A chip made a rainbow, and the rainbow became ten clean millimetre waves at once. Nobody sent a single bit over them. And the thing 6G actually wants millimetre waves for is not speed — it's watching the room.

In partnership with

🌈 A rainbow, and then ten of them

Shine one colour of laser light into a glass ring about a millimetre across, and if you get everything exactly right, something strange happens.

The single colour splits. Not into a smear, the way a prism splits sunlight — into a ladder. Hundreds of separate colours, each one spaced from its neighbour by precisely the same amount, over and over, like the teeth of a comb.

Physicists call it a microcomb. Everyone else, on seeing the spectrum plotted, calls it a rainbow.

Aim that comb at the right antenna and those evenly spaced colours become a radio signal — one so clean that the best benchtop electronics on the market cannot match it. The older, table-sized version of this trick was part of what earned Hall and Hänsch half of the 2005 Nobel Prize in Physics. Shrinking it onto a chip came later, from about 2007.

Last week a team led out of Loughborough University published a version of it that does something new: instead of producing one clean tone, it produced ten at once, spread across more than 600 GHz.

The press coverage said this could supercharge 6G.

Here is what the paper actually did, what it carefully did not do, and the thing about 6G’s plans for millimetre waves that almost none of the coverage mentioned — which is that the marquee new feature of 6G is not speed at all. It is turning the network into a radar that watches the room.

In June, Germany’s data protection authorities wrote that, in depth of intrusion, it bears comparison with an imperceptible video surveillance.

🔬 What the chip actually is

The device is a four-port high-index doped silica resonator — a glass racetrack on a chip, roughly a millimetre across. Its quality factor is about 1.3 million, which means the light field inside oscillates some two hundred thousand times before it decays — a few hundred trips around the ring.

But the interesting part is what the team wrapped around it. The ring isn’t pumped by an external laser in the usual way. It sits inside a fibre laser cavity, so the ring and the loop are one coupled system. That architecture is called a laser cavity soliton microcomb, and it is the reason for the headline property.

The numbers, from the paper itself:

• Comb spacing (resonator FSR): 48.91 GHz

• Resolvable millimetre-wave harmonics: 10

• Measurable bandwidth, within their detection dynamics: > 600 GHz

• Resonator linewidth / Q factor: < 140 MHz / ~1.3 million

• Soliton output power: ~5 mW

• Electrical power for the microcomb laser’s pump diode: < 1.5 W

• Integrated timing jitter, locked (10 Hz–1 MHz): < ~150 fs

• Integrated timing jitter, free-running: ~6.3 ps

That jitter line is the one to hold onto. A hundred and fifty femtoseconds is 0.00000000000015 seconds. Light — the fastest thing there is — travels about 45 micrometres in that time. Roughly half the width of a human hair.

Note the honest asterisk on the power figure, which I’ve worded more carefully than the coverage did: under 1.5 watts is the pump diode for the microcomb laser, not the experiment. The rest of the bench — a second amplifier, a lock-in, a delay stage, a PID controller, a frequency counter — draws what benches draw.

📡 Why anyone wants a very clean millimetre wave

This is the part the coverage skipped, and it’s the part that makes the result genuinely impressive.

To make a high-frequency radio signal today, you start with something slow and stable — a quartz crystal at 100 MHz, say — and you multiply it up. That works. It also has a brutal, unavoidable cost: every time you multiply the frequency, you multiply the jitter with it. Go from 100 MHz to 100 GHz and you have multiplied by a thousand, and the noise floor rises by about 60 decibels for free. It is the frequency-metrology equivalent of photocopying a photocopy.

Optical frequency division runs the same physics backwards. You start with something enormously fast and stable — an optical carrier at around 193,000 GHz — and you divide down. The same arithmetic that punished you on the way up now rewards you on the way down.

The gap this opens is not marginal:

• A commercial benchtop signal generator, referred to a 100 GHz carrier, sits at roughly −95 dBc/Hz at 10 kHz offset.

• An integrated optical-frequency-division source published in Nature in 2024 hit −115 dBc/Hz at the same offset on a 100 GHz carrier. The authors say it plainly: 20 dB better than a standard signal generator — from a chip.

