...
By Catherine Howe
0
Share
Newsweek is a Trust Project member
See more of our trusted coverage when you search.Prefer Newsweek on Googleto see more of our trusted coverage when you search.
Axelera AI is a semiconductor company that develops hardware and software to enable artificial intelligence to run anywhere your data is created, with high performance and power efficiency.
Which applications or markets for edge AI solutions are generating the most interesting conversations for you at the moment, and why?
The conversations I find most interesting are rarely with AI companies. They are with people who run factories, ports, farms, hospitals, rail networks and retail chains. Meaning, companies that are adopting AI to build a better business.
Read More on Created For Sponsor
Industrial manufacturing and quality inspection are furthest in adoption, followed closely by security and public safety, retail and enterprise. We have now shipped to more than 600 customers across those sectors.
Customers across these sectors are looking for AI sovereignty. In Europe, the Middle East and increasingly in the United States, customers want to know exactly where their data is processed and who can reach it. That is both a governance and technical question.
What keeps me engaged is the human specificity of these deployments. One of our partners published a case study on a factory that increased worker safety and protection by roughly a quarter using vision models running on Axelera technology. That is not an abstraction. That is a smaller number of people getting hurt at work.
How is AI architecture evolving? What belongs in the cloud, what belongs at the edge and what increasingly sits somewhere in between?
It really depends on what problem someone is trying to solve. Training large frontier models belongs in the cloud, and I do not expect that to change. The capital intensity and the scale are simply better served centrally. Inference is a different story. Inference is the workload that runs forever and repeatedly, and lifetime inference cost can exceed training cost by a wide margin. That economic reality is what is redistributing the compute architecture.
Where the compute lands should be largely dependent on where the data is created, what is the speed and repetition at which that data needs to be analyzed, and how private is that data. For applications where regulation and privacy are of high concern, or latency really matters, that inference should sit right next to that data input. As an example, if you have a drone out searching for a missing person, streaming back every frame caught on camera will ensure that the battery life of that drone is minimized. But running on-board AI to identify key elements (a person, backpack, etc) first, and then only sending back important notifications will allow longer-life and better search and rescue.
In the context of an enterprise, running their own internal language models, or chatbots to support their business can drastically improve their TCO as hardware options exist now to support those applications, and models that deliver high accuracy can be run locally.
Axelera talks about performance per watt, but what does energy efficiency actually mean at the system level?
I try and look at things in terms of outcomes for a company. A few examples of where energy efficiency really matters include robotics and drones, enterprises with a data center they cannot rebuild and retrofits of installed equipment.
With robotics and drones, on a battery-powered platform, every watt spent on inference is a watt not spent on acting. Efficiency here converts directly into runtime between charges, or how much can be accomplished. For an autonomous mobile robot in a warehouse, an extra ninety minutes per shift is the difference between a fleet of eight and a fleet of ten. For an inspection drone, it is whether the whole span of the bridge gets covered in one flight. Those are real business outcomes, directly impacted by energy efficiency.
The conversation I have most often in the Americas is around enterprises with a data center they cannot rebuild. A company with an on-prem data center has racks, air handling, a floor loading limit and a power allocation already determined. They want AI inference, and the standard answer is a new class of machine that needs liquid cooling, reinforced flooring and a power upgrade none of which they can justify or schedule. When inference fits into the existing air-cooled chassis and the existing power envelope, the project stops being a capital construction program and becomes a procurement decision.
On retrofits of installed equipment—most cameras, gateways and industrial PCs in the field today were specified years before anyone planned to run AI on them. But in many cases, those machines have an extra/empty compute slot that if an AI module can fit both physically and within the power-budget of that machine, that company can upgrade its fleet without starting over.
These are just a few examples, but the more we lean into a better energy-efficiency conversation, the better the outcome for the companies and humans adopting these technologies.
One of the big promises of generative AI is that it will become embedded into physical systems. What has to change in the hardware and software stack before we can realistically have AI everywhere, not just in a chatbot?
On the hardware side, the industry spent several years optimizing for raw compute when the actual constraint at the edge for GenAI is memory. Generative and multimodal workloads are bandwidth and capacity hungry in ways that classic computer vision was not. That is why our next-generation architecture, Europa, put so much emphasis on memory and memory bandwidth. Alongside that, embedded systems need realistic power envelopes, passive or modest cooling, low TCO and supply commitments measured in years, because industrial equipment has a ten-year lifespan.
The harder problem is software. A developer should be able to bring a model, compile it, quantize it without losing meaningful accuracy, deploy it, monitor it in the field and update it later without rebuilding the system. Every one of those steps is where projects stall. It is the reason we invest heavily in the Voyager toolchain; it shortens the path from model to running application. We also have a thriving partner network that supplies the system integration we do not build ourselves.
The last piece is trust. Physical systems act on and in the world. That raises expectations for security, functional safety and verifiability that a chat interface never has to meet. That work is unglamorous, but it is the gate to true scale.
If AI inference becomes dramatically cheaper and more energy-efficient, what applications suddenly become commercially viable that simply aren't viable today?
There are two areas I am most excited to see change: inspection coverage and access.
Today, most companies do a sampling to test product ahead of it coming off a manufacturing line, or a fraction of the video footage, or a section of a structure. Sampling happens because without automation and AI, it is too expensive to process full inspections by hand. But when the costs drop far enough—the cost per image, cost per video stream, cost per weld—now the coverage rate goes to 100 and consumers and businesses produce higher quality products, lower defect rates and hopefully, fewer recalls and consumer concerns.
I believe AI has the potential to be as influential as access to stable electricity has been on a community, company or economy. And yes, I know that’s a big statement, but the productivity gains and ability to drive better outcomes for your citizens or customers is becoming table stakes. But without access, the chasm between the “haves” and “have-nots” just widens.
This article has been paid for by a third party. The views and opinions expressed are not those of Newsweek and are not an endorsement of the products, services or persons mentioned.
View 0 Comments
Submit Correction
View Editorial & AI Guidelines