Separation, Not Ownership

Who should fund AI infrastructure, and the test I watched fail.

Fabrizio de Liberali · fabrizio.deliberali.com

---

In brief

The AI bubble debate is asking the wrong question. Not whether valuations are too high, but whether the party funding the buildout is also the party selling into it.

The clearest case is written into the contract. Alphabet agreed to invest at least ten billion dollars in Anthropic, rising towards forty billion against performance targets, with the funding tied to Anthropic spending on committed compute. Anthropic has reportedly committed around two hundred billion dollars to Google Cloud over five years. The investment buys the chips, the chips run the model, the model pays for the investment.

And the concentration is the real story. Contracts with Anthropic and OpenAI now reportedly account for more than half of the two trillion dollars of revenue backlog sitting across AWS, Azure and Google Cloud. Two customers. Both partly financed by the providers recording the backlog.

We have run this experiment before. Telecom equipment makers lent roughly twenty-six billion dollars to their own customers between 1998 and 2001. The lending was never the problem. The accounting was. Revenue was booked the day the crates shipped while the risk waited on the balance sheet as an asset nobody stress-tested. Lucent took over three billion in provisions and was sold four years later.

The best objection runs out this quarter. Today’s buyers have real revenue, near-zero vacancy and contracts signed before ground is broken. All true. But hyperscaler capital expenditure overtakes operating cash flow around the third quarter of 2026.

The rule I started with was wrong, and the correction is the point. It is not that a funder cannot be a customer. Municipal utilities break that rule and work beautifully. The pathology is funder and customer as separate legal persons with a shared interest in appearing independent.

Then I tested my own five conditions and watched them fail. Finland passes on energy because Finland has hydropower. South Africa fails on hardware because no competitive non-American accelerator exists. The test measures endowment, not intent.

Three questions for any AI number you are shown. Who funded it. Could the funder also be the customer. Would you know from the accounts.

---

One dollar, entered twice

Suppose I sell you a machine, and to help you buy it, I lend you the money.

My revenue rises. Your capacity rises. We both look better than we did last quarter. And nobody outside the two of us has bought anything at all.

That trade sits at the centre of the AI buildout. Nvidia stated in September 2025 that it intended to invest up to a hundred billion dollars in OpenAI, released against deployed gigawatts, as OpenAI filled data centres with Nvidia systems. Reporting this month on a far larger financing guarantee attached to an Ohio site extends the same shape.

But the Nvidia case is the one where you have to infer the loop. There is a better example, and in it the loop is written into the terms.

The version in the contract

In April 2026, Alphabet agreed to invest at least ten billion dollars in Anthropic at a valuation of around three hundred and fifty billion, with the commitment rising towards forty billion as performance targets are met. The funding is tied to Anthropic spending against committed capacity on Google’s tensor processing units, built with Broadcom, amounting to some five gigawatts from 2027.

In May, The Information reported the other side of that arrangement. Anthropic has committed roughly two hundred billion dollars to Google Cloud over five years.

The investment buys the chips. The chips run the model. The model pays for the investment. Nobody has had to arrange this quietly, because there is nothing to hide. It is a commercial structure disclosed to investors on both sides.

Amazon runs the same shape and adds an accounting turn worth understanding. In April 2026 it agreed to invest up to twenty-five billion dollars in Anthropic on top of eight billion already committed, while Anthropic undertook to spend more than a hundred billion dollars on AWS technologies over ten years, including current and future generations of Amazon’s own Trainium silicon, securing up to five gigawatts of capacity.

Then follow the stake itself. Amazon is required to mark its Anthropic investment to fair value whenever a significant financing event establishes a new reference price. Anthropic’s spring round established one. Amazon booked sixteen point eight billion dollars of pre-tax gains on the position in the first quarter of 2026, of which twelve point three billion was revaluation.

So one relationship produces an equity gain landing in reported net income, cloud revenue, custom silicon revenue, and a decade of contracted backlog. Four pieces of good news from a single set of arrangements, three of them visible in earnings.

The arithmetic that does not run one way

Here is where honesty requires slowing down, because the numbers cut against the alarm as well as for it.

