The Real Reason Apple Pays Google $1 Billion a Year · Issue #124

Opening

Hello, subscriber. This is OZ Talking. “Disaster.” In December 2025, that’s the word Wedbush analyst Dan Ives used to describe Apple’s AI strategy. Needham’s Laura Martin said Apple was “1–2 years behind its competitors,” and Siri was still fodder for memes. Then, on June 8, 2026, on the WWDC stage, Apple unveiled 5 of its own foundation models¹ all at once. What matters here isn’t model performance. The real news is how Apple built them and the structure behind them. Personally, I think this is the single most important thing to watch from WWDC 2026, and it seems to be getting buried — so I wanted to bring it up.

Let me give you the conclusion up front: Apple didn’t lose the AI arms race — it was playing an entirely different game.

🏗️ The Company That Sat Out a $600 Billion Arms Race

Right now, Big Tech’s AI infrastructure spending is unprecedented in scale. Looking at annual CapEx² for major companies as of 2026:

Amazon: about $200 billion
Microsoft: about $190 billion
Alphabet (Google): about $180 billion
Meta: about $125 billion

These four companies alone add up to more than $600 billion a year — about ₩830 trillion (~$600B), roughly 1.2 times Korea’s entire annual government budget. Where does this money go? Buying GPUs, building data centers, training bigger models. Microsoft even disclosed in a recent earnings call that about $25 billion of its annual CapEx was attributable to “more expensive memory components.” This is the largest simultaneous infrastructure buildout in history.

But what about Apple? Its total 2025 CapEx was $12.7 billion. The 2026 forecast is about $14 billion — roughly 1/40th of what the four Big Tech companies spend combined. Apple’s annual investment is less than what Google spends in a single quarter.

Wall Street’s response was brutal. Wedbush’s Dan Ives called Apple’s AI strategy a “disaster,” going as far as to say “no one on Wall Street believes Apple is producing meaningful AI results compared to Microsoft, Google, Meta, or OpenAI.” According to Quartz, some analysis suggested that Apple itself internally assessed it was 2 years behind OpenAI and Google.

But once you dig into the WWDC 2026 announcement, the picture changes. It’s not that this company didn’t spend money — it spent it in a completely different way. It headed toward the same destination via an entirely different route.

💰 A ₩1 Trillion Tuition, and What It Actually Buys

In January 2026, Apple and Google officially announced a multi-year AI partnership. According to Bloomberg’s Mark Gurman, the deal is worth about $1 billion a year (roughly ₩1.4 trillion). Apple secured access to Google’s Gemini, a frontier model with 1.2 trillion parameters. Here’s where a lot of people got it wrong. The reaction was, “So Apple just handed its AI over to Google after all?” The actual structure is completely different.

The key technique is distillation³ — training a small, efficient “student model” (AFM) by referencing the outputs of a massive “teacher model” (Gemini). Right after WWDC, Apple’s Senior Vice President Craig Federighi put it this way:

“The amount of Google Assistant we use is zero.”

Amar Subramanya, Apple’s vice president for AI, put it more precisely: “Every model is a custom build for Apple Silicon, refined using outputs from the Gemini frontier model.”

Here’s the structure in summary:

Teacher: Google Gemini (1.2 trillion parameters)
Student: Apple’s AFM 3 series (5 models)
Teaching method: Distillation ≈ using Gemini’s outputs as the training signal
The exam (runtime): AFM runs inference alone, without Gemini

Apple takes the lessons but sits the exam alone. To put it another way: it pays about ₩1 trillion a year for tutoring from the best teacher, but the diploma still bears its own name.

This is becoming a pattern across the AI industry in 2026. Only a handful of companies build frontier models from scratch; everyone else distills the output and reworks it for their own products. Apple is simply the largest and most explicit example of this pattern. Below is what’s essentially become an industry standard for distillation, as referenced in the recent Elon Musk–OpenAI lawsuit.

