When EBITDA Lies

Oracle’s Compute Debt

How the AI build-out turned the market’s favourite metric into its most misleading one -and what the shift from labour debt to compute debt means for reading a P&L.

The technology & hyperscaler paper in the Bloor Finance series: a CFO-perspective benchmark of Oracle (NYSE: ORCL) against the enterprise-software peer set, on the FY2026 audited record and the peers’ latest half-year filings to 3 September 2026.

1 · Where the opening paper left off

The opening paper in this series made one claim and asked twelve questions. The claim: wherever an organisation embeds AI, cost leaves the wage bill -where EBITDA (earnings before interest, tax, depreciation and amortisation) sees it -and reappears as depreciation, interest, leases and provisions, where it does not. Technology was named the epicentre and Oracle the clean case. This is the industry paper that section promised: the company-level analysis, the peer benchmark, and the questions a CFO in this sector should be asking before the next results call. Five weeks is a long time in this cycle. Since the opening paper went to press, S&P has cut Oracle to one notch above junk; Alphabet has posted the first negative free-cash-flow quarter in its history as a public company; Microsoft has closed a year in which capital spending rose by four-fifths and free cash flow fell; and the four largest hyperscalers have lifted their combined 2026 capital plans to roughly three-quarters of a trillion dollars. Every one of those events happened below the line EBITDA draws.

2 · The number that stopped telling the truth

EBITDA was never meant to be a measure of profit. It was a convenience -a way to strip out financing and accounting estimates so two businesses could be compared on the cash-like performance of their operations. For three decades the convenience held, because in most businesses depreciation was a slow-moving echo of capital spending that had already happened. Removing it changed the picture only modestly.

That assumption has broken. In the AI-infrastructure build-out, depreciation is no longer a quiet historical echo; it is the fastest-growing, most economically significant cost in the business -and it is precisely the cost EBITDA removes. The metric is doing exactly what it was designed to do. The trouble is that the thing it discards has become the thing that matters most. Nowhere is the gap clearer than at Oracle.

3. Oracle: the inversion in a single company

Oracle’s FY2026 -the year to 31 May 2026 -is, on the headline metrics, the best in its history. Revenue reached $67.4bn, up 17%. GAAP (Generally Accepted Accounting Principles) operating income was $20.6bn, a 31% margin; add back $7.6bn of depreciation and $1.7bn of amortisation and derived EBITDA is roughly $29.9bn, a margin of 44%. Operating cash flow grew 54% to a record $32.0bn. The contracted backlog -remaining performance obligations, or RPO -stood at $638bn, up 363%, lifted by very large AI-capacity commitments. Guidance for FY2027 is $90bn of revenue. By any EBITDA-centred reading, this is a company accelerating.

And yet the same year consumed more cash than it generated. Capital expenditure rose 162% to $55.7bn -83% of revenue and 174% of operating cash flow -turning that record $32.0bn into free cash flow of −$23.7bn. Oracle raised $43bn of debt and $5bn of equity to bridge the gap; borrowings and finance leases ended the year at $137bn, against $93bn a year earlier. Alongside the build, roughly 21,000 roles -13% of the workforce -were removed and a $1.84bn restructuring charge taken, the reductions attributed in Oracle’s own filings to AI adoption.

Two details in the cash-flow statement deserve a CFO’s attention. First, that record operating cash flow includes $4.6bn of customer prepayments which Oracle itself classifies as carrying a ‘significant financing component’ -customers, in effect, lending Oracle the money to build the capacity they will later consume. Oracle now publishes a ‘net cash outlay for capital expenditures’ of $47.7bn that nets such items against capex; the honest reading nets both sides. Second, depreciation nearly doubled -from $3.9bn to $7.6bn -and still amounts to only 14% of the year’s capital spending. The depreciation that $55.7bn of FY2026 capex implies has barely begun to arrive.

For FY2027 Oracle guides to a net cash outlay of around $70bn, with reported capex $20–25bn higher still, funded by a further $40bn or so of debt and equity including a $20bn at-the-market equity programme. The investment case no longer rests on demand -the $638bn backlog settles that, and Oracle expects 12% of it to convert to revenue within twelve months -but on whether that backlog becomes cash before capital intensity and leverage peak. EBITDA cannot see any of it, because the whole story sits below the line it draws.

