The Ungoverned Void — Why No Board Owns the Risk in Circular AI Financing
Vendor financing in the late nineties was not a scheme. It arrived as a sales instrument, approved deal by deal by boards that examined each one on its merits and were right to approve most of them. The systemic exposure that accumulated — vendors financing customers who existed to buy from vendors, priced against a traffic-growth figure that one employee had set to whatever seemed appropriate — crossed every board’s boundary and sat inside none of them. Twenty-five years later the instruments are far better and aimed at the same boundary. Every party in today’s AI financing loop can answer for its own leg and none can answer for the loop. This essay argues that the gap is not an absence of governance but a jurisdiction with no occupant, and that the fix is narrower than a new regulator: a disclosure trigger built on closure rather than control.
I.
In the late nineties I sat close enough to the telecom equipment business to watch vendor financing become ordinary. It did not arrive as a scheme. It arrived as a sales instrument, discussed in the same register as delivery schedules and service-level terms. A customer wanted to build a network and lacked the capital. We wanted the order. Credit closed the gap. Every deal came to a board that examined it, priced the risk, and approved or declined it on the merits.
Those boards were not careless. Most carried the governance habits of a regulated era --- deliberate, documented, conservative by disposition. That is the part worth sitting with, because it is the reason nobody suspected it.
II.
The number nobody owned
Liberalization opened the market to entrants with no balance sheet and made growth the measure of a firm's prospects. New carriers needed equipment before they had revenue. Equipment makers had capital and wanted share.
The two facts met around a single figure. Internet traffic, the industry agreed, was doubling every hundred days. The claim originated inside WorldCom's UUNET, and its origin is worth knowing precisely. An employee built a spreadsheet translating the sales forecast into expected traffic, and described his method plainly years later: he assigned variables and set them, in his words, to whatever seemed appropriate. It was a best case. Within about a year it had escaped into white papers, analyst notes, a government report, and the mouths of politicians. Andrew Odlyzko, then at AT&T Labs, checked it against measured traffic and found the real rate was roughly a doubling per year. He published. Nobody stopped.
By then the number had become infrastructure. Business cases rested on it. Credit committees priced against it. It had authors but no owner, and no mechanism existed by which anyone could withdraw it.
Now consider a single decision as a director would have seen it. A new carrier wants a network. It has a license, a plan, and no money. Your competitor will finance the build if you will not. The projections are aggressive, but everyone's projections are aggressive, and demand is in fact growing quickly. The credit is secured against the equipment. Declining costs you the order, the reference, and a position in a market that will not stay open.
A board that approves this has not failed. It has done its job.
The common explanation for what followed is that appetite for risk grew. That explanation is too comfortable, because it lets a reader conclude the fault was character and that better people would do better. The mechanism needs no change in appetite at all. Once financing terms become the competitive variable rather than a service, holding your risk standards constant means ceding share to whoever matches the market.
Lucent said as much in its own filings, in language a board had approved. Network operators, the company disclosed, increasingly required their suppliers to arrange or provide long-term financing as a condition to obtaining or bidding on infrastructure projects. That is not a confession of appetite. It is a statement that the option to decline had closed. Its quarterly filings track what followed: credit commitments of roughly $3.19 billion at the end of 1998, $7.1 billion by September 1999, and $8.4 billion by that December. The same standards, applied consistently, quarter after quarter.
The same filings describe the disposal route. Lucent intended to lay these arrangements off to financial institutions and other investors, which would free capacity to write more. Read as treasury management, that is prudent. Read structurally, it is the mechanism by which the exposure left the one balance sheet where somebody was accountable for it.
III.
Where the risk went
The exposure did not disappear. It moved somewhere that had no owner.
Each board's duty is bounded by its own entity. That boundary is not a technicality; it is the substance of the obligation. A director owes care and loyalty to one company and its shareholders, and the exposure they answer for is the exposure on their own books. The systemic exposure --- vendors financing customers who existed to buy from vendors, revenue circulating rather than arriving, all of it priced against a number from somebody's best-case spreadsheet --- crossed every one of those boundaries and sat inside none of them.
This is not an absence of governance. It is a jurisdiction that exists and has no occupant. A director who noticed the pattern was noticing something they had no standing to act on. Acting on it meant unilateral withdrawal from a live market, which is a breach in the other direction. The alarm went unraised not because nobody saw, but because seeing conferred no authority. Odlyzko saw, published, and had no authority either.
When the reckoning came it took the shape available to it. Some of what happened was straightforward accounting fraud and was prosecuted as such, correctly. But the larger phenomenon --- hundreds of individually defensible decisions accumulating into a market that could not clear --- had no defendant, no meeting, no minute. Public accountability attached to the parts where a decider could be named, because a named decider is the only kind of failure our instruments recognize. The rest was written off as a bubble, which is the word we use for outcomes we cannot assign.
