The global edge AI market reached $26.9 billion in 2025 and is projected to hit roughly $128.5 billion by 2032, according to P&S Intelligence, with manufacturing growing faster than any other end market. The reason becomes clear once you look at what modern sensors produce: data that is too large and too time-sensitive to send back to a central cloud.
Alan Benjamin, president and CEO of GigaIO, has built his company around that problem. Alan joined me on “The AI Hedge” at the AI Infra Summit in Santa Clara to talk about why some of the hardest compute challenges in AI now sit outside the data center.
GigaIO launched in the summer of 2017 with a group of Southern California engineers who had started as the interconnect research group at Cray, moved to Sun Microsystems when Cray sold the team, and later scattered to Broadcom, QLogic, Emulex and Qualcomm after Oracle bought Sun. Alan saw them as the group that could pull off his vision.
That vision relied on using PCI Express in different and unique ways. Processors already speak PCI natively, Alan explained, so avoiding a conversion to another protocol delivers faster speeds, and the same interconnect can scale systems both up and out.
His engineers describe their Sun-era work as the NVLink of its day: incredibly performant for its time, very expensive and highly proprietary. GigaIO’s rack-scale approach, which attached 32 or 64 GPUs to a single server, worked well enough that d-Matrix bought that part of the business earlier this year.
GigaIO now takes the same technology into what Alan calls “high data environments” far from any data center. New sensors, including high-frequency vibration sensors and hyperspectral cameras, generate terabytes of data a day, and much of it is perishable. If operators learn what the data is saying within a few seconds or minutes, or 100 milliseconds in the most extreme cases, they can adjust manufacturing equipment in time.
Military and intelligence customers make up GigaIO’s biggest market, with uses that include drone detection, forward operating bases and crowd control, often with little or no cloud connection. Manufacturing, industrial processing, oil and gas, mining and automated agriculture are growing markets too.
For many of these customers, the appeal of AI has less to do with generation than with the pattern matching at the heart of the models. As Alan put it, “The model can say, I have no idea why this is true, but I can tell you from the past when your sensor started to do this, this is what happens.” That signal then feeds a digital twin that works out the cause. Speed matters, because in a drone attack, “you don’t have a minute to wait to transmit the data and get an answer.”
“Compute out at the very edge is getting more intense, but it's not generally an island unto itself,” said Alan. He thinks of the combined footprint as a distributed data center, one that operators must still control, keep updated and eventually pull data from to refine their models.
GigaIO sells two products that share an architecture and software stack in different form factors. The portable version has wheels and a handle and fits in the overhead bin of a commercial airplane. It includes a built-in 100-gigabit Ethernet switch that connects radios, wireline or fiber. The systems can keep multiple groups isolated on shared hardware, which fits the distributed virtual private cloud model Hedgehog has built. Alan also stressed independence from the hyperscalers’ edge offerings: GigaIO plugs into any cloud or all of them, and it can fit up to five petabytes of storage in a box. “So if you really want to store a bunch of data, it’s entirely possible within a small form factor,” he said.
To raise wireless bandwidth, GigaIO works with San Diego-based Mushroom Networks, whose technology bonds Wi-Fi, private 5G, consumer cellular and wireline channels into bigger pipes.
Physical AI may be where AI delivers its clearest return, because the payoff in lower costs, better quality and safer operations is easy to measure. Alan’s argument is that many operators still assume edge compute has to be limited, which leads them to pare down models and then struggle to keep different versions current across their equipment. GigaIO is betting that “bringing real data center class hardware to the very edge” removes much of that friction. The interconnect lessons learned in scale-up now apply in a forward operating base or on a factory floor, and the companies that pair that compute with secure, cloud-independent networking will set the pace.
Listen to the full episode of The AI Hedge here.