A Cloud-Grid Settlement

What this is for

Much of the world now runs on the cloud and AI systems of a few companies, nearly all of them based in one country, and there is still no agreed answer to two plain questions: who pays when that infrastructure fails all at once, and who decides when a country gets cut off from it. This page points to that open space and suggests one way to fill it.

What it proposes is a treaty in which the largest providers pay into a shared fund against their own failure, the countries that depend on them help write the rules of service in advance, any country can join on the same terms, rivals included, part of the money pays for access for countries that are priced out today, and if a provider or a government switches off a member’s cloud and AI services, that member’s hospitals, banks, and public systems keep running, because backups and the ability to move to another provider are built in ahead of time.

Many people now describe the present as a kind of world war fought without armies meeting, carried on through export controls, damaged undersea cables, cyberattacks, drone strikes on data centers, and the power to switch a country off. If that description is even partly right, it is fair to ask what the settlement at the end of such a conflict would look like, and whether some of it can be written down before the fighting rather than after. This is one answer to that question for the part of the conflict that runs through cloud and AI infrastructure.

The questions are pressing now because the dependence has deepened much faster than any rule has formed. This year financial regulators in several places have begun to name the concentration as a risk to the systems they supervise, at least one government has committed tens of billions of dollars to building a system of its own as insurance, and countries that cannot afford to do the same are left with whatever terms they are offered.

The method was to work several steps ahead of events, including the severe cases, which this page does not dwell on, to see what would eventually have to be dealt with, and then to work back to terms the parties could negotiate from where they stand today. The result is written to look like the end product of a hard negotiation, meaning something every party can picture being achieved at scale because it is designed so that each of them leaves with more than it gives.

It is offered as a reference point rather than a final answer. The June 2026 text is fixed and dated so that it can serve as a benchmark, and if someone can do better, the hope is that they will, and soon.

The scale, in dollars

The numbers below are what the proposal is sized against. Each is labeled for what kind of figure it is, because a reported total, a company’s own forecast, an analyst’s estimate, and a number this proposal marks as illustrative are not the same thing.

A few comparisons carry the argument. The entry fee is at most a few hundredths of one percent of what the market takes in each year, small enough that no provider could credibly say it changes whether they build. The full fund, built over a decade or more, would be about the size of what one country has reportedly committed to spend insuring only itself, and it would cover every member. Against the capital four companies plan to spend this year alone, the largest version of the fund is roughly one-twentieth, which is the sense in which the price of governing the risk is small next to the price of the build-out it protects.

The fund amounts are stocks built up over years, while the market and premium figures are yearly flows, so the comparisons above are between orders of magnitude and not a like-for-like ratio.

A two-page briefing sheet sets out the same argument in charts, with its sources on the sheet: Exposure (PDF, 107 KB).

What it would mean, and for whom

The same arrangement looks different depending on who is looking at it, and some of what matters most is not visible from where most people stand.

For Americans

What most Americans see of this infrastructure is the construction: large buildings going up, questions about power bills and water, and tax breaks that local governments grant to win them. Opposition has grown quickly, and in an Annenberg survey released in August 2026, 61 percent of adults said they oppose new data centers in their area, up from 49 percent earlier in the year (Annenberg Public Policy Center).

What is harder to see is where the losses land when these systems fail. After the July 2024 outage, hospitals, airlines, and public agencies absorbed their own losses, most of them uninsured, and the providers’ contracts capped what they owed at roughly the fees their customers had paid. Under this proposal, the providers pay for that risk into a fund, rather than the public carrying it by default. The honest cost to the United States is that its government would accept a capped and declining share of the vote in a body it does not control, in exchange for allies who stay on American infrastructure instead of building their own and a priced settlement of a risk that currently has no owner.

For NATO partners and other allies

Allies see their banks, hospitals, ministries, and in some cases militaries running on American providers. More than 80 percent of Europe’s digital technology and infrastructure is imported (Bertelsmann Stiftung, February 2025), three American firms hold 85 percent of Canada’s public cloud market (Canadian Anti-Monopoly Project, via National Observer), and in July 2026 the United Kingdom placed Amazon, Google, Microsoft, and Oracle under direct financial-stability oversight (Bank of England).