Musk in Court: “Everyone Does It”The same U.S. Big Tech firms that criticized DeepSeek were doing the same thing at home all along.

oztalking.com

There’s one more angle worth noting: what this deal means for Google. According to The Information, Apple secured full access to the Gemini model within its own data centers — not merely calling an API, but pulling the model weights themselves to run its distillation pipeline. For Google, that translates into $1 billion a year in stable revenue, plus AFM 3 Cloud Pro running on top of its own cloud infrastructure (Google Cloud + NVIDIA GPUs). It’s a rational trade for both sides.

Fortune raised an interesting question about this structure: “Will AI models become interchangeable commodities, or a source of durable competitive advantage?” Apple is betting on the former. If models become commodities, there’s no need to spend hundreds of billions of dollars building one from scratch — you can simply take the best one and distill it.

📱 The Real Weapon Isn’t the Model — It’s 2 Billion Pockets

So what did Apple focus its savings on instead? On-device AI.

AFM 3 Core Advanced is the technical highlight of this announcement. It’s a 20-billion-parameter model, and Apple runs it on an iPhone with just 12GB of RAM. Normally, that’s impossible — a typical LLM needs to load its entire set of weights into DRAM. What this model enables — “emotionally expressive, natural text-to-speech, more accurate dictation, multimodal processing that understands images” — features like Siri all run on-device, with no cloud involved. That’s a genuinely huge technical achievement.

When Apple first announced Apple Intelligence in 2024, its on-device model had about 3 billion parameters. In 2 years, that scale grew nearly 7-fold — made possible by a technique Apple’s researchers developed called IFP (Instruction-Following Pruning)⁴. Here’s how it works:

All 20 billion parameters are stored in NAND flash⁵ — the same memory where your photos and apps live.
When a user asks a question, a lightweight prediction block selects only the “experts” needed for that specific request and loads them into DRAM.
Only 1–4 billion parameters are actually activated per prompt; the remaining 16 billion stay dormant in flash.

A typical MoE⁶ model has to swap experts token by token, which turns memory bandwidth into a bottleneck. IFP instead makes its routing decision once per prompt. It’s an architectural workaround for the real-world constraint of slow NAND-to-DRAM bandwidth.

Here’s why that matters. Whether it’s ChatGPT, Gemini, or Claude, cloud-based AI ultimately has to send a request to a server. That requires a network connection, and data leaves the device. Apple’s on-device model skips that entire step. Privacy is guaranteed structurally, by design. Heavier tasks get offloaded to Apple’s own servers, called Private Cloud Compute, where user data still isn’t stored and no one — not even Apple — can access it.

And here’s a number worth sitting with. Apple has more than 2 billion active devices. ChatGPT has 400 million weekly active users (as of December 2025); Gemini has about 350 million monthly active users. Apple isn’t building a single new AI app — it’s shipping AI as an OS update to 2 billion devices already sitting in people’s pockets.

On pure benchmark performance, AFM 3 is unlikely to beat GPT-4o or Gemini 2.5 Pro. But the generational leap in Apple’s own benchmarks is clear. In Apple’s human evaluations, AFM 3 Cloud was preferred over its predecessor 64.7% to 8.7% on text tasks, and 37.8% to 9.6% on image understanding. And for the first time this year, the Foundation Models framework supports image input, letting third-party developers integrate on-device multimodal AI directly into their Swift apps.

What this means is that Apple isn’t chasing the top benchmark spot. It’s trying to deliver a “good enough” model to “the largest possible number of people,” and to build an app ecosystem on top of that model. If ChatGPT or Gemini is a single app, Apple’s strategy is to embed AI into the OS itself — the layer those apps run on.

Oz’s Lens

To be honest, when I watched this announcement, what struck me first wasn’t the technology — it was the distribution strategy.

There’s a pattern I’ve repeatedly seen while building technology management strategy. In tech markets, the “best product” wins far less often than you’d think. What wins more often is “good enough product + overwhelming distribution.” That’s the same structure behind VHS beating Betamax, and Android capturing 70% of the smartphone market.

