The Machines That Teach AI Chips to Behave

A finished semiconductor can look immaculate and still be wrong. A film may be too thick by a fraction of a nanometer. A copper bump may lean. A buried void may wait inside a package like a hairline crack in glass. The trouble is not merely that the defect is small. It is that chipmaking stacks hundreds of precise steps on top of one another, each adding time and value. Find the mistake late and the factory may discard weeks of work. Find it inline and engineers may rescue the process.

This is the unphotogenic corner of the AI boom where Onto Innovation works. The Wilmington, Massachusetts company builds inspection machines, optical metrology systems, advanced-packaging steppers and factory software. In plain English, its equipment looks for flaws, measures structures, prints patterns onto large substrates and turns torrents of production data into decisions. The company is not selling intelligence in a chat window. It is selling the ability to see whether the physical hardware of intelligence has been manufactured correctly.

“We can't control what we can't measure.”

The expensive art of seeing

Onto's portfolio follows a chip through much of its manufacturing life. On unpatterned wafers, systems look for particles, haze, crystalline defects and contamination before valuable circuitry is added. During fabrication, optical tools measure thin films and the critical dimensions of three-dimensional structures. Later, inspection platforms examine wafers, packages and panels for macro defects, while 3D metrology checks the height and shape of the tiny bumps that connect dies.

That last stage has become much more consequential. The industry can no longer rely only on shrinking a single slab of silicon. AI accelerators increasingly combine logic, high-bandwidth memory and other components in 2.5D or 3D packages. The package becomes a system. More chiplets mean more bonds, bumps, redistribution layers and opportunities for error. A defect that once ruined one die can now compromise several expensive ones assembled together.

Four businesses hiding in one toolbox

The company has four main product families. Inspection includes the Dragonfly line for fast 2D defect detection and 3D metrology. Dragonfly G5, introduced for the newest packaging demands, has been qualified at a leading 2.5D logic customer and a high-bandwidth-memory customer. Firefly performs automated inspection on advanced IC substrates and panels. Other platforms examine wafer edges, backsides, transparent materials and defects below the surface.

Metrology is the ruler cabinet. Atlas systems use optical critical-dimension techniques to infer the shape of structures too small and complicated for a literal ruler. The newer Atlas G6 has been selected for gate-all-around transistor metrology, where channels are wrapped by gates and three-dimensional geometry must stay within narrow tolerances. Echo uses an opto-acoustic technique to measure metal films without touching or damaging them. The Semilab product lines acquired in 2025 add electrical and materials analysis, including inline contamination monitoring.

Lithography gives Onto a less common position among measurement vendors. Its JetStep machines expose patterns for advanced packaging on substrates that can be far larger than round silicon wafers. A panel of 650 millimeters by 650 millimeters offers manufacturers more working area, but panels warp and individual dies shift during processing. Onto pairs inspection with StepFAST software, using measured placement errors to adjust exposure. The product is a neat expression of the company's preferred phrase, “connected thinking”: do not merely observe the mistake; feed the observation into the next machine.

The customer buys recovered yield

Onto's customers are the companies that own manufacturing risk: integrated chipmakers, foundries, outsourced assembly and test houses, wafer and substrate manufacturers, and some fabless companies that need visibility into outside production. The company says all ten of the world's largest semiconductor device manufacturers use its systems, along with more than 50 others. Most revenue comes from outside the United States, matching the geographic reality of semiconductor fabrication and packaging.

The business model begins with expensive systems and software. It continues through spare parts, maintenance, applications support and training. In fiscal 2025, systems and software produced about $848 million of revenue; parts contributed about $84 million and services about $73 million. That installed-base income provides ballast, but this remains a capital-equipment company. When customers pause factory spending, orders can move quickly. When a new architecture creates an unsolved measurement problem, the cycle can move the other way.

Selling the first machine is only part of the work. Engineers must tune recipes to a customer's materials and process, prove that measurements repeat, and support the tool in high-volume production. That makes expertise a product of its own. Of Onto's roughly 1,615 employees at the start of 2026, 739 worked in sales, applications and service support, while 419 worked in research and development. The factory relationship is as important as the optics.

A broad specialist among sharp rivals

There is no single Onto competitor because the portfolio crosses several tool categories. KLA and Nova meet it in thin-film and optical critical-dimension metrology. KLA and Camtek compete in advanced-packaging inspection. Canon and Ushio sell packaging lithography. GigaVis appears in panel inspection, PDF Solutions in factory analytics and Nova again in integrated metrology. Some rivals have more resources or a larger installed base. Others concentrate deeply on one measurement technique.

Onto's difference is the connective tissue. The pitch is not that one camera or model defeats every alternative. It is that inspection, metrology, lithography and software can share context across a process. That breadth reaches from bare-wafer quality to transistor geometry, then outward to package interconnects and factory analytics. It can also create a more durable customer relationship: a supplier involved at several control points has more chances to see the next problem forming.

The panel is the plot twist

The company's most visible ecosystem bet is PACE, its Packaging Applications Center of Excellence. Opened in 2024, the facility brings together equipment and materials companies including Corning, Lam Research, ASMPT, MKS Instruments and others to experiment with panel-level packaging. Customers and collaborators get early access to tools and can work through an awkward industry problem: the economics of panels are attractive, but the processes, materials and standards must mature together.

The collaboration is strategically useful. Semiconductor tools are qualified inside tightly controlled production lines, and no vendor can make panel packaging happen alone. Glass suppliers, laser makers, deposition equipment, inks and inspection systems must agree with one another physically, even when the companies do not share a balance sheet. PACE gives Onto a place near the center of those conversations.

The Rigaku partnership extends the same logic into X-rays. Announced in April 2026, the collaboration pairs Rigaku's X-ray technology with Onto's process-control stack, while Onto plans to buy a 27 percent stake for about $710 million. X-rays can reveal composition and buried structures that optical methods cannot easily see. As memory stacks grow taller and novel materials spread, the ability to combine different kinds of evidence becomes more valuable.

What the next machine must notice

Onto crossed $1 billion in annual revenue in 2025 and followed it with a record $291.9 million quarter at the start of 2026. The numbers matter less as a victory lap than as proof of where customers are spending. A leading HBM manufacturer signed a Dragonfly volume agreement estimated above $240 million through 2027. Atlas G6 won another logic customer for gate-all-around metrology. Both developments point to the same pressure: AI chips are creating fresh inspection work at the transistor, memory and package levels at once.

That opportunity is not automatic. Tool qualifications are long, customer concentration can make results lumpy, competitors fight hard for each process step and semiconductor spending remains cyclical. Integrating acquired technology and a large strategic investment adds execution risk. Even a technically elegant instrument must prove throughput, reliability and total cost of ownership on a factory floor where downtime is measured in very expensive minutes.

Still, Onto occupies a useful place in the market. The AI industry is famous for scale: more parameters, more compute, more memory and larger packages. Manufacturing answers with discipline. Every ambitious architecture eventually meets a film, a bump, a bond or a panel that must land in the right place. Onto Innovation makes the machines that ask whether it did.