Edge AI platforms grow

Edge AI has spent years as a promise. A camera in a factory spots defects before a worker can. A sensor on a ship predicts failure before the engine room does. A retail freezer notices it is drifting warm and fixes the problem before inventory spoils. The trouble was never the demo. The trouble was everything after the demo: provisioning hardware in ugly places, updating models over flaky links, locking down devices that sit in stores and oil fields instead of data centers, and proving the whole stack can survive contact with a Fortune 500 change-control board. That is the gap ZEDEDA is trying to close. On March 16, the company launched what it calls an Edge Intelligence Platform, a package meant to build, deploy, secure, and operate AI systems across distributed sites at scale. The pitch is not just inference at the edge. It is the full operational layer around it: orchestration, governance, approval workflows, audit trails, hardware-aware optimization, and the ability to roll out autonomous agents across thousands of locations without treating each site like a special case. ZEDEDA says the platform builds on an installed base already managing tens of thousands of application instances, and it says customers have cut deployment times from weeks to minutes. (businesswire.com) That matters because edge AI is no longer a science-fair project inside big companies. ZEDEDA’s own March survey found that 83 percent of C-suite and IT leaders said edge AI is important to core business strategy, and 86 percent of enterprises with active edge AI deployments are already pursuing agentic capabilities. Half are still researching. But 21 percent are piloting autonomous multi-step agents, and 15 percent say they have already put them into production. The shift here is simple. Companies are moving from models that classify events to systems that take action. (secure.businesswire.com) Once that shift begins, the market stops rewarding point tools. It starts rewarding platforms. ZEDEDA had already moved in that direction last November with Edge Kubernetes App Flows, a managed stack that handles the operating system, cluster lifecycle, and application delivery for edge environments where normal cloud assumptions break down. In March it went a step further with Submer, pairing its software with modular, liquid-cooled infrastructure for high-density GPU inference in places where a normal data center is unavailable or absurd. The story is not that edge AI got smarter. It is that vendors are racing to own the boring parts that make smart systems deployable. (businesswire.com) Google is pushing on the same frontier from the opposite direction. Instead of starting with industrial fleets, it is turning the phone into an edge runtime. In late February, Google updated AI Edge Gallery with on-device function calling, using its compact 270 million-parameter FunctionGemma model to translate natural language into app actions like opening maps, creating calendar events, or toggling the flashlight, all offline. Then on April 2, Google expanded that push with Gemma 4 and a new Agent Skills feature in AI Edge Gallery, describing it as one of the first apps to run multi-step autonomous workflows entirely on-device. The same announcement tied Gemma 4 to Android’s AICore Developer Preview and to Google’s lower-level edge stack for mobile, desktop, and embedded systems. (developers.googleblog.com) This is why “edge AI” now means two things at once. In the enterprise, it means software that can survive warehouses, telecom closets, factory floors, and remote energy sites. In consumer tech, it means assistants that do not need a round trip to the cloud before they can act. The common thread is not model size. It is locality. Compute happens where the data is created, where latency matters, and where privacy or connectivity make cloud dependence expensive. Google’s own materials lean hard on that point, promising instant response, offline operation, and private inference on-device across Android and iOS through the open-source AI Edge Gallery. (developers.googleblog.com) That helps explain the smaller projects in the card, too. Real-time WebSocket updates for distributed intelligence explorers sound modest next to enterprise orchestration and mobile agents, but they solve the same underlying problem: if intelligence is spread across many nodes, users need a live view of what those nodes are doing. The edge stack is filling in from both ends now. At the top are vendors selling governance, deployment, and resilience to large enterprises. At the bottom are developer tools and open projects making local agents feel responsive enough to trust. In between sits the new expectation that distributed systems should stream their state as it changes, not dump logs after the fact. Google’s AI Edge Gallery is on the App Store and Play. ZEDEDA’s next public stop is an AI in Oil & Gas conference in Houston on April 8 and 9. (play.google.com)