I think what’s happening in the AI market right now is at a similar inflection point. The performance gap between frontier models keeps narrowing. GPT-4o, Gemini 2.5, and Claude 4 keep swapping rankings by 1–2 benchmark points — a race where everyone converges near the ceiling. The less users can feel the difference, the more accessibility and integration — rather than raw performance — become the deciding factor. That’s a signal that models are entering the commoditization stage.

If this direction holds — if who deploys the model, where, and how matters more than the model itself — then Apple’s $14 billion strategy isn’t wasteful, it’s foresight. Of course, that’s a big “if.”

There’s a clear risk, too. At launch, AFM 3 won’t be available on iPhones and iPads in the EU or mainland China. Given that a significant share of Apple’s 2 billion active devices sit in those two markets, there’s a hole in the premise of a “maximum distribution” strategy. And whether the “good enough” model is truly good enough won’t be verifiable until real-world use begins this fall.

Closing

To sum up: Apple didn’t lose the AI arms race — it chose a different game entirely: repackaging and distributing the output of the arms race. It spends roughly ₩1 trillion a year distilling Google’s frontier model, and poured the savings into on-device architecture.

For this strategy to work, AI models need to keep commoditizing. If a single model pulls dramatically ahead of the pack, the strategy collapses — but based on the trend so far, that possibility keeps shrinking. The real question in AI competition may not be “who builds the smartest model,” but “who reaches 2 billion pockets first.”

💬 Which do you think matters more in AI — “model performance” or “scale of distribution”? Let me know in the comments.

References & Further Reading

Primary sources

Apple Machine Learning Research, “Introducing the Third Generation of Apple’s Foundation Models”, June 8, 2026. — Apple’s official announcement, and the technical basis for today’s issue.
Bairu Hou et al., “Instruction-Following Pruning for Large Language Models”, Apple AI/ML & UC Santa Barbara, ICML 2025. — The original paper on IFP, showing a 3B-active model approaching the performance of a 9B dense model.
ofox.ai, “Apple’s Third-Generation Foundation Models: A Developer’s Read on WWDC 2026”, June 8, 2026. — The most accurate developer-side breakdown of the Gemini distillation structure.

Background

Fortune, “While big tech burns cash on AI, Apple waits”, February 17, 2026. — A column analyzing Apple’s restrained CapEx strategy through the lens of the “model commoditization” hypothesis.
CNBC, “Apple is facing pressure from Wall Street to figure out its AI strategy”, July 30, 2025. — Shows how Wall Street viewed Apple’s AI ambitions before WWDC.
VentureBeat, “On-device AI agents hit a hard memory limit. Apple’s new architecture routes around it.”, June 9, 2026. — A clear explanation of the technical significance of IFP and NAND-based on-device inference.

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management and business analytics — while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master’s from Korea University’s Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

Foundation Model: An AI model pre-trained on large-scale data that can be applied generally across a wide range of tasks. GPT, Gemini, and Claude are prime examples. ↩
CapEx (Capital Expenditure): The money companies invest in facilities, infrastructure, and equipment. In the AI era, most of this goes toward buying GPUs and building data centers. ↩
Distillation: A training technique that transfers knowledge from a large “teacher model” to a smaller “student model.” By learning to mimic the teacher’s outputs, the student model achieves strong performance despite its smaller size. ↩
IFP (Instruction-Following Pruning): A dynamic pruning technique developed by Apple. It analyzes a user’s input (prompt) and activates only the parameters needed for that specific task. Unlike traditional pruning, which permanently shrinks a model, IFP can activate different parts of the model for each request. ↩
NAND Flash: Non-volatile memory used to store data in smartphones and SSDs. It’s slower than DRAM but has much higher capacity and retains data even when the power is off. Apple repurposed this storage space as a “warehouse” for its AI model. ↩
MoE (Mixture of Experts): An architecture that houses multiple “expert” networks inside a model and activates only a subset based on the input. It reduces actual computation relative to total parameter count, making it useful for running large models efficiently. ↩