4 · What the credit market said in July

On 9 July 2026 S&P Global Ratings cut Oracle’s long-term issuer rating to BBB- from BBB -one notch above speculative grade -and named the concentration of the backlog in a single, loss-making AI customer as a key credit risk. Moody’s carries a negative outlook. Oracle’s five-year credit-default-swap spread widened to its highest level in nearly eighteen years, and its ten-year bonds traded to yield around 6.5%, closer to high-yield territory than to the BBB average. S&P’s own arithmetic puts FY2027 capital expenditure at $90–95bn and warns the free-cash-flow deficit could widen to around $42bn.

Equity followed. The shares fell from above $230 at the start of June to a low near $115 in late July before recovering to $154 at the 3 September close -a market capitalisation of roughly $445bn on 2.88bn shares. The point is not the price. It is that the rating agencies and the bond market read the lines EBITDA discards -leverage, interest cover, cash conversion, counterparty concentration -and acted on them. EBITDA, on the same numbers, hit a record.

5 · The peer benchmark: one company crossed the line first

Set Oracle against the enterprise-software peer set and its distinctiveness is not growth or margin -it is the sign of its free cash flow. The asset-light peers -SAP, Salesforce, ServiceNow, Workday, Snowflake -convert revenue to cash at margins between 11% and 34% on capital expenditure of no more than about 3.5% of revenue. Microsoft is the only peer approaching Oracle’s capital intensity: $116bn of cash capital expenditure in the year to June 2026 -35% of revenue, and roughly $175bn once finance leases are included -yet it still produced $67bn of positive free cash flow. Same spending magnitude; opposite cash sign. Microsoft is spending a third of a much larger revenue base; Oracle is spending more than four-fifths of its own. That is what isolates the Oracle-specific risk.

Fiscal years and backlog definitions differ (Oracle May; Microsoft June; SAP, ServiceNow and IBM December; Salesforce, Workday and Snowflake January); half-year figures are six months to the date shown; Workday is the July quarter only; the comparison is indicative, not exact. Percentages are Bloor arithmetic on reported figures. n/d -not disclosed in the cited release; never treated as zero. Sources: each company’s 10-K, 8-K earnings release or half-year report as listed in Sources & basis.

The Rule of 40 makes the divergence numeric. On a cash basis -revenue growth plus free-cash-flow margin -Oracle scores about −18 (17 minus 35), failing dramatically. On a non-GAAP operating-margin basis it scores about 60 (17 plus 43), passing comfortably. SAP, Salesforce, ServiceNow and Snowflake pass either way; Workday passes on its guided full-year cash flow (13 plus 30) and only narrowly on the July quarter alone. Microsoft, for the first time, does not quite pass on cash: 18 plus 20 is 38. That a single company can both pass and fail the same test, depending only on which line you read, is the ‘EBITDA lies’ phenomenon quantified.

6 · The peers’ own disguises

The asset-light peers do not carry compute debt. They carry other dialects of the same lie, and each is worth naming because each is exactly the kind of below-the-line item the opening paper catalogued.

Salesforce borrowed $25bn in the first half of its fiscal year to fund an accelerated share repurchase. Debt rose from $14.4bn to $39.3bn; quarterly interest expense went from $67m to $473m; full-year free-cash-flow growth guidance came down to 4–5%. Nothing about the operating business changed. The financing line did -and EBITDA cannot see it.

ServiceNow’s non-GAAP operating margin in the June quarter was 29.5%. Its GAAP operating margin was 4%. The gap is stock-based compensation and the amortisation of intangibles from three acquisitions -Moveworks, Veza and $7.5bn for Armis -that took the company from net cash to a modest net-debt position in six months. Deal-amortisation debt, in the opening paper’s taxonomy, hiding in plain sight.

Snowflake guides to an adjusted free-cash-flow margin of 23% and reported a non-GAAP operating margin of 15%; on a GAAP basis it lost $589m in six months, because stock-based compensation ran at 30% of revenue. That is labour debt settled in equity rather than cash, and excluded from every headline metric. Workday’s July quarter shows a milder version of the Salesforce pattern: free cash flow fell 22% year on year to $460m while $1.3bn went to buybacks, and cash and securities have come down from $8.0bn at January 2025 to $3.4bn -capital returned to shareholders, defensibly, but a balance sheet thinner than the margin headline suggests. IBM spent $10.5bn on acquisitions in the first half, taking net debt to around $54bn while free cash flow stayed flat at $4.8bn. SAP is the control case: capital expenditure at 1.8% of revenue, free cash flow at 32%, net liquidity of €2.9bn, and no adjustment large enough to change the story.