What allies cannot count on today is that service continues if a company, or the American government through an export control or a sanctions order, decides otherwise. Their choices are to accept that dependence or to spend heavily building their own systems, as Israel has begun to and as European sovereignty proposals would. This proposal offers a third choice: a guaranteed right to keep running, delivered by the design of the system, and a seat with a guaranteed minimum voice, in exchange for buying into the fund and accepting weighted governance, alongside the fee the providers pay. The countries most exposed are the closest allies, so they are also the ones with the most to gain.

For rivals, providers, and countries priced out

A rival state sees a chokepoint it does not control, and today its only answer is to build its own. Here it can join on the same binding terms as everyone else and lower its own exposure by doing so, while taking on the same obligations. The largest providers see a new fee, and in return they get a pooled fund standing behind them, regulatory and antitrust certainty, and one set of rules in place of a growing patchwork of national ones. Countries that cannot afford a seat at any table today get a funded floor of access. The full ledger appears further down the page.

One case

In September 2025, Microsoft switched off part of an Israeli military intelligence unit’s access to the cloud and AI services it had been running on, saying the unit had used the platform to store mass-surveillance data in breach of the company’s terms of service. Whether the company was right, and what the unit was doing, are questions this page does not try to settle.

There is a smaller question that holds up whatever you conclude about the larger one. A private firm switched off a state under terms it wrote itself, read by its own lawyers, and nobody outside the company reviewed the call. If you think the decision was right, you are counting on a company to land where you would have landed the next time as well, and if you think it was wrong, there is nowhere to take that. The same thing is missing either way, and it is missing for every country that now runs its hospitals, banks, power grids, and defense ministries on the same few systems.

What Israel did next is what any government would do. It already had contract terms, reported from leaked documents on its 2021 cloud deal with Google and Amazon, meant to keep those providers from cutting it off; its data had been backed up in advance; the material was moved out of the country where it had been held, with a reported plan to shift it to another provider; and in June 2026 its government approved a national AI program with sovereign infrastructure among its goals, at a reported cost of $20–30 billion or more. None of that is a rule, none of it covers anyone but one country, and a country without the money or the standing to demand those terms gets none of them.

Sources: Microsoft statement (Brad Smith), September 25, 2025; The Guardian and Times of Israel, September 25, 2025; Israel Democracy Institute, September 28, 2025; The Guardian / +972 Magazine / Local Call, October 2025; Jerusalem Post and Business Standard, October 30, 2025; Ynet, June 25, 2026. Full citations are in the plain-language introduction listed below. For an outside legal reading of the episode, see Lawfare, October 2025.

What has been missing

Over the past year and a half a growing number of serious institutions have argued, each in its own terms, that concentrated cloud and AI capacity should be treated as a commons, a public utility, or critical infrastructure. Their work names the problem well and proposes real remedies, and it is the debate this proposal belongs to. Each is linked here and worth reading on its own terms.

Open Markets Institute, Engineering the Cloud CommonsMay 2025

Utility-style regulation of cloud through a dedicated regulator, access on fair and non-discriminatory terms, and structural separation of cloud provision from the businesses that run on it.

Canadian Anti-Monopoly Project, Parting CloudsJune 2026

Binding interoperability and portability, an end to exit fees, neutrality rules for cloud as utility-like infrastructure, and coordination among middle powers on procurement.

AI Now Institute, Computational Power and AI2023, 2025

Antitrust enforcement, structural separation, and nondiscrimination obligations on the providers that control key layers of the stack.

Mozilla and the Vanderbilt Policy Accelerator, public AI2025

Public options for compute, data, and models, including existing federal authority to build a public cloud.

EuroStack, Bertelsmann StiftungFebruary 2025

A federated European stack built through open standards, procurement, and public investment, framed as digital sovereignty.

Centre for the Governance of AI, Computing Power and the Governance of AI2024

Why compute, because it is detectable, excludable, measurable, and concentrated, is among the most governable parts of AI.

Pope Leo XIV, Magnifica HumanitasMay 2026

The common good, subsidiarity, and solidarity applied to AI, with a warning against the concentration of technological power in a handful of corporations.

What none of these sets out, as far as the author’s own review in September 2026 could find (a search, not an exhaustive survey), is how the shared risk would be paid for, who would govern the infrastructure across borders, and what a country that signs on would actually receive in return. That is the part this proposal works out, as one example of what filling the space would take. It borrows its tools from places where they already work: the fees large banks pay into a fund against their own failure, the voting rule that stopped any single country from controlling the satellite system that once carried the world’s international calls, and the vote floors that give the smallest members of the International Monetary Fund a guaranteed voice.

