The Hidden Winners of Edge AI: Comparing 4 Stocks in Vision, Network, and Test

While robot brains and self-driving chips grab the headlines, I pay closer attention to the companies quietly holding up edge AI behind them. The eyes of the camera, the gateway in your home, the custom chip, and the equipment that tests all of it. Not glamorous — but if this layer breaks, there is no layer above it.

I will walk through four names with sharply different personalities, then compare them in a table at the end.

Ambarella — the purest edge vision

Ambarella is the purest bet there is on edge AI vision — the chip that lets a camera, a drone, or a robot actually see the world and decide on the device itself, with no trip to the cloud.

While almost everyone else treats edge AI as a slice of something bigger, Ambarella is the rare name where roughly 80% of revenue genuinely is edge AI. It has shipped more than 46 million vision chips into the field, an installed base nobody conjures overnight. It just signed a decade-long agreement with Hanwha to build Hanwha's security and robotics products around Ambarella silicon, and its newest CV7 chip pushes 8K vision with more than double the AI performance of the generation before. The honest read on the numbers: this is a pure play still investing hard for growth, so it is not throwing off big profits yet. But revenue re-accelerated about 37% over the past year while operating losses narrow quickly. You own it for where edge vision is heading, knowing earnings are still a step behind.

Calix — the front line inside your home

Ask where the physical edge is, and the answer is the gateway box your home Wi-Fi runs through. Calix builds the software platform that runs on those gateways for internet providers across the country — now with an AI layer that runs intelligence at the edge instead of a distant data center.

The story underneath is the real hook. For its entire first decade as a public company, Calix never turned a profit — it lost money every single year. Then it flipped the whole model, from a box seller into a recurring-revenue software platform that gets paid per subscriber, every month. Since that turn, revenue has more than doubled while gross margin climbed from the mid-40s to a record near 57% — plainly software-company territory. The most recent quarter set yet another revenue record.

Broadcom — the king of custom AI chips

Broadcom is the giant behind the custom AI chips powering data centers — what the industry calls ASICs, silicon built for a customer's exact workload. The only other real player is Marvell, and Broadcom is the clear leader, building around 60% of the world's custom AI chips.

The financial transformation underneath is striking. Over the last five years, profit is up roughly eight times and operating margin has climbed from 17% to 40% — meaning the company now keeps nearly twice as much of every dollar it brings in. And Broadcom just put a dedicated AI engine inside the Wi-Fi router and home gateway: the first Wi-Fi 8 chip with its own neural engine, so inference happens right inside the box in your house, with a 5G platform built alongside Samsung. That edge line is still small next to today's data centers, but it plants Broadcom at the front of the network edge while a profit machine compounds underneath.

Cohu — the name nobody thinks about until a chip fails

Cohu is the company nobody thinks about until a chip fails. Every processor and sensor headed into a car — every edge chip in the world — none of it ships until it has been tested and proven. Cohu sits right at that gate.

It is a cyclical name, and right now you are catching it as the cycle turns. Orders just jumped roughly 57% year over year, the kind of move you only see when the flood of edge and automotive silicon starts rolling across its equipment again. What makes it stick is the recurring side: once Cohu's handlers are installed on a production line, spares and services keep paying for years, and that piece now runs nearly 60% of the business. Its Eclipse platform is winning the test slot for high-performance AI processors at major names, pulling Cohu off the automotive island and straight into the compute story.

The four at a glance

Stock	Role in edge AI	Profitability	The point right now
Ambarella	Eyes of cameras, drones, robots (vision chip)	Still loss-making, losses narrowing	Revenue re-accelerating ~37%, a bet on direction
Calix	Software on the home gateway	Turned profitable, gross margin near 57%	Shifted to recurring revenue, record revenue
Broadcom	~60% of custom AI chips (ASICs)	40% operating margin, profit up 8x in 5 yrs	Neural engine inside Wi-Fi 8
Cohu	Test gate for every edge chip	Cyclical, services ~60% of revenue	Orders up ~57%, cycle turning

How I would combine them

From my seat, these four are an 'edge AI basket' with different risk-reward. Ambarella is a growth stock where the direction is right but earnings must catch up; Calix is a recurring-revenue turnaround from loss-maker to software margins; Broadcom is a stable compounder with an edge option bolted onto a dominant core; Cohu is a cyclical playing the turn.

One thing is clear: while the flashy robot brains (the brains of machines that move) soak up the spotlight, nothing above works without this layer. If you want broad exposure to the edge AI shift, I think holding these hidden winners alongside gives the portfolio balance.

FAQ

Q: Which of these actually makes money today? A: Broadcom (40% operating margin) and Calix (turned profitable, gross margin near 57%) clearly earn. Cohu is cyclical, so it depends on timing, and Ambarella is still loss-making but shrinking its losses fast.

Q: Why look at four stocks together? A: Because their roles do not overlap. Vision (Ambarella), network (Calix), custom compute (Broadcom), and test (Cohu) each hold a different chokepoint of the edge AI supply chain, so owning them together diversifies single-company risk.

Q: What is the biggest risk in each? A: Ambarella risks earnings never catching the story; Cohu risks the cycle rolling over again; Broadcom's risk is an already-elevated valuation. For Calix, the question is whether the recurring-revenue shift keeps holding.