Apple Lost the First Half of the AI Race — But It's Betting on the Next Decade

In November 2022, ChatGPT launched. For the next three years, the same question kept surfacing across tech: where is Apple?

Siri was still asking "would you like me to search the web?" while Bing Chat could draft your weekly status report. Google Gemini was weaving deep into Android, and iPhone users who invoked Siri still got the familiar colorful orb — followed by "sorry, I didn't catch that."

In 2024, Apple unveiled Apple Intelligence, promising to "redefine Siri." Then — delays. More delays. Some features quietly appeared in iOS 18.4, then quietly vanished in iOS 18.5.

Apple did not win the first half of the AI race. That's a fact. But at WWDC26, Tim Cook and Craig Federighi didn't walk on stage with a patch release. They delivered a full counter-offensive — Siri Agent, Foundation Models 2.0, Apple Intelligence going wide, and a long-term bet on privacy-first AI.

This piece walks through why Apple fell behind in the opening stretch, exactly what WWDC26 shipped, and whether the path Apple is charting for the next decade is a retreat into conservatism — or a much larger wager.

1 · The First Half: Why Apple Fell Behind

1.1 Siri's Structural Problem

Siri launched in 2011 — a full eleven years before ChatGPT. But shipping early doesn't mean leading early. Siri's underlying architecture was a stack of rule engines, speech recognition, and API bridges — not a language model. That meant its capabilities were pre-scripted: set an alarm, play music, check the weather. Step outside those lanes and you got the fallback — "here's what I found on the web."

GPT-4 demonstrated something entirely different: intent understanding, reasoning, and cross-context work. This wasn't Siri being "one generation behind." It was two fundamentally different system designs — a state machine versus a language model.

1.2 The Apple Intelligence Delay Mystery

At WWDC 2024, Apple announced Apple Intelligence — on-device intelligence, a rewritten Siri, ChatGPT integration, Private Cloud Compute. The keynote hall erupted. Then most features landed in iOS 18.1, 18.2, 18.4…

Behind the delays sits a structural reason that rarely gets airtime: Apple's way of doing AI is inherently slower than OpenAI's. OpenAI can swap models on the server overnight; users never feel the seam. Every Apple update must pass App Store review, ship inside a new iOS release, and run reliably across four billion devices — the most demanding AI deployment surface on the planet.

1.3 Perception Gap: Losing the Narrative Doesn't Mean Losing the Tech

One detail that's easy to miss: Apple's Neural Engine has been built into silicon since the A11 in 2017 — earlier than most AI players were thinking about edge hardware. The NPU on M-series chips ranks among the best in the industry. The on-device Foundation Models stack at ~3B parameters can do more than many people expect.

What Apple lost was the visible AI product experience — the moments that get screenshotted, shared, and demoed to friends. Losing the first half of the race doesn't mean the hand you're holding is weak.

2 · WWDC26 Breakdown: What Actually Shipped?

2.1 Siri Agent: Finally Does Things, Not Just Talks

This is the most important shift at WWDC26. Old Siri was a Q&A interface — you ask, it answers. The new Siri Agent is an action executor — you state a goal, it gets it done.

The capability gap comes down to two things: cross-app action chaining and multi-step task planning.

Concrete example: you tell Siri Agent, "Turn yesterday's meeting recording into a to-do list, send it to my work group chat, and schedule a follow-up on my calendar for tomorrow." That's a four-step task spanning Notes, Messages, and Calendar. Legacy Siri hears that and offers to search the web. Siri Agent actually runs it.

The mechanism underneath is App Intents 2.0 — Apple exposes hundreds of built-in system intents for Siri Agent to call, and third-party developers can surface their app's core actions through the AppIntent protocol. Siri Agent is, at its core, an LLM-driven intent routing engine — and intents are how it acts on the world.

2.2 Apple Intelligence Goes Wide: From "Preview" to "Product"

The clearest signal at WWDC26: Apple dropped the "Beta" label next to Apple Intelligence. That's not cosmetic — it means Apple believes the stack is stable enough, complete enough, and confident enough to call it a product rather than a promise.

What landed in practice:

• Writing Tools across languages: rewrite, summarize, and tone adjustment now cover 20+ languages including Traditional Chinese, Japanese, and Korean
• Image Playground upgrades: expanded beyond cartoon styles; new personalized Genmoji can generate stickers based on your contacts
• Photo Intelligence improvements: natural-language queries like "photos from last summer at the beach" are significantly more accurate
• Notification summaries refined: fixes last year's widely criticized misreads that turned news into clickbait; adds priority-tiered display
• Screen Awareness: Siri can now read what's on screen and answer questions or take actions based on the current view

2.3 Foundation Models 2.0: Real Developer Firepower

Last year's Foundation Models framework was already impressive — zero token fees, no API key, data never leaves the device. Foundation Models 2.0 at WWDC26 goes further:

import FoundationModels

let session = LanguageModelSession()