Twenty decibels is a factor of a hundred in noise power. And that number is not an abstraction. Phase noise is the floor under two things people care about a great deal:

1. Radar sensitivity. In a Doppler radar, oscillator noise smears the returns from big stationary clutter across neighbouring Doppler bins. A small, slow-moving target hides inside that smear. Cleaner oscillator, smaller thing you can see.

2. Navigation when GPS is gone. Light travels about 30 centimetres in a nanosecond, so nanosecond timekeeping is metre-level positioning. When the satellite signal is jammed or spoofed, everything falls back to how well your local clock holds its own count.

A typical GPS-disciplined oscillator sits somewhere around 10⁻¹¹ to 10⁻¹² at one second, limited by the oven-controlled crystal inside it. Chip-scale optical-frequency-division sources reach the 10⁻¹⁵ range; national-lab optical clock division has gone to 10⁻¹⁸. That is why DARPA is currently paying for shoebox-sized optical clocks intended to hold GPS-grade timing for up to two weeks with no satellites at all.

A wry footnote on that, since I’m holding others to primary sources: the Loughborough system’s own lock is referenced to a GPS-disciplined oscillator, and the authors say so, and say that their Allan deviation figure should therefore be read as a measure of locking quality rather than absolute stability. Good practice, honestly stated, and a reminder that “GPS-free” is a destination rather than a description of this bench.

So: a chip-scale source of many clean millimetre-wave tones is a real and useful thing. Good.

Now the caveats, which are entirely absent from the coverage and entirely present in the paper.

⚠️ Five things the paper says that the headlines didn’t

I read the full paper — it’s open access, so you can too — and the authors are notably more careful than their coverage.

1. Nothing was transmitted. Not one bit. This is a sub-terahertz time-domain-spectroscopy experiment: they generated the tones and measured them beautifully. There is no link, no channel, no data rate, no modulation. The paper’s own forward-looking language is about “opening perspectives for sub-THz communications, spectroscopy, and precision metrology.” Perspectives. Meanwhile one derivative article asserts this enables “practical 6G speeds by 2028.” That date appears in no primary source. It was manufactured somewhere downstream.

2. It is a tabletop, not a chip. The Methods section lists a fibre amplifier, a mechanical delay stage, Teflon lenses, an optical isolator, two tuneable filters, a piezo fibre stretcher wound on a 6 cm cylinder, and a PID controller. The press release says so plainly — “the entire system currently occupies a tabletop laboratory setup” — and to their credit, most outlets carried that line. The chip is one component inside it.

3. “The size of a grain of rice” is not in the paper. No chip dimension appears anywhere in the published text; it comes from the press release. Neither does the widely quoted line about people “jumping up and down next to the system” — that’s a quote given to a press officer, not a documented experiment. Both are perfectly ordinary press-release colour. They are just not findings.

4. The system does change state under real perturbation. This is the honest bit, and I respect the authors for putting it in. They deliberately parked the laser at the boundary between soliton regimes “with relatively loose temperature stabilisation” and left it exposed. It kept running — but it visibly jumped from a two-soliton state with 10 ps spacing, to 6.6 ps, to 5.6 ps, to a single soliton, and “the THz emission is affected by the soliton dynamics, with noticeable changes in both phase and spectral content.” Self-recovering, yes. Indifferent to the room, no.

5. The antennas were the wrong ones. The team used off-the-shelf commercial photoconductive antennas rated for pump rates near 100 MHz, and drove them with a 50 GHz pulse train — “approximately 500 times higher,” as the paper puts it. They call the configuration “notably sub-optimal” in their own text, and estimate that a redesigned antenna could bring “substantial improvements, possibly spanning several orders of magnitude.” The result is a floor, not a ceiling. That’s good news, honestly reported.

Two more, neither of which is a flaw. This is not August research. Nature Communications received the manuscript on 3 October 2025; the preprint went up in December 2025; acceptance landed 4 August 2026 and publication on the 20th — and what’s on the journal site right now is an accelerated preview, not yet the final version of record. The August news cycle is a publishing event. The science is nearly a year old.

