Field notes from European Microwave Week 2026, London.
One display at the Anritsu booth at EuMW 2026 carried the most ambitious headline of the show: "Meet Tensor: The World's First AI-Powered VNA". Under it, the demo was about as unglamorous as millimetre-wave work gets. A two-port VNA drove a pair of WR-3.4 frequency extenders joined by a short waveguide section, sweeping 220 to 330 GHz. In other words, it was a thru measurement.
Putting those two things next to each other turned out to be the most useful part of the demo. It raises a concrete question: at 300 GHz, which parts of the work can an AI assistant help with, and which parts still depend on the engineer at the bench?
The setup
- VNA: an Anritsu MS46642A, the two-port member of the Tensor family. The toolbar shows the "AAI" assistant button.
- Extenders: two VDI modules labelled WR-3.4, 220–330 GHz, each driven by a bundle of cables from the VNA's front-panel ports. Anritsu states that Tensor works directly with its own and third-party mmWave modules, with flexible multiplier and source control for extender chains.
- DUT: a short gold-plated WR-3.4 waveguide section with a flange joint in the middle, connecting the two extender test ports.
The screen showed a single trace: "S21 Trans LogMag RefLvl 0 dB Res 5 dB/Div", on channel 1, from 220 to 330 GHz, at IFBW 1 kHz with averaging off. The correction indicator was active. The trace lies flat on the 0 dB line across the whole band, which is what a calibrated thru should look like.
What the AI is offered for
The table card next to the instrument listed what the Tensor assistant does today. Summarised in my own words:
- It takes natural-language requests in place of menu navigation and manual lookups.
- It guides setup, calibration, troubleshooting and measurement workflows.
- It generates SCPI commands and automation scripts from plain-language descriptions.
- It answers RF concept questions interactively.
- It combines a local real-time agent for instrument control with cloud-connected engineering agents.
The example prompts on the card include "Which Cal-Kit do I need for this measurement" and "Why do I have noise on my S11". A separate line, marked Coming Soon, promises AI-powered analysis of traces, S-parameters, TDR, matching and stability.
Reading the manual versus reading the trace
That split between what is available now and what is coming soon is the line that matters most.
The features available today are about reading the manual. They help you find the right setting, the right calibration procedure and the right SCPI syntax. That is valuable. VNA menu trees are deep, and SCPI generation alone saves real time when you automate a test.
The feature on the roadmap is about reading the trace: looking at the measured data and telling you what is wrong with it. In the WR-3.4 band, that is where most of the hours go.
Take the two example prompts and ask them at 300 GHz.
"Which Cal-Kit do I need?" In waveguide at these frequencies, the choice of calibration (TRL with waveguide line sections, offset shorts, or a load-based method) depends on the band and the accuracy target. A manual-reading assistant can answer that well. But the cal kit choice is rarely what limits accuracy here. The limits are physical:
- Flange alignment, because a few micrometres of lateral offset in a WR-3.4 flange shows up as measurable mismatch.
- Repeatability every time a flange is broken and remade.
- Thermal drift of the extenders after the calibration.
- The LO and IF cable bundles moving between calibration and measurement.
"Why do I have noise on my S11?" At 300 GHz the honest answer is "it depends on what your trace looks like". It could be IF bandwidth and averaging, too little extender drive, a loose flange, drift since calibration, or a standing wave between two flanges. Each cause leaves a different signature: random noise, slow drift, or periodic ripple. You cannot tell them apart without looking at the data. A manual-reading assistant can list the candidates, and only a trace-reading assistant can rank them.
The flat line on the screen
The demo trace itself illustrates the point. A calibrated thru at 5 dB/div will always look perfect, because any real residual error sits far below one division. To judge the calibration you would:
- Zoom S21 to tenths or hundredths of a dB per division and look for ripple.
- Look at S11 and S22 of the thru, which show the residual source match and directivity after correction.
- Break and remake the flange to see how much of what remains is repeatability.
None of that is a menu question. It is exactly the work that "AI-powered analysis of traces and S-parameters" would have to do to be worth its name. When that feature arrives, a sub-THz thru like this one is a good test of whether it can separate a calibration problem from a flange problem.
Takeaway
Building an AI assistant into the instrument is a sensible direction. Anritsu has started with the part that can be done reliably today: setup, calibration guidance, SCPI generation and concept questions. Engineers who spend their days on mmWave and sub-THz benches will judge it on the roadmap item: whether the assistant can look at a trace, compare it with what a well-made measurement should look like, and say why the two differ. Until then, the flat line at 0 dB still needs an engineer to zoom in and check it.
Notes: Observations and interpretations are my own, based on what was running at the booth during the show. Screen values are as read from my photos. Product capabilities and claims are paraphrased from Anritsu's booth material and published information; "world's first" is the manufacturer's claim. The DUT and calibration method used in the demo were not confirmed with the exhibitor.