7 · Why EBITDA lies now: depreciation is not ‘non-cash noise’

The defence of EBITDA has always been that depreciation is non-cash and backward-looking -the money left the building when the asset was bought, so adding it back reveals true operating performance. That logic holds for a factory expected to run for twenty years. It collapses for a fleet of GPUs (graphics processing units) that loses its economic edge in two or three. When the asset must be continually replaced simply to stay competitive, depreciation stops being a historical artefact and becomes a recurring reinvestment requirement -a forward cost as real as payroll. What looks like one-off growth capex is, in substance, permanent sustaining capex.

Treating adjusted EBITDA as earnings for a fast-obsolescing compute fleet is like valuing a steel mill while pretending the furnaces never wear out.

The size of the hidden charge turns on one discretionary assumption: assumed useful life. Shorten it and depreciation rises; lengthen it and reported profit flatters. Goldman Sachs’ sensitivity puts the industry-wide stake in the trillions: shortening assumed silicon life from five years to three lifts cumulative 2026–2031 depreciation from roughly $3 trillion to roughly $4 trillion; lengthening it to seven years cuts it to about $2.2 trillion. One assumption; a swing of close to $1.8 trillion. There is a credible counter-argument -older chips cascade to less demanding workloads and retain resale value -so the honest conclusion is not that one side is right. It is that the true charge sits within a wide, consequential range, and EBITDA hides the entire range from view.

The hyperscalers have shown, in the same twelve months, which way the lever moves in practice. In July Microsoft extended the accounting life of its data-centre and office buildings from 15 to 25 years and said more future data-centre leases would be classified as operating rather than finance leases -a change that trims reported capital expenditure and depreciation while, in its CFO’s words, leaving spending plans unchanged in economic terms. Amazon went the other way in 2025, shortening server lives and absorbing the charge. Same technology; opposite accounting; and only one of them shows up in EBITDA.

CoreWeave is the same story in miniature, and its June quarter sharpened it: a 59% adjusted-EBITDA margin on $2.6bn of revenue -and a $626m net loss, because $1.4bn of depreciation and $640m of net interest sat between the two lines. Its 2026 capital-expenditure guidance is $35–39bn against revenue guidance of $12.4–13.2bn: nearly three dollars of capital for every dollar of sales.

8 · From labour debt to compute debt

Underneath the accounting sits a structural shift that deserves its own name. The classic software business carried what the opening paper called labour debt: its dominant cost was people, and people are an operating expense that sits fully inside EBITDA. When such a company grew, its costs and its reported operating profit moved together, and EBITDA tracked economic reality reasonably well.

The AI build-out swaps labour debt for compute debt. The dominant cost is no longer headcount but capital assets -data centres, chips, power -and the financing behind them. That cost reaches the income statement not as operating expense but as depreciation and interest, both below the EBITDA line. Oracle’s FY2026 is the swap in a single frame: roughly 21,000 jobs removed even as net property, plant and equipment rose from $43.5bn to $100.0bn, depreciation nearly doubled and interest expense rose 29% to $4.6bn. A further round of reductions was reported in August ahead of the second fiscal quarter, which Oracle has not confirmed. Cost did not disappear. It migrated -out of the line EBITDA captures and into the lines it ignores.

9 · The liability the balance sheet does not yet show

The migration is larger than the balance sheet admits. Oracle’s FY2026 10-K discloses $260bn of additional lease commitments -substantially all data centres -that were not on the balance sheet at 31 May 2026, expected to commence between the first quarter of FY2027 and FY2029 for terms of fifteen to nineteen years. Set the $260bn against $137bn of on-balance-sheet borrowings and finance leases and $30bn of operating-lease liabilities, and the gross forward obligation reaches roughly $427bn -against a market capitalisation of about $445bn at the 3 September close. Net of $31bn of year-end cash, the obligation is still close to nine-tenths of the equity value.

That is compute debt made explicit. These leases are contractually senior, long-dated and largely uncancellable once commenced, yet they appear in no reported debt figure and no EBITDA multiple. They are the item a rival CFO has not yet computed -and the reason public financials alone will not price this business correctly. Layer on the concentration -Bank of America estimates that more than half of the $638bn backlog is a single counterparty, and S&P named that dependence in its downgrade -and the asymmetry is stark: Oracle is committing to two decades of fixed lease cost against a backlog that one renegotiation could reshape.

