Vendor Financing in AI: What Lucent and Nortel Taught Us the Hard Way

In the late 1990s, telecom equipment makers lent their customers the money to buy equipment. Revenue looked exceptional until the customers could not pay, and the losses landed back on the suppliers who had financed them. Similar structures now run through AI compute deals. The mechanics rhyme closely enough to be worth understanding, and the differences matter more than most comparisons admit.

What vendor financing actually is

Vendor financing is any arrangement where the seller provides or facilitates the money the buyer uses to purchase from them. It can be a direct loan, an equity investment, a guaranteed lease, or a purchase commitment paired with capital.

The accounting consequence is what matters. Revenue is recognised when the sale completes. The credit risk sits with the seller for the life of the financing, which can be years.

None of this is improper. It is a recognised commercial tool, used legitimately when a buyer has a viable business and a temporary financing gap. It becomes dangerous when it is used to manufacture demand that would not otherwise exist.

What happened in telecom

Through the late 1990s, telecom equipment manufacturers competed to supply a wave of new network operators. Many of those operators were young companies with ambitious plans and no revenue.

To win the orders, equipment makers extended credit directly. Lucent and Nortel both built substantial customer financing books. The equipment shipped, the revenue was booked, and the growth numbers were spectacular.

Then the operators failed to reach the subscriber numbers their plans assumed. They could not service the debt. The equipment makers absorbed the write-offs, and the same revenue that had driven their valuations up reversed into losses.

What made it destructive rather than merely unwise

Three features compounded. The buyers were financially weak, so default risk was genuine rather than theoretical. The financing was extended competitively, so terms deteriorated as rivals matched each other. And disclosure was poor enough that outside analysts could not size the exposure until it crystallised.

Any one of those alone is survivable. All three together turned a downturn into a solvency event.

The structures appearing now

Current AI arrangements are not identical to 1999 telecom, and they are not unrelated either.

The recurring shape is an equity investment or financing commitment paired with a purchase commitment running the other way. Nvidia investing in OpenAI while OpenAI buys Nvidia hardware. Microsoft investing in Anthropic while Anthropic commits to purchase $30 billion of Microsoft cloud capacity.

These are disclosed and they are legal. What they make difficult is separating genuine third-party demand from demand a supplier helped create, which is exactly the analytical problem telecom investors faced.

Where the parallel breaks down

Three differences are substantial and they all point the same way.

Buyer quality is not comparable

The telecom operators being financed in 1999 were pre-revenue companies with speculative business plans. Today's largest AI buyers include companies with enormous operating cash flow and investment-grade credit. That is not a small difference. It is the main one.

The equity structure differs from the debt structure

Telecom vendor financing was largely lending. If the customer failed, the loan was written off. Much of today's arrangement is equity investment, where the supplier owns a share of the customer's upside rather than holding a claim on it.

Equity is riskier in a downside and better aligned in an upside. It also does not create the same forced-sale dynamic that debt does when a covenant breaks.

The product has demonstrable demand

The financed capacity in 1999 was built for subscriber numbers that never arrived. Today's compute is running at capacity, with allocation constraints reported across the market. The demand is not speculative even if the return on it is unproven, a distinction examined in the analysis of measured AI returns.

How large is this actually

Scale determines whether a structural pattern is a curiosity or a problem, and here the answer is genuinely uncertain.

The announced arrangements are large in absolute terms and modest against the revenue of the parties involved. A supplier investing tens of billions in a customer is enormous by historical standards and small relative to a business generating hundreds of billions in annual revenue.

What cannot be established is the aggregate. No published figure sizes total sector revenue supported by supplier-provided capital, because the arrangements are disclosed individually and structured differently. Some are equity, some are purchase commitments, some are deferred payment terms, and none are reported under a common heading.

That absence should temper confidence in both directions. It means alarmist totals circulating in commentary are estimates, and it also means reassuring statements about limited exposure are equally unverifiable.

The tell to watch for

Historical parallels are only useful if they produce a specific thing to observe. Here is the one.

Watch for financing terms deteriorating under competitive pressure. In telecom, the fatal turn was not the first financing deal. It was the point where suppliers matched each other's terms to win orders, and credit standards fell because no one could afford to be the strict one.

In AI, the equivalent would be capacity providers extending increasingly generous payment terms, deferred commitments or guaranteed minimums to win commitments from buyers who could not otherwise afford them. That would be visible in disclosure long before it was visible in revenue.

The second tell is concentration. When a supplier's largest customer is also a company it has invested in, and that relationship represents a material share of revenue, the disclosure obligation exists and the risk is real regardless of how the structure is described.

Why nobody is required to add it up

Individual disclosure without aggregate disclosure is the condition that made the telecom episode a surprise, and it is the condition that exists now.

Each company reports its own arrangements in its own format, under accounting standards designed for a single reporting entity rather than for a sector-wide exposure question. There is no line item called supplier-funded revenue, and no requirement to create one.

The consequence is that even a diligent analyst cannot size the total. They can list the announced deals, which is what most coverage does, and they cannot say what share of any company's growth those deals supported.

That is worth stating plainly because it cuts against alarmism as well as complacency. Anyone quoting a total for supplier-funded AI revenue has estimated it, and anyone dismissing the concern on the basis that the numbers are small is doing the same thing in the other direction.

What it means if you are buying compute or software

Two practical implications, neither dramatic.

The first is that pricing you receive today may reflect a subsidised competitive position rather than a sustainable cost base. That is fine while it lasts and it should not be assumed to last through a full contract term.

The second is counterparty diligence. If your provider depends heavily on financing from a supplier who is also an investor, your service continuity is exposed to a relationship you are not party to and cannot observe. That belongs in a risk register even if nothing ever comes of it.

Why disclosure is the weak point

The reason the telecom episode surprised people was not that the financing was secret. It was disclosed. The problem was that nobody could aggregate it into a picture of total sector exposure, because each company disclosed its own book in its own format.

The same condition exists now. Every major arrangement in AI has been announced. No party publishes a consolidated view of how much revenue across the sector is supported by capital the seller provided, and no regulator requires one.

That gap is why commentary oscillates between dismissal and alarm. Both positions are available because the number that would settle it does not exist in published form.

Regulators eventually closed the equivalent gap in other markets by requiring standardised disclosure of counterparty exposure. Nothing comparable exists here, and nothing suggests it is coming, which means this remains an analytical problem rather than a supervisory one for the foreseeable future.

What to watch

The last row is the most important and the least discussed. A shift in the financing mix from equity toward debt would change the risk profile of these arrangements considerably, and it would appear in filings before it appeared in commentary.

Frequently asked questions

Where to start this week

One question and one document.

Ask your largest infrastructure or AI provider whether any of their capacity is financed by an equity investor who is also a supplier. Most will answer directly, because the arrangements are disclosed and the question is reasonable.

Then read the customer concentration section of any listed supplier in your stack. If their largest customer is also a company they have invested in, you have learned something about how durable that revenue is, and it took ten minutes.

Related reading