← Blog

A Tale of Two Earnings - What the Filings Cannot Show

Lida Liberopoulou ·31 July 2026 · CC BY-SA 4.0

Raw .md ↗ Share on LinkedIn ↗ Share on X ↗

Independent research. Available for advisory, research collaboration, and expert consultation. Get in touch

On July 27, this publication filed six checkable claims about what Microsoft, Apple, and Amazon's earnings releases would disclose and what they would leave impossible to calculate. All six held. The claims tested for the same recurring pattern: investment effects, commercial demand, and infrastructure costs were each disclosed separately, with no bridge connecting them. This article scores the results, introduces a framework of three separations that prevent investors from evaluating the interdependence of these relationships, and asks whether the demand driving the AI infrastructure build is independently financed or ecosystem-financed. The distinction is important because the models, price targets, and portfolio weights set this week already carry assumptions the filings do not contain enough information to verify.

On July 27, in "A Tale of Two Earnings - Microsoft, Apple, and the Circle Between Them" this publication described the financing relationships connecting hyperscalers to the frontier AI labs they invest in, purchase from, and sell infrastructure to. The article specified six checkable claims about what that week's earnings filings would disclose and what they would leave impossible to calculate. Each claim included explicit pass/fail conditions. On Wednesday and Thursday, Microsoft, Apple, and Amazon reported. All six held.

The six results were not independent guesses. They tested for the same thing: a recurring reporting pattern in which investment effects, commercial demand, and infrastructure costs are each individually visible, but the connections needed to evaluate their economic interdependence are not. The result is more important than the score because across every filing examined, the information gap ran in the same direction.

The results

Not all six claims carried equal difficulty. Several tested whether established disclosure boundaries would persist. More specifically, they tested whether Microsoft would suddenly introduce an AI profitability metric and whether Apple would acquire material frontier-lab exposure. Those were the conservative ones. The Amazon mark and the cash-conversion test were genuinely exposed to failure.

They are summarised here briefly:

Claim: Microsoft's free cash flow would deteriorate despite continued revenue growth, as additions to property and equipment outpaced operating cash flow growth.

Result: Microsoft reported revenue of $90 billion, up 18%, and Azure growth of 43%. For the full fiscal year, additions to property and equipment reached $115.9 billion, nearly double the prior year. Full-year operating cash flow was $182.9 billion, up 34%. Free cash flow (defined here as operating cash flow minus additions to property and equipment) was $67 billion, down from $71.6 billion in FY2025. Revenue grew 18% while free cash flow declined.
Claim: Microsoft's reported OpenAI investment effect would not provide an operating bridge to OpenAI's underlying commercial performance.

Result: The reported OpenAI investment effect moved from a $3.6 billion full-year drag in FY2025 to a $5.0 billion contribution in FY2026. The release reconciled the accounting effect but did not explain how much of the reversal reflected OpenAI's underlying commercial performance.
Claim: Microsoft would not disclose how much Azure revenue comes from OpenAI.

Result: Azure growth of 43% appeared without disclosure of how much revenue was recognised from the customer Microsoft funds.
Claim: Microsoft would not introduce an AI-specific operating margin or return measure.

Result: No AI-specific operating margin or return measure appeared anywhere in the completed fiscal year filing.
Claim: Apple would report clean earnings with no frontier-lab investment effect, no AI-adjusted metrics, and continued buybacks.

Result: Apple reported revenue of $109.4 billion and earnings per share of $2.02, driven by a 22% increase in iPhone sales. No frontier-lab investment effect appeared and no AI-adjusted earnings reconciliation was needed. Gross margin was 50.1%. The $100 billion buyback authorisation from April remained active and the income statement reflected the business without adjustment.
Claim: Amazon would fold a material Anthropic gain into headline earnings and would not disclose Anthropic's contribution to AWS revenue.

Result: Amazon reported net sales of $200.6 billion, up 20%. AWS revenue reached $42.2 billion, growing 37%. Earnings per share were $5.75 against a pre-release consensus of approximately $1.82. The release disclosed $53.4 billion of non-operating pre-tax other income, primarily from investments in Anthropic. Trailing twelve-month free cash flow was negative $7.6 billion, driven by a $66.1 billion year-over-year increase in property and equipment purchases that the company attributed primarily to AI investment. Amazon named Anthropic and OpenAI as making multi-year, multi-gigawatt commitments to its Trainium chips. Anthropic's contribution to AWS revenue was not disclosed.

The three separations

These filings do not conceal anything. Investment gains, cloud growth rates, and capital expenditure are reported. Customer relationships are acknowledged, the accounting is compliant and the footnotes contain what they are required to contain.

But the question here is whether the numbers can be connected.

Three kinds of information are disclosed as separate, unrelated facts. And this is where the analytical problem lives.

Investment and operation. Microsoft records a share of OpenAI's economic results through equity-method accounting. Amazon holds convertible notes and preferred stock in Anthropic. Google marks its Anthropic stake to market. Each of these produces income-statement effects that are reported and reconciled. What is not reported is how much of the gain reflects improved operations at the investee versus financing events, valuation adjustments, or mark-to-market movements driven by the broader investment cycle.

