Apple stays out: a contrarian position for the AI era — AlphaSense

1B in, 20B out. Apple stays out of the war.

Apple’s real posture on AI isn’t “buying Google’s models” — it’s using its distribution power (Safari + App Store + Apple Intelligence routing) to negotiate a bundle of multi-supplier procurement contracts: OpenAI for free in exchange for distribution ($0), Google for $1B/yr (starting 2026-04), Anthropic walked away because the price was too high, the China market split between Alibaba and Baidu, and training compute rented from Google TPU (estimated $0.1–0.5B/yr). Meanwhile Google still pays Apple roughly $20B a year for Safari default-search placement — sum every cash flow and Apple’s net inflow on the “AI upstream” is still about +$18–19B / yr. This piece is built around five observations: 1) the multi-supplier graph and reverse cash flow; 2) Mac is the only consumer device that combines “silicon + macOS + your daily data” into a household compute + data hub; 3) macOS is the best vehicle for Computer Use today; 4) Apple’s 2026 capex is just $14B, an order-of-magnitude gap below the $660B+ of five hyperscalers combined; 5) smart glasses launching 2027 could become the next AirPods-class category entry point.

§01 · Reverse flow — $1B in, $20B out

The common simplified narrative is “Apple pays Google $1B to use Gemini.” Reality is much more layered: along the AI axis, Apple simultaneously maintains 5 third-party suppliers + 2 in-house layers + 1 training-compute supplier, and almost none of those relationships involve cash going out. The table below maps the entire 2026 stack.

2026 Apple AI supplier stack · all known relationships

Sources: CNBC, Bloomberg / Mark Gurman, Apple ML Tech Report, 9to5Mac, AppleInsider, official press releases; USD annualized basis.

The OpenAI relationship is a non-cash barter — neither side pays, but OpenAI monetizes via ChatGPT Plus conversions. Anthropic Claude quoted Apple in the “billions per year” range for the Siri job and was rejected; Claude has nonetheless reached the Apple developer ecosystem via Xcode 26.3. The China market requires a local model to obtain generative-AI regulatory approval, so Apple picked Alibaba as primary and Baidu as secondary.

Sum every cash flow · net inflow ~$18–19B / yr

Apple’s real posture on AI isn’t “buying Google’s models” — it’s using distribution power (Safari + App Store + Apple Intelligence routing) to negotiate a bundle of multi-supplier procurement contracts: OpenAI free for distribution, Google paid $1B, Anthropic rejected on price, China split between Alibaba and Baidu, training compute rented from Google TPU. The total cost is still far below the $20B Google pays unilaterally for default search.

— Apple doesn’t have an “AI strategy”; Apple has a roster of “frontier model suppliers”

§02 · Home hub — silicon + OS + your data

Mac’s true position in 2026 isn’t “cheap hardware that can run big models” — it’s the physical base inside the home that holds both compute and data. The truly scarce input for an AI agent has never been the model; it’s your everyday data: 14 years of family in Photos, 12 years of conversations in Messages, tens of thousands of contracts and attachments in Mail, every event in Calendar, every draft in Notes, every contract in Files, every health trace in Health. Apple has been quietly accumulating this for almost two decades — distributed across roughly 1.4 billion iPhones and ~150 million Macs globally, synced via iCloud, with the local copy always present.

To say it plainly: in the consumer market, only one player has all five pieces — silicon + operating system + your daily data + privacy-default architecture + on-device inference framework. That player is Apple.

A typical household’s “data foundation”

Google wants this dataset, OpenAI wants it, Meta wants it — but none of them can get to it without the user actively uploading it, because it lives locally on Apple devices, locked behind sandboxes and Secure Enclave. Apple is in fact the compliance-default gateway to this dataset.

— in 2026 the scarcest agent context is your past, not a new model

Mac local-inference ladder

Reference price = US Apple Store base config · tok/s measured on llama.cpp / MLX, varies by model and quantization.

