The economics of AI, on the record

01 · The problem with the evidence

One filing, two black boxes

The question everyone is asking about artificial intelligence is financial, not technical: is the capital being deployed into AI infrastructure going to earn a return? It is a reasonable question. It is also, for the most part, being argued with numbers that no auditor has ever seen.

Consider what is actually available. OpenAI is valued at $852 billion. Anthropic is reported at around $965 billion. Between them they define the commercial frontier of the industry. Neither has ever filed a financial statement with the U.S. Securities and Exchange Commission. A search of EDGAR, the SEC's filing database, returns no registrant under either name — not zero registration statements, but no corporate entity at all. Every revenue figure, every loss figure, every margin estimate you have read about either company originates in a leak, a briefing, an investor update, or a model built by someone outside the company.

Then, in June 2026, one exception appeared.

SpaceX priced its initial public offering on 12 June 2026 at $135.00 per share and now trades on Nasdaq as SPCX. Because it had acquired xAI in February 2026, and xAI had earlier absorbed X Corp, the resulting company reports a segment called AI. That segment files. It has an audited income statement, a disclosed capital expenditure line, a revenue disaggregation note, a useful-life table for its equipment, and related-party disclosures.

For the first time, one frontier-scale AI operation has to show its books.

02 · The evidence that is audited

What the one audited AI business shows

SpaceX reports three segments: Space, Connectivity (Starlink) and AI. The filing defines the third one plainly: the AI segment "includes our AI compute, Grok, and X." It was formed when SpaceX acquired xAI, effective 2 February 2026 — xAI having itself acquired X Holdings on 28 March 2025. Crucially, the filing accounts for that acquisition as a reorganisation of entities under common control — Elon Musk controlled both sides — and states plainly that "no new goodwill or other intangible assets have been recorded." Prior periods are retrospectively combined at historical carrying amounts.

The practical consequence is unusual and valuable: we do not get one quarter of an AI business. We get a restated multi-year history of one.

Source: SpaceX Form 424B4 (12 June 2026) and Form 10-Q for the period ended 30 June 2026. FY2024 depreciation and amortisation was not captured in our extraction of the segment note and is left blank rather than estimated.

Note the operating loss line before going further: it narrowed sharply in FY2024, to $1,561 million, before quadrupling to $6,355 million in FY2025. This is not a business steadily converging on breakeven, nor one steadily collapsing. It is one whose cost base changed shape when the capital expenditure began.

Read the revenue line alone and you would conclude that this AI business grew from $2,961 million to $3,201 million in two years — sluggish, for a segment absorbing that much capital. That reading would be wrong, and the filing itself tells you why. Under the revenue disaggregation required by ASC 606, the AI segment breaks into two very different businesses.

In FY2025, 57.6% of the "AI" segment's revenue was advertising — the X platform's advertising business, which is still below its 2023 level of $2,323 million. The actual AI business earned $1,357 million. Against $12,727 million of capital expenditure, that is 9.4 times capex per dollar of AI revenue, not the 3.98 times you get by dividing capex by the segment total.

This is the single most important thing in the filing, and it is not a press error. It is what happens when a segment total is quoted without the disaggregation note underneath it.

Now the other half, which is just as important and points the other way. The interim filing shows the composition inverting, fast.

AI revenue grew 7.1 times year on year in the second quarter, from $311 million to $2,194 million. The segment flipped from 58% advertising in FY2025 to 79% genuine AI revenue in the first half of 2026. Company backlog rose from $28,377 million at the end of December 2025 to $47,461 million at 30 June 2026 — a 67% increase in six months, of which 56% is expected to be recognised within one year. Advertising, meanwhile, kept shrinking.

Anyone arguing that demand for AI services is illusory has to explain a sevenfold revenue increase in an audited filing. That is not a survey, a forecast or a briefing. It is a number a Big Four auditor signed.

And yet.

The trajectory is worth stating plainly, because it is not a straight line. In FY2023 the segment spent $0.73 of capital per dollar of AI revenue. In FY2024 that became $6.32, in FY2025 $9.38, and in the first half of 2026 $8.82. Capital intensity did not creep up; it stepped up by an order of magnitude in a single year and has stayed there.