And the work is funded by an unusually long list: an ERC Horizon 2020 grant, three EPSRC grants, five Leverhulme fellowships, a Loughborough fellowship, an Australian Research Council centre — and, last on the list, the US Army Research Office. That is in the public funding statement of an open-access paper. Nothing is hidden. It’s simply the kind of detail that gets sanded off between a funding statement and a headline, and I’d rather you had it.

One quick word from today’s sponsor — then the part about 6G.

Stop making AI decisions in the dark.

Leadership is asking: are we getting value from AI? Which tools are worth the spend? Where are we exposed? Right now, most teams have no idea.

Harmonic Security Usage Explorer changes that.

You get a complete picture of how your organization uses AI, automatically categorized into custom tasks and use cases.

You’ll see the projects being worked on, who’s using what tools, where AI investments are driving value, and where employees are engaging in risky behavior.

CIOs can rationalize spending and cut wasted licenses. CISOs can pinpoint where risk exists and neutralize it. AI committees can show exactly how their efforts are paying off.

See it in action

📶 Now the part about 6G

Every article framed this as a 6G story. So let’s take 6G seriously for a moment, because the reality is stranger and more interesting than “faster phones.”

First, the timeline. 6G does not exist, and will not exist soon. 3GPP agreed the schedule in June 2026: Release 21 is the first normative 6G release, Stage-1 requirements freeze in March 2027, architecture in June 2028, protocol in December 2028, and the final protocol coding in March 2029. Operators — Docomo, the Chinese carriers — target commercial launch in 2030. Anything sold to you as 6G before 2029 is marketing.

Second, millimetre waves already had their moment, and it went badly. Two years into 5G, in a March–June 2021 measurement window, Opensignal found its users spent less than 1% of their time connected to millimetre-wave 5G: 0.7% on Verizon, 0.4% on AT&T, 0.2% on T-Mobile. A 2019 study presented in the Defense Innovation Board’s 5G report estimated it would take roughly 13 million transmitters and $400 billion to deliver 100 Mbit/s to 72% of the US population that way. In 2024 T-Mobile handed 520 millimetre-wave licences back to the FCC, saying deployment was “not feasible… in a way that would benefit the public.”

Be fair about what that does and doesn’t mean: T-Mobile kept 865 licences, and Verizon still runs millimetre wave in stadiums, airports and fixed wireless. It survives as a venue technology. It failed as coverage.

The reason is almost funny. Ordinary clear glass costs a 28 GHz signal only about 3–4 dB. Low-emissivity glass — the energy-efficient kind now in every new building — costs far more, and a 2024 measurement study of high-thermal-efficiency buildings found a median 46 dB penetration loss at normal incidence, with a further 15 dB median penalty at oblique angles. We spent two decades making buildings thermally efficient and, in doing so, made them radio-opaque. No amount of beamforming fixes a Faraday cage you installed to save on heating.

So 6G’s spectrum plan is a retreat: the centre of gravity has moved to the upper mid-band, 7–24 GHz, where signals actually go through things.

Third — and this is the real story — the marquee new capability of 6G is not bandwidth.

When the ITU published the official framework for what 6G must be, it named six usage scenarios. Three carry over from 5G. Three are new: ubiquitous connectivity, AI and communication, and — the one this issue is about — Integrated Sensing and Communication.

📻 The network becomes a radar

ISAC means the network reuses its own radio signal as radar. The same waveform that carries your call bounces off the room, comes back, and is analysed for range, velocity, angle and micro-motion. No new hardware. The base station you already walk past becomes a sensor.

This is not a thought experiment. On 10 July 2026, at AT&T Stadium in Arlington, Texas, AT&T and Ericsson detected, located and tracked multiple drones at 300 to 400 feet using existing cellular towers — trade coverage puts it at three sites, one about a quarter of a mile from the stadium and two around 1.6 miles out.

Precision matters here, so let me use Ericsson’s own words rather than the excitable version: they deployed Massive MIMO radios on existing towers and combined “sensing-enabled radio transmissions with advanced signal processing and AI-enabled sensing algorithms” — achieving this “without requiring additional standalone sensing technology.” So: new software, repurposed radios, no radar dish. Not quite “the towers did it with the signal they were already sending,” which is the version that travelled.

