Edge vs Cloud Processing: Where Family Data Really Goes

On September 22, 2026, Qualcomm announced two chips whose headline capability was not speed. It was location. The Snapdragon 8 Elite Extreme Gen 6 can run a 30-billion-parameter mixture-of-experts model on the phone itself, and the chip’s always-on sensing hub can run models up to 200 million parameters without waking the main processor. That changes the practical meaning of edge vs cloud processing for a family: a feature that used to require a server can now run in a pocket, but only if the manufacturer chose to build it that way.

Key Takeaways

• Qualcomm’s September 22, 2026 launch included the Snapdragon 8 Elite Gen 6 and 8 Elite Extreme Gen 6. The Extreme supports running a 30-billion-parameter mixture-of-experts model locally.
• The sensing hub runs models up to 200 million parameters for always-on tasks including speaker differentiation and local note-taking, without engaging the main chip.
• Apple released a 20-billion-parameter mixture-of-experts model at WWDC in June 2026, so the direction is an industry pattern rather than one company’s bet.
• Every connected device now has three possible processing tiers. The useful parental question is which tier each feature uses, not whether the device is “private.”
• Local processing is a design decision, not a property of modern hardware. Capable silicon in a device tells you nothing about where its data actually goes.

What Qualcomm announced, and why the parameter count matters

TechCrunch covered the launch on September 22, 2026. Two chips: the Snapdragon 8 Elite Gen 6 and the Snapdragon 8 Elite Extreme Gen 6. Motorola announced the Signature 27 using the Extreme version.

Two specifications are worth understanding rather than skimming.

First, the sensing hub runs models up to 200 million parameters. A sensing hub is a small, very low-power processor that stays awake while the rest of the phone sleeps. Running a 200-million-parameter model there enables things like telling two speakers apart in a recording and transcribing locally, continuously, without the battery cost of waking the main chip.

Second, the Extreme supports running a 30-billion-parameter mixture-of-experts model locally. That number sounds impossible on a phone, and the trick is in the architecture.

A parameter is one learned number inside a model. A 30-billion-parameter model has 30 billion of them, and storing them all at a typical compressed precision takes more memory than a phone has spare. A mixture-of-experts model solves this by splitting the parameters into many specialised groups and activating only a few per task. Total parameters stay large; active parameters per token stay small. So memory bandwidth, which is the real constraint on a phone, stays manageable.

For comparison, Apple released a 20-billion-parameter mixture-of-experts model at WWDC in June 2026. Industry figures including Nothing’s Carl Pei and Apple’s chief executive John Ternus have argued that the smartphone, not a dedicated gadget, remains the primary platform for this.

The three tiers, and why the distinction is practical

Tier	What runs there	What leaves your house	Latency
Always-on sensing island	Wake-word detection, step counting, speaker separation, up to ~200M parameters	Nothing, until a gate opens	Milliseconds
Main on-device processor	Photo processing, transcription, translation, local assistants, up to ~30B sparse parameters	Nothing, if the feature was built local	Tens to hundreds of milliseconds
Vendor cloud	Large models, cross-device history, camera search, account features	The input, and often a stored copy	Hundreds of milliseconds, plus a record

The tier boundaries are where privacy actually lives. A wake-word detector in tier one keeps audio in a rolling memory buffer that is continuously overwritten, which is why Schönherr and colleagues’ 2020 study of accidental triggers across 11 smart speakers from 8 manufacturers matters so much: the only audio that reaches tier three is audio that got through the tier-one gate, deliberately or not.

Once something is in tier three, retention becomes a company policy question. The FTC and DOJ’s $25 million action against Amazon in May 2023 concerned exactly that: children’s Alexa voice recordings and geolocation data kept indefinitely and used to train algorithms, with parents’ deletion requests undermined.

The question that replaced “is this device private?”

“Is it private” has no answer, because a single device usually spans all three tiers.

Take a smart display. Wake-word detection: tier one. Local face grouping in photos: tier two. Asking it to summarise your calendar: tier three. Same box, three answers.

So the question becomes per-feature. For each thing a device does, which tier does it use, and does the manufacturer say so in writing?

Google’s Gemini for Home, announced June 25, 2026, illustrates the shape. Basic voice assistance is free on Google Home speakers. Google Home Premium starts at $10 a month and includes features like searching what your cameras are seeing. A feature that searches camera footage is a tier-three feature by definition, and the price tag is a reasonable proxy for where the computation happens: cloud compute costs money every month, so monthly pricing usually means cloud.

That heuristic is imperfect but useful. One-time purchase, works offline: probably local. Monthly subscription, needs an account: probably not.

How to Teach Your Kid About Edge vs Cloud Processing

Ages 5–8: the airplane-mode game

Put a phone or tablet in airplane mode and try things together. The camera works. The calculator works. Asking a question out loud does not. Taking a photo and having it recognise a dog might, depending on the device. Children learn the shape of the distinction long before the vocabulary, and the vocabulary can arrive later.

Ages 9–12: time the difference

Use a stopwatch. Ask a voice assistant a question that needs the internet and count the pause before the answer. Then use an offline function, like a local timer, and count again. The gap you are measuring is a round trip to a building full of computers, possibly on another continent. Then ask: what travelled during that pause?