Investor and customer. OpenAI has committed to purchase at least $250 billion of Azure compute through 2032. Anthropic buys AWS compute and uses Amazon's own chips. Anthropic purchases Google Cloud TPU capacity. Each hyperscaler reports its cloud revenue in aggregate. None separates the revenue recognised from a customer it also funds. In a prior quarter, Microsoft disclosed that excluding OpenAI, commercial remaining performance obligations grew 26% rather than the headline 99%. In Q4, the company repeated the disclosure: 25% excluding OpenAI, against 84% headline growth. That single recurring datapoint is the closest any hyperscaler has come to quantifying the concentration.

Revenue and capital. Microsoft added $115.9 billion of property and equipment during FY2026. Amazon's trailing twelve-month purchases produced negative free cash flow. Google's Q2 capex exceeded operating cash flow for the first time since its IPO. These figures are reported in aggregate. None is matched against AI-specific revenue, AI-specific margins, or AI-specific asset utilisation.

Each relationship is disclosed in its proper accounting category. Investment effects appear in other income or non-GAAP reconciliations, commercial demand appears in segment revenue, and infrastructure cost appears in capital expenditure and the cash flow statement.

Management provides several connections around scale and demand through AI revenue run rates, commitment figures, growth contributions and backlog expansion. But the equivalent connections for concentration, margins, and return on the infrastructure being built are not provided. And the resulting information gap ran in the same direction across every release examined.

Amazon

Amazon is the clearest case because all three separations appear in a single quarter at a scale that is difficult to contextualise as routine.

Amazon reported operating income of $27.5 billion and net income of $62.6 billion. The difference is dominated by $53.4 billion in pre-tax other income from Anthropic investments. That figure exceeded operating income from the entire business.

In the same release, AWS grew 37%. Amazon named Anthropic as a major customer making multi-gigawatt infrastructure commitments. Anthropic's share of that growth was not disclosed. Trailing twelve-month free cash flow was negative $7.6 billion. The company attributed the deterioration primarily to AI-related infrastructure purchases.

All three sides of the relationship were material enough to affect Amazon's reported results. But the filing did not provide the attribution needed to assess them as one economic system.

Amazon disclosed the value created on one side of the Anthropic relationship, the existence of demand on the other, and the broader cash cost of expanding the AI infrastructure in which that relationship operates. What they didn't disclose was how much of the cloud growth came from the company Amazon funds, and whether that revenue covers the infrastructure built to serve it. That is the number required to evaluate the relationship as a whole.

The pattern

In the Q4 earnings release, Microsoft identified several discrete items affecting quarterly comparability. These were a $3.2 billion gain from an investment in Anthropic, lower voluntary-retirement costs, severance expense, and impairment charges in Xbox.

An impairment charge has a specific triggering event, a recognised accounting amount, and a period in which it must be recorded. An AI profitability measure would require management to allocate shared infrastructure, depreciation, research, and sales across an internally integrated programme. These are different accounting objects, and the reasons one is disclosed while the other is not are technically legitimate.

Still, the analytical consequence remains. Microsoft can identify discrete investment gains and charges when they affect quarterly comparability. But it does not provide an equivalent economic bridge connecting the far larger AI capital programme to the revenue and returns it is intended to produce. Investors can see individual valuation events and individual write-downs more clearly than they can see the economics of the programme generating the spending.

This observation extends beyond a single quarter. Over the completed fiscal year, Microsoft reported $115.9 billion in additions to property and equipment, a $5.0 billion net contribution from OpenAI investments, a $3.2 billion Anthropic gain, Azure growth of 43%, a commercial backlog that grew 84%, and an AI revenue run rate that crossed $37 billion. And the last figure is a management calculation that does not appear as an SEC disclosure. At no point did the company connect these figures into a return measure, a margin disclosure, or a customer-concentration range for its AI programme.

To be fair here, management itself identifies AI as the principal driver of revenue growth, capacity constraints, and capital expenditure. But once the same activity is used to explain all three, the inability to connect them becomes material on the company's own account of its performance.

The same separation had already appeared in Alphabet's preceding quarter. It showed infrastructure cash pressure that was severe enough to produce the company's first negative free cash flow quarter since its 2004 IPO. It also included cloud growth, $99 billion in net gains on equity securities including holdings in companies that purchase its compute, and no full attribution bridge. So the pattern was not limited to the three-company scorecard.

The contrast

The original article called Apple the "control group." That language was deliberate, though the comparison is not between equivalent businesses. Apple is not a hyperscale cloud vendor. It does not sell frontier-model training infrastructure, and embeds its AI spending in products and R&D rather than deploying it as a standalone capital programme. It also operates its own hardware and models through Apple Silicon and Private Cloud Compute. So the distinction is one of strategic architecture and degree, and not a clean division between building and renting.

