Apple, Adding It Up: Potential Energy

In July we called Apple the company outside the frame. It had been left off a chart of the AI builders because "hyperscaler" is "a label Apple has not applied to itself," and it was the only one of seven large balance sheets we read that week to shrink. On Tuesday morning the Wall Street Journal's Heard on the Street column reached a similar place from the market's side, calling Apple an AI safe haven with "no big AI bills coming due." (1)

Apple filed its annual report a few hours later, and the bills are in it. They sit in rent and licenses on a schedule that runs to 2030, alongside memory that AI demand has repriced. The rest of Apple's AI position is held in reserve, like potential energy: in its chips, in 2.5 billion devices, in a balance sheet with almost all of its borrowing power unused, and in a margin that AI can lift or wear down. Apple's strategy sits outside the record. What the record shows is its outline, and the outline looks considered. We expect to make out more of it with each filing.

Kinetic and potential

Oracle spent $28,499 million on data centers and equipment in the three months to August 31, more than three times what Apple spent on capital in its whole year. Customers prepaid $11.4 billion of that quarter's cash. On Monday Oracle priced $10 billion of notes at yields of 7.062% to 8.667%, with expected ratings of Baa2 on negative outlook, BBB- and BBB, and filed for $10 billion of new stock alongside them. That is energy in motion: cash turned into buildings, and buildings financed with debt. (2)

Apple spent $9,041 million on capital in fiscal 2026, down from $12,715 million. It produced about $136 billion of free cash flow and returned $102.1 billion of it to shareholders. It issued no new long-term debt and repaid $12.4 billion. It holds $136.9 billion of cash and marketable securities against $80.1 billion of debt, and its debt is rated Aaa by Moody's and AA+ by S&P. (3)

The agencies are moving on AI spending: on October 30 S&P revised its outlook on Meta, a consumer company funding its own build, to negative from stable, keeping the AA- rating and citing "significant uncertainty around capital spending and free operating cash flow." (3)

Eight rating notches separate Apple's credit from Oracle's at both agencies. That gap is stored energy: borrowing capacity at the top of the rating scale, left in reserve.

Stored is a different thing from idle. Apple is spending on AI. It is spending in places other than the capital budget.

What Apple is paying for

Research and development rose 32% to $45,765 million, "primarily due to higher infrastructure-related costs, including investments in artificial intelligence, and headcount-related expenses." (4)

Apple's commitments to buy goods and services rose to $24,844 million from $13,308 million, at about $5.5 billion to $6.7 billion a year through fiscal 2030. Part of that line is content and distribution rights; Apple does not break it out. (5)

A new class of asset, licensed intangibles, rose to $29,582 million from $16,061 million. Amortization of all intangibles was $4.9 billion this year and is scheduled at $6.6 billion next year. Other non-current liabilities rose $12.8 billion in the same year, in a line the filing leaves unitemized. (6)

Siri AI now sits beside iCloud, Siri and Maps in the bundle of services each iPhone, Mac and iPad carries, so part of each device's price is set aside and recognized over time. The chief financial officer said the cloud cost of running it will land in services cost of sales. (7)

And a new risk factor says the AI business "depends on access to sufficient compute capacity," that the market has seen "constrained supply, extended lead times, and increasing costs," and that Apple "relies on third-party cloud service providers." (8)

Apple names none of its providers. On January 12, Apple and Google said Apple's next models "will be based on Google's Gemini models and cloud technology." In the quarter before that statement, Apple's commitments rose by $19.9 billion, and in the quarter after it the licensed intangibles appeared. Sequence, not causation, and no filing on either side names the other. (9)

Updating our July reading

Some of what we wrote in July needs correcting, and some of it looks better now than it did.

The correction first. We wrote that Apple's intangible amortization "had run about $160 million a quarter" and "jumped to $833 million" in the June quarter. Those numbers were changes in accumulated amortization, which rises with the expense and falls when old, fully amortized assets are written off. The annual report now gives the expense itself: $3.7 billion in fiscal 2025 and $4.9 billion in fiscal 2026. The older licenses were amortizing at close to $1 billion a quarter all along. That takes down our timing for the new asset, which we said "switched on" in the June quarter, and our attribution of the whole $5.1 billion current portion to the new purchase. (10)

