensor reading to a remote computing facility.

The broader transition is from devices that simply collect data to machines capable of interpreting that data where it is created.

Europa Takes Axelera Beyond Traditional Edge AI

Axelera made a significant step beyond its original market in September 2026 with the launch of Europa, its second-generation AI Processing Unit.

Europa doubles the architecture to eight second-generation AI cores and delivers up to 629 TOPS, alongside 16 RISC-V vector processing units and up to 64 GB of memory per chip.

The architecture is designed to scale from demanding edge deployments into enterprise servers, extending Axelera's technology into workloads substantially larger than those targeted by Metis.

Metis demonstrated that Axelera could run computer vision and other inference workloads efficiently near the point where data is generated. Europa is intended to carry the same efficiency philosophy into enterprise AI, generative models and increasingly sophisticated Physical AI applications.

Larger AI Models Can Move Closer to the Enterprise

Generative AI makes local inference considerably more challenging.

Running an object-detection model on a camera requires far less memory than operating a modern language or multimodal model.

Europa is designed to address that gap, with enough memory to support models of around 32 billion parameters on a single chip, depending on model configuration and quantization. Larger workloads can be distributed across multi-chip systems.

This could give enterprises another option between small embedded AI systems and enormous GPU clusters.

Organizations may be able to operate capable generative AI models within their own infrastructure, particularly where the largest frontier models are unnecessary.

Many enterprise applications do not require the most powerful model available. They require a model capable of performing a specific business task reliably, securely and at an acceptable cost.

The Location of AI Is Becoming a Business Decision

Cloud AI provides enormous flexibility, but it also creates questions about cost, privacy and control.

An organization using external AI infrastructure may need to send information outside its own systems. For many applications this is acceptable, but healthcare, finance, government, defense and industrial organizations can face stricter requirements.

Local inference provides an alternative.

Sensitive information can remain inside the organization while models operate on infrastructure the company controls. Latency can fall, dependence on external connectivity can decline and recurring cloud costs can become easier to manage for persistent workloads.

This is why the location of AI processing is becoming almost as important as the choice of model itself.

Sovereign AI Is Moving From Governments to Companies

The same issue is usually discussed at national level through the concept of sovereign AI.

Countries want greater control over computing infrastructure, strategic datasets and the technologies supporting critical AI systems.

But a similar requirement increasingly exists within individual enterprises.

A manufacturer may want production data to remain inside its factories. A hospital may need strict control over patient information. A financial institution may face regulatory requirements governing where data can be processed.

Efficient inference hardware makes those architectures more practical because organizations do not necessarily have to choose between using AI and maintaining greater control over their information.

Axelera is positioning Europa around this intersection of performance, economics and data sovereignty.

Dell and Supermicro Bring Europa Closer to Enterprise Deployment

Semiconductor startups face a challenge that software companies can sometimes avoid: an impressive chip is not a complete product.

Customers need servers, accelerator cards, software, integration and technical support before the silicon can become part of production infrastructure.

Axelera's Europa launch therefore included validated systems involving Dell Technologies and Supermicro, alongside a wider network of OEM, ODM and technology partners.

This ecosystem is strategically important.

Enterprise customers are more likely to deploy alternative AI accelerators when those chips are available through recognizable server architectures rather than requiring organizations to construct specialized hardware environments themselves.

For Axelera, success will therefore depend not only on semiconductor performance but on how easily its technology can enter existing enterprise infrastructure.

More Than 600 Customers Are Already Using Axelera Technology

By September 2026, Axelera said it had deployments across more than 600 customers globally, spanning sectors including defense, robotics, drones, retail, security, telecommunications, aerospace and enterprise applications.

The company has also signed supply deals worth tens of millions of dollars and is pursuing a potential sales pipeline exceeding $1.5 billion.

That pipeline should not be confused with confirmed revenue or orders. It represents prospective commercial opportunities rather than completed sales.

Nevertheless, the figures demonstrate the scale of the market Axelera is now attempting to address.

The company's challenge is shifting from attracting interest in an alternative AI architecture toward converting that interest into large, repeatable deployments.

AI Factories Are Becoming Part of the Business

Europa has also pushed Axelera into a market that would have appeared distant from its original edge-computing focus.

The company is participating in European AI Factory projects, including initiatives in Italy and Luxembourg, as the European Union expands access to domestic AI computing infrastructure.

This creates an interesting progression.

