TL;DR: In mid-July 2026, Samsung Health showed users a pop-up asking them to consent to their health data being used for AI training — with a warning that declining would stop account syncing and delete their synced data. Samsung later walked back the scariest reading (only the AI-training copy gets deleted, not your records), but the coercive framing was the point. The lesson isn’t “Samsung bad.” It’s that every cloud service you feed data to will eventually ask for training rights, and the only computer where that question never comes up is one you own.

	Cloud AI (Samsung Health / ChatGPT / Copilot)	Local GPU (used RTX 3090 rig)	Rented cloud GPU (RunPod)
Who trains on your data	The vendor, on their terms	No one — inference stays on your box	You control the container; provider sees traffic
Monthly cost	~$20/mo subscription or per-token API	~$52/mo (year 1–2 amortized) + a few $ power	~$0.30–$0.90/hr, pay-as-you-go
The catch	Consent is a moving target	Upfront hardware cost, you maintain it	Data leaves your LAN; good for burst, not secrets

Honest take: Buy a used 24GB card and route anything you’d hesitate to say to your employer through it locally; keep the cloud for generic, non-sensitive work. Owning the hardware is the only consent setting that can’t be changed on you overnight.

What Samsung actually did

Starting around July 14, 2026, Samsung Health users began seeing a pop-up titled “Consent to the Use of Health Data for AI Training and Modeling.” Agreeing lets Samsung use synced health data — including human review — to improve Samsung Health’s algorithms and AI features.

The problem was the opt-out path. Users who declined got a warning that account syncing would be disabled and that “your health data will be deleted unless retained pursuant to applicable law.” Read cold, that lands as an ultimatum: hand over your steps, sleep, heart-rate, medication, and menstrual-cycle data for model training, or lose it.

Samsung later clarified that withdrawing consent deletes only the separate copy collected for AI development — your existing Samsung Health records stay intact. But syncing to your Samsung account still stops if you refuse, and the consent screen was written to make “yes” the path of least resistance. You can turn it off under Samsung Health → Settings → Privacy, toggling off “Consent to the use of health data for AI training and modeling.”

Whether you read it as a hard ultimatum or a badly-worded prompt, the structure is the same: the data is on Samsung’s servers, so Samsung sets the terms, and the terms changed with a single app update.

This is the pattern, not the exception

It’s tempting to file this under “Samsung overreached” and move on. That misreads what happened. Samsung didn’t do anything structurally different from the rest of the industry — it just made the ask explicit and put it in front of your face.

Every service that holds your data has the same latent option. The default terms of service for most cloud AI products already reserve broad rights to use your inputs to “improve the service,” which is the polite phrasing for training. The Samsung pop-up is what that clause looks like when a company decides to cash it in and needs fresh consent to do it cleanly. Health data drew the outrage because it’s the most intimate category most people sync — but the identical logic applies to every prompt you send to a hosted assistant, every file you drop into a cloud IDE, every voice note a wearable uploads.

The regulatory environment is catching up, slowly, and it’s worth being precise about how. Connecticut’s amended Data Privacy Act took effect July 1, 2026, and it’s often described online as an “AI-training opt-in.” It isn’t — it’s a disclosure mandate: covered businesses must now state clearly in their privacy notices whether they use personal data to train large language models. Affirmative opt-in is required specifically for selling sensitive data, not for training on it. So the law forces Samsung and its peers to tell you they train on your data; it doesn’t stop them. In the EU, GDPR gives users a stronger hand — a lawful basis is required, and consent obtained under duress is shaky — so EU users have a real complaint path through their data protection authority. But “file a regulatory complaint” is not a workflow. It’s a last resort after the data has already been copied.

Disclosure and complaint rights are useful. They are not the same as control. The only place you get actual control is a machine where the training question is never asked — because there’s no server-side copy to train on.

What “private AI” actually buys you

Running a model on your own GPU changes the trust model completely. The weights sit on your disk. Inference happens in your VRAM. Nothing about your prompt has to leave your LAN, and there’s no vendor sitting on a copy waiting for a future terms-of-service update to monetize it.

This is genuinely practical now in a way it wasn’t two years ago. The reason local models got good enough for daily work — MoE sparsity, speculative decoding, and multi-token prediction — is a story we’ve told in detail, and the short version is that a single consumer card now runs models that hold up for coding, writing, research, and — relevant here — private health or legal questions you’d never type into a hosted box.

The hardware math is unglamorous and cheap by comparison. A used RTX 3090 averaged about $1,254 in July 2026 across 319 tracked listings (fair range roughly $1,194–$1,306, with eBay dipping near $1,050), versus about $1,488 new. That’s 24GB of VRAM at 936 GB/s of memory bandwidth. On that card, Google’s Gemma 4 26B-A4B MoE runs at roughly 80–110 tok/s at Q4 — a tight ~22.6GB fit, so budget your context — which is far past reading speed and fine for interactive use. Power draw under load lands around 350W; at $0.12/kWh that’s about $0.042 per hour of active generation, so even heavy daily use adds a few dollars a month.