Two hundred billion dollars out against as much as forty billion in is a ratio of five to one. The overwhelming majority of that compute has to be paid for by customers, not by Google’s capital. Anthropic’s annualised run rate reached thirty billion dollars in April 2026, more than triple where it stood at the end of 2025. A company growing at that rate may well fund its commitments from operations. Fair value accounting is mandatory rather than chosen. Neither Google nor Amazon is doing anything improper, and neither is Anthropic.

Now hold two numbers together anyway.

Committed compute spend of roughly forty billion a year to Google, plus something near ten billion a year to AWS, against an annualised run rate of thirty billion. The commitments already exceed current revenue. That is not a prediction of failure. It is a statement of what has to keep happening.

And then the number that matters most. Contracts with Anthropic and OpenAI now reportedly account for more than half of the two trillion dollars of revenue backlog disclosed across AWS, Microsoft Azure and Google Cloud. Amazon’s own backlog rose forty-nine per cent to three hundred and sixty-four billion dollars in the first quarter of 2026. Anthropic alone reportedly represents over forty per cent of what Google disclosed.

Two customers. Half of two trillion dollars. Both partly financed by the providers recording the backlog.

That is not a story about overvalued startups going to zero. It is a concentration of counterparty exposure at the centre of the largest companies in the index, and when it moves it will be reported as a market event.

What 1999 already knew

The precedent is close enough to be useful.

Between 1998 and 2001, the telecom equipment makers financed their own buyers. Lucent committed around eight billion dollars, Nortel roughly three with under half drawn, Cisco something over two in customer loans. McKinsey later counted about twenty-six billion of exposure across nine suppliers by the end of 2000.

Here is the part worth carrying forward. The danger was never the generosity. It was that a vendor loan let the supplier book equipment revenue the moment the crates left the warehouse, while the matching risk waited on the balance sheet as an asset nobody stress-tested. When outside capital fled in early 2001, the vendors did not retreat. They lent harder, to keep weakening customers buying.

Lucent took provisions of over three billion across 2001 and 2002. The company that housed Bell Labs was sold in 2006.

The best objection, and its expiry date

There is a serious answer to all of this, and it deserves better than a wave.

The competitive local exchange carriers of 1999 had no revenue. Microsoft, Amazon, Alphabet and Meta have search, advertising, retail and enterprise software, and most of their AI capacity serves their own products rather than a third party’s promise. Microsoft is guiding towards something near a hundred and ninety billion dollars of capital expenditure this calendar year. Meta has lifted its 2026 range to between a hundred and twenty-five and a hundred and forty-five billion. Amazon expects to spend roughly two hundred billion. AWS grew twenty-eight per cent in the first quarter, its fastest in fifteen quarters, and Google Cloud sixty-three. They are not building on hope, and on their own account they cannot build fast enough to serve contracted backlog. North American datacentre vacancy sat near one and a half per cent at the end of 2025. Accelerators refresh every three to four years, so temporary excess gets absorbed rather than stranded. Four of the five largest operators carry debt at around one times EBITDA or below.

This is not Winstar. The analogy holds in mechanism and breaks on solvency.

Amazon is the awkward case, because it sits on both sides of the line I am drawing. Its capacity serves its own retail and advertising business, which is ordinary industrial investment, and it also sells compute to a model lab whose valuation it books gains on. A distinction worth having should be capable of cutting through a company rather than tidily between them.

And the solvency argument has a date on it.

Epoch AI’s tracking puts aggregate hyperscaler cash capital expenditure overtaking operating cash flow around the third quarter of 2026. Alphabet raised eighty-four and three quarter billion dollars of equity in June, the largest such transaction ever completed by a listed company. Oracle, whose commitments are contractual obligations to build for one model lab, was downgraded by S&P in July to a single notch above speculative grade. And the depreciation question stays open: hardware written down over five to six years may live two to three, flattering current earnings by something in the region of a hundred and seventy-five billion dollars over the next three years.

So the honest claim is narrower than the bubble commentary allows, and stronger for the narrowing. Capital a company spends from its own cash to serve its own users is ordinary investment, and there may simply be too much of it. The round trip is a different animal. Money leaves a provider, arrives at a customer, and returns as that provider’s revenue and reported gains. The first risk is overbuilding. The second is that nobody left in the system can tell demand apart from its own capital.