IV.
Twenty-five years on
Today's firms sit in the same position, and they say so without being asked. Announcing Nvidia's credit support for an OpenAI data center campus in Ohio, Jensen Huang raised the objection himself and answered it in four words: is this circular financing? No. OpenAI will pay the lease. That is not evasion. It is a chief executive who knows exactly whose obligation it is, volunteers the question, and is willing to be held to the answer. Every party in the loop can do the same about its own leg. Nobody can do it about the loop.
Note also what has happened to the resistance in the system. It is never stripped out by anyone's decision. It gets outrun. Approval thresholds are fixed sums; flows are not. A threshold that constitutes real resistance against one volume of business constitutes none against ten times that volume, and no one has to lower it for it to stop working. Governance also keeps its own clock. Boards meet quarterly; audit committees work annually. When money changes clocks and review does not, the ratio widens without a single control being weakened.
And the boards this pattern has returned to are better equipped than any of their predecessors. The nineties directors came from a regulated monopoly culture. Today's directors come from the reforms built on their failure. Audit committees are independent and must disclose whether they hold a designated financial expert. Enterprise risk management is a formal discipline with a published framework. Board risk committees, mandated after 2008 for large financial institutions, have spread well beyond them. Chief risk officers report to boards rather than through finance. Stress testing and scenario analysis migrated out of banking supervision into ordinary corporate practice.
The instrumentation is genuinely better. It is aimed at the same boundary.
Every one of those instruments answers a question of the form: what is our exposure, and can we survive a shock? A stress test models the effect of a downturn on your own book. It does not ask whether the demand you are stressing is demand you helped manufacture. Scenario analysis, done well, will tell a chipmaker's board precisely what happens if AI capital expenditure halves --- and will not tell them that their own credit support is one of the reasons the expenditure is at its current level.
The closest instrument, related-party disclosure, misses by construction. It triggers on control or significant influence. A web of arm's-length counterparties, each holding minority positions below the thresholds, each negotiating at market terms, generates no related-party disclosure at all, no matter how completely the money circles. The rules were written to catch a transaction between two parties who are secretly one. They were not written to catch six parties who are genuinely separate and collectively closed.
So a director at any node in the loop today can see further and more accurately than a Lucent director in 1999, and remains in exactly the same position. They can see their own book. They cannot see the loop. And if they could see the loop, they would have no standing over it, because their duty runs to their own shareholders and stops at their own boundary.
One thing has changed, and it does not point the way one might hope. In the nineties the diagnosis was published and ignored. Today the diagnosis is continuous. Analysts describe the circularity weekly. The Bank for International Settlements devoted part of its 2026 annual report to it, naming the sustainability of AI investment as one of four pressure points facing the global economy, alongside inflation, public finances and financial vulnerabilities. The debate is public, informed, and entirely correct in its particulars.
Not one of those observers has jurisdiction over any part of it.
That is what the void looks like when it is fully lit: diagnosis without authority. The nineties version was the same shape with worse instruments and a quieter diagnosis. Better instruments have improved our view of the thing and changed nothing about who can act on it.
The answer is not better directors, because the directors were not the problem the first time. Nor is it a new supervisor for the loop, which would mean inventing a jurisdiction and an authority to occupy it, and would arrive years after the cycle it was built for.
There is a narrower fix, and the related-party rule shows where it sits. That rule fails because it triggers on the wrong thing. Control and significant influence are tests of who commands whom, and the risk here comes from closure rather than command. A trigger built on closure would ask something a filer can already answer: what share of your reported revenue comes from customers in whom you hold an equity position, or whose purchases you have financed or guaranteed? Disclose that figure and its counterpart --- the share of your own spending funded by a party that books it as revenue --- and the loop becomes visible from every node on it, in numbers that are already in the general ledger.
The obvious objection is that disclosure is not what is missing. The circularity is described weekly, and describing it has changed nothing. But an analyst's description is an observation, and a line in a filing is an obligation. It is signed by officers, tested by auditors, and defended to a board. That is the conversion the last twenty-five years never made: it turns the loop from something anyone may notice into something someone must state.
None of this requires a new regulator or a new duty. It requires an existing disclosure rule to trigger on the structure that actually produces the risk. And it would put the question inside the boardroom, which is the only place it has ever needed to be --- rather than in the commentary, which has been right and powerless for twenty-five years.
Sources and notes
Mike Edholm has spent three decades advising corporations and institutions on scenario planning and strategic foresight, with assignments spanning telecommunications, computing, industrial automation, and international development. This essay is part of a series on scenario planning methodology.