Apple Spent $6B on AI in 6 Months and the Chart That Matters Isn't Siri

$4.3 billion is practically a rounding error. Apple generated more than $82 billion in operating cash flow in the first half of fiscal 2026, yet its capex across that same period came in at $4.3 billion — roughly five cents deployed for every dollar of cash the business produced, according to CNBC. Everyone’s fixated on Siri.

The real chart is the one showing Apple’s research spending pulling away from its willingness to build anything to run that research on.

Apple executives reportedly called the Siri delays “ugly and embarrassing,” and they weren’t wrong. Internal testing showed the new Apple Intelligence features working correctly only about two-thirds to 80% of the time, which is why Cupertino shelved the rollout indefinitely.

But here’s what the print doesn’t say. R&D hit $11.4 billion in a single quarter — the highest in company history — representing 10.3% of revenue versus 7.6% the prior quarter and 9.0% a year earlier. That acceleration is staggering.

The tell is the gap. Apple isn’t struggling to afford infrastructure — it’s choosing not to build it. When a company separates research spending from deployment spending this dramatically, it is funding ideas it has no infrastructure to serve at scale.

The chart that matters isn’t Siri’s feature roadmap. It’s the widening distance between what Apple is learning and what Apple can actually deliver to a billion devices. Consider what that means for the next three to four quarters: every dollar sunk into R&D without corresponding capex is a dollar that produces shelf-ware, not product.

This is the strongest counterargument and it deserves a direct answer. Apple’s asset-light model — outsourcing manufacturing to Foxconn, chips to TSMC — generated extraordinary returns precisely because it outsourced commodity production while retaining proprietary design. The math worked beautifully for two decades.

The trap in 2026 is that AI inference isn’t a commodity. It’s the interface layer.

When Apple routes queries through Google’s models or OpenAI’s infrastructure, it isn’t outsourcing assembly. It’s outsourcing the customer relationship at the moment of highest intent. Google and OpenAI don’t merely answer the question — they own the context, the session, and increasingly the habit.

Apple’s full-year fiscal 2025 capex of roughly $12.7 billion — already the ceiling, not the floor — looks thin against peers spending multiples of that figure on inference infrastructure alone. The asset-light model works when you control the value-added layer. Apple is currently licensing that layer from the two companies whose explicit strategic goal is to own the AI interface Apple currently controls.

Worth noting: I’ve covered Apple through four platform transitions. This is the first time I’ve seen Cupertino willingly hand the high-margin layer to a competitor and call it a partnership.

Related: what we covered last week

Here’s what most analysis misses entirely. Apple’s R&D surge may be making the dependency worse, not better. At 10.3% of revenue, Apple is running research at a pace its own infrastructure can’t absorb.

Every model trained internally, every capability developed inside Apple Park, hits a deployment ceiling the moment it needs to serve a billion devices at inference scale. The result is a structural perverse incentive — Apple’s researchers produce outputs Apple can’t run, so Apple either shelves them or routes users to third-party infrastructure to demonstrate them. The University of Wisconsin School of Business summarized the dynamic plainly: Apple’s AI reliability gap reflects a deeper organizational challenge, not just an engineering delay.

This is how negotiating use transfers. Quietly.

Google and OpenAI aren’t passive vendors. They’re learning Apple’s users — query patterns, device contexts, language preferences — through every inference call Apple can’t serve itself. The data relationship that made Apple’s privacy narrative valuable is being sublicensed to rivals, not through any single contract but through the compounding logic of capex restraint meeting R&D ambition.

Call it what it is. Apple is paying two companies to learn its customers.

The time-horizon question matters here: the cash pile buys optionality, and Cupertino has surprised before when cornered. But the margin compression risk in Services — if the AI interface layer migrates to Google or OpenAI within three to five years — is not in any model I’ve seen. Watch the capex cadence, not the guidance. That’s the number that tells you whether Apple is solving this or deferring it.

The setup into the September quarter is murkier than the tape suggests. If Apple announces its own inference infrastructure at WWDC, the thesis changes. If it doesn’t, every channel check I run will be measuring how fast Google and OpenAI are learning Apple’s users — on Apple’s dime.