Apple Is the Most Valuable Company Again, and It Spends 6% of What Alphabet Spends on AI

On Monday Apple passed Nvidia to become the most valuable public company on earth. On Tuesday it touched $5 trillion, hitting an intraday high of $342.89 before giving the gain back and closing roughly flat.^[1]Second company ever to get there. Nvidia was the first, in October 2025.

Here is the number that makes it interesting. Apple is up nearly 25% this year. Nvidia is up 2.6%.^[1] And Apple spent $12.7 billion on capital expenditure in fiscal 2025, against roughly $205 billion at Alphabet.^[1]

So the market has decided that the company spending the least on AI infrastructure is worth more than the company selling it. That is not a small rotation. It is a different answer to the question the entire trade has been built on since 2023.

The Question Everyone Has Been Answering the Same Way

For three years the AI trade rested on one premise: intelligence is compute-bound, compute is supply-constrained, and whoever supplies it captures the economics. Every downstream conclusion followed from that. Buy the picks and shovels. Capex is a moat, because a rival who has not spent the $100 billion cannot catch up. Hyperscalers that underspend get left behind.

Apple was the designated loser of that story. It was late, it was cautious, Siri was a punchline, and it refused to announce a nine-figure GPU budget while its peers were competing on the size of theirs. The market treated that as a failure of nerve.

The reading that just won is the inverse: Apple did not fail to enter the arms race, it declined to. It has a billion-plus device install base and an unusually direct relationship with the customer. If AI turns into a feature of products people already own, then whoever owns the product collects, and the party who financed the data center is a supplier with a supplier's margins and a supplier's cyclicality.

That is the entire argument, and the reason it is landing now rather than a year ago is that the depreciation question has become concrete. How long a GPU stays economically useful determines whether that spending is an investment or a recurring cost, which is the thread I pulled on in how long an AI GPU actually lasts. Nothing about that maths changed this week. What changed is how many people are doing it.

The Other Shoe: China Started Building the Machine

In the same week, The Information reported that China has begun manufacturing its own immersion DUV lithography machines, with deliveries expected this year to SMIC, Hua Hong Semiconductor and ChangXin Memory Technologies.^[2] DUV immersion is the tool class ASML has effectively owned, and export controls on it are a central instrument of US semiconductor policy.

Chip stocks sold off, and the Nasdaq slid toward correction territory.^[3]

Be careful here, because the caveats are load-bearing:

• This is DUV, not EUV.Extreme ultraviolet remains a separate and much harder problem, and ASML's monopoly there is untouched by this.
• The nodes come from multi-patterning. Engineers believe the machines could support 7nm and potentially 5nm production by patterning multiple times, which is a known technique with known costs in throughput and defect rate. It is not the same as reaching those nodes directly.
• Yields are below ASML's.^[2] This is the one that matters. A lithography tool that cannot hit competitive yield is a science achievement, not a supply chain. Yield parity is the threshold for actual adoption, and the reporting does not claim it has been crossed.

What Is Actually Being Re-Priced

Both stories are the same story. The market is not deciding AI matters less. It is deciding that AI capital expenditure is a cost of doing business rather than a durable competitive advantage.

Those two framings produce wildly different valuations from identical revenue. Spending that buys a moat gets capitalized: investors look past the cash going out because it buys a defensible position. Spending that is simply the price of staying in the game gets subtracted: it is an expense wearing an investment's clothes, and it recurs every cycle as the hardware depreciates. Same income statement, very different multiple. If the distinction between those two treatments is fuzzy, the difference between market cap and enterprise value is where it becomes concrete.

Apple at $5 trillion on $12.7 billion of capex is the purest expression of the second view that exists. It is the market saying: show us the return, and until you do, we will pay more for the company that never wrote the cheque.

The Case Against This Trade

I think the rotation is directionally right and I also think it is early, and it is worth being specific about how it fails rather than just enjoying the narrative.

The returns may be there and just not visible yet. Infrastructure spending shows up in the cost line immediately and in the revenue line slowly. Amazon, Apple, Meta and Microsoft all report within days of this being published, and a quarter that shows clear AI revenue attribution at the hyperscalers would undercut the entire premise. That is the near-term test, and I would not want to be positioned as though it has already been failed. The question of what those returns look like is the one in hyperscaler earnings and capex ROI.

The frontier may not flatten. The device-centric bet assumes models good enough to run near the user are good enough, full stop. If capability keeps jumping, the useful model stays too large for the phone, and Apple becomes a customer of the infrastructure layer rather than an alternative to it. That is the bear case for Apple and nobody has disproved it.

A 1.8% intraday move is not a thesis. Apple touched $5 trillion and gave it back the same session. The two companies have swapped the top spot repeatedly since mid-July.^[4] Treating a crossover that reverses within hours as a regime change is how people talk themselves into positions. The capex asymmetry is the durable fact here. The ranking is noise.

What I Would Actually Watch

Three things, none of them a stock price.

Capex guidance, not capex. The number already spent is sunk and known. What tells you whether management believes its own moat story is what they guide to for next year, and more precisely how they justify it. A hyperscaler that quietly trims guidance while insisting nothing has changed is telling you something.

Yield reporting out of SMIC. The Chinese lithography story becomes real at exactly one moment: credible evidence of competitive yield on the domestic tools. Until then it is a capability demonstration. After that it is a supply chain, and the entire pricing power of the equipment layer has to be re-examined.

Whether Nvidia's revenue keeps concentrating. The problem with selling picks and shovels to five customers is that your moat is really their conviction, which is the structural risk laid out in Nvidia's customer concentration. If the same buyers who fund the capex start to question its return, the supplier finds out before anyone else does.

If you want the longer version of the bull and bear arguments on the chip side specifically, they are both laid out in AI chip stocks in 2026, and the wider valuation question is in is this an AI bubble.