Both camps have to sit with that. The bear case that AI demand is fake does not survive a 7.1× revenue increase. The bull case that scale brings operating leverage does not survive a capital intensity that has barely moved while revenue quadrupled.

03 · The analogy, tested

The SpaceX analogy, tested against SpaceX

SpaceX has spent a decade as the standing rebuttal to capital-expenditure scepticism. The argument is familiar: serious people said reusable rockets were a fantasy and satellite broadband was a money pit, and they were wrong, so today's sceptics of AI infrastructure are probably wrong too. Neil Armstrong and Gene Cernan testified to Congress against commercial spaceflight. An Arianespace executive called reusability "a dream." The company nearly went bankrupt in 2008 after three consecutive Falcon 1 failures, saved by a $1.6 billion NASA cargo contract.

The analogy is now testable, because the company making it files financial statements. And on its own numbers, it does not support the use it is put to.

The segment everyone points to as vindication — Starlink — runs at 0.37 times capital intensity and a 38.8% operating margin. It took roughly two decades to get there: SpaceX was founded in 2002, and Starlink reached cash-flow break-even around 2023 — a milestone stated by company executives rather than disclosed in a filing, since SpaceX was private at the time. The AI segment in the same company, in the same filing, runs at ten to twenty-five times that capital intensity depending on which revenue line you use, and loses money at scale.

You cannot invoke SpaceX to argue that heavy AI capital expenditure will come good, because SpaceX's own accounts say the part that came good looks nothing like the AI part. If anything the filing offers the opposite lesson: the buildout that worked was comparatively capital-light, took twenty years, and was carried by a company that nearly failed on the way.

Three further disclosures in the same documents deserve more attention than they have had.

A $13.3 billion related-party equipment financing

The filings disclose an equipment lease entered in October 2025 and amended in November 2025 with Valor Equity Partners — whose founder and chief investment officer, Antonio Gracias, sits on the SpaceX board — covering AI infrastructure hardware. The arrangement is accounted for as a failed sale-leaseback, which is the accounting conclusion that the transaction is in substance a financing rather than a sale. The recorded debt was $455 million current and $4,052 million non-current at 31 December 2025. Six months later, at 30 June 2026, it was $2,039 million current and $11,290 million non-current — roughly $13.3 billion. Interest expense on it was $327 million in the second quarter alone.

This matters beyond one company. The Bank for International Settlements warned in its March 2026 Quarterly Review about "shadow borrowing" — "obligations that are economically akin to debt but largely reside outside corporate balance sheets." Here is a named, audited, quantified instance of exactly that pattern — GPU hardware financed through a director-affiliated vehicle, at a scale that tripled in six months.

A take-or-pay cliff in 2027

A risk factor states: "We have non-cancellable, multi-year capacity commitments to cloud compute providers, requiring payment regardless of usage." The contractual commitments note quantifies them as of 31 December 2025 at $25,451 million in total — $2,720 million due in 2026, $21,476 million due in 2027, $1,250 million in 2028 and almost nothing thereafter. Eighty-four per cent of the obligation lands in a single year, and it is payable whether or not the capacity is used.

The risk factor is about power and water, not chips

Much of the bear case concerns chip obsolescence. It is therefore worth noting what the one company legally required to disclose its AI risks actually says. There is no GPU-obsolescence risk factor. The operative heading concerns "the availability of power, water, AI processors, and other critical components," and the first sentence reads: "Our ability to scale our data center infrastructure, which supports our AI segment, is increasingly constrained by the availability of power and water at economically feasible prices, long lead times, availability of materials, and changing regulatory requirements."

The binding constraint, according to the only filer who has to name one, is physical infrastructure. Not depreciation.

The filings also disclose customer concentration without naming the customers. "Customer B", whose revenue "relates to the AI segment", accounted for 19.5% of consolidated revenue in Q2 2026, against less than 10% a year earlier. A fifth of group revenue now depends on one unnamed AI counterparty.

For completeness: total debt rose from $22,049 million at 31 December 2025 to $38,433 million at 30 June 2026 — a near doubling in six months, during which the company also took in roughly $74.4 billion of IPO proceeds. Accumulated deficit stood at $41,852 million at 30 June 2026. Related-party dealings also include $506 million of Tesla Megapack purchases and $131 million of Cybertrucks in FY2025, a further $295 million of Megapacks in the second quarter of 2026, and $1,421 million of common stock purchased from current and former employees by Elon Musk's trust during 2025.

