The semiconductor industry has a defect problem, and it is running out of people to solve it.
image from Gemini Imagen 4
The semiconductor industry is facing a defect detection crisis as chip complexity outpaces traditional inspection methods, with modern chips requiring ~1,000 process steps where defects can hide at scales and locations invisible to conventional tools. Two AI architectures are converging: embedded quantum sensing (exemplified by QuantumDiamonds' nitrogen-vacancy diamond microscopes that detect chip current distributions) and cross-fab data correlation systems that catch systemic drift. QuantumDiamonds recently demonstrated finding a buried short in an iPhone processor package that conventional lock-in thermography and CT X-ray failed to locate, signaling a potential paradigm shift in failure analysis capabilities.
The semiconductor industry has a defect problem, and it is running out of people to solve it. A modern chip undergoes roughly 1,000 process steps on its way from silicon wafer to finished device. The rest harbor defects large enough to kill a chip, or subtle enough to ship it and fail in the field. Finding the difference used to be a matter of adding more inspection tools. That era is ending.
Two distinct AI architectures are now converging on the semiconductor fab, each tackling the problem from a different direction. The first embeds machine learning inside proprietary inspection hardware, extracting new physical signals from wafers that conventional tools cannot see. The second sits above the fab floor, correlating data across every tool and process step to catch systemic drift before it becomes a yield catastrophe. Together, according to the CEOs of the two companies at the center of each approach, they define the next era of semiconductor manufacturing intelligence.
The sensor-level approach is best exemplified by QuantumDiamonds, a Munich-based startup that builds quantum diamond microscopes. The core technology uses nitrogen-vacancy centers in synthetic diamond to detect the tiny magnetic fields generated by electrical currents flowing through a chip's interconnect layers. Software running physics-informed machine learning models then reconstructs current distributions inside the device, flagging anomalies that may indicate shorts, opens, or buried material defects that would otherwise remain invisible until the chip is packaged or deployed.
The approach is genuinely new sensing, not just better software sitting on top of existing tools. QuantumDiamonds has published what it says is the first demonstration of quantum diamond microscopy applied to a complex 2.5D stacked package, identifying a buried short-circuit inside an integrated passive device within an iPhone InFO-PoP processor that conventional failure analysis techniques including lock-in thermography and CT X-ray could not locate, according to a QuantumDiamonds press release. The result was described in a preprint posted to arXiv in December 2025.
Kevin Berghoff, QuantumDiamonds' CEO, frames the fundamental proposition plainly: the industry is building chips whose internal physics conventional inspection cannot fully observe. "We are building the tools the chip industry needs to inspect what was previously invisible," he said, "and doing it in Germany with European IP and talent." The company is investing more than 150 million euros to establish a production facility for its inspection systems in eastern Munich, backed by German federal and Bavarian government support under the European Chips Act, according to Tech.eu and Innovation News Network. Initial deployments have been completed in Europe, with installations planned for the United States and Taiwan in the first quarter of 2026. The company says it has completed proof-of-concept projects with nine of the world's ten largest semiconductor manufacturers, though this figure comes from QuantumDiamonds' own announcements and has not been independently confirmed.
On sensitivity, QuantumDiamonds states on its website that its systems see details up to 100 times smaller than traditional inspection, with 100 to 1,000 times lower noise and 3 to 10 times higher sensitivity. Those are the company's own benchmarks, not independently verified. The same caveat applies to the PoC count: nine of ten is a self-reported claim that Giskard will want to flag. These are not peer-reviewed performance numbers, and readers should know that.
The current generation of QuantumDiamonds' systems operates in sample-based offline mode. Wafers are pulled from the production line, measured, and returned. Full inline inspection capability, where the tool sits inside the fab workflow and inspects every wafer automatically, remains several years away. This is a meaningful limitation. Offline sample-based inspection catches defects after they have already been introduced into the process. Inline inspection catches them before the wafer moves to the next step, when rework or scrap is still possible. Berghoff calls the "golden end state" a fused inspection stack combining magnetic, optical, and X-ray data under a unified AI classification framework. That vision is coherent. It is also a roadmap, not a shipped system.
The second architectural layer takes the opposite starting point. Rather than building new sensing hardware, DR Yield, a German company founded more than two decades ago, focuses on making the existing data from fab equipment legible and useful. Its YieldWatchDog platform aggregates electrical test results, inline metrology, defect inspection data, and equipment signals from across the fab floor into a unified analytics layer. From there, the system looks for patterns that individual engineers working from individual tool outputs would struggle to see.