What holds: Apple added $13.5 billion of licensed assets in the year, the current portion, which Apple classifies as falling within a year, is still rising, now $5,422 million against $2,208 million a year earlier, and nothing filed says what any of it is. Spread over four to five years, the year's additions come to about $2.7 billion to $3.3 billion a year, roughly three times the $1 billion reported for Apple's Gemini license, where we had said five. In July we said either the reported figure understated the arrangement or other licensed property shared the line. The annual report says both, in part. (11)

A test running in the market

AI's demand for memory chips has pushed their price up sharply, and Apple's chief executive called the pricing "unprecedented." Apple's answer so far has been to buy early and charge more. Components inventory stood at $10,447 million at year end, up from $2,124 million. Manufacturing purchase obligations reached $89.7 billion, $82.7 billion of it due within a year. Prices on many products rose at the end of June, by as much as a third according to the Journal. (12)

Whether that comes out ahead is unknown. Products gross margin rose to 39.3% for the year, helped by tariff refunds. The December quarter is guided to 45% to 46% overall, down from 47.9%, with memory most of the decline. Apple's iPhone shipments fell 4% in the September quarter against a 22% fall for the whole market, by IDC's count. The chief financial officer said it is "early to really read price elasticity." The December quarter is the first real read on whether Apple earns more from the shortage than it costs. (13)

The same question runs through services. As a cost, AI arrives there as inference and, possibly, as license amortization; the $1.7 billion step in amortization next year would be a little under 1.4 points of this year's services margin if it all landed there. As revenue, it arrives through the Siri AI deferral, through iCloud plans with "usage limits on features that rely on server-side models," and through what the chief financial officer called "a compelling reason for customers to upgrade." Services earned a gross margin of 76.2% this year, which is the margin of a platform. Google Cloud, selling capacity, earned an operating margin of 36.2% last quarter. A neocloud selling capacity alone, in our September reading of CoreWeave, earned about 9% on its capital. They are three different measures, set side by side for order of magnitude. The point to watch is where AI's cost in Apple's services starts growing faster than the revenue AI brings in. (14)

What "the edge" means

The next two sections step outside the filings. They run on Apple's published research, outside studies and our own arithmetic, and they describe what the filings point toward.

Most writing about AI infrastructure is about training, the months-long work of building a model in a data center. The money that recurs is in inference: running the model each time someone asks it something. Inference can run in three places. It can run in a data center. It can run at the network edge, on servers spread across many cities close to users, which is what Akamai sells Anthropic across "thousands of points of presence." Or it can run on the device in someone's hand or on their desk. Running it on the device is what people mean by moving AI to the edge. (15)

Three things limit how much a device can do. The model's weights have to fit in its memory. The speed of an answer depends on how fast those weights move through the chip, which Apple's own engineers describe as the bottleneck for generating text. And a phone has a battery and no fan. (16)

A fourth thing works in the device's favor: most of what people ask a phone to do sits well short of frontier work. Apple's technical report describes its on-device model as one that "excels at a diverse range of text tasks like summarization, entity extraction, text understanding, refinement, short dialog," and says that "it is not designed to be a chatbot for general world knowledge." Turning "lunch with Sam on Thursday at noon" into a calendar entry is AI, and it is entity extraction, well inside that list. Developers have been able to call that model from their own apps since 2025. The skill in the system is sizing each job to the compute it needs and sending only the hard ones away. (17)

Apple's third-generation models push the device's share further. The smallest has 3 billion parameters. A larger one has 20 billion, but it is built as a "mixture of experts" that wakes only 1 billion to 4 billion of them for any given request, keeps the full model in flash storage and loads pieces into memory as needed. The latest chips add accelerators that Apple says cut the wait for a first answer by up to four times against the prior generation. On a desk, the reach is larger still: in October Apple's researchers ran an 816-billion-parameter model across four Mac Studios at more than 36 tokens a second. (18)

The hard requests go out to a server. Apple's own servers run on Apple chips, and since June some of that work also runs on NVIDIA GPUs in Google Cloud, the first time Apple has used third-party data centers for it. Apple says its aim is to "match every user request to the model that provides the best response at the lowest latency," and its chief financial officer says the company is in "very, very early stages of understanding the kind of mix." Google's Android works the same way, trying the model on the phone first and falling back to the cloud. (19)

The economics follow from where each answer is computed. An answer computed on the device costs Apple no rent; the customer has already paid for the chip. An answer computed in the cloud is the bill in the commitments note. And Apple has opened its developer framework to "nearly any LLM, local or server-based," with Anthropic and Google announced to bring Claude and Gemini into it, which puts Apple's interface between the models and the people using them. (20)

Who pays for the power

A data center's electricity is a contract, a queue and a public argument: a grid connection, a utility tariff or power purchase agreement, often a gas line, and the debt that finances the building. A device's electricity arrives on a household bill. A Mac mini with an M5 Pro chip draws 6 watts sitting idle and 145 watts at full load. (21)

Horace Dediu put the idea in a line in March: "Apple turned 2 billion devices into the data center." It can be given a rough size, in two numbers.

