When EBITDA Lies
The AI-Embedded Organisation
Every sector is crossing the same line. The questions your metrics are not asking.
A cross-sector CFO diagnostic on the fiscal challenge of embedding AI — where the cost goes when it leaves the wage bill, and why the number on the cover of the results no longer sees it. This paper asks the questions. It does not answer them.
The line every sector is crossing
This is the opening paper in a Bloor Finance series, and it makes a single, uncomfortable claim: the AI transition is not a technology story with financial side-effects; it is a financial story wearing a technology costume, and it is playing in every sector at once. Wherever an organisation embeds AI, the same thing happens to its P&L. Work that used to be done by people — a cost EBITDA captures cleanly — is replaced by compute, models, automation and the financing behind them, costs that reach the income statement as depreciation, interest, lease charges and cash provisions, all below the line the metric draws. The disguise changes by industry. The migration does not.
That migration deserves precise language, and this series gives it names. Cost is moving out of wages, which sit inside EBITDA, and into depreciation, interest and leases, which sit below it — and partly off the balance sheet altogether. Where the new cost is capital, chips, power and the financing behind them, we call it compute debt. Where it is the fixed payroll of roles the operating model no longer needs, we call it labour debt. Where it is the estate that IFRS 16 quietly holds up, we call it lease debt. And the stack of accruals, provisions and charges that never touches EBITDA yet settles in cash, we call the Ghost Liability. This paper maps where each one hides, sector by sector. The full case for each sector — the company-level peer analysis, the scorecards, the reset economics — is the subject of the industry papers that follow.
Three accounting facts make the move almost invisible. First, IFRS 16 put operating leases onto the balance sheet as a right-of-use asset and a lease liability, splitting rent into depreciation and interest — mechanically raising reported EBITDA even as economic cost is unchanged. Second, capex-driven depreciation from the AI build-out lands below the line: the cash left the building when the asset was bought, but the recurring reinvestment it implies is real and rising. Third, the Ghost Liability — accruals, provisions, redundancy charges, tax and social-security payables — never touches EBITDA yet settles, relentlessly, in cash.
EBITDA flatters most precisely when it matters most. In the AI transition, every sector is learning its own dialect of that same lie.
The scale is now a governance question, not a curiosity. Goldman Sachs models several trillion dollars of AI-infrastructure capital spending this decade and shows that one discretionary assumption — whether a GPU lasts three years or five — swings cumulative depreciation by well over a trillion dollars, a number that appears on no earnings call. Epoch AI shows aggregate hyperscaler capital spending growing roughly three times faster than the cash generated to fund it, with free cash flow crossing zero around the third quarter of 2026. Moody’s finds hundreds of billions of lease commitments not yet recognised as debt. And the return is elusive: an MIT study of enterprise deployments found the overwhelming majority of generative-AI pilots delivered no measurable profit impact, even as the FinOps community reports that almost every organisation now carries AI spend it struggles to see, name or attribute. Meanwhile the outplacement firm Challenger, Gray & Christmas records tens of thousands of job cuts attributed to AI — and cautions, in the same breath, that “AI” is often the acceptable public reason for an old-fashioned cost problem. That caution is the whole series in one line: the headline inverts, and the truth sits below it.
What follows is a tour of that below-the-line world, sector by sector. Each section names how the standard metric misleads there, which hidden debt it conceals, and the honest gauge that would tell the truth instead. Each closes with the question the sector’s CFOs are not asking, and a marker for where Bloor will take it next. This paper surfaces the blind spot across the whole economy; the forthcoming papers open it up one sector at a time. We name the question. We do not hand over the model. Reform adjusts the past; Reset designs the future — and the reset is a conversation, not a download.
THE HIDDEN DEBT TAXONOMY
Ten terms for where the cost goes when it leaves the wage bill. Two are established Bloor usage; eight are named here for the first time. Definitions only — what each is, not what to do about it.
- Labour Debt. Fixed payroll committed to roles the operating model no longer needs.