Where I was wrong

I began this thinking with a rule I liked very much.

Infrastructure works when the entity that funds it is structurally incapable of being its principal customer.

It lasted about a week.

The counter-examples are embarrassingly ordinary. Municipally owned utilities sell electricity to the residents who own them. State-backed railways run trains on track the state financed. National research clouds fund capacity their own institutions consume. None of these are pathologies. Several are among the best-run infrastructure on earth.

So the rule was not merely imprecise. It was pointed at the wrong thing.

Here is the version that survives.

The pathology is not that funder and customer coincide. It is when funder and customer are separate legal persons with a shared interest in appearing independent of each other.

A municipal utility is not pretending its customers are at arm’s length. The identity of owner and user is the design, it is disclosed, and the accounts get read accordingly. Google’s cloud backlog and Amazon’s reported gains depend on Anthropic reading as an independent buyer, which in most respects it is, and in one respect it is not. The circularity does its damage through the reader, not through the flow of funds.

That gives three questions you can actually use. Is the funder’s own consumption bounded and disclosed? Is the allocation of capacity governed separately from the decision to invest in it? And would an outsider reading the accounts correctly identify who is buying from whom?

Spectrum is the case I know from the inside. I spent years on the operator side of licence conditions. The state grants use, attaches coverage obligations that are demanding and occasionally painful, and takes no position in the traffic. It does not book the minutes. Those obligations are why coverage reached places no commercial case would have reached alone. The arrangement survives because nobody is confused about who the customer is.

Three problems wearing one coat

European sovereignty programmes keep stalling for a reason that has nothing to do with ambition. They treat compute, models and applications as one problem, and the resulting programme is too big to govern and too vague to ship.

Compute, energy and interconnect is a grid problem in a technology costume. High capital intensity, long amortisation, natural concentration, and siting decisions that are really energy and water decisions. In Europe this is EuroHPC and the national datacentre programmes, and it is the layer that has actually worked.

Foundation models are closer to a public research good. Expensive to make, cheap to copy, valuable mainly through what gets built on top.

Applications are commercial, competitive, and properly free to fail. Nothing here argues otherwise. Founders building at this layer are right to.

I should concede something on the middle layer, because the CERN analogy is more comforting than accurate. Models now arrive bundled with proprietary tooling, serving stacks and evaluation infrastructure. A publicly funded open model is not competing against a paper. It is competing against a product.

Gaia-X is the cautionary tale, and I was too dismissive of it in an earlier draft. Yes, governance and specification work outpaced deployed capacity, and the incumbents it existed to counterbalance sat inside its governance. But its data-space work addresses portability, identity and federation, which are real and load-bearing for any practical sovereignty. The lesson is about sequence, not substance. Start with concrete physical capacity, or you end up a standards body with a communications budget.

EuroHPC started there, and it shows. LUMI in Kajaani is owned by the Joint Undertaking, hosted by an eleven-country consortium, and its European share is allocated through open peer-reviewed calls rather than purchase orders. Unglamorous. Delivered.

Rules, not discretion

A government that funds compute and then expects a say in what runs on it has not fixed circularity. It has rebuilt it in a worse material, because a state can enforce its preferences in ways an investor cannot. Trading a vendor who wants your workload for a ministry that wants your compliance is a change of landlord.

But the clean version of that claim does not survive either.

Setting safety and access rules in advance is itself an exercise of influence. A democratic state that funds infrastructure and then declines to constrain, say, its use for population surveillance has not achieved neutrality. It has abdicated. There is no vantage point outside politics from which infrastructure gets built.

So the distinction is not influence against its absence. It is rules against discretion.

Write the conditions in advance. Publish them. Bind the funding authority on exactly the terms that bind everyone else. No reserved capacity, no preferential access, no editorial claim exercised after the fact. Allocation administered at arm’s length against stated criteria. The state builds the road and then queues on it like everybody else.

What corrupts the arrangement is not the existence of rules. It is rules applied case by case to whoever has become strategically important.

