· via dev.to (home feed)
Indian chip startup Bharat Semi raises $24M to run AI inference inside image sensors
Pre-revenue Indian chipmaker Bharat Semi has raised $24M in seed funding to build a single programmable chip that captures images and runs AI inference on the sensor itself.

Bharat Semi, a pre-revenue semiconductor startup based in India, has raised $24 million in seed funding to develop image sensors that run AI inference at the point of capture, according to Agentic Affairs, in a post republished on dev.to. The company's core idea is to merge the three components of a modern vision system into a single programmable chip.
One chip doing three jobs
In a typical camera pipeline, an image sensor records the scene, an image signal processor cleans up the raw output, and a separate system-on-chip runs the AI model, with visual data travelling between them at every stage. According to the report, Bharat Semi collapses that chain so capture, processing and inference all happen inside one component, letting a device interpret what it sees without first shipping images elsewhere for analysis.
A 70% cost claim, still unverified
Co-founder Vrinda Kapoor says the approach could cut system costs by at least 70% compared with conventional setups. The report is careful to frame this as a company projection rather than an independently measured figure: the underlying logic — fewer parts, less data movement, and no need for expensive compute elsewhere in the system — is coherent, but a startup with no revenue cannot yet prove it. The number only becomes meaningful once sensors ship and customers can benchmark it themselves.
Design wins, not sales
General Atomics and Kaynes Semicon, an Indian electronics manufacturer, have selected Bharat Semi's sensors for development programs. Per the report, that is genuine early validation — a defense contractor does not choose an unproven chip casually — but a design win is a selection, not an order. Whether those programs convert into paying customers is the milestone to watch once production begins in 2027.
Privacy built into the architecture
Because inference happens on the sensor, raw images never have to leave the device. The company's examples include a doorbell that recognizes a delivery, a drone that spots power lines, and an air conditioner that detects an empty room. As the report notes, in an environment where every connected camera is a potential data-governance liability, keeping the interpretation on the device may end up being the strongest selling point of all.
Roadmap and team
Bharat Semi has roughly 75 employees and operates a fabless model, concentrating on chip design and software while outsourcing fabrication. Its CTO previously led image sensor work at OmniVision, according to the report. Three things will determine whether this becomes a real product business: production starting in 2027, second-generation sensors reaching prototype stage by mid-2027, and the 70% cost claim holding up under real customer measurements.
Why it matters
The round is a sizable bet on a purer form of edge computing than most: rather than moving sensor data to where the AI runs, it moves the AI to where the data originates. If the single-chip approach delivers anything close to the claimed cost reductions, it would reshape the bill of materials for cameras, drones, defense systems and smart-home hardware — categories currently assembled around separate capture, processing and inference silicon. The privacy architecture matters too, since on-sensor inference removes the image-transfer step that makes connected cameras a data-exposure risk. And for India's semiconductor ambitions, a fabless startup landing defense design wins and eight-figure seed funding before shipping a single chip is a signal in itself — though, as the report itself stresses, every headline number here remains a promise until hardware is in customers' hands in 2027.
- #ai-chips
- #edge-computing
- #semiconductors
- #image-sensors
- #india