Principles it holds to

Price the risk rather than tax the industry.

The fee exists because concentration creates a risk the public now carries for free. It is sized to that risk and pays for protection against it, not for general spending.

Write the rules before the crisis.

Conditions of service are set in advance and apply automatically, so that nobody has to reach agreement in the middle of an emergency.

Protect continuity by design rather than by permission.

A country that depends on the infrastructure keeps running because of how the system is built, not because someone decided to let it.

Open the door to everyone on the same terms.

Membership turns on accepting the obligations, not on which side a country is on.

Put the least powerful first in line.

Money for universal access is the first call on the fund, so that it cannot be drained by a crisis elsewhere.

Say plainly what is not yet known.

Numbers are marked illustrative, detailed calibration is left to the work that would follow, open problems are listed, every claim is labeled as provable, derived, or inference, and the dated text is not revised after the fact.

What this settlement would do

This proposal is written as a term sheet: indicative heads of terms for a treaty, with a standing council to administer it and a fund it collects into. It is pitched at the level of architecture, and where a number is needed to make a mechanism testable the number is marked illustrative. It has four moving parts.

A price on the risk

Concentration creates a risk that the public currently carries for nothing. No insurance market can price a multi-day failure across the major clouds, and no government has said it would cover the gap, so when one vendor’s bad software update took down millions of machines in July 2024, every hospital, airline, and public agency ate what landed on it. This settlement would charge the largest providers a fee for that risk, the way large banks are made to pay into a fund against their own failure, and the fund it builds is what pays for the guarantees below.

Mechanism. An assessment in basis points on providers above a size threshold. Illustrative entry band 1–5 basis points, about $25–130 million a year, rising over a 10–15 year glidepath toward a first-loss layer on the order of $15–40 billion. Modeled on FDIC deposit-insurance assessments and the EU Single Resolution Fund.

Shared governance

The rules of service are written jointly by the states and the providers inside the settlement rather than by a few firms alone, and they are written in advance. A body that tried to judge each crisis as it happened would deadlock, which is why the council’s job is to set the conditions of service ahead of time so that they apply automatically, the way a bank’s capital rule applies during a run without anyone calling a meeting.

Mechanism. A council with weighted voting in which any single party’s weight is capped and declines over time, on the model of the INTELSAT agreements that capped the United States at 40 percent; a basic-vote floor for every member, on the Bretton Woods model; supermajorities for reserved matters; and a recusal bar listing the matters on which providers may not vote.

A continuity guarantee

A country that depends on this infrastructure keeps running even if its service is cut. That protection is what Israel arranged for itself, through contract terms and its own backups, in the case described above; this settlement would make it public and give it to every party on the same terms, delivered by the design of the system rather than by anyone’s permission.

Mechanism. Escrowed continuity, mandated portability, and stand-by capacity. The grounds on which service may lawfully be interrupted are listed in advance; an interruption whose grounds are not sustained on a fast review is reversed automatically, with a remedy priced and set aside ahead of time.

A floor of universal access

Part of the fund pays for access for countries that would otherwise be priced out entirely and have no leverage to bargain with, and that money is set aside as a first call on the fund so that a crisis somewhere else cannot drain it.

Mechanism. A ring-fenced access floor measured against the Broadband Commission’s affordability target of basic access at no more than 2 percent of monthly gross national income per person in developing countries, with a named deliverer.

The door

Any state can join, rivals included, on the same binding terms as everyone else.

Several of the proposals now in circulation for governing compute are built as coalitions of allied democracies designed to keep a rival out. This proposal starts with the provider country and its allies, because that is where the dependence runs deepest and where the trust to begin already exists, but it is written so that entry is gated by accepting the rules rather than by belonging to a side. The arms-control and monetary institutions of the last century were built the same way, admitting states that were not allies on condition that they accepted the same constraints.

The reasoning is that a rival joins because a guarantee it can rely on is worth more to it than anyone’s freedom to cut it off, and that a country which lowers its own exposure by joining has less reason to build a competing system that splits the grid in two. Whether a rival would actually walk through the door is a judgment rather than a forecast, and the sourcing index states plainly what evidence would prove it wrong. That open door is what separates a governed commons from a defensive bloc, and it is the part of this proposal that no other proposal found in the review has put forward.

