← All tech signals
Infrastructure

Intelligence That Stays Home

A model small enough to fit inside a phone now answers without a byte leaving the device, ending AI’s need for someone else’s data centre.

What is it?

A model small enough for a phone now answers without a byte leaving the device — Apple and Google both shipped runtimes for exactly that. It reads as hardware, but it’s really about where trust in AI sits. Whether people will ever notice, or care, stays open.

Why now?

Two pressures arrived together: phones finally have headroom for real inference, and buyers increasingly read outbound data as risk, not convenience. Much of the groundwork sits in open developer communities, not finished products. Trust standards for a local model are being set before most organisations take a position.

Who’s building it?

App-platform makers compressing frontier models onto phone-sized chips — while the trust standards form elsewhere entirely.

AppleGoogleChip makersOpen-source runtimesEnterprise IT
The full read
Why now, who is building it, where the market goes, and what stays unsettled.
Get weekly →
01 · The market
Who’s in it
Builders — phone makers and the open-source runtime projects beneath them.
Buyers — privacy-wary consumers first, followed by enterprises wary of data leaving their control entirely.
Gatekeepers — chip makers, app stores, and the forums setting update norms.
What’s at stake
The privacy pitch is real, but so is lost visibility — nobody outside the device can audit what it did.
Battery life and inference quality trade off against each other.
Standards are forming in open forums, well ahead of vendors.
What’s unsettled
Does local really mean private, or just private from one company now?
Who controls the update path once a model is running on-device?
Whether compression narrows the capability gap enough to matter for real, demanding tasks.
02 · How the demand matureda progression, not a timeline
Capability
Sparse models finally fit inside a phone’s memory and battery limits.
Adoption
Offline features ship quietly into devices, felt more than announced.
Standard-setting
Open developer forums decide update, audit and trust norms before vendors settle on their own.
Default
On-device inference becomes the assumed mode, not the fallback option for most tasks.
03 · Where the market goesthe forks, not a prediction
→Convergence or fragmentation — whether one open runtime wins out, or every vendor keeps its own compression stack and update path.
→Enterprise IT may end up setting the real trust bar before any regulator writes one.
→The gap between phone-sized and data-centre models keeps narrowing, changing what “good enough” now means.
→Value could sit in the chip, the runtime, or the model weights themselves.
Underneath every fork sits one open premise: that running locally is actually more trustworthy, not merely less visible. Nobody has proven that yet.
04 · The Edge Shift deskadjacent markets, moving now
Hardware Before Habit
Chips built for AI, not humans.
A chip built for the PC.
AP and Reuters via Al Jazeera
Data Stays Home
Local-first apps keep a full copy on the device.
Local-first software makes each device a full replica of the app’s state.
Martin Kleppmann
Small Models, Big Intimacy
Compact models bring real AI into daily habits.
IDC forecast gen-AI smartphone shipments growing over 363% in 2024.
IT Reseller (citing IDC)
05 · The Wire
01
Apple’s on-device model loads only 1 to 4 billion parameters per prompt
The Next Web ↗
02
Google’s AI Edge Gallery app runs Gemma 4 fully offline on phones
Android Authority ↗
The panel on this signal
They agreed on
–The pull is real on both sides — hardware finally has the headroom, and users are wary of sending data outward.
–Standards are forming in open forums, not vendor roadmaps.
–Compression is steadily narrowing the gap between phone and data-centre models.
The privacy counsel disagreed on
–Local processing removes the network transmission step but not necessarily later on-device logging, retention or a subsequent sync; without independent verification of what a device stores, trustworthiness is asserted, not demonstrated.
Know more about the method →
✦How this signal was built
01 · Panel
A Delphi panel of 4 AI analysts, each reasoning independently.
→
02 · Debate
They challenge and revise each other’s reads over two rounds.
→
03 · Evidence
Every claim is tied to a cited, linked source on The Wire.
→
04 · Synthesis
A moderator writes it into one read, disagreement and all.