The numbers tell a story nobody at Apple's PR team would dare write down. Amazon is planning roughly $200 billion in capital expenditures for 2026. Alphabet sits between $175 and $185 billion. Meta is somewhere around $115 to $135 billion. Microsoft is pushing $145 billion. Together, four companies are about to spend close to $700 billion in a single year chasing AI dominance.
Apple's number? Around $14 billion.
That is less than ten percent of what Alphabet alone plans to spend. It is the kind of gap that gets a CEO fired at most public companies. At Apple, it has coincided with returning $106 billion to shareholders in a single fiscal year and shrinking the share count by almost a third over the last decade. The stock has wobbled. Analysts have grumbled. Executives have retired or walked across the street to Meta. And yet I keep coming back to the same uncomfortable conclusion for anyone cheering on the AI capex boom.
Apple might already be winning.
The shape of a commodity
There is a bet buried inside Apple's non-spending, and it is worth stating plainly. The bet is that foundation models will become commodities. Not in five years. Not eventually. Soon enough that it would be foolish to sink $100 billion a year into building your own.
Look at what has happened in the last eighteen months. GPT-class models from four different labs now cluster within a narrow band on most benchmarks. Open weights from Meta and Mistral have caught up to closed frontier models with a lag measured in months, not generations. Chinese labs have shipped competitive models at a fraction of Western training costs. The "moat" that OpenAI had in 2023 has been colonized by Google, Anthropic, xAI, DeepSeek, and a dozen others.
If models are becoming interchangeable, then the economic winner is not the company with the biggest training run. It is the company that sits closest to the user. The company whose logo is on the glass when you ask the question.
That is Apple.
The Gemini deal is not what it looks like
On January 12, 2026, Apple and Google announced a multi-year partnership. Google's Gemini would power the rebuilt Siri and the next wave of Apple Intelligence features. A custom 1.2 trillion parameter model, eight times larger than Apple's internal cloud models, built to Apple's specifications. Reported price: around $1 billion a year.
The market reaction was a mix of relief and pity. Wedbush called it a stepping stone. Ming Chi Kuo called it short-term pressure relief rather than strategy. Fortune framed it as Apple conceding it could not build a competitive frontier model. Every take assumed Apple had lost something by writing a check to Google.
Consider the other side of that check. Google is spending on the order of $85 billion this year on AI infrastructure. Apple is paying roughly 1.2 percent of that to rent the output. Google does the training. Google owns the data centers. Google depreciates the GPUs that become obsolete within eighteen months. Apple gets a state-of-the-art model delivered to private cloud compute with the exact privacy properties it demanded, plus full access to distill smaller student models for on-device use.
Who is the customer in this relationship? The company writing the billion-dollar check, or the company that just locked itself into supplying one of the most demanding buyers in software?
I think Apple knows exactly what it is doing. It has done this before. Apple does not manufacture its own displays at scale. It does not run its own foundries. It designs chips and pays TSMC to build them. It demands terms that would bankrupt smaller customers, and it walks away when suppliers get too comfortable. The Gemini deal fits this pattern perfectly. When Gemini 4 arrives, Apple can renegotiate. When Anthropic or a Chinese lab ships something better, Apple can swap. The Siri user will never know the model changed underneath.
Google, meanwhile, has $85 billion in annual infrastructure it needs to service. It cannot walk away.
The App Store quietly collects either way
Here is the piece of the puzzle that analysts keep missing. Whichever AI app wins on iPhone, Apple takes its cut. If it is ChatGPT, Apple takes a cut. If it is Gemini, Apple takes a cut. If it is some startup nobody has heard of yet, Apple takes a cut. The company spent two decades building the toll booth. It does not have to guess which carriage wins the race.
The Epic ruling and various regulatory actions have chipped at the thirty percent, sure. Apple is not going to collect what it used to on pure commissions. But the ecosystem effects run far deeper than a percentage point on an in-app purchase. Every AI product built for consumers still has to be demoed on someone's phone. The default phone in the United States and much of Europe is still the iPhone. The defaults Apple ships, the entitlements Apple grants, the permissions Apple gates, shape which AI products get usage and which do not.
If you are running a Series C AI company and your product does not work well on iPhone, your product does not work. That is distribution power. It does not show up on a capex line.
The real risks are real
I am not going to pretend this position is bulletproof. There are two failure modes that genuinely worry me, and hand waving them away would be dishonest.
The first is execution on Siri. Apple has promised a rebuilt Siri for years, and the 2025 version was delayed, then delayed again. Even with Gemini doing the heavy lifting, Apple still has to build the orchestration layer, the guardrails, the on-device distillation pipeline, and the integration with messages and calendar and mail. If that ships and feels like a worse ChatGPT, the ecosystem advantage erodes fast. iOS 27 in September 2026 is the real test, not the announcement in January.
The second is the shape of the next device. Jony Ive, the man who designed the iPhone, is now at OpenAI running design for a device explicitly meant to reduce screen dependency. The project has slipped to 2027, and the form factor keeps changing. A screenless companion, then a pen, and now reportedly a smart speaker priced between $200 and $300 with a camera. Most of these bets will fail. Humane's AI Pin is already in the graveyard. But Ive is not humane, and the Apple alumni queuing up to join him are not amateurs.
If a post-phone device ever lands, Apple's distribution moat drains overnight. That is the scenario that should keep Tim Cook awake.
Here is why I am still not panicked on Apple's behalf. The device that replaces the iPhone, if it exists, will need processing power. It will need chips designed for on-device AI inference at low power. It will need a local compute substrate that does not round trip to a data center for every query. Apple already makes the best mobile silicon in the world. The M series and A series chips are, by most neutral measurements, two to three years ahead of Qualcomm. Developers running local models today buy Macs, not Windows boxes, because nothing else runs a 70B parameter model at reasonable speed on a laptop. That is not marketing. That is what the inference benchmarks show.
Whoever builds the next dominant AI device will either buy silicon from Apple, license Apple's IP, or spend a decade trying to match it. The $500 billion that Amazon and Google are collectively pouring into GPU farms does not buy them a single node of on-device efficiency advantage.
The privacy card nobody else can play
There is one more piece worth naming, and it is the one that has aged best over the last two years. Apple bet on privacy as a brand position in 2013, and the bet looked a little corny for about a decade. Then OpenAI trained GPT on basically the entire internet. Then Meta started running ads against your AI conversations. Then every news cycle carried a fresh story about chat logs, training data, leaked prompts, or a data broker.
Ask a random iPhone user in Seoul or Munich or São Paulo whether they want Siri to send their messages to Google or keep them on device, and the answer is not subtle. Private Cloud Compute, for all the engineering theater around it, is a genuinely differentiated product. Apple can ship AI features that its competitors literally cannot ship because those competitors' business models depend on harvesting the data Apple refuses to touch.
This is the rare case where a decade of apparent stubbornness turned into a product feature overnight.
I keep thinking about the railroad boom in the 1880s. The companies that built the tracks mostly went bankrupt. The companies that sold tickets, ran the hotels at the stations, and owned the land next to the depots did fine. The capex winners lost. The distribution winners won. Tim Cook does not need to out-spend Sundar Pichai or Satya Nadella. He needs to own the last mile, keep the margins intact, and wait for the model layer to commoditise.
A billion dollars a year to Google buys him all the time in the world.
And if the rebuilt Siri ships in September and actually works, I suspect the Wall Street note titled "Apple's Lazy AI Strategy" from February 2026 is going to read very differently by Christmas.