Oracle is not alone in the structure, only in its proportion. Alphabet’s June-quarter filing shows $85.2bn of data-centre leases signed but not yet recorded, alongside a debt balance that has passed $100bn and a $49.6bn equity raise in June. The difference is scale of cover: Alphabet’s operating cash flow was $39bn in a single quarter, and even so its free cash flow went negative for the first time since its 2004 listing.

10 · What CFOs and boards should track instead

The remedy is not to abandon EBITDA but to refuse to read it alone. A depreciation-aware dashboard restores what EBITDA removes. Applied to Oracle’s own FY2026 numbers, it reads as follows.

-Free cash flow and its conversion -the hardest number to flatter. Oracle: −$23.7bn, or −139% of net income.

-Capex as a share of revenue and of operating cash flow, and the moment the second exceeds the first. Oracle: 83% and 174%. FY2026 was the year of the crossing.

-Depreciation as a percentage of capex, so a falling ratio reads as the warning it is. Oracle: 14%. A low ratio is not efficiency; it is a wave that has not yet landed.

-Assumed useful life, benchmarked against the real refresh cadence -and disclosed, so that a change in it is visible when it happens.

-Gross debt to EBITDA alongside interest cover. Oracle: about 4.6× on borrowings and finance leases (roughly 3.5× net of cash); interest expense of $4.6bn, up 29%, against derived EBITDA of $29.9bn -cover of about 6.5×, and falling.

-Off-balance-sheet lease commitments counted as senior forward debt. Oracle: $260bn.

-Backlog-to-cash conversion -the bridge between a $638bn promise and the cash that validates the capital spent to serve it. Oracle: 12% of RPO, about $77bn, due within twelve months against $90bn of guided FY2027 revenue.

Used together, these turn EBITDA from a headline into a footnote.

11· The Bloor lens

For finance leaders living through the build-out, the work is concrete: construct the depreciation-aware dashboard; stress-test useful-life assumptions against the chip cycle; model exactly when free cash flow crosses zero and how the gap is funded; quantify the off-balance-sheet lease book as senior forward debt; and, on that basis, tell a credible capital-structure story to debt and equity markets before the depreciation wave -arriving in force from 2027 -tells it for them. Oracle’s July shows what happens when the market tells it first.

EBITDA is not dead. But in the AI-infrastructure era it has become the number that flatters most exactly when it matters most. The companies that keep the market’s confidence will be those whose CFOs get ahead of the metric -reframing the conversation around cash, capital intensity and the honest economic life of the assets -rather than letting a single, increasingly misleading line do the talking.

Watch for the When EBITDA Lies industry papers as they publish. For boards that want an early diagnostic read of their sector’s honest gauges before their industry paper lands, the Bloor Finance Desk is available for a confidential conversation.

12 · Sources & basis

Oracle FY2026 figures are drawn from the audited Form 10-K (filed 22 June 2026) and the Q4/FY2026 earnings release and call (10 June 2026). Peer figures are from each company’s most recent annual or half-year filings to 3 September 2026: Microsoft FY2026 8-K and call (29 July 2026); SAP Half-Year Report 2026 (23 July 2026); ServiceNow Q2 2026 8-K (22 July 2026); IBM Q2 2026 release (22 July 2026); Salesforce Q2 FY2027 release (26 August 2026); Workday Q2 FY2027 release (27 August 2026) and FY2026 annual report; Snowflake Q2 FY2027 8-K (2 September 2026); CoreWeave Q2 2026 release and call (11 August 2026); Alphabet Q2 2026 results and Form 10-Q (22 July 2026). Rating actions are per S&P Global Ratings (9 July 2026) and Moody’s; concentration estimates (Bank of America) and useful-life sensitivities (Goldman Sachs) are attributed and directional; the reported further round of Oracle job reductions is unconfirmed. Share-price and market-capitalisation figures are as at the 3 September 2026 close. Fiscal years and backlog definitions differ and are noted. Derived EBITDA, ratios and Rule-of-40 scores are Bloor arithmetic on reported figures. A full source note listing the origin, pinpoint and status of every figure accompanies this paper. This is analysis for a CFO audience and is not investment advice.