// New: vision understanding (pass an image directly)
let image = UIImage(named: "receipt.jpg")!
let result = try await session.respond(
    to: "Extract the line items from this receipt into JSON",
    including: [.image(image)]
)

// New: structured output (returns a Swift Codable type directly)
struct Invoice: Codable {
    let vendor: String
    let total: Double
    let items: [InvoiceItem]
}

let invoice = try await session.respond(
    to: "Parse this receipt",
    including: [.image(image)],
    generating: Invoice.self
)

Core upgrades:

• Multimodal support: pass images directly; the model runs vision + language reasoning on-device
• Structured output: returns Swift Codable types directly — no manual JSON string parsing
• Streaming responses: token-by-token output for real-time UI
• Tool Calling: the model can invoke functions you define mid-reasoning for agent-style tasks
• Python SDK + fm CLI: non-Swift access for scripts and backend tooling
• Framework open source: the Foundation Models framework itself is on GitHub for community contribution

2.4 Private Cloud Compute 2.0: Cloud Inference With Verifiable Privacy

For tasks that need more than the on-device 3B model, Apple's answer isn't "just send it to OpenAI." It's Private Cloud Compute (PCC) — Apple-operated cloud inference clusters built for AI workloads, with the same privacy commitments as on-device.

PCC 2.0's headline improvement: Security Research Virtual Machine — any security researcher can request a VM replica of a PCC node to verify whether Apple's privacy claims hold up in practice. This is "trust but verify" taken seriously: don't take our word for it — audit the code yourself.

2.5 macOS 26 Tahoe × iOS 26: AI Woven Through Every Layer

WWDC26 also announced macOS 26 Tahoe and iOS 26. AI is no longer a standalone "feature module" — it's embedded across the OS:

• Xcode 27 on-device completion: multi-line code completion runs locally on Apple Silicon, no cloud round-trip
• Safari smart summaries: web page summaries generated on-device, never uploaded
• Finder semantic search: "find last month's Excel file about the earnings report" — natural-language search across local files
• Mail smart compose: replies drafted in your writing style from historical mail, fully offline
• Health app AI coach: personalized guidance from your health data; data never leaves the device

3 · Strategy Analysis: What Is Apple Betting On?

3.1 Privacy as a Moat, Not a Marketing Line

The key to understanding Apple's AI strategy is treating "privacy" as a business barrier, not a brand slogan.

OpenAI and Google lead with data volume, compute scale, and iteration speed. Apple can't close those gaps — it doesn't have comparable AI user data or fleets of A100s and H100s. It chose a different curve: push the strongest AI it can onto the device, and make "data doesn't need to leave" a feature rather than a limitation.

The side effect is a moat OpenAI struggles to replicate: you can't deliver true "user data never leaves the device" on rented servers. That architectural advantage only grows more valuable as AI regulation tightens — GDPR enforcement, data-sovereignty laws worldwide.

3.2 Ecosystem Lock-In: AI Features × Apple Silicon × App Ecosystem

One deliberate design choice at WWDC26: nearly every major new AI feature requires A17 Pro or later, or an M-series chip, to run in full. That's a clear upgrade driver — want Siri Agent? You'll want an iPhone 17. Want Foundation Models 2.0 running locally on your Mac? M-series performance advantage is obvious.

Meanwhile, deep App Intents integration means the entire iOS and macOS app ecosystem has to keep pace — apps that adopt App Intents get natural Siri Agent exposure; apps that don't slowly get edged out. Classic Apple ecosystem governance: guide developers with capability, not mandates.

3.3 The Long Game: OS-Level AI vs API-Level AI

OpenAI, Anthropic, and Google DeepMind are essentially selling AI as a service — call their API, pay per million tokens, get frontier model capability. That model crushed 2025–2026 commercially, but it has a structural weakness: anyone can swap the API — including Apple.

Apple's bet is different: make AI part of the operating system, not a replaceable service. Siri Agent's awareness of on-device context, Foundation Models' deep NPU integration, PCC's dependency on Secure Enclave architecture — all of this makes "Apple's AI" harder to rip out and replace with a third party.

4 · Developer Perspective: What WWDC26 Changes

4.1 App Intents: From Optional to Essential

If you maintain an iOS app, one item belongs at the top of your backlog after WWDC26: audit your core features and decide which ones should be exposed as App Intents.

Siri Agent's capability ceiling equals the set of actions reachable through App Intents. Every intent you expose is an action Siri Agent can complete for your users. When someone says "do X in [your app]" and you have no intent, Siri can only reply, "sorry, this app doesn't support that yet."

import AppIntents

struct CreateNoteIntent: AppIntent {
    static var title: LocalizedStringResource = "Create New Note"
    static var description = IntentDescription("Create a new note in the app")

    @Parameter(title: "Content") var content: String

    func perform() async throws -> some IntentResult {
        // Your business logic
        let note = NoteService.create(content: content)
        return .result(value: note.id)
    }
}

4.2 Foundation Models 2.0: Practical Use Cases

With multimodal input, structured output, and Tool Calling, Foundation Models 2.0 opens significantly more ground:

4.3 New Build Environment Pressure: Xcode 27 + iOS 26 SDK Pinning

All of these features depend on Xcode 27 and the iOS 26 SDK. That means your build environment has to keep up — and that's where things get messy.

