Three lines converged in an evening digest that are usually read separately. Nvidia guarantees up to $105 billion for OpenAI's data center lease in Ohio. SoftBank postpones the IPO of SB Energy — the very company building that data center — due to SEC questions and dependence on a single customer. Oracle cites force majeure in a dispute over its New Mexico campus, and this is the first clear signal that a construction contract can turn into a contingent liability.
The scheme itself is neither new nor invented in 2026. A supplier giving the buyer credit to purchase its own product — this is exactly what Lucent, Nortel, and Cisco did in 1997–2001, and how that ended is well known. What's far more interesting: in 2001 the money burst, but the physical asset outlived its owners by twenty years. Why — and whether this will repeat with GPUs that have a six-year term written directly into the credit agreement.
I checked the archive: out of 585 published Curiosity pieces, vendor financing has never been examined. Neither Lucent, nor Nortel, nor the words "vendor financing" appear in the texts.
The setup is simple to the point of indecency. An equipment manufacturer gives a startup telecom operator a loan, the operator spends it on equipment from that same manufacturer, and pays it back from revenue that equipment is supposed to generate. The vendor gets revenue immediately, risk — later. The young company gets networks that the capital market wouldn't have given it.
In the late 1990s this was considered not a trick but a competitive weapon. Lehman Bros. analyst Steven Levy explained in 2000 that in the international wireless market, large firms capture share precisely through "deep pockets," and warned this would hit margins: "the question is revenue sustainability. When growth comes from better financing rather than technology, profits ultimately suffer because the R&D payoff doesn't arrive."
The scale was significant. Lehman Bros. estimated that as of March 31, 1999, Lucent had $7.2 billion in customer obligations on its balance sheet — roughly 19% of sales for the fiscal year. By December 2000, Light Reading compiled different numbers from reports: Alcatel — $1.5 billion, Cisco — $2.5 billion, Nortel — $2.1 billion, Lucent — $7 billion. Nortel's obligations, however, were already shrinking: from $4.1 billion at the end of 2000 to $2.8 billion at the end of March 2001.
Then begins the part that almost no one could spot in 2000 reports. As CFO.com wrote, American standards required almost no disclosure of such financing details, and "companies are stingy about this in financial reports." Yankee Group analyst Nancy Ruzicka put it more directly: "Nobody will talk about this. Nobody wants to discuss it — ever. Which is itself strange."
The perfect illustration — the Winstar Communications case. Lucent in 1998 lent the company over $700 million to build a network and under the agreement terms was obligated to give more. In March 2001, Winstar didn't receive from Lucent about $90 million that it was supposed to transfer on March 30, and couldn't service its own debts, including debt to Lucent.
The mechanics became self-enclosed: to pay Lucent, Winstar needed Lucent's money. The company missed interest payments of $75 million, filed for bankruptcy in Delaware, and on the same day rolled out a $10 billion lawsuit against the supplier, claiming that breach of agreement was precisely what drove it into Chapter 11. Lucent declared the lawsuit groundless and stated that it was Winstar who owed money.
Formally it all ended in May 2004 — and ended unpleasantly. The SEC accused Lucent of accounting fraud: the company improperly recognized about $1.148 billion in revenue and $470 million in pre-tax profit in fiscal 2000, with the scheme's foundation being "extra-contractual obligations" — discounts and side agreements used to lure buyers, then hidden from the company's own finance department. A separate item was precisely that Winstar deal: a $125 million software purchase at the end of the fourth quarter and a signed document masking the date of the side agreement. Ten people received charges, nine of them Lucent employees.
The fine was $25 million, imposed not for the fraud itself but for lack of cooperation: the SEC specifically listed that the company sent documents incompletely, lost materials from subpoenas, and after a settlement in principle, its former chief and outside counsel gave an interview to Fortune calling that very Winstar deal "a communications failure." Lucent also expanded the circle of employees entitled to indemnity — the SEC noted that too.
While lawyers sorted through papers, what papers don't describe remained in the ground. In two years the world laid about 100 million miles of fiber, spending $35 billion. In the U.S. alone, under railroads, cornfields, and gas pipelines lay 39 million miles of cable — about 1,566 Earth circumferences at the equator, $90 billion over four years. Merrill Lynch estimated utilization at 2.6% of capacity, other estimates gave 5% lit fiber, still others — 97.5% dark routes.