The first is capacity. In laboratory tests, an iPhone running a small language model drew up to 13.8 watts while writing an answer, and phones spent roughly 0.3 to 0.9 joules on each token, the word fragment these models produce one at a time. Apple counts more than 2.5 billion active devices. If one in ten were writing an answer at the same moment, at 6 watts each, the draw would be 1.5 gigawatts, in bursts, since batteries drain and chips heat up. (22)

The second is use, and that number stays private. The arithmetic runs in steps. At 0.5 joules per token, a device averaging a tenth of a watt of AI work produces about 17,000 tokens a day, and 2.5 billion of them add up to 250 megawatts. A quarter of a watt is about 43,000 tokens and 625 megawatts. Epoch AI estimates that Colossus 2, SpaceX's AI campus in Memphis, runs on 946 megawatts. The device base matches it at about 65,000 tokens a day per device, close to 100 pages of generated text. That matches the power draw, a different thing from matching the machine: Colossus 2 trains and runs frontier models, work beyond a phone at any wattage. Only recent models run Apple Intelligence (the iPhone 15 Pro and later, and Macs and iPads with M1 chips or later), and Apple gives no count of them, so each figure here is an upper reading. (23)

Apple got paid to build it. Its customers bought the chips, the memory and the batteries, at Apple's margin, and they pay for the electricity. Apple can tap that capacity for its own features or route the work to its servers and Google's. Either way, the construction was paid for at the register.

All of it is potential. How much gets used, how efficiently, and whether it is switched on for this work at all remain open. Apple's researchers are testing one version of the idea, a system called Minions that "aggregates idle Apple Silicon GPUs across macOS devices to execute large-scale machine learning inference workloads." It is a research paper, and Apple describes no deployment. (24)

Whether a question costs more energy to answer on a phone or in a data center is unsettled too. One recent study found phones about three times less efficient than a busy server; another, backed by Qualcomm, found savings of 75% to 95%. What holds either way is where the bill lands. It sits inside household utility payments, it appears in no filing, and the capacity exists the day the device is sold, with no wait for a substation. (25)

The schedule ends in 2030

Both of Apple's long-term schedules end in the same place. The commitments fall from $5.5 billion in fiscal 2030 to $178 million in fiscal 2031. Amortization falls from $3.8 billion to $0.8 billion over the same step. (26)

Not starting has a cost, and some of that cost is the price of seeing more before committing. Working through the options and watching how things develop is part of what management is paid for. Apple's record has that look. It rents compute and licenses models during the most expensive stretch of the industry's build, keeps its credit almost untouched, and puts its own money into chips, devices and the software that routes work between them. The familiar line that Apple arrives late reads, on this schedule, as sequence.

That is a reading of structure and schedule. It claims nothing about intent or about whether the plan works, and the record can test it. It weakens if the commitments are renewed past 2030 at scale, if Apple's capital spending steps up into data centers, if the cloud share of Siri AI keeps rising, or if large license additions recur every year.

What comes next

The July piece set several things to check against this report. Revenue growth of 12% beat guidance of 9% to 11%. Gross margin of 47.9% landed inside guidance of 47% to 48%, and the decline excluding tariff refunds was, in the chief financial officer's words, "driven entirely by memory impacts." The tax note now names Ireland and Singapore. The intangibles received a table and a schedule, and still no sentence. One July line reversed: Apple's total liabilities rose for the year, to $292.5 billion. (27)

The December quarter answers the memory test and gives the first full quarter of Siri AI in the deferral and in services cost. The first-quarter report updates the commitments and the licensed balance. The Claude and Gemini integrations are announced and not yet shipped. And the line that carries amortization will be named when Apple adopts the new expense disclosure rules in the fourth quarter of fiscal 2028. (28)

Free to read, no paywall. Readers who find the work useful can support it at the foot of the page; payment offsets the time and buys no access. Corrections run in daylight, dated, in a comment or in the following piece, and a reader who finds one is doing us a favor.

Analysis: Cape Fear Advisors.