Apple Just Made Server-Side AI Free for Any App Under 2M Downloads. If You Ship iOS, WWDC Handed Solo Builders a Real Inference Budget.

At its Platforms State of the Union on June 8, Apple said developers with fewer than 2 million first-time App Store downloads can call its Foundation Models on Private Cloud Compute for free. No server to stand up, no per-token bill, no infrastructure to babysit. If you ship an iOS app and you are under that threshold (which is almost everyone reading this), Apple just gave you a server-side AI budget you don't pay for.

The framing matters. For two years the iOS AI story was "on-device or nothing." The on-device model was free but small, and the moment you needed real reasoning you were back to renting an API and eating the bill. Apple just removed the cliff. You get a managed cloud tier, free, as long as your app is small enough to count as indie. The cutoff is the news: this is a perk that big apps explicitly do not get.

What you actually get

Three things shipped together, and only the first one made headlines.

The free cloud tier is Foundation Models running on Private Cloud Compute: Apple's own infrastructure, with the privacy posture it has marketed for two years. You call it through the same Swift API you'd use for the on-device model. Your code doesn't know or care whether the inference happened on the phone or in Apple's data center.

The second thing is a new LanguageModel protocol: a public Swift interface that cloud providers implement so their models sit behind the same API as Apple's. At launch, Anthropic and Google have implemented it. That means you can route a request to Claude or Gemini without rewriting your session logic, your tool calls, or your context management. You add a Swift Package Manager dependency for the provider and the rest of your app stays the same.

The third thing is the framework picking up image input, server-side calls to those third-party models, and a "Dynamic Profiles" system for building multi-agent workflows. Apple also said the Foundation Models framework goes open source later this summer.

Why the 2M cutoff is the whole story

Read the threshold as a clock, not a limit.

Right now you are under 2 million first-time downloads, so the cloud tier is free. The entire point of building an app is to cross that line eventually. When you do, the free inference goes away and you're paying, either Apple, or whichever provider you routed to. So the free window is precisely the window you're in today, and it's the window where every extra dollar of margin matters most.

That changes what's worth building. A feature that was too expensive to ship (summarize every note, draft every reply, analyze every uploaded photo) is now free to run at the exact stage where you have no revenue to cover it. The calculus that used to be "is this AI feature worth the API bill against my zero dollars of MRR" is, for the cloud tier, just gone.

What I'd build now, and what I'd wait on

Build the AI feature you've been postponing because of cost. If you had a summarize, rewrite, classify, or extract feature sitting in a backlog because the per-call price didn't pencil out at your user count, that constraint is lifted for the cloud tier. Ship it now while it's free and let it be a reason people upgrade.

Use the LanguageModel protocol from day one even if you only call Apple's model today. The whole value of an abstraction layer is that you adopt it before you need it. Wiring your app to route through the protocol now means that when you cross 2M downloads (or when Apple's model isn't good enough for one specific feature), switching to Claude or Gemini is a package line, not a refactor. The cost of adopting it early is close to zero. The cost of not adopting it is a rewrite under deadline.

Wait on betting your core product on Private Cloud Compute capacity or latency guarantees. iOS 27 ships to everyone this fall; the developer betas are out now. Free managed inference is a gift, but you don't yet have production numbers on throughput under load or what happens at peak. Use it for features that degrade gracefully (where a slow or queued response is annoying, not broken) and keep the on-device model as the fallback for anything that has to work offline or instantly.

The honest catch

Free has a ceiling, and the ceiling is named "2 million downloads." If your app actually works, you graduate into paying, and at that point you're either paying Apple's cloud rate or a provider's API rate, neither of which Apple has fully detailed for the post-threshold world. So treat this as a subsidy on the early stage, not a permanent zero. It lowers the cost of finding out whether an AI feature matters to your users. It does not lower the cost of running that feature at scale.

There's also a quieter dependency here that deserves its own conversation: when you call the cloud tier, you're routing through Apple's infrastructure and, for the third-party providers, through Google or Anthropic. The on-device model stays Apple's. The cloud tier does not.

But for where most solo iOS builders actually are (pre-revenue or barely past it, with a feature idea blocked on cost), this is the most useful thing Apple has shipped for indie developers in years. The infrastructure bill that used to gate your AI feature is, for now, zero. Build the thing while that's true.

Author

Lukas

@lukcombinator

Sources

• Apple unveils next generation of Apple Intelligence, Siri AI, and more (Apple Newsroom)
• Apple Outlines Major AI and Developer Tool Updates at 2026 Platforms State of the Union (MacRumors)
• WWDC 2026 Developer Tools: Foundation Models Now Swaps AI Providers Without Code Changes (TechTimes)
• Foundation Models (Apple Developer Documentation)