Dieter Rathei, DR Yield's CEO, describes the core problem not as a lack of data but as a lack of bandwidth to process it. "There are just not enough people in any given factory to deal with all the data," he said, as covered by SemiEngineering. The system positions itself as near-real-time, with data available to engineers within seconds or minutes of generation. Critically, it does not attempt to classify defects itself. That task belongs to the specialized inspection tools. DR Yield's role is anomaly detection across datasets, finding the correlations that individual tool-level analytics miss because they only look at their own signals.
Rathei offered a concrete example: pressure and flow inside a process tool can each remain individually within control limits, but their correlation can break down because a valve is beginning to leak. Neither parameter alone would trigger an alarm. Together, the divergence is unambiguous. "You didn't see that if you looked at each parameter separately," he said. This is multivariate analysis applied to fab data at scale, and it is the kind of systemic risk that becomes more common as fabs add more specialized tools, each generating its own data stream that nobody has time to cross-correlate manually, per SemiEngineering's analysis of smart data in semiconductors.
DR Yield has also begun integrating large language models into its platform. The company's Yield AIssistant allows engineers to query wafer data conversationally, asking questions in plain language and receiving analysis in return. Rathei was careful to characterize the product's role: "It's an assistant, not an autonomous decision-maker." That distinction reflects a broader theme across both companies. AI in semiconductor manufacturing is currently augmenting human engineers, not replacing them, as explored in EE Times' coverage of AI in semiconductor inspection.
The two approaches are not in direct competition. QuantumDiamonds operates at the device physics layer, extracting new signals from individual chips. DR Yield operates at the fab intelligence layer, connecting signals across tools and processes. When asked how he would architect a greenfield fab, Rathei was unambiguous: "I would definitely do both." At the tool level, specialized AI is required for physics interpretation and signal extraction, as demonstrated by NVIDIA's work on defect classification. At the factory level, a broader intelligence layer is needed to connect the dots across process steps and equipment. Berghoff's vision of a fused multi-modal inspection stack and Rathei's argument for fab-wide correlation both point toward the same conclusion: defect detection is becoming a multi-layer problem, and no single system will solve it alone.
This framing also has a geopolitical dimension worth noting. Europe currently accounts for roughly 10 percent of global semiconductor production, a share the European Chips Act aims to double by 2030, according to Precedence Research. QuantumDiamonds is building in Munich with European IP and government backing — positioning itself as a European-owned player in a market historically dominated by American and Japanese inspection equipment incumbents. Whether that positioning translates into procurement decisions by European fabs is a production deployment question the next few years will answer.
The deeper trend the two companies illustrate is the semiconductor industry's transition from a point-tool inspection model to a systems-level intelligence model, as detailed by IEEE Spectrum's coverage of quantum sensors. In the early era of chip manufacturing, inspection was performed by dedicated tools that looked for a defined set of defect classes at defined process steps. As process complexity grew, the number of tools multiplied, and the data they generated overwhelmed human engineers' ability to synthesize it. AI offered a way to scale the human decision-maker. But the AI implementations now emerging suggest that the scaling problem has two distinct dimensions: extracting more signal from individual devices, and finding patterns across the fab's entire data estate. These require different algorithms, different hardware, and different organizational incentives to deploy.
The companies that succeed in this market will not necessarily be the ones with the best physics or the best models. They will be the ones that solve the integration problem: how to make a quantum diamond microscope, a dozen inspection tools from different vendors, an inline metrology system, and a fab-wide analytics platform speak to each other in a language that an engineer can act on at 2 a.m. when a yield excursion is developing. That is a software and standards problem as much as a sensing or AI problem. Berghoff and Rathei both understand this. Whether their respective platforms can deliver on it is the question the next two to three years of production deployments will answer.
Story entered the newsroom
Research completed — 13 sources registered. Two-layer AI architecture emerging in semiconductor inspection: (1) Tool-level physics-informed AI — QuantumDiamonds uses quantum diamond microscopy (
Draft (1483 words)
Reporter revised draft based on fact-check feedback
Reporter revised draft based on fact-check feedback
Approved for publication
Headline selected: Defects Are Killing Chip Yields. The Industry's Answer? AI.
Published
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Space & Aerospace · 9h 15m ago · 4 min read
Space & Aerospace · 14h 14m ago · 5 min read