04 · The provenance problem

Everything else you have read is a leak

Outside that one filing, the evidential quality collapses. Here are the figures that drive the public argument, with what they actually are.

None of the first eight is dishonest. Several come from the companies themselves. But not one is an audited financial statement, and a confidential S-1 submission is by definition not publicly verifiable — you are being asked to accept a report that a document exists which you cannot read.

The reason this matters is not abstract. When we checked the most-quoted numbers in this debate against primary documents, four of them turned out to mean something materially different from the meaning they carry in circulation. Three overstate; one understates.

1. OpenAI's "$39 billion loss" is mostly a non-cash accounting charge

OpenAI's FY2025 GAAP net loss of roughly $38.5–39 billion is real and was verified by the Financial Times from documents. But it includes a one-time non-cash charge of approximately $41.5 billion arising from the conversion of the non-profit into a public benefit corporation. FT reporting puts the underlying cash loss at roughly $8 billion. The headline figure is routinely deployed as evidence of catastrophic burn. It is not a burn figure. The separately reported cash burn — $3.7 billion in the first quarter of 2026 — is the number that speaks to burn, and it is large enough without help.

There is an irony worth recording. The documents the FT verified were obtained by Ed Zitron, the most prominent bear in this debate. The correction to the bear case's most-quoted number came from the bear case's own reporting.

2. Nvidia's "$105 billion for OpenAI" is not an investment

In August 2026 Nvidia was reported as backing $105 billion of financing for an OpenAI data centre in Pike County, Ohio. This is a contingent residual-value guarantee supporting lease obligations, disclosed in an SEC filing. It is not a cash outlay, and it is not equity. Nvidia's actual cash investment in OpenAI is $1.5 billion. The headline overstates the cash committed by roughly seventy times. Whether a $105 billion contingent guarantee is more or less alarming than a $105 billion investment is a fair question — but it is a different question, and it cannot be asked while the two are conflated.

3. The SpaceX AI segment's revenue is 58% advertising

Covered above. Quoting the segment total of $3,201 million as AI revenue overstates the AI business by 2.4 times for FY2025. Unlike the others, this one is nobody's error but the reader's: the disaggregation is right there in the note.

4. Microsoft did not cut its capital expenditure — it reclassified it

This one runs the other way. Microsoft's capital expenditure guidance for calendar 2026 moved from roughly $190 billion to roughly $175 billion, which was widely read as a pullback. It was not. The company extended the useful life of data-centre and office buildings — not servers — from 15 to 25 years effective FY2027, alongside a reclassification of certain finance leases to operating leases. The mechanical effect is to move spending out of the reported capex line. The physical buildout did not shrink by $15 billion. Anyone reading that as evidence of hyperscaler retrenchment has misread an accounting change.

One further note on the private-capital machinery. OpenAI has publicly warned that investors should be "extremely cautious of any firm that purports to have access to OpenAI equity, including through an SPV," because it runs no official secondary programme. Meanwhile, retail vehicles offering indirect exposure to these companies — the Destiny Tech100 closed-end fund among them — have traded at large premiums to their own stated net asset value. Price discovery in this asset class is not merely opaque; parts of it are actively distorted.

05 · The case against, at its strongest

The bear case, scored

The most systematic bear is Ed Zitron, arguing since "The Subprime AI Crisis" (September 2024) that the buildout requires end-customer revenue that does not exist, that the appearance of demand is manufactured by circular transactions in which suppliers fund their own customers, and that the unwind propagates through the credit that financed it. In "The AI Demand Bubble" (4 August 2026) he put the requirement at $2–3 trillion of annual AI revenue by 2030. The record is mixed — but the central structural claim has held up better than his critics allow.

What has been vindicated

Circular financing is now official-sector documented. The claim most often dismissed as conspiratorial is quantified in a BIS working paper: $46 billion of direct equity alongside $879 billion of multi-year compute purchase commitments linking chipmakers, hyperscalers, labs and neoclouds — compiled from SEC filings, not speculation. The structures are on the record: Nvidia's equity in OpenAI, CoreWeave and xAI; the AMD warrant giving OpenAI rights over roughly 10% of AMD, vesting per gigawatt deployed; Nvidia's take-or-pay obligation to buy CoreWeave's unsold capacity through April 2032; the residual-value guarantee described above. Money leaves a chipmaker, returns as a customer's purchase order, and is recognised as revenue.