What each party gives and receives

A settlement only holds if every party at the table leaves with more than it put in. The term sheet’s own ledger is reproduced here, lightly shortened.

Lowering the stakes of the conflict

This infrastructure is already a front. In March 2026 drone strikes damaged three Amazon Web Services facilities in the United Arab Emirates and Bahrain, the first known military strikes on an American hyperscaler (CNBC; Rest of World). Researchers counted 44 incidents of damage to undersea cables in 2024 and 2025 (Recorded Future). And the case described above shows that the power to cut a country off already exists and has already been used.

In a conflict, dependence becomes a weapon and the backup systems countries build become targets. A settlement of this kind cannot stop anyone from fighting, but it can lower what is gained by fighting in this particular way. Service interruptions limited to grounds listed in advance, and reversed automatically when those grounds are not met, make the switch less useful as a weapon. Continuity built into the system, with portability and stand-by capacity, makes a single strike or a single cutoff less decisive. And an open door lowers the pressure on every major power to build a separate, competing system, which would split the world’s infrastructure into rival blocs, each more fragile than the whole.

The institutions this proposal borrows from were themselves written in the middle of great-power conflict. The Bretton Woods agreements were signed in 1944, before the Second World War had ended; the INTELSAT agreements were negotiated during the Cold War; and the Nuclear Non-Proliferation Treaty was opened for signature in 1968. None of them ended the conflict it was born in, but each gave states on different sides a set of rules to live under, and several outlived the conflicts that produced them.

What it does and does not answer

The current alarm about AI has several sources at once, and a proposal that claimed to answer all of them would not be believable. Here is where this one stands on each.

Is the build-out a bubble?touches

Planned spending by four companies has nearly doubled in a year, and the borrowing needed to finance it through 2028 is estimated at about $1.5 trillion (see the dollar figures above). This proposal does not set the pace of investment and would not prevent a financial bust. What it would do is keep a provider’s failure, whether from an outage or from its balance sheet, from taking down the countries that depend on it, because continuity is guaranteed by portability and stand-by capacity rather than by the provider staying solvent. That link between financial failure and continuity is a judgment, not something the June documents tested.

Too much power in too few hands?addresses in part

A widespread concern is that a small number of executives, and the politics around them, now hold power over public life that no one voted for. This proposal moves the rules of service out of company contracts and into a treaty body where no single party controls the outcome and providers cannot vote on specified matters. It does not break companies up; that is the work of antitrust, which others cited above are pursuing, and the two approaches are compatible.

Is the industry regulating itself?addresses

Since July 2026, leading AI companies have discussed an industry-funded standards body of their own, modeled on the financial industry’s self-regulator (see Lawfare, “Designing a FINRA for Frontier AI”). A treaty body with capped votes, a recusal bar, and an open door is a different arrangement from a standards body funded and run by the companies it would oversee.

The dangers of AI systems themselves?outside its scope

This proposal governs the infrastructure, not the models that run on it, and it takes no position on how fast AI should advance or how its behavior should be controlled. Any rule about models, however, would eventually have to be enforced where the models run, and a governed infrastructure layer is one place such rules could be attached.

Jobs, power bills, and the communities where this is built?addresses in part

The work began from the people physically building this infrastructure, who hold real leverage while construction lasts, and from the question of whether that leverage gets spent on wages alone or on a say in who owns and governs what gets built. The access floor covers affordable connection, not electricity bills, so ratepayer costs remain a separate fight that this proposal does not settle.

The best argument against it

The strongest objection is that a body containing the United States, China, Israel, and the Gulf states will never agree to cut anyone off, and that deadlock would not be an occasional flaw because hard cases are the only cases such a body would ever see. On that view a settlement like this one takes the power to cut service away from private companies and quietly retires it.

The answer is that the council’s work is writing rules rather than judging incidents. Banking supervision does not function because a committee meets during the run; it functions because the capital rule was set years earlier and applies without a vote. A settlement that tried to decide each September as it happened would fail, while one that sets the conditions of service in advance and lets them bind automatically does not need agreement in the moment.

The political condition this depends on is not hidden; it is set out on its own near the top of this page.