Mainstream 2026 Apple Silicon configurations · M4 Ultra has not yet shipped, the flagship remains M3 Ultra (819 GB/s). Equivalent 70B inference on the x86 path requires an H100 ($30K+ per card) and 5–10× the power.

Four pillars of the “home hub”

1. Silicon: unified memory + high bandwidth. CPU / GPU / Neural Engine share the same RAM pool, so model weights don’t need to be ferried over PCIe. M4 Pro 273 GB/s, M4 Max 546 GB/s, M3 Ultra 819 GB/s — data-center-GPU-class memory bandwidth, packed into a desktop, ~30W-idle chassis.
2. OS: macOS Tahoe (26) is now a competent home server. launchd daemons let agents start at boot (before any user login) with auto-restart on crash; Tailscale mesh VPN lets your iPhone / iPad / out-of-home MacBook connect directly to the Mac mini at home with zero exposed ports and zero public IP. Caffeinate / pmset to prevent sleep is a known engineering issue, not a blocker.
3. Data: all your everyday data is already on the Mac. Photos / Messages / Mail / Notes / Calendar / Files / Health all have full local copies on the Mac (or “Download Originals” as an option). An agent can read, index, and run RAG without uploading anything to any cloud; ollama / mlx embeddings hit the file system directly. This is structural — Google / NVIDIA / Meta cannot buy their way into this layer.
4. Framework: Apple Intelligence + App Intents are the compliance channel. The Apple Foundation Model on-device version is ~3B parameters with 2-bit QAT + LoRA adapters; the server version uses PT-MoE on Apple’s own M-silicon data centers (Private Cloud Compute, encrypted memory, no human visibility). App Intents let any third-party app expose features, entities, and queries to Siri / Apple Intelligence — from iOS 26.4, Personal Context already spans the six system domains of Calendar / Files / Mail / Messages / Notes / Photos.

One line · who actually has all five

The only player with silicon + OS + data + privacy architecture + on-device inference is Apple.

• Google: has the data (Search / Gmail / Drive / Photos), but no consumer hardware distribution; data lives entirely in the cloud, the user controls no local copy.
• NVIDIA: silicon only. No consumer OS, no consumer data, no household form factor.
• Meta: has social data + Ray-Ban glasses + Quest, but no consumer OS, no on-device inference hardware, and most of the data still lives in the cloud.
• Microsoft: has enterprise OS + Office data; consumer Surface has been peripheral for a long time, and Copilot+ PCs are still chasing Apple Silicon’s memory bandwidth.
• Apple: M-silicon + macOS + 1.4 billion devices’ worth of Photos / Messages / Mail / Notes / Calendar / Health + sandbox / Secure Enclave / Private Cloud Compute + 3B on-device foundation model + App Intents. All five.

This isn’t the “Mac inference is cheap” story. This is the first time household compute and household data have been collected by a single system.

§03 · Computer Use — the agent’s preferred OS

On 2026-04-16, OpenAI shipped the largest update ever to its Codex Mac client — adding Computer Use: Codex can see the screen, control the cursor, and click and type across any macOS app. MacStories’ Federico Viticci, in his first-hand review, gave the verdict directly: “the best computer use feature I’ve ever tested”.

That same week OpenAI also upgraded the underlying model to GPT-5.5, with SWE-bench Verified at 88.7%, Terminal-Bench 2.0 at 82.7%, and a hallucination rate down 60% from GPT-5.4. In other words, Computer Use as a capability has — for the first time — both a model layer that’s strong enough and a system layer that’s smooth enough.

