Samsung wants you to believe your Galaxy S26 Ultra keeps your data on your phone. That’s a bold claim in a market where “on-device AI” has become the most overused phrase in tech marketing. The S26 Ultra series brings a new privacy display and a redesigned camera bump, along with the usual wave of AI feature announcements [1]. But here’s the question nobody at the launch event wants to answer directly: how much of that AI actually runs on the phone, and how much just phones home to the cloud the second you turn your back?
What Samsung Actually Built Into the S26 Ultra
The privacy display is a real hardware feature. It narrows the viewing angle so the person next to you on the train can’t read your screen. That’s a physical, verifiable upgrade you can test in about five seconds without trusting a single marketing slide.
The AI story is murkier. Samsung has leaned hard into what it calls hybrid AI, meaning some tasks run locally on the phone’s chip and others get shipped to Samsung’s servers or a partner cloud. The company frames this as the best of both worlds: fast, private processing for simple stuff, and heavier cloud compute for complex requests.
That’s not a new idea. Samsung’s own executives were making this exact case back with the Galaxy S24 Ultra, when the Snapdragon 8 Gen 3 chip delivered somewhere in the range of 10 to 15 TOPS of NPU performance to power features like Live Translate, generative photo editing, and Circle to Search. Independent testing at the time showed on-device features ran faster and more reliably offline than cloud-dependent tools, especially on mid-range phones, though the more demanding generative fills still fell back to the cloud when the results needed to look good.
That’s the pattern to watch with the S26 Ultra. The easy stuff (basic translation, simple search, quick edits) tends to stay local. The impressive stuff, the kind that ends up in a keynote demo, often still needs a server somewhere doing the heavy lifting.
The NPU Arms Race Nobody Outside Tech Media Cares About
TOPS numbers are climbing fast across the entire industry. Counterpoint Research projects that more than 70% of premium smartphones will ship with hardware-accelerated AI chips by 2025[2], and flagship NPUs are now hitting 10 to 20 TOPS ranges that make real-time translation and low-latency assistants actually usable[2] instead of laggy party tricks. Qualcomm’s Snapdragon X Elite platform, aimed more at laptops than phones, pushes as high as 45 TOPS, enough to run local language models with 7 to 13 billion parameters at usable speeds[2].
I’ve tested enough of these chips generation over generation to tell you the raw TOPS figure is a starting point, not the finish line. What matters is what the software actually does with that horsepower. A chip capable of 20 TOPS is useless to you if the app still routes your voice command through a server in another state.
IDC projects that AI PCs will account for nearly 60% of all PC shipments by 2027[3], which tells you the entire industry, phones and laptops alike, is betting the farm on local AI processing becoming standard [3]. Whether that bet pays off for the average person buying a phone is a separate question entirely.
Why “On-Device” Doesn’t Automatically Mean “Private”
This is where I get skeptical, and you should too. Samsung’s TM Roh has said the company built Galaxy AI around a hybrid approach so users get experiences “you can trust and rely on,” combining on-device and cloud processing. Dr. Won-Joon Choi, who leads mobile R&D at Samsung, has made a similar case: on-device AI means lower latency and better privacy because your data stays put, while cloud AI handles the tasks that need more horsepower.
Apple’s Johny Srouji has said basically the same thing about Apple Silicon, and Qualcomm’s Kedar Kondap makes the identical pitch about reduced latency and reduced cloud costs alongside better privacy. Notice a pattern? Every chipmaker and every phone maker says the same three sentences with different names swapped in.
Cryptography professor Matthew Green has a more careful take, and it’s the one I trust most. Doing more processing on the device can meaningfully cut privacy risk, since less of your data ever leaves your control. But you still have to trust the vendor not to quietly ship that data off for analytics anyway. On-device processing reduces exposure. It doesn’t eliminate it, and it definitely doesn’t replace the need for a company to be transparent about what it collects.
The Electronic Frontier Foundation makes the point even sharper: features billed as private or on-device can still involve heavy data collection and profiling unless you get real transparency and real control over the settings. A privacy display is easy to verify. A privacy claim about background AI processing is not. You can’t see it happening, so you’re stuck trusting a settings menu you probably never open.
What History Tells Us to Expect From the S26 Ultra
Nobody has independent benchmark data on the S26 Ultra yet, so let’s be honest about that upfront. But the generational trend gives us a reasonable way to predict what’s coming. Apple’s M1 to M3 transition on the Neural Engine delivered 20 to 60% ML throughput gains per generation on tasks like image classification, speech recognition, and local transformer inference, and Johny Srouji has said the M3 Neural Engine runs up to 60% faster than its predecessor for exactly those kinds of tasks.