The loss scale was not exaggerated. $3.7 billion of cash burn in a single quarter, at a time when the company was publicly framed as approaching sustainability.

Enterprise repricing happened. Priority tiers requiring upfront throughput guarantees, tightened free tiers at both OpenAI and Google, and long-context surcharges of roughly twice list price.

What has not survived

The Anthropic call was wrong by a wide margin. Zitron cited roughly $4 billion of annualised revenue against $3 billion of expected losses, as a business that could not grow into its costs. The reported run-rate went from about $9 billion at the end of 2025 to $65 billion by July 2026. Those are leaks rather than audits, but the direction is not in dispute, and it is not the direction the thesis required.

Capability stagnation did not hold. Kelsey Piper's April 2026 rebuttal in The Argument: "AI progress from 2024 to 2026 has been much faster than 2022 to 2024." She also fairly catches arguments advanced as "while I don't have proof, I would bet…" sitting alongside documented claims without the distinction being marked.

The collapse timing is unfalsified, not confirmed. Predictions that OpenAI exhausts cash by 2027 and triggers a cascade are forecasts. This article does not score forecasts before their test dates arrive.

06 · The case for, at its strongest

The bull case, scored

Stated at its strongest: demand is real, growing faster than any enterprise software category in history, and the capital is not a bet on a research breakthrough but a response to demand that already exists and cannot currently be served.

What is documented

Revenue growth is extraordinary, and in one case audited. The SpaceX AI segment's 7.1× year-on-year growth in Q2 2026 is in a filing. Anthropic's reported move from roughly $9 billion to $65 billion of run-rate in twelve months, and OpenAI's reported $40 billion, are leaks — but three independent trajectories pointing the same way is evidence, even when each is individually unaudited.

Prices for equivalent capability fell sharply. Anthropic's Opus tier went from $15 and $75 per million input and output tokens to $5 and $25 — a two-thirds cut at constant capability tier, published on the company's own pricing page. Not an estimate; a price list.

Inference overtook training. Gartner puts 2026 as the first year in which inference spend ($23.3 billion) exceeds training ($19 billion). The capital case has stopped depending on the next training run and started depending on serving demand that already exists.

Hardware efficiency is improving fast. MLPerf shows Blackwell at up to 4× Hopper's inference performance and roughly 2× the performance per dollar on large models. Nvidia's position that CUDA keeps the installed base productive well beyond its book life, with A100s "mission-capable" from 2020 into 2029, is corroborated by CoreWeave contracting A100 capacity into 2029.

Where it rests on estimate rather than disclosure

Margin is the weak link. Gavin Baker's 60–70% inference gross margin is an estimate, and it sits well above the alternatives: ICONIQ puts AI-native gross margins near 52% in 2026 against 41% in 2024, and Bessemer's February 2026 work puts them at 50–60% against 80–90% for traditional software. Nobody outside these companies knows, because none of them discloses it. For context on what margins software businesses normally carry, and how AI pricing models depart from them, see the SaaS pricing playbook. Marc Andreessen's framing that scaling laws make capital expenditure "a known function, not a research bet" is an assertion about the future stated with the confidence of a measurement.

Ben Thompson is the useful barometer: "we've obviously crossed the line into bubble territory" in November 2025, reversed to "I don't think we're in a bubble" by March 2026. That is what updating looks like, and a fair indicator of how far the evidence moved in nine months.

07 · The argument that misses

Depreciation is the wrong argument

The most technical front in this debate concerns how quickly GPUs should be written down. In November 2025 Michael Burry argued that hyperscalers understate depreciation by extending useful lives beyond economic reality, putting the cumulative overstatement at $176 billion across 2026–28, with Oracle's earnings overstated by about 27% and Meta's by about 21% by 2028. He called it "one of the more common frauds of the modern era."

The claim, as stated, does not survive contact with the filings.

The decisive row is the fourth. In January 2025 Amazon shortened the useful life of a subset of its servers, at a cost of $1.4 billion in additional depreciation and $1.0 billion in net income, explicitly "due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning." A company inflating earnings through depreciation policy does not voluntarily take a billion-dollar hit and explain why in its accounting note.