Axelera began by arguing that AI should move out of centralized data centers and closer to physical devices.

It is now taking the efficiency principles developed for those constrained edge environments and applying them back inside larger enterprise and AI Factory infrastructure.

Rather than abandoning edge computing, the company is attempting to scale its architecture upward.

Software May Decide Whether the Chip Succeeds

The history of semiconductor computing shows that hardware performance alone rarely determines the winner.

Developers need tools that allow existing models to run on the processor without forcing them to rebuild entire applications.

Axelera addresses this through its Voyager SDK, which provides the software environment for deploying AI models across its hardware.

The same software environment spans Metis and Europa, allowing applications to move across different levels of Axelera hardware without requiring developers to begin again for each processor generation.

Companies evaluating a new accelerator will consider not only performance and power consumption, but how much engineering work is required before an existing AI application actually runs.

Axelera Is Using AI to Make Its AI Chips Easier to Use

The company added another layer to its software strategy in July 2026 with the public release of Voyager Wingman, an AI assistant designed to accelerate application development.

Deploying models onto specialized AI hardware can involve model export, compilation, pipeline configuration, optimization and performance analysis. Wingman is intended to assist developers through parts of that process.

The product illustrates an important change in semiconductor competition.

Hardware companies increasingly need to reduce the technical friction surrounding their processors, and generative AI can itself become part of that developer experience.

The easier it becomes to move an existing model onto a new architecture, the lower the barrier facing an alternative semiconductor platform.

Physical AI Could Be Axelera's Biggest Opportunity

The emergence of Physical AI may ultimately make edge computing far more important than it is today.

The first wave of generative AI largely transformed software. Physical AI takes machine intelligence into robots, drones, industrial systems, autonomous vehicles, medical devices and other machines interacting directly with the real world.

These systems have different requirements from a cloud chatbot.

A robot navigating a warehouse needs to process sensor information quickly enough to react to obstacles and changing conditions. A drone may operate with limited network connectivity, while industrial machinery may need to make safety-related decisions locally.

The closer AI moves to physical action, the more important latency, energy efficiency and local processing become.

Robots Have a Computing Energy Budget

Power consumption becomes particularly important in mobile machines.

A robot or drone has a finite energy supply, and computing competes with motors, sensors, communications and other systems for that power.

Higher computing consumption can also create additional heat, requiring more cooling and potentially adding size and weight.

This means the most powerful processor is not automatically the most useful processor for Physical AI.

Performance per watt becomes a fundamental design variable.

Axelera's edge-first architecture was created around precisely this constraint, giving the company a potential advantage if AI adoption accelerates across autonomous and robotic systems.

Billions of Edge Devices Could Create a Second AI Energy Challenge

The enormous electricity consumption of AI data centers has become a global infrastructure issue, but widespread edge AI could create another energy challenge.

A single smart camera consumes little compared with a data center. A factory containing hundreds or thousands of intelligent cameras, sensors and machines changes the calculation.

Scale that across manufacturing, transportation, retail, cities and logistics, and the collective electricity demand of edge intelligence becomes significant.

Efficient inference therefore matters not only for individual device economics. It could determine how extensively AI can be deployed across physical infrastructure without creating unsustainable energy and cooling requirements.

Titania Reveals How Far Axelera Wants to Go

Europa is not the end of the company's roadmap.

Axelera is also developing Titania, a chiplet architecture intended to extend its Digital In-Memory Computing approach toward high-performance computing, supercomputers and large AI infrastructure.

The project has received support of up to €61.6 million through the EuroHPC Joint Undertaking and participating states as part of Europe's wider efforts to develop greater technological sovereignty.

If Metis represents the edge and Europa connects edge computing with enterprise infrastructure, Titania points toward the highest end of the market.

The roadmap suggests Axelera is attempting to create an architecture capable of scaling from compact Physical AI systems to major computing facilities.

Europe Wants More Control Over the Chips Behind AI

This gives Axelera strategic significance beyond its individual products.

Europe remains dependent on overseas suppliers for much of the advanced semiconductor technology driving artificial intelligence.

The AI boom has made that dependency increasingly important because access to computing infrastructure now influences industrial competitiveness, scientific research, defense capability and national technology strategy.

Axelera is one of a growing group of European semiconductor companies attempting to challenge that dependence.

Its position is particularly interesting because the company is not attempting to reproduce the entire GPU market. It is concentrating on inference and energy efficiency, where changing workload requirements may create room for alternative architectures.