The ratchet, corrected

Infrastructure outlives the governments that fund it. The characteristic failure of public digital infrastructure is not opening day. It is the day eleven years later when access conditions get quietly relaxed for a partner who now matters.

In my partnership governance work I use the term Ethical Ratchet for a rule that standards tighten and never loosen. Applied carelessly here, it breaks. Sometimes loosening is right. Initial risk assessments are often too cautious. New classes of benign workload appear. Access priced to protect scarce capacity becomes a barrier once capacity is abundant.

The repair is to be precise about what the ratchet grips.

Standards of protection tighten only. Scope of permitted use may widen. Those are different operations, and conflating them is what makes the principle look unworkable. Admitting a new category of research workload weakens no safeguard. Relaxing an audit requirement does.

Where protection genuinely needs relaxing, the ratchet turns procedural rather than absolute: a presumption against, overcome only by a higher hurdle than tightening requires. Published reasoning. A supermajority. A sunset date. The aim is not to make relaxation impossible. It is to make quiet erosion expensive, because quiet erosion is the failure that actually happens.

The arithmetic of continents

The case for continental scale is economic, not metaphysical. Training costs are fixed and large, serving costs are marginal and small, and a general-purpose model serving sixty million speakers cannot amortise a national compute programme. Four hundred million can.

Europe is not special in that arithmetic. The generalisation is that each continent funds and governs its own, which is the only version of the argument that is not simply European exceptionalism.

Two things complicate it.

The arithmetic only bites at the frontier. Specialised models at regional or even municipal scale are entirely viable for healthcare, agriculture, land management and local administration, which happen to be the domains where local data and local accountability matter most. The continental case is about general-purpose frontier capacity. It is not about all useful AI. That distinction matters enormously to anyone who cannot afford the frontier.

The second is harder, and I will not pretend it resolves. An argument for sovereignty that stops at Europe’s borders is enclosure with better manners. I am arguing for blocs, and blocs exclude.

I ran my own test and watched it fail

An earlier draft of this essay listed five conditions a non-extractive programme would have to meet, then asserted that nothing meets them. Assertion is cheap, so let me actually run it.

The conditions. Weights transferred outright rather than access granted through an interface. No dependency on a single vendor’s stack to keep running. Governance held where the capacity physically sits, with the legal power to refuse. Energy and water accounted locally and charged to the operator. Annotation and evaluation labour treated as a capital contribution to the asset rather than a commodity input bought at the clearing price.

EuroHPC, taking LUMI as the instance. Governance passes cleanly. Environmental accounting passes emphatically: entirely hydro-powered, waste heat covering around a fifth of Kajaani’s district heating, and a successor facility under construction in a former paper mill since January 2026 that will recover considerably more. Stack independence fails. The machine runs AMD accelerators and an American interconnect, and the point is not that AMD is the wrong supplier. It is that European processor work stayed immature enough that there was no European option to choose. Weights and labour barely apply, since the programme allocates compute rather than producing models.

Two clear passes out of five.

Cassava Technologies in South Africa, the most substantial attempt at African AI infrastructure and explicitly framed as sovereign. Data stays on the continent and the operator is African-owned, so local governance partially passes. Stack independence fails: Cassava is Nvidia’s first African cloud partner, and Nvidia has since taken an equity position in Cassava. The weights condition fails, because the multi-model exchange distributes models from Anthropic and Google through an interface. Environmental accounting is not reported in terms anyone can check. Labour is not addressed.

Now look at what those results actually say.

Finland passes on energy because Finland has hydropower, a cold climate, an existing district heating grid and a municipal energy company willing to sign. South Africa fails on hardware because nobody sells a competitive accelerator that is not American.

Neither outcome is about virtue. Both are about endowment.

Which means my test largely measures wealth. It hands a pass to jurisdictions that were always going to pass, and issues a fail where no alternative exists. As a benchmark it is close to useless, and worse than useless if it becomes a reason to do nothing while the extractive version proceeds.