That demonstration is impressive, useful, and close to the limit of what 3GPP has taken on so far. The current cellular ISAC work is deliberately narrow: base-station sensing, for detecting drones. It is a work item in progress, not a finished standard.

Release 21 is where it opens up. And the research literature has spent a decade showing what radio sensing can do once you point it at people instead of drones:

Read that last row again, because it is the one that changed things.

Earlier RF-identification work needed modified router firmware and a research lab. BFId does not. It uses beamforming feedback information — routine data that ordinary Wi-Fi routers already exchange with devices, by design, all day. It worked across different walking styles, from multiple viewing angles, and at reduced sampling rates. It beat the older techniques while assuming a weaker attacker.

The bar for identifying a specific human being by the way their body disturbs radio just dropped from “custom hardware” to “any router.”

🚨 The problem nobody has solved

Here is the structural issue, and it is not a bug that a patch fixes.

Sensing works by reflection. Reflection does not check whether you’re a customer.

A camera has a lens you can see and a red light you can notice. An app asks permission. A cookie banner, for all its faults, at least appears. Radar sensing has none of these. It senses everyone in range — subscribers, houseguests, the neighbour through the party wall, the delivery driver on the step, a child who has never owned a phone. There is no device to configure, no toggle to find, no way to know it happened.

In June 2026, Germany’s federal and state data protection authorities — the DSK — did something unusual. Rather than waiting for deployment and then complaining, they looked at where the standardisation work was going and published a formal position paper while the standard is still being written.

Their conclusions, translated from the document itself, with the hedges the original actually contains:

• “In principle, ISAC can be compared, in terms of possible depth of intrusion, with an imperceptible video surveillance” — one which additionally captures what optical sensors cannot see. They are drawing an analogy about potential intrusiveness, not issuing a legal classification, and the distinction is worth keeping.

• Radar returns plus machine learning permit inference of gait, gender, facial features, gestures, respiration and heartbeat — a set which therefore includes Article 9 GDPR special categories. Elsewhere the paper is careful to write “insofar as Article 9 is engaged.” So: conditional, not automatic.

• Because the signal penetrates walls, it potentially engages Article 13 of the German constitution, the inviolability of the home — which “would additionally constitute a particularly serious interference with fundamental rights.”

• And the finding with teeth: consent is unsuitable as a legal basis for area-wide sensing, because you cannot obtain it from people you cannot identify, address, or even detect in advance. They do leave a door open — consent remains conceivable in scenarios bounded in time and space, and they give a medical Wi-Fi-sensing example where it would work.

ETSI, the European standards body itself, reached compatible conclusions. Its February 2026 report on ISAC lists 19 key issues — 15 on security and privacy, four on sustainability — and among them names one that has no technical fix: “unawareness and unintervenability.” The people being sensed don’t know, and can’t stop it.

What exists to fix this today? Being honest: very little.

3GPP’s service requirements do contain privacy language — encryption, integrity protection, exposure limited to authorised parties. But almost every operative clause is qualified “subject to regulation,” “subject to operator’s policy,” “subject to user consent.” The standard defers the hard question to regulators. The security study on ISAC privacy is still a study, not a specification.

IEEE’s Wi-Fi sensing standard, published in September 2025, does require a sensing session to be set up and negotiated between the participating devices — a consent surface of sorts, for devices that are associated with the network. What it has no mechanism for at all is the person standing in the room who owns none of them.

The research countermeasures are real but all lab-stage: full-duplex jamming relays, reflecting surfaces that scramble the channel, differential privacy applied to the signal features. Every one needs hardware you cannot buy. There is no sensing opt-out. No geofenced exclusion. No standards-track mechanism by which an unassociated person can decline. If someone tells you otherwise, ask them for the spec number.

🏠 This is already in houses

It would be comfortable to file all of this under 2030. It isn’t.

In August 2026, Comcast rolled Wi-Fi Motion sensing out to millions of its newer Xfinity gateways. It detects movement in the home using the Wi-Fi signal already there. The subscriber opts in — genuinely, it is opt-in, and it is useful. Verizon offers a comparable service on Fios.