- Ghost Liability. The stack of accruals, provisions, redundancy charges and payables that never touches EBITDA yet settles, relentlessly, in cash.
- Compute Debt. Capital, chips, power and the financing behind AI infrastructure, reaching the P&L as depreciation and interest rather than wages.
- Lease Debt. The IFRS 16 estate — rent reclassified as depreciation and interest, mechanically lifting reported EBITDA against unchanged economic cost.
- Deal-Amortisation Debt. M&A amortisation and impairment stripped out of “Core” earnings, obscuring whether acquisitions created or destroyed value.
- Model-Risk Debt. Unpriced exposure from AI-driven decisions — pricing bias, reserving inadequacy, governance liability not yet reflected in capital.
- Stranded-Capex Debt. Long-life capital committed against a demand forecast that may not hold, leaving stranded-cost risk today’s depreciation schedule assumes away.
- Content-Amortisation Debt. Cash content spend that outpaces the amortisation charge flowing through the P&L, letting adjusted profit rise while cash leaves.
- Maintenance-Backlog Debt. Deferred maintenance and asset-renewal obligations that grow as the income available to meet them shrinks.
- Maintenance-Backlog Debt. Deferred maintenance and asset-renewal obligations that grow as the income available to meet them shrinks.
- Restructuring Debt. Severance and asset-retirement charges that settle in cash but are routinely excluded from “adjusted” earnings.
1 · Technology & hyperscalers – the compute-debt epicentre
This is where the lie is loudest, because the numbers are largest. The four biggest hyperscalers plan on the order of $725bn of capital spending in 2026, up roughly three-quarters on 2025. Record revenue and record backlog coexist with collapsing free cash flow. Oracle is the clean case: FY2026 delivered record revenue and a contracted backlog of $638bn (up 363%) — yet capital spending of $55.7bn turned that into free cash flow of about −$23.7bn, funded by tens of billions of fresh debt and equity, with more financing flagged for FY2027.
CoreWeave is the same story in miniature: a recent quarter paired a headline 56% adjusted-EBITDA margin with a net loss, because depreciation and interest between them absorbed almost the entire revenue line. The hidden liability is compute debt in its purest form — and the amplifier is the useful-life assumption. Depreciate a GPU over six years while a new architecture ships every twelve months, and reported operating profit becomes an accounting choice rather than an economic fact.
THE QUESTION CFOs AREN’T ASKING
If your GPUs are written down over six years but obsolesced in two, what is your true maintenance capex – and what happens to reported operating income the day the useful-life assumption resets? When exactly does your free-cash-flow line cross below your capex line, and what is the funding plan on the far side of that crossing?
→ Technology & hyperscalers industry paper + peer scorecard to follow.
2 · Banking & financial services – the category error
For a bank, EBITDA is not distorted – it is meaningless. Interest is revenue and cost, not something to add back; depreciation is trivial against the balance sheet. The honest gauges are return on tangible equity against the cost of equity, the cost:income ratio and jaws, net interest margin and the structural hedge, and CET1 capital generation converting into distributions. Here the AI story is a labour-debt story dressed as transformation. Barclays lifted its RoTE ambition above 14% and committed to more than £15bn of distributions, citing AI in its efficiency case; Lloyds is opening more than a thousand AI-related roles while closing branches and banking tens of millions in generative-AI “value”; peers across the sector are confirming thousands of support-function cuts. The question is not whether AI removes cost. It is whether the removed cost is really gone, or merely reinvested and renamed.
THE QUESTION CFOs AREN’T ASKING
How much of your wage bill is labour debt – fixed payroll for processed, transactional work whose demand has already left? If AI genuinely takes a third out of servicing cost, does that show up as a lower cost:income ratio and positive jaws within two reporting periods — or does it quietly disappear into reinvestment? And where, exactly, does your model-risk and AI-governance provision sit?
→ Banking industry paper + peer scorecard to follow.
