The iPhone's Last Stand | Stratechery by Ben Thompson

ELI5/TLDR

Microsoft unveiled a flashy concept for a future where you talk to AI “agents” living in the cloud through a swarm of thin, dumb devices — no phone required. The same week, Apple finally showed off a rebuilt Siri that actually works, but it’s noticeably behind the cutting edge: it can find things and set reminders, but it won’t go do tasks for you. Ben Thompson’s argument is that Apple’s “behind-ness” doesn’t matter, because the people who pay for AI agents are companies trying to make employees productive, and regular consumers don’t want to work — they want to watch short videos on a phone they already own. So Siri being merely good enough is, for now, enough.

The Full Story

Two visions, one week

Two of the biggest names in computing laid their cards on the table within days of each other in June 2026, and they pointed in opposite directions.

Microsoft, at its Build developer conference, announced Project Solera (Thompson alternately calls it Solara) — a moody, music-heavy concept video for “a new chip-to-cloud platform for an agent-first world.” The pitch: in the future you’re surrounded by a halo of small devices, none of which really does anything on its own. They’re just portals — windows you talk through — to AI agents that live in the cloud and do the actual thinking.

Thompson is openly skeptical, and he reaches for a word Apple fans used to throw at Microsoft: vaporware — products a company talks about but may never ship. After Apple’s own embarrassing 2024 launch of Apple Intelligence and a new Siri that didn’t materialize, he figures the jab is fair game again.

“Apple fans would, for years and years, sneer at Microsoft’s pension for talking about products that may or may not ship, deriding them as vaporware. After Apple’s bungled 2024 launch of Apple Intelligence and new Siri, however, vaporware is fair game.”

Thin is in — the idea behind Solera

To understand why Microsoft’s vision isn’t crazy, you need one concept: the thin client. Think of it like a TV remote versus a games console. A console does the heavy lifting itself; a remote does nothing but send your wishes to a box that does the work. A “thin client” is a device that’s deliberately dumb — all the real computing happens on a server somewhere else.

Thompson had argued months earlier that AI pushes computing toward the extreme thin-client end. His reasoning: an agent — software that goes off and completes a task for you — doesn’t need your device to be smart at all. Everything between your request and the finished result is meant to be invisible.

“It’s not just that you don’t need any local compute to get an answer from a chatbot. You don’t need any local compute to accomplish real work. The AI on the server does it all.”

But there’s a catch that doomed earlier “post-phone” gadgets like AI pins and smart glasses: interacting with them sucks. For all of computing history, using a computer and interacting with it were the same thing — which is why every new input method (mouse, touchscreen) triggered a revolution.

“The problem with wearables as the paradigm beyond the iPhone is that interacting with them generally sucks… anything longer than a few seconds is much less convenient than simply swiping on your phone.”

Agents are the one thing that escapes this trap. Because an agent works on your behalf while you do nothing, a few seconds of talking can, in theory, buy you hours of work. That’s the use case where a thin, voice-first device finally makes sense — and it’s what made Microsoft’s demo interesting to Thompson despite the eye-rolling.

Apple’s answer — Siri AI

Apple walked into its WWDC keynote facing a different question. Microsoft can dream up devices; the doubt about Apple was the reverse — can the company actually make AI?

By the standard of mid-2024, the answer was a solid yes. Apple showed real, working demos (spinning loaders and all) of what it now calls Siri AI. In the marquee demo, a user asks when a concert is, gets a grounded answer pulled from current world knowledge, then says “remind me to sign up when the lottery opens,” and Siri sets the reminder.

Here’s Thompson’s sharp observation — that demo simultaneously reveals how far behind Apple is:

“What would have been state-of-the-art would have been asking Siri to enter the lottery on his behalf when the time came. In other words, to act outside of the interaction paradigm that has traditionally defined computing.”

Setting a reminder is Siri telling you to do something. The cutting edge would be Siri doing it for you. Apple stopped at the first.

Why being behind might not matter

This is the heart of the essay, and it rests on two ideas.

First, Apple’s structural advantage: personal context. Because Siri lives on your iPhone, it knows more about you than any rival AI ever could — your messages, email, voicemail, what’s on your screen right now. Apple deliberately aims Siri at a constrained, grounded problem space, the kind where it’s hard for the AI to embarrass itself.

“Apple is addressing a space that is very useful that only they can address and which also happens to be safe in terms of reputation risk. Honestly, it almost seems unfair… it speaks to what a massive advantage there is for a trusted platform.”

Second — and this is the contrarian core — consumers don’t want to be productive. Thompson argues Silicon Valley has to relearn this every decade. His example is Dropbox: a product that spread virally among consumers (refer a friend, get free storage) but could only make real money once it gave up and sold to enterprises. Why? Companies pay for their employees’ time, so they’ll happily pay for tools that save it. Consumers, by contrast, are mostly trying to kill time.

“Consumers, on the other hand, are mostly looking to waste time, which is why attention-harvesting advertising is the only software business model that works at scale for consumer services.”

