On August 24, IEEE Spectrum published a piece signed Ritu Favre, president of Emerson's Test & Measurement group. At the top of the page sits the detail that changes everything: Sponsored Article.1 Corporate messaging bought from a serious publication, then. The normal reflex is to scroll past.
That would be a shame here, because the essay makes a claim that independent sources largely confirm: the central difficulty of modern engineering has moved. It is no longer mainly designing and building complex systems; it is verifying that they behave as intended. The marketing text tells the truth. It simply doesn't tell all of it, and certainly not the part that matters to anyone who actually builds things.
An interested argument
Start with self-interest, because it shapes the whole text. Favre has run Emerson's Test & Measurement segment since October 2023. The segment used to be National Instruments, bought that year for an equity value of $8.2 billion at $60 per share.2 The pursuit began in January 2023 with a $7.6 billion offer that business press described as near-hostile, NI having dodged serious negotiations since early 2022.3 Emerson expects $165 million in cost synergies by year five.3
The essay argues exactly what that acquisition purchased: test should stop being an end-of-line checkpoint and become a connected platform spanning the whole design cycle. When the segment's president writes that disconnected tools get in the way, she is either describing the market or justifying her employer's product basket. Very likely both.1
So the question isn't whether the text is neutral. It isn't, and barely pretends to be. The question is whether its claims survive contact with sources that sell nothing. We checked them one by one.
Nobody probes anymore
First claim: every part of a modern system can pass every one of its tests alone, then behave strangely once assembled with the rest. In semiconductors, Favre points to chiplet-based designs: a compute die from one supplier, an interposer from another, a packaging process from a third.1
This checks out down to the letter of the standard governing those assemblies. UCIe, the open die-to-die interconnect standard published since 2022, explicitly includes compliance testing in its specification, precisely because mixing chiplets multiplies failure points at the interface.4 The most telling detail is physical. In advanced packages, microbump pitch drops to between 25 and 55 micrometers. At that scale you cannot land an external probe on the links. The chip must carry its own built-in self-test, its own signal-integrity monitors, and redundant lanes able to replace a failed link in service.5 Typically only the data channels plus one sideband remain to do all of this: there is no dedicated test port.5
The industry knows this problem so well it gave it a brutally economic name: Known Good Die. A chiplet that passes its own tests can still doom an entire package once integrated, and packages are expensive.6 The failure didn't go away; it moved. It now lives between two suppliers, where nobody is really looking.
Twenty-five years of bottleneck
The essay's title speaks of a "new era of test". History is a little flatter than that. Verification has been documented as electronics' main bottleneck since at least December 1999, when an EE Times op-ed already estimated it routinely consumed 50 to 70 percent of a project's engineering manpower, with complexity growing exponentially with design size.7
The macro numbers say the same thing. Back in 2001, the ITRS roadmap declared design cost "the greatest threat to continuation of the semiconductor roadmap", as Professor Andrew Kahng reports. And without tool progress between 1993 and 2009, a consumer SoC designed for roughly $40 million in 2011 would have cost $7.7 billion.8 Verification tooling is not administrative overhead; it is what made the complexity created by everything else bearable.
Even tool vendors admit it, each in their own way. In early 2025 Siemens EDA published a white paper on that same gap 2.0, driven it says by chiplet architectures, 3DICs and software-defined functionality; it also claims first-silicon success rates have hit record lows. A vendor's claim, take it with the usual pinch of salt.9 More interesting: a Semiconductor Engineering survey from June 2026 concludes AI eases the pain without closing the bottleneck, and that tool chains remain poorly connected to each other.10 The "new era" is twenty-five years old, and nobody claims to have closed it.
Born in a rainstorm
The essay's most concrete example fits in one image: adaptive cruise control on a bumpy road in a rainstorm, or an airliner adjusting flaps and engine speeds in turbulence. Systems steered by thousands of interactions per second, whose real-world behavior keeps getting harder to understand.1
What she doesn't say is that this exact example was enough to spawn an entire standard. ISO 26262, the automotive reference, handles malfunctions: a component fails, the risk must be controlled. But some real hazards owe nothing to a failure. The cruise control works exactly as designed; the design is simply insufficient for a situation anticipated nowhere. That territory is called SOTIF, Safety Of The Intended Functionality, covered by ISO 21448 and created precisely because the existing framework didn't address it.11 Its practice calls for hardware-in-the-loop rigs and millions of simulated scenarios to hunt down those edge cases: testing integrated into the development cycle, not an end-of-line pass/fail.12
In other words, Favre's own example confirms her thesis better than she frames it. When automakers had to admit a final test bench can say nothing about system behavior in conditions never met during development, they didn't buy a platform: they wrote a new definition of risk. The problem is architectural, not merely instrumental.
Under the AI veneer
Which leaves the commercial core of the text: AI can only reach its potential when connected to a structured, traceable data stream; without context it confuses normal hardware variance with critical failure.1
True, and almost banal. What deserves attention is what "connected platform" actually means once you scrape off the varnish. Take the concrete example, SystemLink, the software suite inherited from NI: centralized software deployment across a fleet of test systems, machine health monitoring, instrument management including calibration schedules, central storage of results and waveforms, historical search, dashboards, access control, cloud or on-premises deployment.13 No artificial intelligence anywhere in there. Plumbing: every measurement carries its context, which instrument, which calibration, which device under test, which day. The Nigel AI assistant remains a promise on the 2026 roadmap.13
That is the real mechanism of this whole story. Not the model, the ledger. A measurement without provenance is a rumor; a timestamped, calibrated measurement tied to its unit under test is usable evidence, for a human or a model. The essay's AI argument stands because of that plumbing, not because of the model it showcases.
Same idea, no vendor
Last point, the one the essay will never mention: none of these ideas belong to a proprietary platform. Instrument control through open standards has existed since the nineties with VISA, the software layer normalizing how you talk to GPIB, serial, USB or Ethernet instruments. PyVISA exposes it in Python under an MIT license, publicly maintained and citable in scientific journals.14
Closer to embedded test benches, LabGrid, developed at Pengutronix, abstracts remote board control: power, serial consoles, multiplexers, with a pytest plugin for writing automated tests and a client-exporter-coordinator architecture for sharing a fleet of boards across machines.15 And on the research-lab side, ScopeFoundry structures complete experiments in Python, hardware plugins and logged measurements, BSD-licensed.16 Traceability, centralized data, tests woven into the workflow: it's all there, with no platform vendor at the end of the chain.
A maker who versions their test scripts alongside firmware and writes down their multimeter's calibration state next to their measurements already practices the shift-left Favre describes. That's the expensive name for an old habit: test early, keep provenance.
What survives the pitch
An honest recap. The essay stays vague about proven AI benefits in test; it cites no independent figures, and its argument maps onto a specific business case, $8.2 billion and $165 million in expected synergies.23 Nothing illegal about that, but it rules out relaying the text as-is.
Yet its three foundations hold against primary sources. Failures now live at interfaces, down into micrometers probes can no longer reach.5 Verification has absorbed the bulk of engineering resources for twenty-five documented years.710 And some hazards structurally escape end-of-line testing, to the point of forcing a dedicated standard into existence.11
For anyone building anything, the transferable lesson fits in two gestures: treat integration points as first-class test targets, and attach provenance to every measurement from day one of a project. The rest, platforms, AI, dashboards, is the commercial packaging of those two habits.