The Money Goes Round

Correction, 15 October 2026: Several telecom figures in this piece were wrong or unsupported. Nortel's $3.1 billion was its remaining unfunded customer financing at 30 September 2000, not its total commitment, and no source has been found for the $1.4 billion given as drawn or for credit "reported at some 130% of equipment cost". Lucent's $8.1 billion was committed but not substantially drawn: $2.0 billion had been drawn at 30 September 2000. The forty-seven competitive local exchange carriers were those that had "filed for bankruptcy or left the market" since 2000, not forty-seven bankruptcies. McKinsey's $25.6 billion was an estimate of what nine suppliers had extended, not their combined exposure. The statement that until August nobody had been required to hold a view on what a fleet of accelerators is worth also needs qualifying, because Broadcom took a view on a default during a lease in June. The argument that staging separates today's commitments from the telecom era is weaker as a result; revenue separates them more. The details are in In Exchange for the Guarantees, dated 12 October 2026.

Interest disclosure: T-OpenAI is a tokenized loan participation right providing economic exposure linked to OpenAI's pre-IPO valuation, issued through a dedicated issuing subsidiary, as are T-SpaceX and T-Kalshi. Following the SpaceX listing, the investment underlying T-SpaceX is being divested. Redemption opens only once a Redemption Start Date is announced, and as of 28 August 2026 none has been. Tessera's founder, and entities he controls, hold related exposure. Read what follows with that in mind.

Last of three on the capital structure of the AI build. Monday: who holds the science risk. Wednesday: what the collateral is worth. Today: whether the revenue paying for all of it is what it appears to be.

Wednesday's piece argued that the demand-forecast question was not where the telecom money was lost. Today's spends a third of its length on the most famous demand forecast in the sector's history, so the distinction is worth setting out first.

Over-forecasting is ordinary. Markets price it, and being wrong about growth is survivable. What happened in 1998 was different in kind: a number describing supply was put into circulation as a number describing demand, then quoted for three years without anyone re-deriving it. That is a category error rather than a forecast error, and it is harder to detect, because the number is accurate. It is simply about something else.

The Metric That Was Substituted

In April 1998, the US Department of Commerce published a white paper on the emerging digital economy. On page eight it repeated an estimate from UUNET that internet traffic was doubling every hundred days. That sentence appeared in analyst notes, earnings calls, and prospectuses for the next three years, and justified a great deal of the capital described on Wednesday.

Andrew Odlyzko, then at AT&T Labs, measured it. He and Kerry Coffman found US backbone traffic doubling roughly once a year: fast, but an order of magnitude off.

The sourcing is the interesting part. Odlyzko traced the Commerce paper back through an Inktomi white paper to UUNET's Mike O'Dell. When O'Dell spoke at Stanford in May 2000, what he described doubling every four months was the capacity of the UUNET network, measured in OC12-miles. By late 2000 he was describing traffic doubling annually while capacity grew eightfold.

A supply-growth figure was reported as a demand-growth figure, and once in circulation it was quoted rather than checked.

The Modern Version

The equivalent today is the annualized run-rate.

The criticism needs aiming carefully. Annualizing the most recent period is a reasonable operational metric, and private-round valuations are not generally struck as a multiple of it. The damage is done in public commentary: an annualized consumption figure gets divided into a headline valuation, and the resulting multiple is set against a trailing multiple on somebody else's contracted revenue, as though the two were comparable. One is an extrapolation from the most recent month of usage that customers can stop next month. The other is booked revenue under signed contracts.

The underlying growth is real in a way the CLECs' revenue never was. On Wall Street Journal reporting in May, Anthropic moved from $4.8 billion of revenue in Q1 2026 to a projected $10.9 billion in Q2, with compute cost per revenue dollar falling from around 71 cents to 56 cents and a first operating profit of $559 million. Monday's piece cited the same company's 2024 cash burn as equity-tranche spending. Both are true, and the distance between them in eighteen months is the strongest evidence against the case this series has made.