Two things about it are worth sitting with.

First, Comcast’s own support documentation states that, “subject to applicable law,” it may disclose information generated by your Wi-Fi Motion to third parties without further notice to you in connection with law-enforcement investigations, disputes, court orders or subpoenas. Those first four words matter and I’m not going to quietly drop them — that clause is standard legal boilerplate and Comcast is being upfront. It is also a description of a new category of record about the inside of your home that did not exist five years ago and has no camera anywhere in it.

Second, and more simply: the subscriber consented. The houseguest didn’t. Nor did the cleaner, the carer, the babysitter, or the person on the other side of a thin wall. That is the ISAC bystander problem, arriving early and at consumer scale, in a product with a perfectly reasonable feature list.

✅ What to actually do about it

For most readers this is not a “buy a thing” problem. It’s a “know the thing exists” problem. But there are real moves.

1. Check whether your router is already sensing. If you have a recent Comcast/Xfinity gateway or Verizon Fios service, look for Wi-Fi Motion in the app. Decide deliberately rather than by default — and if you keep it, tell people who stay in your home that it’s on. That last part costs nothing and is simply decent.

2. If you run a workplace, treat sensing data as personal data now. Motion, occupancy and presence records derived from RF are not “just network telemetry.” Under GDPR reasoning they can carry gait and vital-sign inferences. Put them in the retention schedule and the DPIA before someone else puts them in a subject access request.

3. The real lever is the calendar, not the settings menu. 6G Stage-1 requirements freeze in March 2027. Privacy properties written into a standard before freeze become defaults for a decade; privacy properties bolted on afterwards become optional features nobody enables. The German DPAs understood this exactly, which is why they filed in June 2026 rather than 2031. If your organisation participates in 3GPP, ETSI or a national regulator, that window is open now and closes in about seven months.

4. Be suspicious of the word “passive.” Sensing is often described as passive because it doesn’t require you to do anything. That’s precisely what makes it not passive at all, from the perspective of the person being sensed.

5. Don’t accept a “6G” claim before 2029. The specification will not be frozen until December 2028. Any product, tender or roadmap using the term today is describing an aspiration.

🎯 The takeaway

The Loughborough result is lovely physics, honestly reported by the people who did it, and mildly oversold by everyone downstream of the press release. Ten coherent millimetre-wave tones from one comb, with jitter measured in femtoseconds. That is a genuinely useful building block for radar, for navigation when satellites fail, and eventually for communications.

But the story isn’t “6G will be fast.” Millimetre waves already tried that, as coverage spectrum, and lost to double-glazing.

The story is that the same physics that makes a clean radio signal makes a good radar — and the next generation of mobile networks has written sensing into its founding requirements as a core capability, not an add-on. The standard that decides whether that sensing has consent built into it, or bolted on later, is being drafted right now.

Germany’s data protection authorities have looked at where it’s heading and said that, in depth of intrusion, it bears comparison to a video surveillance you cannot perceive. Almost nobody reported that.

Which is why, if you only remember one thing from today: the interesting question about 6G isn’t how fast it will be. It’s who gets to see through the wall, and whether anyone in the room is allowed to say no.

📚 Sources and how this was checked

Every claim above was checked against the primary document rather than the coverage. Where I could not verify something, I’ve said so.

• Microcomb paper — 48.91 GHz, 10 harmonics, >600 GHz, Q ~1.3M, <150 fs jitter locked / ~6.3 ps free-running, doped silica, tabletop setup, "sub-optimal" antennas, soliton state transitions, GPS-disciplined reference: Peters et al., "Millimetre-wave comb generated by an optical microcomb," Nature Communications (2026), DOI 10.1038/s41467-026-76747-2 — open access, currently an accelerated preview

• Received 3 Oct 2025 · accepted 4 Aug 2026 · published 20 Aug 2026; preprint arXiv:2512.05005, Dec 2025; funding incl. ERC, EPSRC, Leverhulme, ARC and the US Army Research Office: same article record and funding statement