Here I must admit that numbers in sources don't match and can't be reconciled. In 2001 at the OFC conference, OFS Fitel listed five "analyst misconceptions" and explained that the famous 3% utilization is an average figure from a Merrill Lynch study applied to all operators at once, though network designers calculate peak load, and new networks aren't supposed to be loaded in the first years. It also emerges that 5% lit fiber and 3% utilization are two different metrics that by then had become completely confused. So the most quoted number from telecom crash history can't be used: it doesn't mean what everyone thinks.
What's genuinely unambiguous — prices. Long-haul bandwidth capacity fell 60% in 2001 alone, following a similar drop the year before, and dark fiber cost about $1,200 per mile versus $5,000 in 1997. And against this backdrop came what's usually left out of the bubble story: fiber from bankruptcies was bought cheap by those still standing. Level 3 assembled a network from others' assets, at its peak carried 70% of world internet traffic, and in 2017 went to CenturyLink for $34 billion. Zayo built forty-five acquisitions on the same move: the company's founder called them "optical orphans" held by "accidental owners," while the market after the crash considered the entire sector toxic.
That is, infrastructure didn't fail — the budget under it did. Fiber considered dead in 2002 carried streaming, video calls, and cloud ten years later. The irony is it was dark not only from excess: the bottleneck remained the last mile to the consumer, while long-haul routes were what got rebuilt.
Today's version of the scheme is called differently and doesn't look like a loan. Google committed to cover up to $44 billion in lease payments for data centers it doesn't own and doesn't occupy: if the tenant stops paying, Google pays. At the end of September last year, the same figure was $6.5 billion — nearly sevenfold growth in nine months. The point — chips: a guarantee from a company with Google's balance sheet makes the loan cheaper for the developer, the project becomes profitable, the project buys TPUs, Anthropic rents them.
Nvidia went more direct. Its 8-K from August 17, 2026 concerns residual value guarantees on leasing approximately 4.25 GW of IT load in Pike County, Ohio, with the option to add another 3.8 GW. The commitment is capped at $105 billion and kicks in only upon OpenAI's insolvency or non-payment. Then a queue works: the lessor must try to re-lease the facility at the same price, then sell, and only the shortfall falls on Nvidia. The guarantee applies to completed buildings, not construction, and terminates under three conditions — OpenAI receiving an acceptable credit rating, voluntary exit from the lease, or after twenty years.
A separate detail is worth a whole page of analysis. The company promised support has committed to reimburse Nvidia for everything it actually pays. A claim against a bankrupt debtor is worth little, but this is the sign of what such a guarantee is: not a gift but a bridge.
Scale, to avoid trivialities: the five largest U.S. tech companies hold about $1.65 trillion in obligations not captured in the "debt" line, of which roughly $821 billion are signed but not yet commenced leases, and $848 billion are contracts to purchase chips, energy, and cloud capacity. Goldman Sachs expects capex around $750 billion for 2026 and nearly $1.2 trillion for 2027, with the debt financing share growing from 26% in 2025 to 35% in 2027. For Alphabet this already produced the first quarter with negative free cash flow since going public in 2004.
I deliberately won't draw a bubble verdict here: the dispute is honest, and both sides have numbers. The defense emphasizes that the four largest hyperscalers covered 2025 capex from operating flow, that about 40% of investments are power, land, and buildings with 15–40 year service lives, not chips, and that AI capex constitutes 0.8% of global GDP versus 1.5%+ at past technology build peaks. Objections are also substantive: at Oracle the gap between a $50 billion plan and $20.8 billion operating flow is closed by debt and equity issuance, and tokenized revenue of all participants is still an order of magnitude less than commitments.
Here 2026 genuinely diverges from 2001. Back then the collateral was fiber with a 25–30 year service life — it outlived the debtors. Today the collateral is GPUs, and their service life is already written into credit documents.
In CoreWeave's DDTL 2.0 term loan agreement, the "GPU Depreciated Amount" is calculated on a straight line based on a six-year useful life. This isn't an analyst's estimate, it's the condition on which money was given. The loan itself is structured through a separate legal entity: the borrower is a subsidiary owning servers, parent guarantee is limited to a narrow set of violations, collateral includes practically all assets, 100% of borrower shares, rights under the customer contract, and control over every bank account. The borrower must report serial numbers of hardware purchased with each tranche. Ratings of the $8.5 billion issuance at A3 and A (low) levels are given not to CoreWeave itself with its speculative rating, but to Meta's $14.2 billion contract on the other side of the deal.