It deserves an answer rather than a concession. Margin improvement at one lab does not establish that the demand behind it is independent, which is the claim under test: unit economics can improve on revenue that suppliers and investors are funding. What it does establish is that the economics can move faster than the skeptical case assumes. That is a real hit against this series' timeline, if not against its structure.

One further disclosure: Anthropic and Nvidia are commercial counterparties, with a reported deployment of Nvidia systems at gigawatt scale. This piece uses one side of a supply relationship as the example of real demand and the other as the example of circularity risk. That is not a charge against either company. It is the ordinary structure of this industry, and it is why the question below is hard to answer from outside.

The question is not whether the revenue exists. It is what share of it is independent of the capital that created it.

The Circle, Then and Now

Between roughly 1998 and 2001, the dominant equipment makers lent money to the carriers buying their equipment and booked the resulting shipments as revenue. Lucent committed on the order of $8.1 billion of vendor financing. Nortel committed around $3.1 billion, of which about $1.4 billion was drawn, at one point extending credit reported at some 130% of equipment cost. Cisco promised around $2.4 billion. McKinsey later put combined exposure across nine suppliers at roughly $25.6 billion by the end of 2000.

The recipients were disproportionately competitive local exchange carriers, with little revenue and wholly dependent on continued capital-market access. Forty-seven went bankrupt between 2000 and 2003, against an overbuild estimated at around $60 billion across 1996-2001. The funding did not taper. Billions were available in January 2001 and essentially nothing by April.

Nvidia's commitments into OpenAI and across the neocloud layer are the structural descendant. The comparison is reflexive enough now that it needs disciplining rather than repeating.

What does not transfer. Lucent's $8.1 billion was committed, substantially drawn, and largely written off. Nvidia's headline commitments are reported as staged rather than drawn, which makes them a materially different instrument if the reporting is right. The conditional is deliberate: the terms are not public. And the recipients then had no revenue, where the recipients now have a great deal of it.

On proportion. In the quarter ended 26 July 2026, reported on 26 August, Nvidia's data center revenue was $89.0 billion, up 18% sequentially and 117% year on year, out of $96.2 billion in total. A vendor-financing headline is often juxtaposed against that total in a way implying the whole is suspect, which is not supportable. But the circular portion cannot be called small either, because the disclosure that would establish it does not exist; the terms, as noted above, are not public, and that cuts both ways. There is no public basis for treating the circular share as large, and none for treating it as negligible.

That print did address concentration. Nvidia's chief executive framed the quarter as a broadening of the buyer base, noting that a year ago "one lab alone was driving the buildout" and describing multiple frontier labs now scaling in parallel. The source is worth weighing: this is the party with the strongest commercial interest in the concentration critique being wrong. It is a testable claim, and customer concentration disclosures will test it. The same caution applies to the $108.0 billion guided for the current quarter, which is a forecast, in a piece about forecasts quoted as facts.

What does transfer is the mechanism, at some proportion of a much larger business. A supplier's balance sheet underwrites part of its own demand, and those shipments are recognized as revenue. That is worth monitoring because the proportion is a choice the supplier makes each quarter rather than a fixed property of the business.

It also complicates Wednesday's conclusion. That piece argued that stress concentrates in the credit and SPV layer rather than at the sponsors, with the caveat that the sponsors' cash cushion is thinner than headline balance sheets suggest. If circularity in the revenue base is material, the most consequential sponsor in this build is the supplier itself, and that separation is less clean than Wednesday made it sound. What is currently visible does not overturn the conclusion. It is the variable that would.

Contracted Offtake, and the Instrument It Resembles

The standard rebuttal is that today's build is pre-sold: long-term compute contracts, take-or-pay, contracted offtake, nothing like the speculative fiber laid along duplicated routes.