Foundation Models 2.0 APIs behave differently on simulator versus hardware. Siri Agent's App Intent integration needs a specific Xcode build to index correctly. PCC integration testing requires specific entitlements. If your CI runs on GitHub-hosted runners, you're waiting on an uncertain Xcode 27 support timeline — whereas on a Cloud Mac, you can be on the Xcode 27 beta within hours of WWDC26 ending.

5 · The Next Decade on the Board

5.1 Regulatory Tailwind: Privacy Architecture Gets More Valuable

In 2026, global AI regulation is accelerating: the EU AI Act enters enforcement, multiple US states pass AI transparency bills, and China's generative-AI governance keeps evolving. In that environment, "privacy architecture you can audit" becomes scarcer — and more valuable.

Apple's verifiable PCC stack, on-device Foundation Models design, and differential privacy in Health — these are brand bonuses in the consumer market. In enterprise, healthcare, and finance, they're procurement requirements.

5.2 Hardware × Software Flywheel: Hard to Copy

A reality that doesn't get enough discussion: Apple is the only company on the planet that designs AI silicon, an AI operating system, AI app frameworks, and AI endpoint devices under one roof. Google designs TPUs but Pixel isn't the mainstream device line. Qualcomm designs NPUs but doesn't own the software stack. Microsoft builds AI software on third-party hardware.

That vertical integration enables optimizations others can't match: Foundation Models inference paths tuned directly to Neural Engine instruction sets; Siri Agent response latency pushed under 50ms because the hardware and software teams share a building.

5.3 The Open-Source Signal: Trust Building and Ecosystem Gravity

Open-sourcing the Foundation Models framework is one of WWDC26's most underweighted signals. Apple isn't a company that open-sources its core advantages — its edge has always been tight control of a closed ecosystem. Choosing to open Foundation Models now is primarily a trust-building move: external researchers and enterprise security teams can audit it, not just read a white paper.

It's also ecosystem gravity: an open framework attracts researchers, researchers publish, publications drive developer adoption, adoption enriches third-party apps, richer apps make Siri Agent more capable. It's the Apple Silicon + Swift playbook replayed — this time at the AI layer.

6 · Honest Assessment: Did Apple Win the Second Half?

I'm not going to hand you an overly bullish or bearish verdict. Let's look at the real challenges first.

6.1 Real Challenges Apple Faces

• The on-device model ceiling is real: a 3B-parameter model structurally trails GPT-5.5 and Claude Opus 4 on complex reasoning, code generation, and long-context work. Siri Agent can turn a meeting recording into a to-do list. It can't refactor your Swift project's architecture.
• Third-party app integration takes time: App Intents ecosystem depth depends on developer adoption. After WWDC26, it may take 6–18 months before enough apps support deep Siri Agent integration.
• Non-English experience still lags: Traditional Chinese, Japanese, and Korean NLU improved at WWDC26, but conversational fluency still trails the English build noticeably.
• User habit migration needs education: most people already reach for ChatGPT or another AI app for complex tasks. Getting them to trust Siri Agent for those workflows takes behavioral change — and time.

6.2 Where Apple May Actually Win

That said, Apple has built genuine advantages on several axes:

Privacy-sensitive workloads: healthcare, legal, finance, internal enterprise data — users in these domains will often choose a slightly less capable AI that guarantees data never leaves the device over sending sensitive input to OpenAI's servers. As AI adoption grows in these segments, Apple's share should grow with it.

High-frequency, low-complexity daily tasks: summarization, translation, rewriting, classification — these dominate AI usage volume but don't need GPT-5.5-class intelligence. On-device Foundation Models handle them well, with lower latency and zero marginal cost. For everyday users, "good enough and free" often beats "strongest but metered."

Depth of OS integration: Siri Agent's ability to read your calendar, mail, messages, and photos is something no third-party AI app can replicate — Apple doesn't open those APIs to competitors. That systems-integration moat is hard to copy on any short horizon.

7 · FAQ

Closing: The War Apple Wants to Win Isn't the One You Think

In the first half of the race, Apple lost a contest over "whose chatbot is smartest." It couldn't win that contest — and chose not to try.

The second half it's playing for asks a different question: when AI is everywhere and capable enough, whose AI do you trust most — and which one is most deeply woven into your life?

There's no fast answer and no clean leaderboard score — nothing as simple as "whose MMLU is higher." That's precisely why Apple picked this battlefield: in a competition without quick scoring metrics, early-mover hype gives way to long-term trust — and trust is what Apple knows how to build.

Apple didn't win the first half of the AI race. WWDC26 at least shows it knows which game it's actually playing. Whether that bet pays off over the next five years — we'll be back to talk about it.