How realistic these six years are is an open question, and public data here argue with each other. Silicon Data indices show Hopper and Ampere rental rates declining. Meanwhile, Nebius in an earnings call said customers either renew Hopper or move to Blackwell, and in both cases hardware goes "at a better price than before." CoreWeave reported that two quarters before a 10,000+ H100 contract expired, the customer voluntarily renewed it at a price within 5% of the original. And in Oracle's thin margin report it emerged that old 2020 Ampere actually helps margins, while new chips squeeze them.
I can't resolve this contradiction: aggregated indices and corporate self-reports describe different universes, and I don't know which of the two instruments is lying. An indirect clue exists in actual prices: in US East, H100 rental from a hyperscaler fell from $7.25 in Q1 2025 to $5.40 in Q3 and has sat exactly there for a year, while in Asia-Pacific the drop was 54%. Meanwhile, 2017 V100 still rents at $3.06 an hour — the price hasn't been revised down in nine years, the chip simply stopped being the first choice.
It's worth looking separately at how service life appears in hyperscaler reports. Amazon and Meta in the same month of 2025 drew directly opposite conclusions from the same technology cycle: Meta extended server life to five and a half years and expected about $2.9 billion in annual savings from this, Amazon shortened part of its fleet from six years back to five, citing acceleration of AI technology development, and projected expense growth of $0.7 billion. Alphabet holds at six years. Two accounting departments, two auditors, the same equipment — and three different answers.
This is precisely why insurance began moving to the product itself. Since July 1, 2026, Nvidia offers partners a buyback of unused capacity at a fixed rate — that is, guarantees hardware utilization directly. The first such deal was a six-year deployment of 40,000 GPUs with Sharon AI. This structure underlies a $500 billion platform assembled with six institutional partners, where Nvidia can act as a backstop up to $125 billion. The mechanics are honest: it makes credit cheaper. And it also means vendor-financed capacity and independent capacity are no longer comparable products — the former's price includes Nvidia's share.
Looking at two eras side by side, the difference comes down to one date. In 2001 cheap fiber outlived its owners because its service life was many times greater than the debt term under it. In 2026 the six-year service life isn't just assumed — it's written into credit terms, and the same six-year line appears in hyperscaler reports as the most common estimate. Debt and collateral are matched on paper, while Nvidia's annual chip release rhythm has held for several years.
What follows practically? The debt will probably be repaid: behind borrowers stand contracts with the largest buyers, and the borrower itself is a bankruptcy-isolated entity with locked accounts. But whether collateral returns at the same price is another question, and the answer will be given by the price at which the same rack rents three years out.
A far more probable scenario is redistribution, already described by the previous cycle. Back then survivors weren't those with money but those buying others' assets after the collapse — Level 3 and Zayo. Today's analog of this move is visible in energy: Nvidia invested $3 billion in Lancium, meaning money went into energy infrastructure: land and grid connection. The durable part of the structure is the substation, building, and fiber inside the campus.
And speaking of fiber. Demand for fiber from data centers grew about 76% year-over-year in 2025, by 2027 this segment should deliver 30% of global demand versus under 5% in 2024, Chinese manufacturers are loaded with orders through early 2027, and cable delivery times have stretched to a year: expanding preform production takes up to two years. One AI cluster requires roughly 36 times more fiber than racks with ordinary servers. That is, the bottleneck construction hits today is the thin gut laid into the ground a quarter century ago by people who laid it in the wrong place and wrong volumes. Zayo, that same collector of "optical orphans," in September 2026 was signing long-term fiber supply contracts precisely because supply ran out.
A question remains that I have no answer for. If Nvidia guarantees residual value of buildings, buys back unused capacity, and simultaneously sells chips that load that capacity, then in a bad scenario it ends up at the center of three things at once: collateral producer, lease guarantor, and holder of a stake in the property. Infrastructure won't disappear from this — just as fiber didn't disappear. Next time it will just probably be sold by someone else.