Why macOS is the better CUA system

1. A battle-tested Accessibility API. The macOS Accessibility framework was designed for screen readers, automation, and assistive input — VoiceOver, Switch Control, and Dictation all rely on it, and the API surface has been stable for twenty years. The cost of getting screen semantics for an agent is far lower than the pixel + OCR scaffolding required on Windows.
2. The AppleScript / Shortcuts heritage. From AppleScript in 1993 to Shortcuts in 2018, a system-level automation layer has always been present. Codex / Claude Code can have an agent invoke Shortcuts the user has already written rather than simulate clicks by hand. Windows has no equivalent.
3. The developer’s machine is the user’s machine. Codex / Claude Code / Cursor / Zed — this generation of agent clients is all Mac first: not “also supports Mac” but “Mac is priority one.” The reason is simple: the engineers writing these tools all use Macs themselves.
4. A clean single-window model. macOS’s window model is relatively unified (one main window per app, a global menu bar) and Mission Control’s hierarchy is visually clear; Windows still mixes MDI / SDI / floating panels / taskbar paradigms, making UI-tree parsing harder for an agent.
5. Local visual encoders are affordable. Running CLIP / SigLIP / GPT-4V-class visual encoders on Apple Silicon is within consumer-hardware budgets. The screenshot → element-recognition step in Computer Use, if billed as cloud vision tokens, would spiral; on a Mac, most of it can run locally.
6. A security model that fits. The macOS sandbox, the Privacy & Security pane, and the TCC (Transparency, Consent, Control) framework — explicitly granting an agent capability over a specific app or directory is OS-native. On Windows that takes additional policy plumbing; on Linux it’s stitched together with capabilities + AppArmor.

Computer Use is the “mouse and keyboard” of the agent era — once it stabilizes, agents no longer need dedicated APIs; any GUI app can be automated. The first system where this is production-ready is macOS, and the strategic implications for Apple have not been fully priced in by the market.

— Codex × macOS, 2026-04-16 update

§04 · Capex contrarian — $14B vs $660B

In 2026, the five hyperscalers’ combined capex lands in the $660–690B range, with ~75% tied directly to AI infrastructure (GPUs, halls, power) — see Futurum Group, CreditSights, and Trefis for the breakdown. Apple’s capex the same year is $14B (about 3.4% of revenue), off by an order and a half.

Why Apple isn’t following

Outsource the depreciation cycle to someone else.

Data-center depreciation runs 5–6 years, so Amazon’s $200B capex implies $33–40B in new annual depreciation hitting the income statement; rolled forward to 2028, the depreciation line alone eats tens of billions of profit.

Apple chose another route: $1B/yr to license Gemini. If foundation models continue to commoditize (price wars, open-source catch-up, falling inference costs), Apple is the buyer, with maximum negotiating leverage; the hyperscalers are the sellers who already poured the cash into concrete and can only spread it across scale.

This matches Apple’s historical posture on hardware: don’t make panels (buy OLED from Samsung / LG), don’t make basebands (buy 5G modems from Qualcomm for years), don’t make NAND (buy from Kioxia / Micron) — only do system integration + the entry point + the brand. LLMs now go in the same drawer.

The counter-intuitive truth of the AI era: not burning capex is actually the safer position. Foundation models are commoditizing fast; whoever pours tens of billions into H100 / Blackwell clusters first is also first to carry the depreciation. Apple trades distribution for access fees — effectively transferring AI-era capacity risk to the shareholders of Google / Microsoft / Amazon.

— 2026 capital cycle, priced in reverse

§05 · Smart glasses — ambient computing

In 2025, Meta Ray-Ban smart glasses shipped 7M+ units for the year, taking 82% of the global smart-glasses market — large enough that Apple made a reverse decision in October 2025: pause Vision Air (the lightweight headset), redeploy engineers onto smart glasses, target 2027 launch, with a possible end-of-2026 first cut. (Bloomberg / Gurman; Ming-Chi Kuo.)