Christopher Tsai's rebuttal makes the rest of the case: the extensions largely occurred in 2020–2022, before the AI capital expenditure surge that supposedly motivated them; no hyperscaler has ever used a two-to-three-year server life; every change is disclosed in a 10-K and audited by a Big Four firm; and Meta's depreciation growing 88% while capital expenditure grew 132% is the arithmetic of a young asset base, not of manipulation.

But the bull conclusion does not follow either

That GPUs remain physically productive is well evidenced. CoreWeave has A100 capacity contracted into 2029, nine years after that chip launched. A batch of H100 contracts was rebooked at roughly 95% of original pricing. Jensen Huang's position that the installed base stays useful for the better part of a decade is supported by these observations.

What none of this addresses is what a GPU earns over its life. Cloud rental rates for H100 capacity fell from roughly $8 per hour at launch in 2023 to somewhere in the $1.70–2.85 range at the 2025 trough, before recovering to around $2.35 by March 2026. We label that a market-consensus estimate rather than a measurement, and deliberately so: GPU spot pricing is not filed anywhere. No regulator collects it. The pricing trackers that publish it agree on direction and disagree on level.

The one filing that does split the asset base is instructive. SpaceX discloses servers and networking equipment at 5–6 years and data-centre infrastructure at 20–25 years — the short-lived and long-lived halves that everyone else discusses qualitatively. Microsoft has said roughly two-thirds of its capital expenditure goes to short-lived assets and one-third to long-lived. Beyond those two disclosures, the split is not public for any major operator.

The argument, in short, is being conducted about the wrong variable. Physical obsolescence is measurable and is not happening on the bear's timetable. Economic obsolescence is the real risk and is not measured at all.

08 · The measurable change

The change nobody disputes

Set aside the question of whether AI will pay off. Something else happened in 2026 that is not a matter of opinion, is visible in every relevant earnings release, and has changed the risk profile of the buildout regardless of who is right about the technology.

AI capital expenditure stopped being funded out of operating cash flow.

The individual figures, all from primary sources: Alphabet reported free cash flow of −$5.9 billion in Q2 2026 on capital expenditure of $44.9 billion. Meta's free cash flow was $784 million, down 91% year on year. Oracle's fiscal 2026 closed with $32.0 billion of operating cash flow against $55.7 billion of capital expenditure — free cash flow of −$23.7 billion. Amazon's trailing-twelve-month operating cash flow of $148.5 billion almost exactly matched capital expenditure of $147.3 billion. Only Microsoft retains real headroom.

Guidance suggests this widens rather than closes: Alphabet raised FY2026 capital expenditure guidance to $195–205 billion, Amazon to about $220 billion, Meta to $130–145 billion, and Oracle guided to roughly $70 billion of net cash outlay for FY2027.

Where the difference is coming from

Debt, and structures that are not obviously debt. The BIS recorded hyperscaler gross bond issuance above $100 billion in 2025, most of it at maturities beyond five years, and projected net supply up 30–50% in early 2026. Goldman Sachs expects big technology companies to fund more than a third of AI investment with debt by 2027.

Alongside it sits a category the BIS calls "shadow borrowing" — in its words, "obligations that are economically akin to debt but largely reside outside corporate balance sheets." Meta's Hyperion data centre is the canonical example: a roughly $27 billion debt financing in a vehicle 80% owned by Blue Owl-managed funds and 20% by Meta. Moody's put total undiscounted future lease commitments across five hyperscalers at $969 billion at end-2025, of which $662 billion related to leases not yet commenced — and therefore not yet on the balance sheet. Private credit is lending against GPUs directly — a $2.4 billion Blue Owl-led facility for IREN, a $1.4 billion Pimco and Blue Owl loan to Nscale — and data-centre asset-backed securities have grown from about $4 billion outstanding in 2020 to roughly $61 billion, with Moody's assigning its first AAA rating to a data-centre securitisation in February 2026.

Credit markets are pricing the concentration. CoreWeave's credit default swaps reached roughly 855 basis points in late July 2026 — implying, on standard recovery assumptions, something close to a 50% cumulative five-year default probability. That is a market price, not an opinion, and it applies to a company whose contracted backlog stood at $99.4 billion and whose debt is secured substantially against GPUs.