3 · Pharma & life sciences – the adjustment that hides the answer
Pharma’s distortion is self-inflicted and respectable. “Core” or “adjusted” earnings strip out precisely the M&A amortisation and impairment that reveal whether serial dealmaking created or destroyed value. The majors say so themselves: guidance is given on a Core basis, and companies openly note they cannot reliably forecast the Reported number because they cannot forecast the impairments. Yet 2025 supplied plenty – a £471m write-off on one terminated GSK programme; multi-billion-dollar research impairments elsewhere; sector R&D spending falling even as headline pipelines expand. Layer AI-driven drug discovery on top and a new question appears: is that spend research expense, capitalised asset, or an off-balance-sheet commitment to a discovery platform – and over what life is it carried?
THE QUESTION CFOs AREN’T ASKING
Over a full cycle, does the amortisation and impairment you exclude from Core roughly equal the acquisition premiums you paid – that is, has your serial dealmaking created value or merely recycled it? And what is AI drug-discovery actually costing you, on which line, over which useful life?
→ Pharma & life sciences industry paper + peer scorecard to follow.
4 · Professional services, recruitment & consulting – labour debt in the open
In a people-business, revenue is a vanity number; net fee income is the honest top line and the conversion ratio – operating profit ÷ NFI – is the honest margin. Robert Walters is the exhibit: NFI down 15% and the group into an operating loss, with the IFRS 16 estate quietly holding EBITDA up as the business broke beneath it. The signal event, though, is Accenture – an $865m business-optimisation programme and thousands of roles removed “where reskilling is not a viable path,” language that concedes the point: this is labour debt being written off, not a market being served. When the archetypal AI-transformation adviser is itself restructuring around AI, the demand-side question lands hard on every services CFO.
THE QUESTION CFOs AREN’T ASKING
What is your conversion ratio — and what has it done over three years, not one? How much of your fixed fee-earner payroll is labour debt for work a machine now does? And what is your IFRS 16 lease liability doing to EBITDA while your NFI trend does the opposite?
→ Professional services industry paper + peer scorecard to follow.
5 · Retail & consumer – the lease-debt engine
Thin margins and lease-heavy estates make retail the sector where IFRS 16 flatters EBITDA most – adoption lifted reported assets across the sector by double digits, and every point of rent reclassified as depreciation-and-interest is a point EBITDA no longer sees. On top sits a labour-intensive operation carrying its own labour debt, and an AI-personalisation and supply-chain build whose cost hides as generic compute on the cloud bill. The automation numbers are real – Amazon now runs more than a million warehouse robots and models avoiding six-figure future hiring – but the attribution is where the headline inverts: analysts already forecast that half the firms blaming headcount cuts on AI will later admit other causes. Strip the estate out and the honest question is stark.
THE QUESTION CFOs AREN’T ASKING
Strip IFRS 16 out entirely – what is your EBITDA really, and what is your lease-adjusted net debt? Is your AI-personalisation spend generating attributable incremental margin, or is it ungoverned compute on the cloud bill with no measurable P&L impact – the very outcome the MIT study found in the overwhelming majority of cases?
→ Retail & consumer industry paper + peer scorecard to follow.
6 · Industrials & manufacturing – short-life assets, long-life schedules
Capital-heavy and automation-led, industrials face the GPU problem in a factory setting: AI-optimisation and robotics assets with genuinely short economic lives are frequently depreciated on the long-life schedules built for plant and buildings. The gap flatters current operating profit and defers a reinvestment wave that is as real as payroll. Restructuring is already running — sizeable cuts announced across industrial-goods and chemicals, with automation, tariffs and AI variously cited — which raises the same attribution question the series keeps returning to: how much is genuine operating leverage, and how much is a cost problem wearing an AI label?
THE QUESTION CFOs AREN’T ASKING
Are your predictive-maintenance, vision and factory-optimisation assets depreciated over their true economic life, or over the plant’s? If you shortened those lives to match reality, would this year’s operating margin survive the change?