The telecom era had contracted offtake too. It was called an indefeasible right of use: a long-dated contractual right to specified capacity, sold between carriers. IRUs looked exactly like contracted revenue. Some were swapped between carriers on roughly matched terms, each side recognizing revenue on the sale while capitalizing the purchase.

That practice did not end quietly. The SEC opened a formal investigation into Global Crossing's concurrent purchases and sales of capacity in February 2002, resolved in April 2005 by a consent order under which the company agreed to a cease-and-desist on certain reporting violations, without a civil penalty. Qwest, on the other side of some of those swaps, paid $250 million to settle SEC fraud charges without admitting wrongdoing. A congressional examination found that Global Crossing's pro forma reporting had included virtually the full sale side of the swaps in its cash revenue and earnings figures.

The point is not the enforcement history. It is that these contracts looked like third-party demand and, in the swapped cases, were closer to a mirror.

So the test is not whether a contract exists. It is three questions, addressed to anyone holding paper anywhere in the structure.

Is the customer independent of the supplier? An offtake agreement with a customer the supplier has also funded is not third-party demand. It is the supplier's own capital returning under a different heading.

Is the customer's own funding secure for the contract's duration? A ten-year commitment from a pre-profit business dependent on continued capital-market access is a ten-year commitment in name and a funding-cycle commitment in substance. That is the distinction between a compute contract and a utility power purchase agreement.

Are the minimum commitments and termination rights disclosed? For most large reported arrangements in this sector, they are not. Whether a contract is genuinely take-or-pay is a question filings can answer and press releases cannot.

CoreWeave's investment-grade facility in March is the good case: rating agencies looking at a named, disclosed, investment-grade counterparty behind a specific backlog, which answers the first two questions and is why the paper priced where it did. The same borrower's August facility, without that counterparty behind it, met resistance in syndication. The structure is not what earns the rating. The customer is.

Where the answers are not good (an unnamed counterparty, an undisclosed term, a customer whose funding runs shorter than the contract), it is an IRU.

What the Three Pieces Add Up To

The science tranche sits on balance sheets built for appropriable returns, with no statutory containment of the kind that sustained Bell Labs or sustains pharmaceuticals, and the containment that exists protects deployment rather than research. The collateral securing the build secures very little on its own: the credit rests on the offtaker, the hardware is a recovery floor, and until this month nobody had been required to hold a view on what a fleet of accelerators is worth at the end of a contract. The revenue behind those offtake contracts is real, fast-growing, and of unverified independence.

Each layer rests on the one below it, and the bottom layer, the durability and independence of the end customer, is the one nobody has had to evidence. The defensible position these businesses occupy is narrower than the aggregate capital committed to it assumes. Not absent. Narrower.

That is not a forecast of collapse, and none is being made. The fiber got used. Level 3 survived to be acquired. The technology thesis of 1999 was, in the end, correct, and its equity holders were still wiped out, because being right about the technology and being right about the security are different questions, answered by different documents.

The second kind takes longer to read.

Sources are as cited in the text: US Department of Commerce, April 1998; Odlyzko and Coffman on backbone traffic measurement, and Odlyzko's reconstruction of the doubling claim's sourcing; vendor financing figures per contemporaneous reporting and McKinsey analysis to end-2000; CLEC bankruptcy counts and overbuild estimates per retrospective compilations; Anthropic revenue and margin figures per Wall Street Journal reporting, May 2026; Nvidia figures per the company's second quarter fiscal 2027 results, reported 26 August 2026, for the quarter ended 26 July 2026; Global Crossing enforcement history per company filings and SEC records; Qwest settlement as reported; CoreWeave facility terms per company filings. Figures are as of the dates and reports stated.

This is market commentary, not investment advice. It is not a recommendation to acquire, hold or redeem any Tessera product, or to take or avoid exposure to any company mentioned.

T-Tokens are tokenized loan participation rights, not equity. High risk. DYOR. Not financial advice. Not available in the US or other restricted territories. tessera.pe/terms