The official sector has stopped hedging

This is the strongest evidence in the entire debate and it is largely absent from the popular version of it. Four institutions with no commercial stake have now published:

And the historical framing, from Stijn Van Nieuwerburgh's March 2026 Columbia work, is that this buildout is larger relative to the economy than any of its usual analogies:

The same paper records that "in the fourth quarter of 2025, investment associated with AI infrastructure accounted for essentially all of the observed growth in U.S. GDP." That is the sentence that makes this a macroeconomic question rather than a sector one.

The fibre comparison is the one worth dwelling on, because it is the closest in time and the least flattering. More than $500 billion of largely debt-financed telecom capital expenditure between 1996 and 2001 produced fibre that was roughly 97.5% unlit by 2002, WorldCom's $41 billion bankruptcy and Global Crossing's $12.4 billion one. The fibre itself was not wasted — it carries traffic today. It was simply bought by different owners, after the original investors were wiped out. Note also that fibre has a 15–40 year asset life. Servers have five or six.

09 · The record

Every figure in this article, graded

This table is the most useful thing here. It is what the article's method amounts to: the same discipline applied to figures that support the bull case and the bear case alike. We applied it to the adoption statistics in Enterprise AI adoption, where four widely quoted numbers turned out to be counting four different things.

10 · The forward view

What to watch, September 2026 – March 2027

The first finding here is a negative one, and it is the most useful thing in this section. No forcing event falls inside this window. We could locate no relevant debt maturity for Oracle or CoreWeave before 2031. The EU AI Act's high-risk obligations were deferred to 2 December 2027, removing the one regulatory cliff. SpaceX's own $21.5 billion take-or-pay compute commitment falls due in 2027, not now.

Nothing compels a reckoning in the next six months. The period will be decided by voluntary disclosure and by financing conditions — which means the honest forward-looking answer is that the situation probably does not resolve before spring, and here is precisely what would tell you it was resolving either way.

11 · The practical part

How to read this as the buyer

If you run a mid-sized European company, none of the above is directly actionable as an investment view, and you should not treat it as one. What it does change is how you contract.

Assume vendor pricing is unstable in both directions. List prices for equivalent capability have fallen by roughly two-thirds at the frontier. At the same time, free tiers have tightened, long-context surcharges of about two times list have appeared, and priority tiers now ask for upfront throughput commitments. Negotiate price protection and an exit, not a rate.

Do not sign multi-year take-or-pay compute. SpaceX's filings show what that obligation looks like when demand assumptions move: $21.5 billion payable in a single year "regardless of usage." You are a smaller counterparty with less leverage. Match commitment length to the confidence interval on your own demand forecast, which is usually one year, not five.

Budget on cost per outcome, not price per token. This is the most common budgeting error we see. Agentic and reasoning workloads consume an estimated 5–30 times the tokens of a comparable chat interaction. A 67% price cut against a tenfold volume increase is a cost increase. Instrument token consumption per completed task before you scale anything. The AI measurement crisis works through what that instrumentation actually looks like, and why most cost figures in circulation do not survive it.

Know who funds your vendor. The circular-financing map is now documented by the BIS: $46 billion of equity and $879 billion of purchase commitments linking suppliers to their own customers. If your provider's revenue depends on a chipmaker's investment, and that chipmaker's revenue depends on your provider's orders, you have concentration risk that does not appear in either company's marketing.

Keep an open-weight fallback, whether or not you use it. Hosted open-weight models run roughly five to ten times cheaper for many workloads. The value is not primarily the saving; it is that a credible migration path is the only real leverage you have in a renewal conversation. The build, buy and open-source trade-offs are set out in our agent platform playbook.

Use the regulatory window deliberately. The EU AI Act's high-risk obligations now apply from 2 December 2027. That is not a reprieve from doing the work; it is time to do it properly rather than in a scramble. Classification and documentation take longer than teams expect; our compliance playbook covers the classification work in detail.

And the general point. The buyers who will do well over the next eighteen months are not the ones who correctly call the capital cycle. They are the ones who structured their commitments so that the call does not matter much either way.

If you want to go deeper on a specific part of this: