Solving AI's energy problem isn't just about building more power plants—it's about deploying more efficient compute inside existing facilities. Our CMO & GM of the Americas, Alexis Crowell, took the stage at @Newsweek's #ClimateWeekNYC alongside leaders from @AWS and @QTSDataCenters to discuss how high-efficiency inference hardware can help meet surging AI demand without sacrificing sustainability. Great insights and incredible views from 72 floors up at One World Trade! Read more on it here: eu1.hubs.ly/H0yBypW0 #GreenAI #SustainableAI #AIInfrastructure #AxeleraAI

Most edge AI deployments hit an invisible wall: vendor lock-in. When teams build custom AI models, they often wait months for hardware vendors to support them. Today we are introducing AxeleraScript (AxScript): an industry-first programming model delivering full silicon programmability at maximum performance. AxScript lets you compile and run any custom AI workload directly on Axelera hardware. Key benefits for enterprise leaders: • Accelerated Time-to-Market: Port complex architectures in days using AI-assisted workflows. • Zero Vendor Lock-In: Deploy proprietary models without waiting on vendor roadmaps. • Future-Proof Investment: Write once to target Metis, Europa, and future silicon like Titania. The future of edge AI belongs to teams who control their own models. Read the full blog post by Alexis Crowell: eu1.hubs.ly/H0yyJq60 (eu1.hubs.ly/H0yyJpl0) #EdgeAI #EnterpriseAI #AIInference #Semiconductors #AxeleraAI #modelzoo

"The conversations I find most interesting are rarely with AI companies. They are with people who run factories, ports, farms, and hospitals." — Alexis Crowell in Newsweek Discover how Axelera AI is delivering high-performance Edge AI to more than 600 customers worldwide: Read the full interview in Newsweek: eu1.hubs.ly/H0yny8h0 #EdgeAI #AI #Semiconductors #TechNews

Day 1 at #AIInfraSummit was incredible! CEO @fdelmaffeo took the stage to announce Europa, our next-gen AIPU architecture delivering up to 6x more tokens/s/watt than GPUs, featuring validated systems with @DellTech and @Supermicro! We also highlighted AxeleraScript (AxScript), our breakthrough model compilation approach simplifying deployment from edge to server. Great to see the coverage in @Reuters, @HPCwired and @YahooFinance! The booth was buzzing all day. Come see Europa live in action at Booth 930 for Day 2!

We are launching Europa, our next-gen AIPU built for Physical and Enterprise AI. Europa delivers up to 6x more tokens/sec/watt than GPU-based solutions, bringing high-performance inference to infrastructure you control. Available as standalone silicon, HHHL PCIe (Edge 232p), and FHFL PCIe (Server 250p), with validated OEM systems from Dell, Supermicro, HPE, Lenovo, and more. We are also introducing AxScript, a first-of-its-kind DSL that gives developers direct kernel-level control over our silicon. Instead of waiting for a vendor to update a supported model list, AxScript lets you compile custom architectures—from LLMs to VLMs—at full hardware performance. In fact, an engineer used AxScript to port Qwen3.5 in a single day. Read the full announcement: eu1.hubs.ly/H0ykNwF0 #AI #DeepTech #Hardware

About 12,900 active satellites are in orbit, with 70,000 more low Earth orbit launches forecast within five years. Most spend their time out of contact with the ground, which makes onboard autonomy a requirement. Our work with AIKO: eu1.hubs.ly/H0xLpVW0

Proud to see Axelera AI on the Defence Drone Initiative (DDI) Marketplace list via CanadaBuys. Bringing high-throughput, low-power AI acceleration onboard uncrewed platforms for the Canadian Armed Forces and Coast Guard. Details: eu1.hubs.ly/H0x_5Jd0

Compiled your model and the output shapes look wrong? They're probably right. The AIPU needs specific dimensions, so the compiler pads them and separates the post-processing into a graph you run yourself: eu1.hubs.ly/H0xLqkK0

WG Tech Solutions did not deploy one model. They deployed around 45, each tuned to its own camera angle, station layout and workflow. Across an ODM's factories it cut worker safety violations by 28%: eu1.hubs.ly/H0xLq7y0

Voyager ships with 100+ pre-trained models. Fine, until the one you need isn't in it. So which model do you wish is optimized on your hardware out of the box? Tell us, or log it on the community Launchpad: eu1.hubs.ly/H0xLpRm0

Multi-modal vision and agentic models are pushing edge hardware to its limits on power and thermals. In 2 weeks at AI Infra Summit 2026 (Santa Clara, Sept 15–17), see how Axelera AI solves real-world edge constraints at Booth #930: • Multi-tier high-throughput inspection • Real-time video analytics without cloud latency • Voyager Wingman for fast model compilation & optimization Catch CEO Fabrizio Del Maffeo's talk on "The Physical AI Inflection Point" to see how dedicated silicon scales real-world intelligence. Book a 1-on-1 demo: eu1.hubs.ly/H0x-ZmQ0 #AIInfraSummit #PhysicalAI #EdgeAI #ComputerVision

What was the least glamorous part of getting your AI pipeline running? Kernel versions, camera timing under load, buffer sizes, the config file that took three attempts. Tell us how you built around it. Someone in the community has hit the same thing: eu1.hubs.ly/H0xLq4N0

You don't have to adopt our pipeline to use Axelera tech. AxRuntime for model-level inference in C++ or Python. AxInferenceNet for custom pre and post-processing. InferenceStream for the full YAML route. Pick your level: eu1.hubs.ly/H0xF4zb0

A card that won't enumerate. A throughput ceiling you can't account for. A blank YAML file. Voyager Wingman writes the pipeline from a plain-language description, runs it on your device, and helps you dig into what's blocking you. Free: eu1.hubs.ly/H0xHqhQ0

Person-of-interest ID and weapon detection running simultaneously across two 8K feeds. Three Axelera Server 150p cards, 48 AIPU cores, 1,440+ model inferences a second at 30-58W per card. Full build: eu1.hubs.ly/H0xF4Bb0

Model accuracy dropped after compiling? In our community threads it usually isn't the quantization. More often it's RGB where RGBA was expected, or post-processing the compiler detached. Run the same dataset through torch, torch-aipu and gst to find out: eu1.hubs.ly/H0xF5tn0

Medical imaging data is exploding, leaving trained radiologists facing severe burnout and long backlogs. AI doesn't replace doctors, it empowers them. Our partner @MicronTech highlights how Axelera Metis® AIPUs and Micron LPDDR4 memory deliver real-time scan analysis right at the edge. Read more eu1.hubs.ly/H0xHqR20 #EdgeAI #HealthTech #Radiology

Someone lost days to PCIe enumeration failures on an RK3588 board. Another community member on the same board posted what actually worked: mainline kernel 6.18, no kernel parameters, 165 FPS on YOLOv8s at 23.5ms. Thread: eu1.hubs.ly/H0xF5zw0

At AI Infra Summit 2026 (Santa Clara, Sept 15–17), explore how physical AI architecture handles natural language video search and multi-tier edge workloads. Don’t miss CEO Fabrizio Del Maffeo's talk: "The Physical AI Inflection Point." Booth #930 | Book a demo: eu1.hubs.ly/H0xDstK0

One camera, no stereo rig, no lidar, and a dense depth map at 30 FPS on 1080p. Sanket Shah deployed ZipDepth (CVPR 2026) on an Axelera AI accelerator. Four stages, two on the AIPU at 110 FPS, two on the host. Same frame rate the paper gets from a 15W Jetson. How he did it: eu1.hubs.ly/H0xw1V00