The shift in processing from the Cloud to the Edge has signified a major shift in how data is processed, making way for higher bandwidth, lower latency, and real-time analytics; helped by the application of AI models to quickly and effortlessly sort through large amounts of data. The subject of Edge computing and all of its forms was discussed in length at Advantech’s Edge Computing Summit held in Munich last month.
How tech architecture has fundamentally changed was the guiding topic of IoT Analytics’ talk as Satyajit Sinha, Principal Analyst at the company took to the stage to talk through the evolution of Edge AI and key technology trends in the market.
Tech architecture has changed
Tracking the history of tech architecture, micro and mini computing emerged in the 1980s, spanning to the 2000s, at which point Cloud computing picked up, and 2020 was registered as the period in which Edge computing began to increase in popularity.
There are six key reasons why this has happened, Sinha said: demand from IoT devices, cost-effective computing, low latency requirements, data sovereignty and privacy, network constraints and autonomy, and AI/ML inference at the Edge.
An explosion in the number of IoT devices has been a major driver of processing data on the Edge, as the amount of data they generate can be challenging to process properly and draw insights from.
“We are right now running at 18 billion [IoT] connections, and all this sensor data … went to the Cloud,” said Sinha. “You can do that, but you’re not able to make all the decisions you want, and you’re not able to use all of the sensor information at the optimum [level].”
The definition of Edge was split into three key characteristics, in a nod to the fact that Edge computing is not all the same; thick Edge, which refers to the Edge of the network where a lot of computing power is applied; thin Edge which relates to network equipment; and micro Edge, at the lowest level, which can be attached to a sensor.
“One of the interesting analyses that we do in IoT Analytics … is what CEOs talked about, where we analyse all [of] the earnings calls and see what CEOS are mainly talking about,” said Sinha. “From 2021 to now, there is a surge in Edge AI … more than 488%.”
IoT Analytics’ tracking of the evolution of Edge AI began with TinyML, which can be run at the CPU level and doesn’t require NPU or GPU capabilities, followed by the cellular world entering into Edge AI, demonstrated by Qualcomm introducing its integrated GPUs and NPUs, as companies moved towards the thin Edge.
From Q1 2024 to Q1 2025, there has been a 77.8% increase in the total shipment of AI models in cellular IoT modules.
Key Edge AI trends
There are three key reasons why AI is coming to the Edge: cost, latency, and security.
“One of the key trends that we saw in the market is [that] hardware alone will not solve the Edge AI issue,” noted Sinha. “If you want to do Edge AI, your software needs to be … tightly integrated with your hardware.”
As a result of this, hardware vendors are focusing on developing their own Edge AI software stacks, with some of the vendors Sinha named including SIMCom, STMicroelectronics, NXP, Infineon, and more.
Another trend observed by the market research firm was that LLM inference is coming to the Edge, with real use cases spanning from real-time assistants to contextual search, without the need for hyperscalers.
The third and final trend is that zero-shot vision language models (VLMs) are enabling vision intelligence at the Edge, operating well in use cases such as video surveillance and quality control in food manufacturing.
“With this technology, it’s going to become so easy that [I can ask], ‘hey, I wanted to find out who’s wearing a white hat or white helmet in sector B,’ and I’ll get all the details of which people are wearing this hart, and … the timestamp,” explained Sinha.
In his closing remarks, Sinha acknowledged that Cloud computing is not going to disappear completely as a result of Edge processing’s popularity, and that it will still have a part to play, but critical decision-making will happen on the Edge.
“AI is rapidly advancing and moving towards the Edge,” he said. “Hardware vendors have clearly realised this, that … producing good hardware is a yes, but it will not solve the whole Edge AI problem.
“You need software attached to it, tightly integrated to it, and [to] develop the whole stack.”
There’s plenty of other editorial on our sister site, Electronic Specifier! Or you can always join in the conversation by visiting our LinkedIn page.
