Most restaurant kitchens already have cameras in them, and Wobot.ai, a company in New Delhi, India, sells software that reads what those cameras see. It checks whether staff are following the chain's own rules, such as wearing gloves, cleaning a station on schedule or packing an order the way the brand requires. Until now that checking has happened on distant servers, with the footage uploaded first.
It will now also run on a chip, inside the restaurant. Wobot has put its software on Modalix, a processor built by SiMa.ai of San Jose, California, so the video can be examined on the premises instead of being sent anywhere. The two companies announced the arrangement on 6 October and are showing it at VISION 2026 in Stuttgart this week.
Customers can take Modalix as a card that slides into a server they already own, or as a small board built into a purpose-made box for the back office. Either way, the computer doing the looking sits in the restaurant.
SiMa.ai also supplies the software that sits around the chip, called Palette Neat, which it describes as a ready-made environment for building this kind of application without months of engineering.
"Our collaboration with SiMa.ai makes our reasoning-based video analytics available on the power-efficient Modalix platform," said Will Kelso, president for revenue and growth at Wobot.ai. Durga Peddireddy, who runs product management and partnerships at SiMa.ai, pointed to the network of system builders around the company as the reason customers "can bring deployment-ready systems to market quickly."
Reasoning Replaces Retraining
The word carrying the weight in that announcement is "reasoning", and it marks a real change in how camera software gets built. The older approach had to be taught by example, one example at a time.
To spot a worker without gloves, an engineer gathered thousands of photographs of gloved and bare hands, labelled every one of them, and trained a detector on the set. Each new rule meant another dataset, and changing a rule meant collecting the pictures again.
Instead of being trained on examples, a vision language model takes in the whole scene and answers a question asked in ordinary English. A new check can then be written as a sentence rather than assembled from a fresh set of labelled pictures.
For Wobot's customers that difference is close to being the product itself, because a restaurant chain changes its procedures constantly. A new menu item, a hygiene step or a packaging rule each arrived with a collection and training bill under the old method, often costing more than the check was worth.
What neither company has said is how well any of it performs. There is no word on which model runs on Modalix, how large it is, how many camera feeds a single chip can handle, or how often it gets the answer right.
Why The Video Stays On Site
Camera work is the clearest case for doing the computing in the building rather than shipping it off, and the reasons arrive in the same order every time. Bandwidth comes first: a chain with dozens of cameras at each of hundreds of sites pays every month to push that video upstream. The bill grows with every camera added, while a processor is bought once.
Then there is what happens when the line goes down. A restaurant whose internet has dropped keeps its checks running if the analysis sits on the premises, and loses them entirely if it does not.
The footage in question shows staff at work and customers at the counter. Video that never leaves the building takes a whole set of questions about who stores it, and who can look at it, off the table.
Billing follows the same logic, since cloud analysis is charged by use and the cost climbs with every camera and every rule checked. A box bought outright turns that into a single purchase, and it is the step from a sensor that reports what it sees to a system that acts where it stands. SiMa.ai builds for that power budget, aiming at devices drawing 5 to 25 watts, roughly what a machine can use inside a ceiling box or a small appliance without a fan.
The Partners Are Wildly Different Sizes
The two companies are nowhere near the same size, and that shapes what each of them takes from the deal. SiMa.ai is based in San Jose, California, and was founded by Krishna Rangasayee, formerly chief operating officer at the chip company Groq.
It raised $150M in September in a round co-led by Fidelity and Amplify at a $1.45B valuation, bringing its total funding to $500M. Bosch, Emerson, Micron, Synopsys and L&T Technology Services already appear on its partner list.
Wobot is a far smaller business, founded in the Indian capital in 2017 by Adit Chhabra with Tapan Dixit and Tanay Dixit, and funded with under $6M from Titan Capital and Peak XV. Revenue in the year to March 2025 was ₹13.3 crore, about $1.4M, which sits oddly beside a customer list carrying Whataburger, IRCTC, the catering arm of Indian Railways, cult.fit, Rebel Foods and the Gulf restaurant operator Kitopi.
SiMa.ai gains a working application with recognisable brands attached, which is what a chip company needs when it sells to buyers who have never heard of it. Wobot gains hardware it did not have to design itself, and a second way to sell the same software.
The gap in size also marks the limit of the news. A chip company announcing a software partner has not announced a customer, and nothing published says how many Modalix units Wobot expects to ship, or when the first restaurant group will switch one on.
The Real Contest Is With Nvidia
Whatever Wobot brings, SiMa.ai is selling into a category where most buyers reach for Nvidia first. Nvidia's Jetson line sits inside a large share of camera and robotics products and carries a developer community nobody else matches, while Hailo, Ambarella, Qualcomm and Axelera all sell rival chips into the same cameras. SiMa.ai's case against them rests on how much work it gets out of each watt of power, plus software that shortens the build.
That is the kind of case a market settles with measurements, and this announcement carries none. No throughput figure, no power draw under load, no frames per second, no accuracy on the tasks Wobot actually sells, which leaves it a statement that something is available rather than a result.
What the week does show is the direction the camera software business is taking. The analysis is moving out of the cloud and into the building, and models that had to be trained on examples are giving way to models that can be told in plain words what to look for. The companies writing that software are now choosing which silicon it runs on, and Stuttgart will show whether this pairing works.