And it is proceeding. Siting follows cheap power, available land, tax incentives and lighter enforcement. Annotation and moderation work is contracted in Nairobi and Manila at rates that would be unlawful where the resulting model is sold. Rest of World reported this year that the five largest African markets together hold less datacentre capacity than France had in 2024, while Nigeria, Egypt and Kenya have all published draft strategies naming dependence on American platforms as a security concern. The Tony Blair Institute’s warning about North African facilities is the sharpest formulation I have read: capacity can sit inside your borders while a third party holds the operational keys.

Presence is not control.

And notice what Cassava demonstrates. The chip supplier holds equity in the provider that buys its chips. The structure I opened this essay with has already been exported, and it arrived speaking the language of sovereignty.

What to measure instead

Not the state achieved. The direction of travel, and the terms of the contract.

Is the dependency shrinking or deepening? Is there a documented exit, with a cost somebody has actually calculated? Can the operator refuse a request from the vendor who financed it, and has it ever done so? Does the contract vest anything durable locally, or only access? Are the environmental and labour terms measured at all, given that a condition nobody reports against is not a condition?

That is a weaker test than five absolutes, and it is the one worth using, because it separates a programme that is dependent and reducing its dependency from one that is dependent and comfortable.

Development finance has funded pipes for decades and deferred the governance question, which is how you arrive at capacity that is physically present and functionally rented. The repair is not a purer standard. It is treating governance and labour terms as part of the capital stack rather than as conditions bolted on afterwards.

Where this goes

Who funded it. Could the funder also be the customer. Would you know from the accounts.

Three questions, and they are not rhetorical. They are the ones I now ask in every governance conversation I am part of, and they are cheap enough that anyone can ask them.

The circular financing story matters because it is the small, legible version of a much larger design failure. We are building the substrate for a technology that will shape behaviour at population scale, and financing it through structures that cannot tell demand from their own capital. When the correction lands it will be reported as a market event. It will be a governance event that happened to show up in the accounts.

The people who can change that are not the ones writing about it. They are the ones drafting the contracts, sitting on the boards, and setting the licence conditions. If that is you, the three questions are yours to use, and I would rather be argued with than agreed with.

---

Sources

Google, Amazon and Anthropic. Alphabet’s investment of at least ten billion dollars, rising towards forty billion against performance milestones at a valuation of around three hundred and fifty billion, and the Broadcom TPU capacity: The Tech Portal, 6 May 2026, and BigGo Finance, 9 May 2026. The conditionality of that investment on committed TPU spend: Let’s Data Science, 6 May 2026, drawing on Reuters and CNBC. Anthropic’s reported commitment of roughly two hundred billion dollars to Google Cloud over five years, and the estimate that this represents more than forty per cent of Google’s disclosed cloud backlog: The Information, 5 May 2026, relayed by Reuters. Amazon’s agreement to invest up to twenty-five billion dollars on top of eight billion already committed, Anthropic’s undertaking of more than a hundred billion dollars of AWS spend over ten years, five gigawatts of capacity and Amazon’s roughly two hundred billion dollars of 2026 capital expenditure: CNBC, 20 April 2026, and Capacity, 21 April 2026. Project Rainier scale and Trainium generations: Anthropic’s own announcement, and StartupHub, July 2026. Amazon’s sixteen point eight billion dollars of pre-tax gains including twelve point three billion of mark-to-market revaluation, and the fair value accounting requirement: The Next Web, 3 May 2026, and Yahoo Finance, July 2026. Anthropic’s thirty billion dollar annualised run rate disclosed 6 April 2026: Let’s Data Science, 6 May 2026. The estimate that Anthropic and OpenAI contracts represent more than half of the two trillion dollars of backlog across AWS, Azure and Google Cloud, and Amazon’s backlog rising forty-nine per cent to three hundred and sixty-four billion: Resultsense, 6 May 2026, and BigGo Finance, 9 May 2026.

Nvidia and OpenAI. The September 2025 letter of intent to invest up to a hundred billion dollars against deployed gigawatts, and July 2026 reporting on the Ohio financing guarantee: Benzinga and Michel Johannsen, “Vendor Financing Loops: What 1999 Telecom Tells Us About 2026 AI”, March 2026.

