Apple spends $14B on AI while competitors spend $650B. Is it losing or playing a smarter game? The data tells a complicated story.
Get new posts delivered to your inbox. No spam, unsubscribe anytime.
Apple made $900 million in App Store fees from other companies' generative AI apps in 2025. ChatGPT alone accounted for roughly 75% of that. Let that sink in: the company most accused of "losing the AI race" is collecting a 30% tax on the winners. Meanwhile, Apple plans $14 billion in capital expenditure for 2026 while Amazon, Microsoft, Meta, and Alphabet are collectively spending $650 billion on AI infrastructure. That's a 46:1 spending ratio. Either Apple is the dumbest company in the room, or it's playing a completely different game.
The answer is more complicated than either the doomsayers or the Apple apologists want to admit.
Let me lay out the uncomfortable numbers first.
| Metric | Apple | Google/Alphabet | Microsoft | Meta |
|---|---|---|---|---|
| 2026 AI/capex spending | $14B | $92B+ | $94B+ | $65B+ |
| Own frontier LLM | 3B parameter on-device | Gemini Ultra (frontier) | GPT-5 via OpenAI | Llama 4 (open source) |
| AI chatbot market share | 0% (integrates ChatGPT) | 21.5% (Gemini) | ~64% (ChatGPT) | N/A |
| Stock gain (past year) | 11.95% | 71.06% | ~30% | ~40% |
| Senior AI researchers lost (2025) | 10+ | Gaining talent | Gaining talent | Gaining talent |
By almost every conventional metric, Apple is losing. No frontier model. No chatbot product. A fraction of the spending. Bleeding top researchers. And the stock reflecting it -- Alphabet gained 6x more than Apple over the past year.
But conventional metrics might be measuring the wrong thing.
Let me steelman the bears first. Because they have real evidence.
In testing with 25 common voice requests, Siri could only accomplish 13 tasks consistently and returned incorrect or useless results in at least 7 cases. The "Apple Intelligence"-enhanced Siri was supposed to ship in 2025. It didn't. The overhaul got pushed to spring 2026 because it only worked properly about two-thirds of the time during internal testing.
Two-thirds accuracy for the company whose brand promise is "it just works." That's bad.
Users noticed. Class-action lawsuits were filed alleging Apple "deceived millions of consumers into purchasing new phones based on features that do not exist." The 2025 iPhone cycle -- the first sold explicitly on AI features -- delivered the features late, buggy, or not at all.
This might be the most alarming indicator. Apple lost 10+ senior AI/ML team members in the past year. The head of foundation models, Ruoming Pang, was poached by Meta for reportedly $200 million+. The head of Siri's "Answers, Knowledge, and Information" team left for Google DeepMind. Multiple researchers followed.
When your top people leave, they're not leaving because of the cafeteria food. Reports suggest falling morale on the foundation models team, linked to Apple's reliance on third-party AI and the perception that Apple isn't serious about building its own frontier models. Apple is only "marginally increasing" pay for AI teams -- not matching Meta's offers.
Here's the fact that should worry Apple fans most: to upgrade Siri, Apple had to partner with Google. The new Siri will be powered by Google's Gemini under a deal costing Apple roughly $1 billion per year.
Apple already integrates ChatGPT as an option within Apple Intelligence. Now it's adding Gemini for Siri's core functionality. That means Apple's flagship AI features -- the ones they'll market on billboards and keynotes -- run on other companies' models. Fortune called this a "worrisome sign" that Apple cannot build its own competitive LLM.
The company that built its own chips (M-series), its own operating systems, its own search engine foundations, and its own payment system... can't build its own AI model? Or won't? The distinction matters.
Now the bull case. And it's not as crazy as the scoreboard suggests.
Apple's core bet is that AI models will become interchangeable commodities, not proprietary moats. If that's true, spending $92 billion building your own frontier model is like spending $92 billion building your own microprocessor in 1999 -- impressive, but unnecessary once you can buy equivalent capability off the shelf.
Look at what's happened since 2024. OpenAI released GPT-4. Google matched it with Gemini. Meta released Llama and made it open-source. Anthropic released Claude. The performance gap between frontier models has been shrinking every quarter. If models commoditize, the value shifts to distribution and integration. And nobody integrates AI into consumer products better than Apple.
This isn't a new playbook for Apple. They didn't build their own cellular modem for years -- they used Qualcomm's. They didn't build their own GPU -- they used PowerVR's. In each case, they eventually built their own when the technology matured enough that they could catch up. The pattern: rent first, understand the technology, then build when it makes strategic sense.
Apple's 3-billion-parameter on-device LLM runs entirely on Apple Silicon. Your data never leaves your device. For heavier tasks, Private Cloud Compute processes requests on Apple's servers without storing or sharing the data, with cryptographic verification through Secure Boot.