Apple's filing provides a useful contrast because its AI strategy does not combine frontier-lab investment exposure, cloud infrastructure sales, and hyperscaler-scale capital deployment inside the same corporate structure. Its income statement contained no frontier-lab investment effect and no AI-adjusted earnings measure was needed. There was also no intertwined relationship between investment, revenue, and infrastructure spending in the filing.

Apple's capital expenditure remains small enough that it does not compete with the company's cash returns. Apple generated $110.2 billion of free cash flow during the first nine months of FY2026. The $100 billion buyback authorisation from April remains active, and Apple spent approximately $26.1 billion repurchasing shares during the quarter alone.

Apple demonstrates that the reporting complexity visible in the other filings arises from a particular set of strategic positions in the AI supply chain. It doesn't arise from the technology itself, from AI adoption in general, or from the requirements of being a large public technology company.

What the missing bridges prevent

A sceptical reader may reasonably ask: so what? The hyperscalers can afford the spending. Their balance sheets are enormous, AI demand is growing and the eventual winners will own the infrastructure. If the losses fall on frontier labs, venture investors, suppliers, and weaker competitors while the largest companies consolidate the market, that may be rough but economically rational.

The answer to this fair objection is that the current disclosures do not contain enough information to determine whether the reading just described is correct.

Analysts already distinguish recurring revenue from transactional revenue, organic growth from acquired growth and gross retention from net retention. AI infrastructure demand may now require an analogous distinction: independently financed demand versus ecosystem-financed demand. Ecosystem-financed demand is still real demand for a real service. And the distinction concerns the source and durability of the customer's purchasing power, not the validity of the transaction. But the two types carry different assumptions about durability, margins, customer credit quality, capital intensity, and required future support.

So the three separations leave investors unable to answer three questions that bear directly on valuation.

The first is whether the demand is independent. Would the customer purchase the same capacity without capital support (direct investment, equity stakes, compute commitments, or favourable terms) from the supplier ecosystem? The filings report demand volume but they do not report demand provenance.
The second is whether the demand is profitable at the infrastructure level. Does the recognised revenue cover the infrastructure, depreciation, and operating costs required to serve it? The filings report aggregate capital expenditure and aggregate cloud revenue, but do not connect them at the level of the AI programme driving both.
The third is whether the demand is reproducible outside the current ecosystem. Is the broader economy generating enough independent purchasing power to eventually replace ecosystem-financed demand with arm's-length customers paying market rates from their own revenues?

The third question is where the financial problem meets the wider economy. The same technology is being sold partly on its ability to reduce labour, software, and service costs across the broader economy. Whether those savings generate new investment and demand, or primarily compress the revenues and incomes of future customers, remains unresolved. The filings reveal no mechanism for testing which of these is occurring.

This is a forward analytical problem: the capital commitments extend years, the assets depreciate over decades, and the market valuation capitalises distant returns. If the broader economy's capacity to pay is deteriorating while the infrastructure is being built, that duration mismatch matters regardless of how the next quarter's earnings read.

What this shows

What the results establish is narrower and more important than a prediction score. Across the filings examined, the accounts recorded the components of interconnected commercial and investment relationships. They did not disclose the relationships between them.

In Amazon's quarter, Anthropic appeared simultaneously as an investee whose rising valuation produced $53.4 billion of pre-tax other income and as a named customer making multi-gigawatt infrastructure commitments to AWS. Alongside those relationships, broader AI-related capital spending pushed trailing free cash flow to negative $7.6 billion. These facts appeared in the same filing, with no disclosed connection between them.

That is not an abstract accounting concern. A revenue model that cannot separate ecosystem-financed demand from independent demand will treat both at the same growth rate and the same margin assumption. A credit assessment that cannot distinguish non-cash investment gains from operating cash generation will overstate cash flow durability. A portfolio that treats the investment gain, the cloud growth, and the infrastructure spending as three independent signals may be tripling its exposure to a single relationship. The three separations do not merely leave a gap in disclosure. They have already entered the models, the price targets, and the portfolio weights that were set this week. They are carrying assumptions about demand independence, cash flow durability, and signal diversity that the filings do not contain enough information to verify.

And yes, you can say that AI is not a separable business, that any AI-specific margin would require arbitrary allocations, and that matching current-period spending against current-period revenue would misstate the economic life of the assets. All of that may be true. But management itself presents AI as the principal cause of revenue acceleration, backlog growth, capacity constraints, and capital deployment. Once the same programme is used to explain all of these, the absence of attribution connecting them becomes material by the terms of the company's own performance narrative.

The prevailing interpretation often treats investment gains, cloud growth, and accelerated infrastructure deployment as mutually reinforcing evidence that the AI programme is scaling successfully. The filings do not contain enough information to determine whether that interpretation is correct.

The accounts show the pieces. They do not show the system.

Primary sources

This analysis is part of an ongoing series examining the structural economics of AI infrastructure. For the financing relationships described here, see The Silicon Industrialists — Part 1: AI's Gilded Age. For the historical conditions that made the software economy's escape from industrial economics possible, see Part 2: Tech's Great Escape.