UnbiasedAI

Blog

Apple

As of November 3, 2026 (fiscal Q4 2026 results and the 10-K). Updated when Apple reports or signs something that changes the bet. [All companies](/blog/companies).

Apple barely builds at all: $9.0B of capex last fiscal year against $102B handed back to shareholders, and its most capable server model runs on Nvidia GPUs it rents in Google Cloud.

The bet

Apple's bet is that it doesn't need to own the buildout. Capex for fiscal 2026 (the year ended Sep 26) was $9.04B, down 29% from the year before, 1.9% of revenue, and less than its $13.4B of depreciation and amortization. Revenue grew 15% to $478.8B over the same year.1

The AI money shows up in other lines. R&D rose 32% to $45.8B for the year, and Apple's July 10-Q put the increase mainly down to infrastructure costs, including AI, plus headcount.2 Server-side Siri inference is booked in Services cost of sales, not capex, according to the CFO on Monday's call.7

A lot of the compute is rented. Apple's next foundation models are "based on Google's Gemini models and cloud technology," per a January joint statement that disclosed no financial terms.3 Since June, Private Cloud Compute also runs in Google's data centers, and Apple's top server model, AFM 3 Cloud Pro, runs on Nvidia GPUs there; pre-training used Google's cloud TPUs.45 Apple's own servers use Apple silicon, and four of its five third-generation models are built for it. A new risk factor in the July 10-Q says Apple relies on outside cloud providers alongside its own data centers, and that compute supply is tight and costs are rising.2 The 10-K adds a line: investing in data-center capacity means long-term commitments that "increase the Company's capital expenditures, liabilities and obligations."8 It's a risk factor, not a plan, but it's the first time Apple has written down that the cheapest line in this section might not stay cheap. The December-quarter guide fits the cost story too: gross margin of 45–46%, mostly on higher memory costs.7 Apple is buying ahead: component inventory went from $2.1B to $10.4B over the year, and manufacturing purchase obligations from $56.2B to $89.7B.8

The product is Siri AI, announced in June and rolling out in beta since Sep 14, in English only, with daily limits on server-side features and a paid tier promised for later.6 John Ternus, CEO since Sep 1, described the iPhone on his first call as "the intelligent personal hub."7

How it's paying for it

Company-reported, from the fiscal 2026 10-K (filed Nov 3; every figure matches the Nov 2 release). Quarterly figures are our arithmetic: full-year totals minus the nine-month figures.
Fiscal Q4 (Jun 28–Sep 26)Fiscal 2026
Capex$2.24B$9.04B
Operating cash flow$28.1B$145.1B
Buybacks$24.3B$86.4B
Dividends$3.9B$15.7B
Debt issued$0$0
Stock sold$0$0

Apple paid for all of it from operating cash flow and still returned about $11 to shareholders for every $1 of capex. It also spent more on dividends than on capex, a sentence nobody else on these pages can write. It issued no debt or stock, and total debt fell from $98.7B to $80.1B, leaving net cash of about $56.8B (our arithmetic).1 It added another $100B buyback authorization in April, has $114.0B of room left, and has dropped its net-cash-neutral target.87 Leases are still small: $14.4B of lease liabilities, $1.5B of them finance leases, the kind Microsoft and Meta count as capex.8

The commitments are growing faster than the capex. Unconditional purchase obligations, which Apple says are mostly supplier arrangements, distribution rights, and licensed IP and content, nearly doubled in a year, from $13.3B to $24.8B, at about $5.5–6.7B a year through 2030. Licensed intangible assets went from $16.1B to $29.6B gross, amortizing at about $6B a year over the same stretch.8 The filing doesn't say what was licensed. It's tempting to call that the Gemini deal. Nothing in the filing says so, and content rights fit the description just as well.

Our read

Louis's frame (after Ben Thompson on Sharp Tech): if the hyperscalers end up funding the whole buildout and the money runs out before revenue catches up, it ends like the railroads, and the survivors come out with leverage over Nvidia and TSMC. Apple sits outside the first half. It isn't funding the buildout, so it can't run out of money paying for it; it's a tenant, paying Nvidia by way of Google's bill. That cuts both ways. A bust that leaves cheap, overbuilt capacity is good for tenants. A funding stall before the overbuild makes capacity scarce and expensive, and Apple's own risk factor already describes that version. On the second half, any leverage the survivors win over the chip suppliers would sit with Apple's landlord, not with Apple.

What we're watching

Changelog

Disclosures

Louis previously worked at Google. Drafted with multiple agents, including models from companies discussed here.

Every relationship we have →

Notes

  1. Apple, fiscal Q4 2026 results (8-K exhibit 99.1), Nov 2, 2026. ↩
  2. Apple, 10-Q for the quarter ended Jun 27, 2026, filed Jul 31, 2026. The buyback authorization is in this filing and the Q2 10-Q. ↩
  3. Google, joint statement from Google and Apple, Jan 12, 2026. ↩
  4. Apple Security Research, Expanding Private Cloud Compute, Jun 8, 2026. ↩
  5. Apple Machine Learning Research, the third generation of Apple's foundation models, Jun 8, 2026 (updated Sep 9). ↩
  6. Apple Newsroom, Siri AI is here, Sep 14, 2026. ↩
  7. Six Colors' transcript of the Nov 2 call; we haven't checked it against Apple's webcast. ↩
  8. Apple, Form 10-K for fiscal 2026, filed Nov 3, 2026. Year-earlier figures from the fiscal 2025 10-K. The data-center sentence isn't in the July 10-Q. ↩