Apple Hit $5 Trillion Without Big Tech's AI Capex Bill

On July 27, Apple closed the trading day as the world's most valuable public company, pushing past Nvidia with a market cap around $4.94 trillion. The next day it briefly touched $5 trillion intraday, making it only the second company ever to cross that line, nine months after Nvidia got there first in October 2025. Apple did it on the back of iPhone sales, not a foundation model. Its total capital expenditure for fiscal 2025 was $12.72 billion.

Compare that to the four companies actually building AI infrastructure. Amazon, Microsoft, Alphabet, and Meta are now guiding toward a combined $725 billion in 2026 capex, up 77% from roughly $410 billion in 2025. Microsoft and Meta report earnings later today, so those numbers could move by tonight. But the direction has been one-way all year: Alphabet raised its 2026 guidance twice, most recently to $195-205 billion on July 22; Meta raised twice too, landing at $125-145 billion; Amazon is holding at roughly $200 billion with speculation it goes higher when it reports tomorrow; Microsoft is at about $190 billion. Add it up and Apple spent less than 2% of what its closest AI-era peers spent, and the market just made it the most valuable company on earth anyway.

What Apple actually built instead

Apple didn't sit out AI. It made a specific bet: run inference on-device, on its own silicon, and rent frontier capability instead of training it.

The clearest evidence came at WWDC in June, when Apple announced that the rebuilt Siri would run on Google's Gemini models via a paid licensing deal, not an Apple-trained frontier model. Apple also shipped a Foundation Models framework that lets third-party developers call Apple's on-device models directly, for free, with no per-token API bill. The on-device models handle summarization, writing tools, and routine Siri requests locally on the Neural Engine in your iPhone or Mac. Anything that needs frontier-level reasoning gets routed out to Private Cloud Compute, Apple's own server silicon, or increasingly to Gemini.

None of that requires Apple to buy hundreds of thousands of GPUs. It requires Apple to have very good chips in very many pockets, plus a services business generating $109.2 billion a year at close to 75% gross margins that can absorb a slower AI product cycle without anyone at the company panicking about payback periods.

The capex-to-value math nobody wants to say out loud

Here's the uncomfortable part for the four hyperscalers: their combined capex is now compared against depreciation schedules and demand curves that have to hold for years to make the math work. Alphabet's cloud revenue grew fast enough this quarter to justify its raise, and Google stock still dropped about 5% on the announcement because investors are nervous about the size of the number, not the direction of growth. Meta's CFO has now raised 2026 guidance twice in a single year, from $115-135 billion to $125-145 billion, citing component pricing and data center costs. That's not a company confidently amortizing a known bet. That's a company revising a bet in real time because the underlying cost curve keeps moving.

Apple has none of that exposure. It isn't racing to build enough training capacity to justify a valuation multiple. It's shipping phones people already buy, adding AI features that make the phones marginally better, and paying Google a fee for the parts it doesn't want to build. The market spent 2026 rewarding the company that treated AI as a feature investment instead of an infrastructure bet, which is exactly the opposite of the message the capex numbers from the other four are sending.

The solo-operator parallel

If you're building a product by yourself or with a small team, you are structurally much closer to Apple in this story than to Microsoft, and it isn't close. You don't have a depreciation schedule to justify, a board asking about GPU utilization, or a services business to cushion a bad quarter. What you have is distribution to a specific audience and the ability to ship product improvements faster than a company managing a $200 billion capex line.

That means the winning move is Apple's move, scaled down: integrate frontier capability instead of building your own inference stack. Call Claude or GPT-5 or Gemini through an API. Let the frontier labs eat the capex risk and the training cost. Your job is the thing Apple's job also is: the product surface, the workflow, the thing that makes the model useful to a specific person doing a specific task. I've watched founders burn a year fine-tuning open-weight models and standing up their own GPU boxes for a problem an API call would have solved in a weekend, and the fine-tuned version wasn't meaningfully better. It was just theirs, which felt like progress and wasn't.

What I'd actually do

Rent the model. Build the integration, the workflow, and the specific judgment about your users' problem that a general-purpose model doesn't have. That's the whole playbook, and Apple just proved at trillion-dollar scale that it beats the alternative on a stock chart, not just in a solo operator's cost spreadsheet. I'd rather be the smallest line item in someone else's capex budget than the biggest line item in my own, and every solo AI product I've looked at closely this year confirms that's the cheaper way to still be around next year.

But be honest about the part of the analogy that doesn't transfer. Apple can afford to integrate rather than build because it has 2.5 billion active devices and $109.2 billion in annual services revenue sitting underneath the AI bet. If Siri-on-Gemini underwhelms for two years, Apple's core business barely notices. You don't have that cushion. If the only thing your product does is wrap an API call around a prompt, you have no moat and no downside protection if the frontier lab changes pricing, changes its terms of service, or ships a feature that makes your wrapper obsolete overnight. "Don't build your own infrastructure" is not the same claim as "don't build anything defensible." The lesson from Apple isn't to avoid capex: you were never going to out-capex anyone anyway. It's to spend your scarce capital on product judgment and distribution instead of infrastructure you can rent for a fraction of the cost, and then make sure the thing you build with those savings is actually hard to copy.

Author

Lukas

@lukcombinator

Sources

• Apple ends day as world's most valuable company, passing Nvidia
• Apple touches $5 trillion market cap for first time
• Nvidia becomes first company to top $5 trillion market cap
• Google, Microsoft, Meta, and Amazon capex spending to hit $725 billion in 2026, up 77% from last year
• Apple Avoided the AI CapEx Spending Trap: Now the Bill May Be Coming Due