• "Grain of rice," "jumping up and down," "shoebox," "tabletop laboratory setup": Loughborough University press release, Aug 2026 (syndicated verbatim by ScienceDaily and SciTechDaily) — the first three are not in the paper

• −115 dBc/Hz at 10 kHz on a 100 GHz carrier, 20 dB better than a benchtop generator: Sun et al., Nature 627, 540–545 (2024)

• 20 GHz OFD oscillator, −96 dBc/Hz @ 100 Hz → −135 @ 10 kHz: Kudelin et al., Nature 627, 534–539 (2024)

• Release 21 timeline — agreed 10 June 2026, Stage-1 Mar 2027, Stage-2 Jun 2028, Stage-3 Dec 2028, ASN.1 Mar 2029: 3GPP, "Timeline for Release 21."

• Six IMT-2030 usage scenarios, three of them new (ubiquitous connectivity, AI, ISAC): Recommendation ITU-R M.2160-0, Nov 2023

• mmWave under 1% of connected time (VZ 0.7 / ATT 0.4 / TMO 0.2), Mar–Jun 2021, per-carrier: Opensignal, 13 July 2021

• 13 million transmitters / $400bn to deliver 100 Mbit/s to 72% of the US population: 2019 study presented in the Defense Innovation Board 5G report

• T-Mobile returns 520 mmWave licences, "not feasible"; retains 865: FCC filing, July 2024

• Median 46 dB penetration loss at normal incidence into high-thermal-efficiency buildings, +15 dB oblique: arXiv:2401.03357 — figure from the abstract, not a full-text read

• Clear glass ~3–4 dB at 28 GHz: NYU 28 GHz penetration measurements — cited from the secondary literature, not read in full

• Drone detection, 300–400 ft, 10 July 2026, Massive MIMO radios plus AI sensing software, "without requiring additional standalone sensing technology": Ericsson and AT&T press releases. The three-site and distance figures come from trade coverage, not the releases.

• Wi-Vi, WiTrack, Vital-Radio, EQ-Radio, RF-Pose3D: SIGCOMM 2013 · NSDI 2014 · CHI 2015 · MobiCom 2016 · SIGCOMM 2018

• BFId — 99.5% ± 0.38 identification across 197 people from beamforming feedback: Todt, Morsbach & Strufe, ACM CCS 2025, DOI 10.1145/3719027.3765062

• Comparison to imperceptible video surveillance, Art. 9 GDPR, Art. 13 GG, consent unsuitable for area-wide sensing: DSK 111th conference, Stuttgart, ISAC position paper adopted 17 June 2026 (German; translations mine)

• Keber quote: DSK conference press release, 18 June 2026 — not the position paper

• 19 key issues (15 security/privacy, 4 sustainability), "unawareness and unintervenability": ETSI GR ISC 004 V1.1.1, Feb 2026

• Wi-Fi sensing standard — session negotiation for associated devices, nothing for bystanders: IEEE 802.11bf-2025, published 26 Sept 2025

• Xfinity Shield / Wi-Fi Motion launch 18 Aug 2026, XB7 and newer, opt-in; "subject to applicable law" disclosure clause: Comcast support documentation and Aug 2026 reporting

Named limits, because you should hold me to the same standard. I could not obtain a phase-noise figure in dBc/Hz for the Loughborough work — the paper reports integrated jitter and plots the curve but states no number, so I have not quoted one. I read the December 2025 preprint in full; the published version is an accelerated preview and may differ in wording, so I have avoided quoting anything I could not find in a version I actually read. I could not fetch the Loughborough press release directly; its wording is reconstructed from two syndications that reproduce it word-for-word and credit it. The glass attenuation figures come from an abstract and from secondary citation rather than full-text reads. And the claim, repeated in one downstream article, that this enables “practical 6G speeds by 2028” appears in no primary source — it was manufactured somewhere in the chain, and I’ve excluded it rather than repeat it.

An earlier draft of this issue got several things wrong, including the meaning of the resonator’s Q factor and the number of new ITU usage scenarios. Both were caught in checking, before you read them. That process is the product.

Written and checked in Dublin. If something here is wrong, reply and tell me — corrections run in the next issue with your name on them if you want it.