Three time windows on Apple’s glasses roadmap

1. Gen 1 · 2027 launch · no display, multi-frame / multi-material. Form factor benchmarked to Ray-Ban Meta: camera + microphone + health tracking + Apple chip, leaning on the iPhone for compute. Fashion-accessory route, multiple materials and frame styles. Repeatedly confirmed by Bloomberg / Gurman.
2. Gen 2 · 2028 (originally later → pulled forward) · with display, directly benchmarked to Meta Ray-Ban Display. Originally slated post-2030 but pulled forward after Meta released Ray-Ban Display in 2025. Adds a micro-OLED single- or dual-eye HUD that can show notifications, navigation arrows, and Siri / Gemini visual answers.
3. In parallel · Vision Pro 2 · 2026 H2 · M5 Neural Engine + control upgrades. Headset form factor preserved, continuing in the high-ASP professional content / spatial computing / enterprise segment. The M5 refresh strengthens eye-tracking, gesture recognition, and spatial-content creation. Vision Pro doesn’t retreat; smart glasses launch alongside it.

Entry-point logic

Smart glasses = AirPods moment × 10.

When AirPods launched in 2016, they were mocked as “expensive wireless earbuds.” Three years in, they took 60%+ of the category and contributed more than half of Wearables revenue. Smart glasses sit in a similar position — but their entry-point properties are even stronger:

• Visual input: camera + vision model = “what you’re looking at” becomes agent context directly. Siri × Gemini becomes meaningful on this hardware.
• Voice output: bone-conduction / micro-speakers + all-day wear = the complete loop of ambient computing — the iPhone is “pick up,” the glasses are “always on.”
• First-party distribution: not building for Meta / Samsung — Apple’s own hardware + Apple Intelligence + App Store cut.
• Installed-base conversion: with 1.4 billion iPhone users, even 5% buying a pair within 3 years is 70 million units — already 10× Meta’s full-year volume.

This combination of hardware + entry-point distribution is something Google / OpenAI / Anthropic cannot buy with any amount of capex.

§06 · Bottom line — the contrarian position

Stack the four observations together and Apple’s position in the 2026 AI ecosystem can be described as follows.

Five structural advantages

1. The entry point, not the model. App Store + default search + Apple Intelligence routing — control of all three entry-point layers is unchanged. Foundation models become a licenseable component; a $1B/yr access fee is just 1% of Services’ $109B base.
2. Household compute + data hub. The Mac is simultaneously the home’s compute (M3 Ultra at 819 GB/s runs 600B+ on one box; Mac mini M4 Pro runs 70B on a desktop) and its data foundation (Photos / Messages / Mail / Notes / Health all local). The only player with all five — silicon + OS + your data + privacy architecture + on-device framework — is Apple. This isn’t a model story; it’s a structural story.
3. The best system for Computer Use. Codex / Claude Code / Cursor / Zed are all Mac first; Accessibility + Shortcuts + the sandbox model make agents stable and usable on macOS — other systems are still chasing.
4. No capex arms race. $14B vs $660B+ — an order-of-magnitude gap. Outsource depreciation pressure to the shareholders of Google / Microsoft / Amazon; keep buyer-side leverage for yourself. Once foundation models commoditize, this position is the most comfortable in the room.
5. Smart hardware, the second curve. Smart glasses 2027 + Vision Pro 2 in 2026 + AirPods continuing. 1.4 billion iPhones as the install base + Apple Intelligence + first-party distribution — a hardware + entry-point combination nobody else can buy.

Bottom line

Apple’s real edge in the AI era isn’t “we can also build models” — it’s “we don’t need to build models.”

Everyone else is racing on GPUs, on data centers, on frontier models. Apple isn’t. It’s doing something slower but more structural:

• Procuring LLMs the way it procures OLED panels and 5G basebands
• Turning the Mac into the household’s compute + data hub — silicon, OS, daily data, all five pieces in one place
• Turning macOS into the first system where Computer Use is mature
• Turning smart glasses into the next ambient-computing entry point

$1B in, $20B out, $14B capex, 1.4B installs — this combination is one of the few “win without burning cash” positions inside the 2026 AI capital cycle.

The scarcest thing in the AI era isn’t models — it’s an entry-point distribution system that users already open every day, installed on 1.4 billion devices, and that even antitrust action can’t dismantle. Apple happens to own that thing. The rest is just patience and the discipline to stay out of the war.

— Contrarian research · 2026-04-25