The Outsourced Mind: When $3 Trillion Can't Buy Intelligence

In August 1981, IBM — then the most powerful technology company on Earth — signed a deal with a 16-person firm called Microsoft to supply the operating system for its new personal computer. IBM kept the hardware. Microsoft kept the right to license the software to anyone. Within five years, the standard had outlived the standard-setter.

In January 2026, Apple — the most valuable company on Earth at $3.4 trillion — confirmed a deal with Google to supply the intelligence layer for Siri and its entire AI stack. Apple keeps the hardware. Google keeps the model. Apple can distill smaller versions for on-device use, but every distilled model is a derivative of Google's Gemini.

The parallel is obvious. Whether it's accurate is the harder question.

The Dependency Ledger

The Apple-Google relationship is lopsided in a way that's easy to misread. The money flows both directions, but the dependency flows only one way.

Apple has $130 billion in cash. Its total capital expenditure — everything, not just AI — was $12.7 billion in fiscal 2025. Google spent roughly $90 billion on AI alone. The gap isn't closing. It's widening.

And Apple didn't arrive here by choice. It arrived here by failure.

The Failure That Made It Necessary

Apple tried to build its own intelligence layer. The attempt is ongoing, and it is not going well.

The Apple Foundation Models team — roughly 16 people, led by Ruoming Pang, a 15-year Google veteran — built Ajax GPT, a 200+ billion parameter model trained on Google Cloud TPUs. Their internal framework, AXLearn, is optimized for Google's hardware. Even the attempt at independence runs on the supplier's infrastructure.

Apple Intelligence launched with Private Cloud Compute: custom Apple Silicon servers, stateless processing, end-to-end encryption. The architecture was beautiful. The adoption was not.

The M2 Ultra chips in those idle servers aren't powerful enough to run frontier models like Gemini. Notification summaries were disabled after fabricating news stories. ChatGPT on iPhone outperforms Apple's own AI. Apple built the infrastructure. Users want intelligence, just not Apple's intelligence.

This is worse than the IBM parallel. IBM outsourced the OS because it thought the operating system was a commodity — a reasonable bet in 1981. Apple outsourced the intelligence layer because it tried to build it and couldn't. It knows the layer matters. It simply can't produce it fast enough.

It's Not Just Apple

The dependency pattern is everywhere. Apple is the most visible case because of its market cap and the IBM echo, but the same structural dynamic plays out across the industry.

Four of the five most valuable technology companies on Earth rent their intelligence from someone else. The fifth — Google — is the landlord. That concentration should concern anyone who remembers what happened when the world's computing infrastructure depended on a single operating system.

And 42% of companies abandoned most AI initiatives in 2025 (S&P Global). Building your own is failing at scale. The rent-or-build question isn't theoretical — it's being answered empirically, and the answer is: rent.

The 1981 Parallel

IBM's mistake is well-rehearsed, but the timeline matters more than the narrative. The point of no return arrived faster than anyone expected.

Eighteen months. That's the window between IBM's deal and the moment the industry recognized the standard had escaped its creator. Paul Allen later said: "No one, including us, foresaw that the IBM deal would ultimately make Microsoft the largest tech company."

Apple's deal was announced in January 2026. Eighteen months from now is July 2027.

Where the Analogy Breaks

The IBM parallel is seductive but incomplete. Apple's situation differs in ways that matter — and in ways that might not.

The strongest counter-argument is Apple's exit plan. The Baltra chip enters mass production in the second half of 2026. Data center construction starts in 2027. "Ferret-3," Apple's next-gen proprietary model, is reportedly planned for 2026-27. WWDC in June is expected to showcase "Campos" — an internal chatbot intended to compete with ChatGPT and Gemini — and a new Core AI framework replacing Core ML.

But here's the problem with the exit plan: Apple's own model, even its "proprietary" Apple Foundation Model v11, is described as running locally on Apple Silicon while simultaneously being "based on Gemini" according to multiple reports. The ambiguity of what counts as "Apple's own" is itself the story. When the teacher is Gemini and the student is Apple's on-device model, the boundary between owned and rented intelligence dissolves.

The Question That Actually Matters

The IBM parallel is a frame, not an answer. The real question is structural:

Can the intelligence layer be rented, or must it be owned?

IBM proved the operating system layer must be owned — the OS was a platform that defined what software could exist. But is the AI model layer a platform in the same sense? Or is it more like electricity — a utility that powers the house but doesn't define its architecture?

If models are a utility: Apple's strategy is brilliant. Rent the cheapest frontier model, differentiate on privacy, hardware, and UX. The intelligence layer becomes a commodity, and the company with 2.5 billion devices and a closed ecosystem wins through distribution.

If models are a platform: Apple is IBM. The intelligence layer will define what applications are possible, and whoever controls it controls the ecosystem. Distillation buys time but not independence.

RAND's analysis of AI market concentration identifies four conditions for natural monopoly — homogeneity, economies of scale, network effects, economies of scope — and calls the case "relatively strong." GPT-4 training cost ~$78M. Gemini Ultra: ~$191M. Those are sunk costs that can't be recovered, and they're growing. Only about five entities on Earth can train frontier models.

But Goldstein and Salib argue the opposite: the reasoning revolution (reinforcement learning replacing data-driven scaling), fast-following dynamics (training a model "just as good, six months later" is cheap), and DeepSeek's $5.6M training run all suggest the barrier is temporary. They may be right — at the reasoning layer. At the scale layer, where trillion-parameter models require $700 billion in collective capex, the barrier only grows.

The Layer That Completes the Stack

This is the Concentration Trap operating at a new altitude. In Post #16, I mapped how concentration closes at the processing layer across eight domains — minerals, batteries, chips, pharmaceuticals. The trap isn't at the raw material layer. It's at the refining layer, where the value is extracted.

The model layer is the refining layer for intelligence.

At every layer, the same pattern: raw materials are globally distributed, but the processing step concentrates in a handful of actors. Lithium is everywhere; battery-grade lithium hydroxide comes from China. Silicon is sand; advanced chips come from TSMC. Training data is abundant; frontier models come from five labs.

Apple sits at Layer 5. It tried to build at Layer 4. It's renting while it builds. The question is whether the build arrives before the dependency becomes structural — before the distilled models, the developer tools, the user expectations, and the integration depth make Gemini as irremovable as Windows was from the PC.

Eighteen months passed between IBM's deal and the moment the standard escaped. Apple's clock started in January 2026. Its exit — Baltra, Ferret-3, full deployment — targets 2027-2028.

Whether that's fast enough depends on which kind of layer intelligence turns out to be: utility or platform. The answer isn't knowable yet. But the bet is already placed.

Sources: 9to5Mac, WinBuzzer, Fortune, AppleInsider, SamMobile, CNBC, NextPlatform, RAND, Goldstein & Salib, 9to5Mac (servers), CIO Dive / S&P Global, TechSpot, AppleInsider (Kuo), AppleInsider (WWDC), Wikipedia (IBM PC), BSC Capital