→ Industrials industry paper + peer scorecard to follow.
7 · Insurance – the liability the loss ratio doesn’t show
Start with what’s not priced. AI underwriting is quietly manufacturing a liability most insurers’ capital bases do not yet reflect: model risk, pricing bias and reserving inadequacy accumulating beneath a loss ratio that looks like it’s improving. High-risk AI underwriting also carries compliance obligations from August 2026 — a cost line that, for most carriers, does not yet exist as a line at all.
EBITDA is largely irrelevant to an insurer; the combined ratio is the truth-teller, and AI is genuinely moving it — analytics-led carriers are reporting materially better loss ratios than slower adopters. That improvement is real. What’s being missed is that a better loss ratio bought with AI underwriting can be financing an unpriced liability at the same time — the honest reading nets the model-risk cost against the loss-ratio gain, rather than banking the gain alone.
THE QUESTION CFOs AREN’T ASKING
If AI is improving your loss ratio, is it simultaneously creating a model-risk, bias and reserving liability your capital and governance do not yet price? And what is your AI-Act compliance cost – a line item, or a surprise?
→ Insurance industry paper + peer scorecard to follow.
8 · Energy & utilities – the picks-and-shovels capex trap
Start with the exposure. Twenty-to-thirty-year capital is being committed against a data-centre load forecast that depends entirely on hyperscaler capex holding — the same hyperscaler capex whose free-cash-flow crossover this paper has already flagged. If the AI build plateaus, or GPU useful lives reset and the fleet shrinks, the rate base carries stranded-cost and recovery risk that today’s depreciation schedule assumes away, and that risk is not yet being asked about at board level.
Utilities are cast as the picks-and-shovels of the AI build-out — and the casting hides the risk. Aggregate sector capital plans now run well over a trillion dollars this decade, underwritten on a forecast that data-centre electricity demand roughly doubles by 2030. That is the scale of the bet; the exposure above is what most rate-base planning has not yet stress-tested.
THE QUESTION CFOs AREN’T ASKING
If hyperscaler capital spending plateaus, or GPU useful lives reset and demand softens, what happens to the load forecast underpinning your rate base – and who absorbs the stranded cost, your customers or your equity?
→ Energy & utilities industry paper + peer scorecard to follow.
9 · Telecoms & media – content amortisation and the labour cut
Two migrations run side by side here. Telecoms are cutting labour debt hard – carriers announcing their largest-ever reductions, booking large severance charges, and trimming capex while attributing the shift to AI in customer operations. Media is where the accounting hides the cash: content amortisation divorces reported profit from cash content spend. At more than one major studio, the content amortisation and impairment charge for 2025 exceeded the segment’s entire adjusted EBITDA; at the streaming leaders, capitalised cash content spend runs billions above the amortisation flowing through the P&L. Adjusted earnings can rise while cash goes out the door.
THE QUESTION CFOs AREN’T ASKING “
For the telco CFO: how much of this year’s saving is real operating leverage, and how much is a one-off severance charge that hits cash but not “adjusted” earnings? For the media CFO: does your adjusted EBITDA bear any stable relationship to the cash you spend on content – or does capitalisation let you report profit while the cash leaves?
→ Telecoms and media industry paper + peer scorecard to follow.
10 · Healthcare providers – AI on an inflexible base
Healthcare provision is labour-intensive, capital-equipment-heavy and reimbursement-constrained – and 2026 opened with the sharpest provider job-cut readings in years, driven by inflation, high labour cost and reimbursement pressure. Onto that fixed, often unionised base, AI diagnostics and admin automation add capital-equipment depreciation and model-governance compliance cost that are easy to approve and hard to unwind. The honest question is whether AI reduces cost-to-serve, or simply layers new fixed cost on an old fixed cost.
THE QUESTION CFOs AREN’T ASKING
Is AI genuinely lowering your cost-to-serve, or adding capital and compliance cost on top of a labour base you cannot flex? Over what life is the AI-enabled capital equipment carried – and does that life match its clinical and technical obsolescence?