No other company offers this. Google processes your data on its servers. OpenAI processes your data on its servers. When a user asks Gemini to summarize their emails, Google sees those emails. When an Apple user asks the on-device model to summarize their emails, nobody sees them.
In a world where AI is processing our most intimate data -- messages, photos, health records, financial transactions -- privacy isn't a nice-to-have. It's a product feature that a significant segment of users will pay for. Analysts estimate Apple's privacy-first AI approach could add $10-15 billion in annual revenue by 2027 through hardware upgrade cycles driven by AI features.
And here's the angle almost nobody discusses: regulatory tailwinds. The EU AI Act, GDPR enforcement, and increasingly strict data protection laws worldwide all favor on-device processing over cloud-based AI. Apple's privacy approach isn't just a consumer feature -- it's regulatory compliance built into the architecture. As AI regulation tightens (and it will), Apple's approach becomes more valuable, not less.
Apple doesn't need to win the AI model race. It needs to own the platform where AI models run. And it does.
$900 million from generative AI apps in 2025, on track for over $1 billion in 2026. ChatGPT, Claude, Gemini, Midjourney, Perplexity -- every AI app on iPhone gives Apple 15-30% of subscription revenue. Apple makes money whether OpenAI or Google wins the model war. That's not losing. That's being the casino.
Consider the economics. OpenAI reportedly burns billions per year on compute. Google spent $92 billion on AI infrastructure. Meta spent $65 billion. And a chunk of their consumer revenue flows through Apple's App Store regardless. Apple earns AI revenue with essentially zero AI infrastructure cost. The margin on platform taxes is approximately 100%.
Apple has $130+ billion in cash reserves. If the AI market shifts in a direction that requires Apple to build its own frontier model, it can do so. If an AI startup develops a breakthrough, Apple can buy it. (They already acquired Q.ai for $2 billion and are in late talks for Prompt AI.)
The company announced a $500 billion U.S. investment over four years, creating ~20,000 jobs in R&D, software, silicon engineering, and AI/ML. And they hired Amar Subramanya, a 16-year Google veteran, as VP of AI in December 2025.
Apple isn't out of money. It's out of urgency. Whether that's strategic patience or dangerous complacency depends on what happens next.
The smartphone AI race tells the story in miniature.
| Feature | iPhone 16 (Apple Intelligence) | Samsung Galaxy S25 (Galaxy AI) | Google Pixel 10 Pro (Gemini) |
|---|---|---|---|
| Photo AI editing | Writing tools, basic enhancement | Object removal, shadow/reflection removal, generative expansion | Magic Editor: object removal, backdrop changes, noise reduction |
| AI assistant | Siri (limited, overhaul pending) | Gemini (full integration) | Gemini (native, best-in-class) |
| On-device AI | 3B parameter model, privacy-first | Hybrid (on-device + cloud) | Hybrid (on-device + cloud) |
| Cross-app AI actions | Limited | Real-time summaries, cross-app actions | Deep Google ecosystem integration |
| AI chip | A18 Pro / M-series | Snapdragon 8 Elite | Tensor G5 (purpose-built for AI) |
Google's Pixel 10 Pro XL is rated the best AI smartphone in 2026. Samsung is close behind. Apple is third and not close. In AI photo editing comparisons, Tom's Guide found a clear winner -- and it wasn't Apple.
But here's the nuance: Apple ships fewer AI features, but the ones it ships work within its privacy framework. Google's Magic Editor sends your photos to Google's servers. Apple's on-device processing keeps them local. For users who care about that -- and Apple's user base tends to -- that's a trade-off, not just a deficit.
Everything converges this year. Here's what's on the line.
Spring 2026: The Siri Overhaul. Project "Campos" transforms Siri into Apple's first proper AI chatbot -- longer conversations, complex queries, multi-command processing, on-screen awareness, contextual understanding across apps. Powered by Gemini. Tim Cook has confirmed it's on track. If it delivers, it changes the narrative overnight. If it's buggy again, the "Apple lost AI" story becomes permanent.
WWDC 2026: Core AI and the AI App Store. Apple is replacing Core ML with "Core AI" -- a new developer framework for iOS 27 that formalizes generative AI integration. They're also expected to announce "Siri Extensions" and an AI App Store, letting third-party apps plug directly into Siri. If Apple becomes the platform where AI apps plug into the user's OS-level assistant, that's a very different competitive position than "Apple doesn't have AI."
This is the move that could redefine the whole conversation. An AI App Store would make Apple the distribution layer for AI agents and agentic systems -- not the company building the models, but the company deciding which models get to 2 billion users.
The investor inflection. Apple's stock was the second-worst performer among the Magnificent Seven through mid-2025. Then it surged 35% in the second half as the market grew weary of massive AI spending by competitors. Wedbush Securities called the sell-offs "unwarranted". The Motley Fool framed Apple as an "AI bubble hedge" -- the stock you buy if you think the AI spending party ends badly.