More Than $450 Million Is Backing the Strategy

Semiconductor development requires levels of capital rarely seen in conventional software startups.

Chips need to be designed, manufactured, tested and integrated into complete systems, while each new generation can require years of engineering before commercial deployment.

Axelera announced more than $250 million in new financing in February 2026, led by Innovation Industries and involving new investors including BlackRock and SiteGround Capital alongside existing backers.

That took the total capital attracted by the company through equity, grants and venture debt to more than $450 million since its incorporation in 2021.

The financing gives Axelera greater capacity to expand production, develop its software ecosystem and commercialize Europa internationally.

The Middle East Is Becoming Part of Axelera's Expansion

The company's Expand North Star appearance also comes as it strengthens its position in the Middle East.

In September 2026, Axelera announced a collaboration with PwC Middle East aimed at helping organizations in the region move from AI experimentation toward practical deployments.

The partnership is strategically relevant because Gulf countries are investing heavily in artificial intelligence, sovereign computing, smart cities, robotics and industrial digitalization.

Much of the region's AI investment has focused on large-scale cloud and data-center infrastructure.

As artificial intelligence moves deeper into factories, transport systems, energy infrastructure and autonomous machines, demand for local inference could create a second layer of the market.

Why Expand North Star Matters for Axelera

Axelera is being highlighted through ScaleX, the Expand North Star platform focused on technology companies moving beyond the earliest startup stage.

That description fits the company increasingly well.

Axelera is no longer raising capital to prove a laboratory concept. It has commercial hardware, hundreds of customers, a second-generation architecture, major technology partners and a growing international sales pipeline.

Expand North Star can therefore serve a different purpose for Axelera than it does for a seed-stage startup.

The opportunity lies in meeting enterprise customers, system integrators, government organizations, investors and regional partners capable of accelerating commercial deployment.

Dubai also places Axelera inside a market where both sides of its strategy are relevant: enormous investment in centralized AI infrastructure and growing demand for intelligence at the edge.

What to Watch at Expand North Star 2026

Axelera has not publicly announced a new processor specifically scheduled for launch at Expand North Star 2026, so any additional hardware announcement should remain unconfirmed until formally disclosed.

The recently launched Europa architecture is therefore the most important technology to watch.

Its significance lies in whether Axelera can successfully move beyond specialized edge inference and establish a meaningful position in enterprise AI infrastructure.

Metis remains important for computer vision, industrial systems and Physical AI. Europa expands the company toward larger generative and multimodal workloads, while Voyager and Wingman address the software barrier surrounding alternative processors.

Titania, meanwhile, indicates that the company's ambitions extend further into high-performance computing.

Together, these technologies reveal a company attempting to build a complete inference architecture rather than a single chip.

Axelera Does Not Need to Beat NVIDIA Everywhere

Any company entering the AI accelerator market inevitably faces comparison with NVIDIA, whose hardware and software ecosystem has become central to the global AI boom.

For Axelera, however, success does not require replacing NVIDIA across every AI workload.

The market is becoming too large and too diverse for every application to have identical computing requirements.

Training frontier models inside enormous data centers is fundamentally different from analyzing camera feeds in a factory, operating a robot or running a specialized enterprise model on private infrastructure.

Axelera's opportunity lies in those differences.

Its architecture focuses on inference, energy efficiency and local processing, particularly where power, cost, latency and data control matter as much as absolute computing performance.

If AI continues spreading from data centers into physical infrastructure, those constraints will become increasingly important.

The Next AI Infrastructure Race Will Not Be Confined to Data Centers

Generative AI taught millions of people to think of artificial intelligence as a service delivered from the cloud. A request travels to a remote data center, powerful processors run the model and the result returns to the user.

Physical AI requires a different architecture.

Factories, robots, drones, vehicles, cameras and industrial equipment generate enormous amounts of information that cannot always be transmitted elsewhere before a decision is made. Enterprises also increasingly want greater control over where sensitive information is processed and how much continuous inference costs.

That means the next stage of AI infrastructure will develop in two directions simultaneously.

Data centers will continue becoming larger and more powerful, but intelligence will also move outward into the devices and machines interacting directly with the physical world.

Axelera AI is betting that this second transformation will create room for a new generation of semiconductor architecture.

The next AI chip race will not be decided only by who can build the most powerful processor. It may also depend on who can make powerful intelligence efficient enough to run almost everywhere.