→ Healthcare providers industry paper + peer scorecard to follow.
11 · Public sector, charity & non-commercial — where EBITDA is N/A
Here EBITDA does not apply at all; the honest gauges are reserves, restricted versus unrestricted income, and the maintenance and asset-backlog liability. The Canal & River Trust is the reference case — a shrinking government grant set against ten thousand ageing assets and tens of millions of unavoidable annual maintenance, an obligation that grows precisely as the income to meet it falls. AI-adoption pressure lands on a fixed, constrained budget with no equity buffer to absorb it. For the public and third sectors, the below-the-line liability is not depreciation; it is the backlog — and it is already on the books.
The disguise here is unusually complete, because there is no P&L for the backlog to distort in the first place. A commercial CFO at least has a metric that lies; a public-sector or charity finance director often has no equivalent single number the board is watching, so the maintenance-backlog liability simply doesn’t appear anywhere a governance committee would naturally look. Restricted funding adds a second layer: AI-adoption cost is rarely a purpose a donor or grantor has restricted funds toward, which means it either competes directly with the mission activity the funding was raised for, or it doesn’t happen. Both outcomes are currently invisible to the same reserves reporting that would catch a straightforward income shortfall
THE QUESTION CFOs AREN’T ASKING
With a shrinking grant and a growing maintenance-backlog liability, where does AI-adoption cost come from without eroding restricted reserves – and is the backlog you carry counted as the liability it is, or deferred as a hope?
→ Public & non-commercial industry paper + peer scorecard to follow.
12 · Transport & logistics – the cash the “adjusted” number hides
Asset-heavy and labour-heavy, logistics is automating fast — one major carrier eliminated tens of thousands of roles, closed scores of facilities and guided to tens of thousands more cuts, alongside asset-retirement charges on ageing fleet. The distortion is familiar: severance and asset-retirement provisions settle in cash but are excluded from “adjusted” figures, so the headline improves while the balance sheet pays. The question is how much of the automation dividend is structural, and how much is a one-off charge relabelled as progress.
The sector carries a second, quieter version of the same distortion. Automation capex here is split across warehouse robotics, route-optimisation software and fleet telematics — three asset classes with genuinely different economic lives, frequently collapsed into a single depreciation schedule for reporting simplicity. That collapse hides which part of the automation spend is actually paying back and which part is being carried at a life it will not reach. Combined with restructuring charges that repeat year over year rather than landing as a genuine one-off, the “adjusted” trend can show durable margin improvement for several consecutive periods while the underlying capital base is quietly being replaced faster than its own books admit.
THE QUESTION CFOs AREN’T ASKING
How much of your automation saving is durable operating leverage, and how much is a severance-and-asset-retirement charge that hits cash this year but never touches adjusted earnings? Put differently: is the saving realised, or merely announced?
→ Transport & logistics industry paper + peer scorecard to follow.
13 · One table, every sector – the same line, twelve disguises\
The pattern is not twelve problems; it is one problem in twelve costumes. The metric on the cover of the results conceals a below-the-line liability, and a sector-honest gauge would reveal it.