Ages 13+: work out why a 30-billion-parameter model fits

Give your teenager the actual numbers. A phone might have 12 to 16 gigabytes of memory shared across everything. A parameter at common compressed precision takes roughly half a byte to a byte. Have them calculate how many parameters would fit if all had to be active at once, then look up mixture-of-experts routing and explain how 30 billion total parameters coexist with a small active subset. This is a genuine systems-engineering problem with real arithmetic, and it is far more interesting than any explainer about “AI.”

The question to ask: “If this device never connected to the internet again, which of its features would still work?”

The honest list is the local feature set. Everything else is a service, and services can change, be priced, or be logged. Our explainer on what edge AI means on a device goes further into the hardware side.

What to do at home

Run the airplane-mode audit on every device once

Not just phones. Televisions, speakers, cameras, watches, the doorbell. Disconnect the internet for thirty minutes and write down what each one can still do. That list is your actual local capability inventory, and it takes less than an hour for a whole house.

Treat monthly pricing as a cloud signal

If a feature costs money every month, the computation almost certainly happens on someone else’s servers, because that is what recurring costs pay for. This is not a reason to avoid it. It is a reason to know it, and to decide whether that specific feature is worth sending that specific data.

Prefer the local option when there is one

Many devices offer both. On-device transcription instead of cloud transcription. Local face grouping instead of cloud photo analysis. The local version is usually slightly worse and considerably more private. For a child’s device, that trade is almost always worth making, as our piece on edge AI and family data argues in more detail.

Read the settings menu for the word “improve”

Toggles labelled “help improve this service” or “share usage data” are the ones that move your inputs into tier three for training purposes. They are usually on by default and almost always optional. Turning them off takes two minutes per device.

Ask about the sensing hub specifically

The always-on tier is the least discussed and the most continuous. When you buy a device with a microphone, the question worth asking is whether the always-on detection runs locally and whether muting cuts power to the microphone in hardware. Manufacturers who do it properly usually say so, because it is a selling point.

What not to do

Do not assume a device with a powerful chip processes locally. Capability and architecture are separate decisions. Plenty of devices with capable silicon send everything to a server because the company’s product works better that way, and the silicon is there for the camera.

What to Watch For Over the Next 3 Months

• Week 4: Watch for phone features advertised as working offline. That phrasing appeared in marketing during 2026 precisely because the hardware made it possible, and it is a genuinely useful buying signal.
• Month 2 red flags: A feature that worked offline starting to require a connection after an update. A new “premium” tier that includes something you already had. Either suggests a feature moved from tier two to tier three.
• Month 3 self-check: Redo the airplane-mode audit. If fewer things work offline than three months ago, something moved to the cloud, and it is worth finding out what and why.

Frequently Asked Questions

What does edge processing actually mean?

Processing on the device where the data was created, rather than on a remote server. In 2026 that includes genuinely large models: Qualcomm’s Snapdragon 8 Elite Extreme Gen 6 supports running a 30-billion-parameter mixture-of-experts model locally, and its always-on sensing hub handles models up to 200 million parameters.

Is local processing always more private?

It is more private for the data that stays local, which is the main point. But a device can process locally and still upload summaries, metadata or logs. “Runs on device” and “sends nothing” are different claims, and only the second one is a privacy guarantee.

How can a 30-billion-parameter model run on a phone?

Through mixture-of-experts routing. The model’s parameters are divided into specialised groups and only a small subset activates for any given task, so total size can be large while the memory bandwidth needed at any instant stays within a phone’s budget.

Does my smart speaker process the wake word locally?

On mainstream devices yes. A small model runs continuously on low-power hardware, scoring a rolling audio buffer, and audio is streamed to servers only after it fires. The 2020 Bochum study on accidental triggers is useful context on how often that gate opens unintentionally.

What is the fastest way to find out where a feature runs?

Turn off the internet and try it. The answer takes ten seconds and is more reliable than any marketing page. Our microphone and camera count pairs well with this test, because it tells you which devices are worth testing first.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. TechCrunch. (2026, September 22). “Qualcomm launches two new smartphone chips with emphasis on AI.” TechCrunch. https://techcrunch.com/2026/09/22/qualcomm-launches-two-new-smartphone-chips-with-emphasis-on-ai/
2. Google. (2026, June 25). “100 new ways to make your day easier with Gemini for Home voice assistant.” Google Blog. https://blog.google/products-and-platforms/devices/google-nest/gemini-home-voice-assistant-tips/
3. Federal Trade Commission & U.S. Department of Justice. (2023, May 31). “FTC and DOJ Charge Amazon with Violating Children’s Privacy Law by Keeping Kids’ Alexa Voice Recordings Forever and Undermining Parents’ Deletion Requests.” FTC Press Release. https://www.ftc.gov/news-events/news/press-releases/2023/05/ftc-doj-charge-amazon-violating-childrens-privacy-law-keeping-kids-alexa-voice-recordings-forever
4. Schönherr, L., Golla, M., Eisenhofer, T., Wiele, J., Kolossa, D., & Holz, T. (2020). “Unacceptable, where is my privacy? Exploring Accidental Triggers of Smart Speakers.” arXiv:2008.00508. https://arxiv.org/abs/2008.00508
5. Fagan, M., Megas, K., Watrobski, P., Marron, J., & Cuthill, B. (2022). Profile of the IoT Core Baseline for Consumer IoT Products (NIST IR 8425). National Institute of Standards and Technology. https://csrc.nist.gov/pubs/ir/8425/final
6. Federal Communications Commission. “Cybersecurity Labeling for Internet of Things (U.S. Cyber Trust Mark).” FCC. https://www.fcc.gov/CyberTrustMark