Hailo

What edge AI is, and why it needs its own chip

Most of what is called AI today runs in data centers: your device sends a picture or a sentence to a remote server, the server computes, and the answer comes back. Hailo, founded in 2017, is built entirely on the opposite assumption: that the computing should happen where the data is collected, inside the camera, the car or the industrial equipment itself. This approach, called edge AI, has three practical advantages: faster response, because there is no waiting for the network; better privacy, because the video does not leave the device; and continuous operation even without a stable internet connection. The price is that the computing has to fit into the power budget of a small device, sometimes a battery-powered one, rather than the budget of a server room. A general-purpose processor or an ordinary graphics card is not suited to this, hence the need for a chip designed from the start for one task: efficiently running a trained model on a small device. There is also a privacy aspect worth knowing about. A security camera that sends video of people to the cloud raises questions that local processing inside the camera avoids to a large extent, simply because the image does not leave the device and what comes out of it is a conclusion, not a picture.

The products: the chip family and the software around it

Hailo's chips are tuned mainly for computer vision tasks (object detection and real-time video processing) and are aimed at makers of cameras, semi-autonomous vehicles and industrial systems. The product family includes several generations: Hailo-8 and Hailo-8L for especially low power use, and Hailo-10H and Hailo-15 for more advanced tasks. What surrounds the chip matters just as much as the chip. A customer who has already developed a detection model is not willing to rewrite it, so the company provides software tools such as the Dataflow Compiler that translate existing models into a format its hardware can run. This is in fact the real barrier to entry in AI chips: you can always design a faster chip, but if you have no tool chain that developers already know, you will not sell it. According to reports around its acquisition, Hailo had gained more than a hundred customers and a developer community of about ten thousand users. It is worth noting that the models these chips run are not large language models but compact detection networks trained in advance for a narrow task, which is exactly what makes it possible to run them on a small device. According to reports around the acquisition, Hailo's customers are concentrated mainly in East Asia and China.

From 2017 to unicorn: the funding rounds and the peak

Over the years Hailo raised about $340 million from investors that included, among others, the Zisapel family, OurCrowd, Alfred Akirov and Delek Motors. In April 2024 it completed a Series D round reported at about $120 million, which valued it at more than $1 billion and put it in the club of Israeli unicorns in AI chips. The gap between these two numbers is worth noticing, because it is a common source of confusion: the billion dollars is a valuation, the price investors agreed on in a particular deal, not money that came into the company. The money that came in over its whole life is the much smaller sum of about $340 million. At that peak, Hailo looked like one of the proofs that Israel could produce not only software but also original AI hardware, in a field dominated by very few players worldwide. A point worth making here: in chips, a high valuation does not necessarily indicate revenue. Investors price a chip company by the size of the market it targets and by the quality of its design, sometimes years before there are significant sales. That is exactly the gap that came to light later.

The fall: cash crunch, layoffs and a merger that collapsed

From that peak the company fell sharply, and the fall is a central part of the story, not a footnote. Hailo ran into a cash crunch, took emergency loans from its largest investor, Delek Motors, and cut about half its workforce. A planned merger with a special purpose acquisition company, a mechanism called a SPAC that serves as an alternative route to an IPO, collapsed when the company's value dropped below half a billion dollars; Delek Motors announced it would write off most of its investment in its financial statements. The basic reason, as described in reports around the acquisition, is not that the technology did not work but that it failed to generate significant revenue. This is a typical failure in chips: developing a new chip generation costs tens to hundreds of millions of dollars and takes years, while the customers, camera and car makers, decide slowly and order in quantities that justify the investment only after more years. A company with a good product can run out in that gap. It is also worth saying what did not happen here. There is no claim that the chips did not work, and no report of a technological failure. The gap was between the time the market took to mature and the time the company could fund, and that is a gap even a good product does not close.

2026: the acquisition by Microchip

On 24 July 2026 a definitive agreement was announced under which Hailo would be acquired by the American chip company Microchip Technology. The terms were not disclosed, and industry estimates speak of only a small fraction of the value the company reached at its peak. An indirect indication of the scale can be found in the buyer's own announcement, which said the deal was not expected to have a material effect on its financial results. Microchip stated that it intends to integrate the technology into drones, industrial robots and smart cameras, and to take on the existing customer base and developer community; at the time Hailo employed about 110 people. On 21 September 2026, earlier than the date set in the original announcement, Microchip announced that it had completed the acquisition. The completion announcement also did not disclose the terms, and again said the deal was not expected to have a material effect on its financial results. For Microchip, the acquisition also gives it a second development center in Israel, alongside the one it has had here since an earlier acquisition in 2018.

What this story teaches

It is easy to read Hailo as a story of failure, and that is a partial reading. Its technological bet proved right: running models on edge devices, without going through the cloud, has since become a central field. Security cameras, drones, home robots and cars all want local, real-time AI. What did not prove right was the assumption that a small independent company could fund this race on its own against corporations that also have other product lines to cover the losses. Hailo's approach is also fundamentally different from that of the big chipmakers. They specialize mainly in powerful, power-hungry accelerators meant for training models in huge data centers, while Hailo focused entirely on the other side of the equation: running an already-trained model on a small, power-limited device. Both fields are growing, but only one of them rewarded those who entered early. For the Israeli industry this is a two-sided story. On one side, the ability to design an original AI chip here and reach more than 100 customers worldwide with it is a rare achievement that few countries can show. On the other side, the unavoidable conclusion is that even such an achievement does not guarantee economic independence.