Put the cost side by side. Amortize $1,254 over two years and the card is ~$52/month, dropping toward the cost of electricity alone after that. A cloud AI subscription runs about $20/month indefinitely, and per-token API billing scales with use. The local box is not always the cheapest option on a pure dollar basis in year one — but the thing you’re actually buying is that the consent screen never appears. No vendor can retroactively decide your two years of private prompts are now training data.

If you want the privacy without owning hardware, renting a GPU on RunPod is a reasonable middle path for burst work — you control the container and tear it down when you’re done — but be honest that your traffic still crosses someone else’s infrastructure. Rented compute is good for a one-off heavy job; it’s not where you’d keep your most sensitive stack.

A practical routing guide

You don’t need to move everything local. The useful discipline is deciding, per prompt, whether it belongs on hardware you own.

Route to your local GPU:

• Health questions — symptoms, medications, mental-health, anything you’d tell a doctor under confidentiality
• Legal and financial research tied to your actual situation
• Private codebases and anything under NDA (a local model wired into your editor keeps proprietary code off cloud servers — see the aicoderscope.com guides on pointing Cline and Cursor at a local endpoint)
• Anything you’d hesitate to say in front of your employer or insurer

Cloud is fine for:

• Generic knowledge questions with no personal data attached
• Public-facing writing and boilerplate
• Throwaway brainstorming that reveals nothing about you

The dividing line is simple: if a future headline reading “[Company] now trains on your [X]” would make you wince, keep X local.

The honest caveat: local isn’t automatically private

Owning the GPU is necessary but not sufficient. Your stack can still leak if you don’t check it.

The clearest recent example is Ollama’s own default. Its 0.32.0 interactive agent mode launches with a banner model of glm-5.2:cloud — a cloud model, not a local one. If you assume “I’m using Ollama, so I’m private” and start typing health questions into that default, you’ve routed them to a remote server. The fix is to run an explicit local tag (ollama run gemma4:26b) and to treat the built-in web-search feature as unverified for privacy until Ollama documents exactly what it sends where.

The broader point is to audit before you trust. We did a tool-by-tool privacy audit of the common local stack — Ollama, LM Studio, ComfyUI, and others — covering which ones phone home with telemetry and how to shut it off. For the FOSS-first angle on self-hosting the surrounding tooling (dashboards, health trackers, note apps) without a cloud account, aifoss.dev is the companion resource. Run through both once, confirm your default model tag is local, and then you can honestly say the data stays on your machine.

FAQ

Did Samsung actually delete health data for opting out? Not your existing records. Samsung clarified that withdrawing consent deletes only the separate copy it collected for AI-training purposes; your Samsung Health data stored for the service itself stays intact. What you lose by declining is account syncing. The original pop-up was worded ambiguously enough that many users read it as a threat to all their data, which is why it blew up.

Is running a local model actually cheaper than a cloud subscription? Not necessarily in year one. A used RTX 3090 (~$1,254 in July 2026) amortized over two years is about $52/month plus a few dollars of electricity, versus ~$20/month for a typical cloud AI plan. The local box wins on privacy immediately and on cost only over a longer horizon or at heavy usage. Buy it for control first, savings second.

Which local model should I run for private health or personal questions? On a 24GB card, Gemma 4 26B-A4B (MoE) is a strong default — ~80–110 tok/s at Q4 on an RTX 3090, well past reading speed. If VRAM is tighter, step down to a smaller Gemma 4 or a Qwen 3.6 variant. The model choice matters less than confirming it’s a local tag and not a :cloud one.

Does GDPR protect EU users here? More than US law does. GDPR requires a lawful basis for processing and treats consent obtained under coercion as invalid, so EU users have a genuine complaint path through their national data protection authority. US protections are thinner — Connecticut’s July 2026 law, for example, only forces companies to disclose whether they train on your data, not to stop.

Isn’t a rented cloud GPU just as private as a local one? No. Renting (e.g., on RunPod) gives you control over the software container, which is better than a hosted assistant, but your traffic still crosses a third party’s hardware and network. It’s a good fit for burst compute or one-off heavy jobs, not for your most sensitive, always-on stack. For that, the data should never leave your own machine.

Sources

• Samsung Health won’t delete all of your data if you opt out of AI training after all — 9to5Google
• Samsung won’t kill health data if you opt out of AI training (Updated) — Android Authority
• Samsung clarifies what happens to your data if you withdraw Samsung Health AI consent — SamMobile
• Opt out of Samsung AI training, lose health data — Cybernews
• Connecticut Enacts Comprehensive AI Legislation: Key Obligations — Holland & Knight
• RTX 3090 Price Tracker US — Best Value GPU (Jul 2026)
• Gemma 4 26B A4B on RTX 3090 24GB? Tight Fit — WillItRunAI

Prices and market averages are as of July 2026 and move week to week; treat them as a snapshot, not a quote.