The Hybrid AI Stack: From Your Pocket to the Base Station

The economics of owning AI inference extend beyond the desktop to every layer of the network, from device silicon to telecom base stations. A new investment in EdgeCortix by Axiro Semiconductor reveals the four-layer hybrid architecture that will replace cloud-only AI deployment for enterprise workloads.

I'm writing this on my Google Pixel. Its Tensor chip is making routing decisions right now, choosing what to process locally and what to send to the cloud. I don't think about it. It just works. That invisible orchestration is not a smartphone trick. It is the template for everything that comes next in AI infrastructure.

Last week I wrote about how a $2,000 mini PC can now run 200 billion parameter models locally. The economics of ownership were the point. But there is an entire infrastructure layer between that box on your desk and the hyperscale cloud, and a recent investment from someone I know made it click.

A connection, an investment, and a thesis

Jeff Grossman is at EdgeCortix. We go back to Network Solutions days. So when The Economic Times covered Axiro Semiconductor and MPower Partners investing in his company, I paid attention.

EdgeCortix builds AI inference processors. Their SAKURA-II chip delivers 60 trillion operations per second at 8 watts. That is up to five times more energy efficient than conventional edge GPUs for running generative AI. Their next-generation SAKURA-X platform, backed by ¥7 billion in Japanese government subsidies from NEDO, targets another five-fold efficiency gain and integrates radio access network acceleration for 5G and 6G.

Read that again. AI inference and telecom RAN processing on the same chiplet. That is convergence.

Axiro already ships millions of RF and connectivity ICs to Ericsson, Nokia, Hughes, and Siemens. CG Power invested $36 million to formally establish the subsidiary in December 2024. This is not a speculative bet. It is a connectivity company buying a seat at the AI inference table.

Why this problem matters

The cloud-for-everything approach that seemed inevitable two years ago is proving impractical for production AI workloads. Cloud metering models are economically painful for continuous 24/7 inference. Enterprises that moved aggressively to cloud AI are discovering the same thing I wrote about for individuals last week. You can own this. And the economics demand that you do.

But enterprise does not just mean a box under someone's desk. It means AI running in the 5G base station serving your building, AI running on the factory floor without data leaving the premises, AI running in the autonomous vehicle that cannot tolerate a 50ms cloud round-trip, and AI running in your home gateway keeping your data local by default.

The industry calls it the "Inference Flip." Capital is shifting from cloud-centric training compute toward power-efficient edge silicon. Training is periodic and bursty. Inference runs continuously, is sensitive to latency and cost, and increasingly needs proximity to the data it processes.

How the technology works

EdgeCortix built SAKURA-II as a purpose-designed inference processor, not a repurposed GPU. It achieves 60 TOPS within an 8-watt thermal envelope by optimizing every transistor for one job: running trained AI models in production. The architecture uses a proprietary compiler that maps neural network graphs directly to silicon, eliminating the overhead that general-purpose GPUs carry when they run inference workloads designed for training hardware.

SAKURA-X, the next generation, is a chiplet-based platform. It integrates AI inference with RAN signal processing on the same package, targeted for manufacturing at TSMC/JASM facilities in Japan. For telecom operators, this means a single piece of silicon handles both network workloads and AI inference at the base station. No separate accelerator card. No additional power draw. No extra rack space.

The four-layer hybrid AI stack I see emerging:

Layer 1, your device. Tensor chip, Apple Neural Engine, Snapdragon NPU. Real-time tasks. Voice, photos, predictive text. 10 to 15 TOPS. Milliwatts.

Layer 2, your home. The AMD Halo, the NVIDIA DGX Spark, the $2,000 box I wrote about. 200B parameter models running continuously. Data never leaving your house. 100 to 200 watts.

Layer 3, the telecom edge. This is where EdgeCortix sits. SAKURA-II at 60 TOPS in 8 watts, deployed in base stations and network infrastructure. Running AI inference for workloads that need lower latency than the cloud but exceed what your home box handles. Autonomous vehicles, industrial robotics, smart city systems.

Layer 4, the cloud. Training. Complex multi-step reasoning. Massive context windows. Periodic bursty workloads. Still essential, but no longer the only game.

Each layer handles what it is thermally, economically, and architecturally best suited for. The intelligence is in knowing what runs where.

What the numbers actually show

The edge AI accelerator market was $7.5 billion in 2024. It is projected to hit $35 to $38 billion by 2030, roughly 30 percent annual growth. Ultra-low-power devices, those drawing 1 to 3 watts, already account for over 80 percent of market shipments by volume. The market is not voting for raw performance. It is voting for efficiency at the edge.

The capital flows in 2026 tell the story. AI chip startups raised $8.3 billion globally. NVIDIA acquired Groq for $20 billion. The training compute king spent that much to buy an inference company. They see what is coming. EdgeCortix itself closed a Series B exceeding $110 million, oversubscribed.

The broader semiconductor industry could reach $1.6 trillion by 2030, up from $775 billion in 2024. AI and edge devices are the primary drivers. The money is moving to Layer 3.

The orchestration layer completes the stack

I have been covering Mimik for a while. They build the orchestration layer, the software that decides what runs where in this distributed mesh. If EdgeCortix is the silicon, Mimik is the routing logic.

Hardware alone does not solve the hybrid problem. You need a software stack that can discover compute resources across devices, edge nodes, and cloud. Route workloads based on latency requirements, power constraints, privacy needs, and cost. Handle failover seamlessly when one layer is unavailable.

My Pixel does this between device and cloud. Mimik does it across everything in between. As these layers multiply, home boxes, telecom edge, industrial gateways, the orchestration becomes the product.

Enterprise implications

Enterprise AI deployment is moving from cloud-only to hybrid cloud-edge-on-premises architectures. Agentic AI infrastructure demands distributed, low-latency compute at the edge. The procurement conversation is shifting from "which cloud provider" to "which layers do we own."

For telecom operators specifically, the Axiro-EdgeCortix alignment is a signal. Axiro already supplies RF silicon to the operators building 5G networks. EdgeCortix supplies inference silicon that sits in the same base station rack. Convergence at the silicon level means convergence at the procurement level. One supplier relationship, one thermal envelope, one maintenance window.

The pattern repeats at every scale. Mainframes became PCs. Centralized compute became distributed. Cloud-only is becoming hybrid. The economics always win eventually. By 2029, having an AI inference box will be as unremarkable as having a WiFi router. But the router will also have inference silicon in it. So will the base station serving your building. Every layer of the network becomes a compute layer.