Wrong #1: "Apple has no AI." Apple has a 3B parameter on-device model, a server-side model using a novel Parallel-Track Mixture-of-Experts architecture, and active research publications at NeurIPS, ICML, and ICLR. Their paper "The Illusion of Thinking" on reasoning model limitations was widely cited. Apple has AI. It just doesn't have a frontier AI model.
Wrong #2: "Apple needs to spend $100B+ to compete." Apple's strategy is that you don't need to own the model to own the experience. If Gemini, Claude, and GPT become interchangeable commodities, the value is in integration, distribution, and user trust. Apple excels at all three.
Wrong #3: "Apple is too late." The AI market is two years old. Two years. Declaring a permanent winner in a market this young is like declaring the search engine war over in 1997 because AltaVista existed. Google Search launched in 1998 and took until 2004 to dominate. Apple's history is full of entering markets "late" and winning: iPod (not the first MP3 player), iPhone (not the first smartphone), App Store (not the first app marketplace).
Wrong #4: "The talent loss is fatal." It's serious, not fatal. Apple hired a VP of AI from Google, announced 20,000 new tech jobs, and is acquiring AI startups. Talent flows both ways. Google lost plenty of people to OpenAI and Anthropic. Meta lost people to Google. The AI talent market is musical chairs, not a one-way exodus.
I've laid out both sides. Now let me identify the genuine risks that aren't just media narrative.
Risk 1: The Gemini dependency becomes structural. If Siri's AI backbone is Google's, and Google decides to prioritize Pixel's Gemini integration over Apple's, Apple is in trouble. Dependencies create leverage. Apple has historically hated being dependent on anyone -- that's why they built their own chips. Being dependent on Google for AI is a strategic contradiction.
Risk 2: Developer loyalty erodes. If Google and Samsung offer better on-device AI capabilities, developers will prioritize those platforms for AI-native apps. The Foundation Models framework with free inference is smart, but it needs to be competitive in capability, not just privacy.
Risk 3: The commodity thesis is wrong. If AI models don't commoditize -- if there's a persistent performance gap between the frontier and everything else -- then Apple's strategy of renting instead of building collapses. So far, the commoditization trend holds. But one breakthrough in reasoning, multimodality, or AGI could change that overnight.
Risk 4: Users stop waiting. Apple's brand survives on delivering eventually. But "eventually" has a shelf life. If the Siri overhaul disappoints in spring 2026, that's two years of promises and zero years of delivery. Users who switched to Pixel or Samsung for AI features may not come back.
Risk 5: The "good enough" bar keeps rising. What counts as "good enough" AI on a phone in 2024 is very different from 2026. Google and Samsung are shipping AI features monthly. Apple ships annually at WWDC. That cadence mismatch means Apple is always comparing this year's promise against the competition's current reality. Every month that Siri stays broken while Gemini gets better, the "good enough" target moves further away.
Apple isn't losing the AI game. Apple is playing a different AI game. And whether that game is brilliant or suicidal depends entirely on one question: do AI models commoditize?
If models commoditize -- and the evidence so far supports this -- then Apple's strategy is arguably the smartest in the industry. Let Google, Microsoft, and Meta spend $650 billion building models. Then integrate those models into the best consumer hardware ecosystem on the planet, with privacy guarantees nobody can match, and charge a 30% platform tax on every AI app sold. Apple spent decades building the distribution channel. AI apps are just the latest thing flowing through it.
If models don't commoditize -- if frontier capability keeps mattering, if the gap between GPT-5 and everything else stays wide -- then Apple is in serious trouble. You can't integrate what you can't access, and if one company's model is dramatically better than the rest, that company controls the experience. Apple becomes a dumb terminal for someone else's intelligence.
I think the truth is somewhere in between, and it favors Apple more than the current narrative suggests. Here's why.
The Siri overhaul doesn't need to be the best AI assistant. It needs to be good enough that iPhone users don't switch to Pixel. "Good enough" is Apple's historical sweet spot -- they rarely ship the first or best version of a technology. They ship the version that works seamlessly within their ecosystem. The iPod wasn't the best MP3 player. The iPhone wasn't the best smartphone in 2007. They were the most integrated.
The brain drain is concerning but not fatal. The $900 million in AI app revenue is being underappreciated. The $130 billion in cash provides optionality that no other company has to this degree. And the privacy angle is a genuine differentiator in a world where every AI interaction involves sending your most personal data to a server.
My actual prediction: Apple will be fine. Not dominant in AI research. Not the company building frontier models. But the company that makes AI useful, private, and integrated for 2 billion device users. That's not "losing." That's a different kind of winning -- the kind Apple has always been best at.
But if the Siri overhaul ships broken in spring 2026? Then we're having a very different conversation. And those AI engineers Apple is trying to hire will have their pick of companies to work for instead.