Since June

This proposal has not been revised since it was written. The following developments came after it, and they are listed because each one confirms part of the problem it was written to address. None of them adopts its remedy.

1. June 22–24, 2026The term sheet and its companion documents are written.
2. Late June 2026The term sheet and several companion documents are circulated privately to a short list of people in labor, policy, and church institutions.
3. July 13, 2026The United Kingdom begins direct oversight of Amazon Web Services, Google Cloud, Microsoft, and Oracle as critical third parties to its financial system (Bank of England).
4. July 23, 2026The International Monetary Fund writes that the main risk from AI in finance “stems from concentration of critical services and shared dependencies” (IMF).
5. August 11, 2026A national survey finds 61 percent of American adults oppose new data centers in their area, up twelve points since spring (Annenberg).
6. August 2026The chair of the Financial Stability Board warns G20 finance ministers about risk spreading through highly concentrated third-party service providers.
7. September 19, 2026The documents are published here and deposited, unrevised, in a permanent public archive (Zenodo).

this work    outside events

Status and method

Everything here is a draft and is marked as one. The term sheet is non-binding and pre-decisional, offered to be tested and argued with rather than adopted. The June 2026 documents are published exactly as they were written; three load-bearing figures were corrected before the documents were first circulated, and those corrections are recorded openly in the sourcing index rather than made silently in the text.

The work was produced by one person in a compressed window, using AI research tools (Anthropic’s Claude) to gather and summarize sources, draft and revise text, and run internal consistency checks. It was built under a discipline meant to let it be checked rather than trusted: every load-bearing claim is labeled as provable, derived, or inference, the weak points are named in writing, and the open problems are listed rather than smoothed over. The framing, the judgments, and the conclusions are the author’s, and the author stands behind them.

No outside expert has reviewed this work. Earlier rounds described in the documents as adversarial review were consistency checks run through the same system that helped produce the drafts, which is useful but is not independent review.

The documents

Five documents, in reading order. The first is new for this publication; the other four are the June record. All five are deposited as a single permanent record at Zenodo, where the files cannot be altered; the copies below are for convenience.

1. Public summary The whole proposal in brief, written in September 2026 for a first-time reader. v2026-09-19a · 3 pp · PDF 52 KB
2. A Prospective Settlement for the AI Grid A plain-language introduction through the single case above, with the best argument against the proposal. v2026-06-24e · 4 pp · PDF 67 KB
3. Term sheet The instrument itself: recitals, the fee and fund, covered providers, the continuity guarantee, governance, the access floor, what each party gives and receives, what is open for negotiation, and what is deliberately deferred. v2026-06-22c · 9 pp · PDF 171 KB
4. Provenance and horizon How the proposal was reasoned from evidence, including the claims that earlier review killed, and what an achievement of this kind would open. Published without its final page, an internal appendix. v2026-06-24a · public edition · 4 pp · PDF 104 KB
5. Sourcing and receipts index, public edition Every load-bearing outward claim, tagged provable, derived, or inference, as verified on June 24, 2026, with the three corrections and the weak points named. v2026-09-19a · register verified 2026-06-24 · 5 pp · PDF 87 KB

Two items from the June private set are not published: a separate note on domestic political analysis, and the internal appendix to the provenance note. Apart from those two items, the June documents are published in full and unchanged.

Later material, not part of the deposited record. Exposure: who is exposed, who holds the switch, and who says so — a two-page briefing sheet in charts, September 2026 · 2 pp · PDF 107 KB. Its figures come from the sources named on the sheet.

Authorship, citation, and license

Dan Sullivan has spent over thirty years in labor, politics, and nonprofit work.

Permanent record

All five documents are deposited, unrevised, in a single record at Zenodo, the public research archive operated by CERN. Files in that record cannot be altered; any later changes would appear as a new, separately dated version.

DOI (this version)

10.5281/zenodo.22850312

DOI (all versions)

10.5281/zenodo.22850311

License

Creative Commons Attribution 4.0 International (CC BY 4.0). Anyone may use, adapt, and build on this work, including commercially, with attribution.

How to cite

A note on limits

This is not offered as a finished answer. It surely has mistakes, and other people are likely to do better work on the same questions, which is as it should be. What is here is fixed and dated, so that what comes after it has something to be measured against.