The telecom precedent. Vendor financing figures for Lucent, Nortel, Cisco and the McKinsey nine-supplier tally: Michel Johannsen, March 2026. Lucent’s bad debt provisions: Tomasz Tunguz, “Circular Financing: Does Nvidia’s Bet Echo the Telecom Bubble?”, October 2025. Lucent’s SEC revenue recognition case and loan commitment disclosure: Fortune, October 2010.

Hyperscaler finances. Capital expenditure against operating cash flow: Epoch AI, “Hyperscaler Capex to Exceed Cash Flow by Q3 2026”, June 2026. Alphabet’s equity raise, the Oracle downgrade, Meta’s capital expenditure guidance and leverage ratios: FactSet Insight, July 2026. Microsoft’s calendar 2026 figure, datacentre vacancy, refresh cycles and the depreciation estimate: AL Capital Advisory, “AI Capex Cycle 2026”, July 2026. AWS first quarter revenue and growth, and custom silicon run rate: StartupHub, July 2026.

EuroHPC. LUMI ownership, peer-reviewed allocation, hydropower and district heating: EuroHPC Joint Undertaking and CSC. Successor facility and the Loiste heat recovery agreement: Finnish AI Region, April 2026.

Africa. Capacity comparison, national AI strategies and platform dependence: Rest of World, “Africa’s AI plans still depend on Google, Microsoft, Nvidia, Meta”, May 2026. Nvidia’s equity position in Cassava, the AI Factory rollout and the multi-model exchange: TelecomTV and Cassava Technologies, March 2026. The operational control point attributed to the Tony Blair Institute: Silicon Canals, May 2026.

A note on sourcing. The two hundred billion dollar Google commitment and the two trillion dollar backlog share both rest on The Information’s reporting, relayed onward. They are attributed here as reported rather than disclosed, and they have not been confirmed in either company’s filings at the time of writing. Financial figures are as reported at the end of July 2026. Check against current filings before citing.

---

On my use of AI

The thinking here is mine. The reading, the connections, the argument that the problem is the funding structure rather than the valuation, the line I draw from coverage obligations attached to a spectrum licence to how public compute should be governed, the three-layer separation of compute, models and applications, and the position that a ratchet should bind standards of protection while leaving scope of use free to widen, all came out of my own reflection, my own reading history, and the governance work I do with clients.

English is my second language. I use AI the way I would use a good editor: to structure an argument I have already formed, to tighten sentences, to catch the errors that a non-native writer makes and often cannot see, and to check facts, dates, and attributions against sources before anything is published. Where a claim in this piece is checkable, it has been checked, and the sources are listed above so you can check them yourself.

This piece went through a harder review than most. An earlier draft stated the separation principle as an absolute, and it did not survive the counter-examples: municipal utilities and state railways fund infrastructure they also consume, and they work. The restatement in this essay, that the pathology lies in funder and customer being separate parties with a shared interest in appearing independent, is what came out of that challenge. The Cassava case surfaced during fact-checking rather than from my own reading, and it changed where the essay ends. I would rather tell you that than let the structure of a disclaimer imply otherwise.

One further disclosure, which matters more than the others. I drafted and fact-checked this piece with several AI models, including Claude, a model built by Anthropic. Anthropic is the central example in the section on circular financing, and Google and Amazon, the two investors I examine most closely, are its largest. I asked the model to make that case as hard as the evidence allowed, and it did, including surfacing the mark-to-market detail and the backlog concentration figure that became the strongest claims in the piece. You should still weigh the passage on Google and Amazon knowing how it was assembled. I do not think it reads soft. But an essay arguing that shared interests must be disclosed would be a poor place to leave one unstated.

What I do not do is ask a model to have the idea. The thesis is mine, the framework is mine, and the tension I decline to resolve is one I am genuinely stuck on rather than one staged for effect. If a sentence here makes an argument, I made it. Where evidence or an objection came from the process of checking, I have said so.

I take full responsibility for everything published under my name, including any error that survived the process. If you find one, tell me and I will correct it in the open.

---

Fabrizio de Liberali is a Partnership Architect and strategic advisor. He writes here about the questions he has not finished answering.

This essay was written on Fab Campaigns time. Fab Campaigns is where this way of thinking becomes something a client can buy. This site is where it gets worked out in public, before it is tidy.