He claims OpenAI repeated Dropbox’s mistake — betting it could fund itself on consumer subscriptions — while Anthropic quietly went where the money was, selling AI’s productivity gains to enterprises.

The punchline: agents are a productivity tool, and consumers don’t want productivity. They want short-form video, and an iPhone is the best short-form-video machine ever built. So Siri merely being “good enough” clears the only bar that matters.

“Agents help you do work and be more productive, and consumers don’t want to work or care about being productive. What they do want to do is watch short-form video.”

The iPhone at the center

Under the hood, Apple rebuilt Siri with some genuinely clever engineering — extending its private cloud compute to Nvidia chips running inside Google’s data centers, plus a 20-billion-parameter on-device model built as a “mixture of experts” that picks the right specialist per query (rather than per word) so it can squeeze into an iPhone’s tight memory.

But the strategic takeaway is simpler: everything orbits the iPhone. Microsoft’s cloud-and-devices vision suits a company that missed mobile entirely. Apple, by contrast, is fiercely motivated to keep the iPhone central — which conveniently also spares it from spending billions on data-center capital expenditure (CapEx).

Thompson notes this isn’t just cynical self-interest; both strategies make sense on their merits. He finds Solera most interesting precisely because Microsoft frames it as a pure enterprise play — and enterprises have the rich context about their own work that makes long-running agents viable and worth paying for. Consumer context is far harder to assemble, because it’s scattered across countless apps and services. The only two companies that could stitch it together are Apple and Google — and Google will always favor its cloud over the device.

“That leaves Apple as the only company truly, dare I say it, thinking differently.”

The closing read: rivals will spend the billions on AI infrastructure, and you’ll get to use all of it on your iPhone — you’ll just have to open an app. But for anything personal, or anything that has to work across your apps, Siri will be the only one who can pull it off. “As long as it’s not vaporware. And it appears the second time is the charm.”

Key Takeaways

Thin client = a deliberately dumb device that offloads all real computing to a server. AI agents push computing toward this extreme, because nothing useful has to happen locally.
Agents break the “interaction sucks” curse of post-phone hardware. Voice gadgets fail because anything beyond a few seconds is worse than swiping a phone — but an agent works while you do nothing, so a few seconds buys hours of work.
Apple’s Siri is one step behind the frontier on purpose-or-not: it can find your info and set a reminder to do a task, but it won’t autonomously do the task. The state of the art would act on your behalf; Siri stops at telling you.
Apple’s real moat is personal context. Because Siri sits on the iPhone, it can search your messages/email/voicemail and see your screen — a “trusted platform” advantage rivals can’t match without huge security sacrifices.
The Dropbox lesson: productivity tools make money from enterprises, not consumers. Companies pay for employee time; consumers want to waste time. Advertising is the only consumer software model that scales.
Agents are a productivity product, so they’re an enterprise product — which is why Thompson thinks Apple’s agentic weakness doesn’t matter for consumers (yet) and why Microsoft framing Solera as enterprise-only is the smart move.
The strategies map onto incentives: Microsoft (missed mobile) goes cloud-and-devices; Apple (owns the iPhone) keeps everything device-centric — and avoids billions in AI CapEx by doing so.
Technical nugget: Apple’s on-device Siri uses a 20B-parameter mixture-of-experts model that selects an expert per query rather than per token, a memory-saving trick to fit in an iPhone.
Anthropic vs OpenAI framing: Thompson credits Anthropic for going straight to enterprise productivity, and dings OpenAI for over-betting on consumer subscriptions while declining to build advertising.

Claude’s Take

This is classic Thompson: a tidy strategic frame, one genuinely good insight, and a couple of sweeping claims stated with more confidence than they’ve earned.

The good insight is real — the observation that agents are the first post-phone paradigm that escapes the “interacting sucks” problem is a clean, useful way to think about why AI pins and glasses flopped but agents might not. Likewise the “consumers don’t want to be productive” point is an old truth worth re-hearing, and the Dropbox parallel lands.

Where I’d push back: the consumer-vs-enterprise dichotomy is doing a lot of load-bearing work. “Normal people don’t want agents” is asserted, not demonstrated — plenty of consumers would happily let an AI book a restaurant, fight a parking ticket, or untangle a billing dispute. The claim that advertising is “the only” scalable consumer software model conveniently ignores Netflix, Spotify, the App Store, and, yes, the iPhone itself. And the OpenAI dig is more vibes than evidence; “couldn’t make money on consumer subscriptions” is a strong factual claim dropped without a number behind it. Note also this is a paid-newsletter author whose whole worldview (“Aggregation Theory,” platform moats) predisposes him to crown Apple — so the “Apple is the only one thinking differently” flourish reads as much as house style as analysis.

Net: a sharp, well-argued piece that’s worth the 15 minutes for the agent/thin-client framing and the productivity-vs-attention lens. Just hold the tidy conclusions loosely — Thompson is excellent at making a contrarian read feel inevitable, which is exactly when to keep one hand on your wallet. A 7: high signal, clean thinking, but built on a few assertions that deserve more scrutiny than the essay gives them.