If Samsung’s chip suppliers follow a similar curve from the Snapdragon 8 Gen 3 era, expect real, measurable gains in on-device speed for the S26 Ultra, particularly for translation and camera processing.
Here’s what I want you to watch for once real units ship: does the improved NPU actually reduce how often features fall back to the cloud, or does it just make the cloud round-trip feel faster? Those are two very different kinds of progress, and Samsung’s marketing won’t tell you which one you’re getting. I plan to test this directly with airplane mode the moment I get a unit in hand, the same way I tested Live Translate on the S24 Ultra. If a feature breaks the second you lose signal, it was never really on-device to begin with.
-
Turn on airplane mode and try the marquee AI features. If they stop working, they weren’t running locally.
-
Check your phone’s data usage settings after a week of AI feature use. Heavy upload activity is a tell.
-
Look for a published, specific TOPS figure from Samsung rather than vague “AI-powered” language.
-
Compare response latency between simple tasks (translation) and complex ones (generative image edits). A big gap usually means the complex task is cloud-routed.
Do Buyers Even Care About This?
Here’s the uncomfortable truth for every phone maker pouring billions into AI silicon: most people aren’t buying phones because of AI. Consumer tech analyst Carolina Milanesi put it plainly. People buy a phone for the camera, the battery, the price, and maybe some cool features that happen to be powered by AI, not because a company slapped “AI phone” on the box.
Recent consumer surveys back her up. Only a minority of buyers, often under 20%, name AI features as a primary reason for their purchase[2]. Camera quality, battery life, and price still run the show.
Marques Brownlee has made a similar point in his reviews: a lot of these AI features are neat demos you use once or twice, not reasons to actually upgrade your phone. PCWorld’s Gordon Mah Ung goes further, calling “AI PC” mostly a marketing term right now. The hardware is real, he says, but the software ecosystem that would make these features essential is still young.
I’d argue phones are a step or two ahead of PCs on that front, since translation and camera AI have had a few generations to mature. But the core criticism still lands. A chip capable of 45 TOPS doesn’t matter if the one feature you actually use is Circle to Search.
My Take: Test Before You Trust the Marketing
The privacy display on the S26 Ultra is a legitimate, physical win you can verify with your own eyes. The on-device AI story is more complicated, and I’d bet real money that a chunk of the flashiest features still lean on the cloud when the task gets hard. That’s not necessarily a dealbreaker. Cloud processing isn’t evil, and hybrid AI is a reasonable engineering compromise. The problem is marketing language that blurs the line between “processed locally” and “processed locally most of the time, except when it isn’t,” without ever telling you which mode you’re in.
Analyst Anshel Sag has warned about vendors slapping “AI” on every new device without delivering experiences that actually justify the upgrade. That risk applies just as much to phones as it does to laptops.
My advice: don’t buy the S26 Ultra because of an AI feature list. Buy it if the camera, screen, and privacy display genuinely fit how you use your phone. Then run your own airplane mode test the day it arrives. That’s the only way to know what “on-device” really means on your unit, in your hands, not in a press release.
Frequently Asked Questions
Does the Galaxy S26 Ultra’s privacy display actually work?
It narrows the screen’s viewing angle so people beside you can’t easily read it. This is a hardware feature you can test immediately, and it doesn’t depend on AI processing of any kind.
Is on-device AI on the S26 Ultra actually private?
Some features likely run locally, reducing how much data leaves your phone. But experts like Matthew Green note that on-device processing lowers risk without guaranteeing privacy, since vendors could still collect data for analytics unless they’re fully transparent about it.
What’s the difference between on-device AI and cloud AI?
On-device AI processes data using the phone’s own chip, which is faster and keeps data local. Cloud AI sends data to remote servers for more complex tasks that need more computing power, as Samsung’s Won-Joon Choi has described.
How can I tell if a feature is really running on-device?
Turn on airplane mode and try the feature. If it stops working without an internet connection, it’s relying on the cloud, no matter what the marketing says.
What to Do Next
Skip the keynote hype and wait for hands-on testing before you decide the S26 Ultra’s AI is worth paying for. Once units ship, run the airplane mode test yourself on the features that matter to you. If you want a second opinion grounded in actual testing rather than spec sheets, keep checking back here for the full review once the device is in hand.
Sources
-
Samsung Galaxy S26, S26+, and S26 Ultra: Specs, Features, Price, Release Date | WIRED (wired.com)
-
AI In The Consumer Electronics Industry Statistics 2026 (wifitalents.com)
-
Personal Computers and Smartphones: Processors, Accelerators, and Connectivity (my.idc.com)
Researched from 16 vetted sources · average source authority DR 77