| Sector | Where the metric lies | The hidden debt | The honest gauge |
|---|---|---|---|
| Technology & hyperscalers | Record revenue and backlog mask deeply negative free cash flow | Compute debt | FCF vs capex; depreciation ÷ capex; useful-life |
| Banking | EBITDA is undefined — a category error for a bank | Labour debt | RoTE vs CoE; cost:income & jaws; CET1 generation |
| Pharma & life sciences | “Core/adjusted” strips out the amortisation and impairment that grade the deals | Deal-amortisation debt | Reported (not Core) over a cycle; cash conversion |
| Professional services | Revenue flatters; fixed fee-earner payroll stays after the demand goes | Labour + lease debt | Conversion (OP ÷ NFI); FCF after leases |
| Retail & consumer | IFRS 16 lifts EBITDA while the estate and the wage bill stay fixed | Lease + labour debt | Lease-adjusted net debt; FCF after leases |
| Industrials | Short-life automation depreciated on long-life schedules | Compute debt | True-economic-life depreciation; unit cost |
| Insurance | EBITDA is meaningless; model risk sits off the page | Model-risk debt | Combined ratio; reserve adequacy; AI-governance cost |
| Energy & utilities | 20–30-yr capex underwritten on a data-centre load forecast | Stranded-capex debt | Rate-base risk; load-forecast sensitivity |
| Telecoms & media | Content amortisation and “adjusted” figures divorce profit from cash | Content-amortisation debt | Cash content spend vs amortisation; FCF |
| Healthcare providers | AI capital and compliance land on an inflexible labour base | Labour + compliance debt | Cost-to-serve; capital-equipment life |
| Public & charity | EBITDA is N/A; the backlog liability is invisible | Maintenance-backlog debt | Reserves; restricted vs unrestricted; backlog |
| Transport & logistics | Severance and asset-retirement charges hit cash, not “adjusted” | Restructuring + labour debt | Cash cost vs structural saving realized |
Honest gauges: RoTE – return on tangible equity; CoE – cost of equity; NFI – net fee income; FCF — free cash flow; CET1 — common equity tier-1 capital. Sector characterisations are analytical, not exhaustive; figures throughout are drawn from the most recent company disclosures and named third-party analysis, and forward figures are directional.
14 · The Bloor lens – and the questions that remain
One migration, every sector: cost leaving the wage bill EBITDA can see, and reappearing as depreciation, interest, leases and cash provisions it cannot. The organisations that keep the market’s confidence through the AI transition will not be those that cut hardest and announce loudest. They will be the ones whose CFOs read the honest gauge for their sector, count the hidden debt in full, and redesign the cost base rather than merely relocating it – because AI done carelessly does not retire labour debt and lease debt, it mints compute debt in their place. That is the difference between efficiency theatre and real financial engineering. It is the difference between Reform and Reset.
THE BLOOR TEST
EBITDA is not dead — but in the AI-embedded organisation it flatters most exactly when it matters most. The winners will get ahead of the metric: reframing the story around cash, capital intensity, conversion and the honest economic life of the assets, before the below-the-line wave tells the story for them.
Reform adjusts the past. Reset designs the future.
This paper has asked the questions on purpose. The answers — the depreciation-aware dashboard, the conversion diagnostic, the labour-, lease- and compute-debt map for your numbers, and the reset that pays the invisible balance sheet down — are the work that follows: a programme of dedicated industry papers, company-level peer analyses and scorecards, sector by sector, and a Finance Desk consultation for boards that cannot wait for their sector’s turn. If your board is being asked to accept a restructuring charge as strategy, or to treat a below-the-line liability as though it were free, that is the conversation to have first.
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.
15 · Sources & basis
All figures are drawn from the most recent annual and interim results, investor guidance and regulatory filings available to July 2026, supplemented by named third-party analysis; forward-looking and estimated figures are directional and are labelled as such. Company examples include Oracle, CoreWeave, Amazon, Microsoft, Alphabet and Meta (technology); Barclays, Lloyds and NatWest (banking); GSK, Roche and AstraZeneca (pharma); Robert Walters, Hays, PageGroup and Accenture (professional services); and named carriers, studios, utilities, insurers and logistics operators across the remaining sectors. Cross-cutting evidence is attributed in text to Goldman Sachs (GPU useful-life sensitivity), Epoch AI (operating-cash-flow-versus-capex crossover), Moody’s (unrecognised lease commitments), MIT’s State of AI in Business 2025 (pilot outcomes), the FinOps Foundation and CloudZero (AI-cost governance), Challenger, Gray & Christmas (AI-attributed layoffs), and the IFRS 16 standard on lease accounting. Attribution of job reductions to AI reflects employer statements, not established causation — a distinction this series treats as material. This is analysis for a CFO audience and is not investment advice.
